personal wireless communications: pwc 05 - proceedings of the 10th ifip conference

Personal Wireless Communications

PWC'OS

This page intentionally left blankThis page intentionally left blank

Personal Wireless Communications

PWC'OS

editor

Pascal Lorenz University of Haute Alsace, France

Imperial College Press

Colmar, France 25-27 Augusat 2005

Published by

Imperial College Press 57 Shelton Street Covent Garden London WC2H 9HE

Distributed by

World Scientific Publishing Co. Pte. Ltd. 5 Toh Tuck Link, Singapore 596224 USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601 U K office: 57 Shelton Street, Covent Garden, London WC2H 9HE

British Library Cataloguing-in-Publication Data A catalogue record for this book is available from the British Library.

PERSONAL WIRELESS COMMUNICATIONS Proceedings of the 10th IFIP International Conference

Copyright 0 2005 by Imperial College Press All rights reserved. This book, or parts thereoj may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.

For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.

ISBN 1-86094-582-1

Printed in Singapore by B & JO Enterprise

Welcome to theloth IFIP International Conference on Personal Wireless Communications hosted in Colmar, France. PWC provides an international forum for discussions between researchers, practitioners and students interested in new developments in mobile computing and wireless networks. PWC’2005 is the tenth conference of this series and is sponsored by IFIP WG 6.8.

As other PCW events in the past, this professional meeting continues to be highly competitive and very well perceived by the international networking community, attracting excellent contributions and active participation. This year, a total of 192 papers from 31 countries were submitted, from which 57 have been accepted. Each paper has been reviewed by several members of the PWC’2005 Technical Program Committee. We were very pleased to receive a large percentage of top quality contributions.

The topics of the accepted papers cover a wide spectrum: wireless sensors, signalization, traffic and QoS in wireless networks, Ad-Hoc, IEEE 802.11, cellular and mobiles networks. We believe the PWC’2005 papers offer a large range of solutions to key problems in wireless networking, and set challenging avenues for industrial research and development.

We would like to thank the PWC’2005 Technical Program Committee members and the referees. Without their support, the creation of a very broad conference program would not have been possible. We also thank all the authors that dedicated a particular effort to contribute to the PWC’2005. We truly believe that thanks to all these efforts, the final conference program consists of top quality contributions.

We are also indebted to many individuals and organizations that made this conference possible, specifically, IFIP, IEEE, ARP, Conseil Gtnkral, France Telecom and University of Haute Alsace. In particular, we would like to thank the members of the PWC’2005 Organizing Committee for their help with all logistic aspects of organizing this professional meeting.

We expect the 10th International Conference on Personal Wireless Communications to be an outstanding international forum for the exchange of ideas and results between academia and industry, and provide a baseline of further progress in networking area.

We hope you will enjoy your stay in Colmar and be able to spend some time to visit various points of interest on this lovely city.

Pascal LORENZ Conference Chair

V

Preface

International Scientific Committee

Arup ACHARYA, IBM TJ Watson Research Center, USA Villy Baek IVERSEN, Technical University of Denmark, Denmark Sathish CHANDRAN, RF Consultant, Malaysia Prosper CHEMOUIL, France Telecom R&D, France Pedro CUENCA, Universidad de Castilla La Mancha, Spain Franc0 DAVOLI, DIST - University of Genoa, Italy Silvia GIORDANO, ICA-DSC-EPFL, Switzerland Cambyse Guy OMIDYAR, Oman Veikko HAKA, Telecom Finland, Finland Takeshi HAlTORI, Sophia University, Japan Sonia HEEMSTRA de GROOT, Ericsson EuroLab Netherlands, The Netherlands Ousmane KO&, UniversitC Paul Sabatier - IRIT, France Sanjeev KUMAR, University of Texas, USA Pascal LORENZ, University of Haute Alsace, France Damien MAGONI, University of Strasbourg, France Gerald MAGUIRE, KTWInst. for Teleinformatik, Sweden Zoubir MAMMERI, University of Toulouse, France Olli MARTIKAINEN, Helsinki Univ. of Tech., Finland Ignacious NEMEGEERS, Delft University of Technology, The Netherlands Algirdas PAKSTAS, London Metropolitan University, UK Guy PUJOLLE, University of Paris 6, France Pierre R. CHEVILLAT, IBM Zurich Research Laboratory, Switzerland Pierre ROLLN, France Telecom R&D, France Nikola ROZIC, University of Split, Croatia Debashis SAHA, Indian Institute of Management (IIM) Calcutta, India Tadao SAITO, Toyota Infotechnology Center, Japan Dilip SARKAR, University of Miami, USA Jan SLAVIK, TESTCOM, Czech Republic Otto SPANIOL, Aachen University of Technology, Germany Samir TOHME, ENST, France Adam WOLISZ, Technical University Berlin, Germany Jozef WOZNIAK, Technical University of Gdansk, Poland Jun ZHENG, University of Ottawa, Canada

vii

Contents

Wireless Sensors

Energy-Efficient Application- Aware Communication for Wireless Sensor Networks

R.M. Passos , C. J. N. Coelho, A.A.E Loureiro, R.A.E Mini (Federal University of Minas Gerais, Brazil)

3

SDMA in Connections between Wireless Sensors and Wired Network K Hasu, H. Koivo (Helsinki University of Technology, Finland)

11

A Reliable and Energy-Efficient Routing Protocol for Wireless Sensor Networks

19

K. K. Loh, S. H. Long (Nanyang Technological University, Singapore); Z Pan (Georgia State University, USA)

Analysis of Coverage and Connectivity in Wireless Ad Hoc Sensor Networks

27

J. Wang, L. Wang (Central South University, China); R. Xiao (National Nature Science Foundation of China, China)

MANET

Cross-layer’s Paradigm Features in MANET Benefits and Challenges 37 L. Romdhani, C. Bonnet (EURECOM Institute, France)

Two Bandwidth-Violation Problems and Bandwidth-Satisfied Multicast Trees in MANETs

C.C. Hu, G.H. Chen (National Taiwan University, Taiwan); E. H. K. Wu (National Central University, Taiwan)

51

LAP-RRP: A Reliable Routing Protocol with Link Availability Prediction in MANET

59

J. Wang, J. He (Central South University, China); X . Lu, (National University of Defense Technology, China)

ix

X

An Efficient Load-Balancing Algorithm for Supporting QoS in MANET M. Brahma, K.W Kim, A. Abouaissa, f? Lorenz, (University of Haute Alsace, France); M.M.O. Lee (University of Dongshin, Korea)

A Bandwidth-Efficient Cross Layer Probability Routing €or MANEiTs X. Wang, H. Yu, C. Ran, X. Zhang, W Qi, (Information Science and Engineering Institute, China)

Ad Hoc (I)

Efficient Bandwidth Allocation for Basic Broadcast and Point-to-Point Services in the ADHOC MAC Protocol

J. R. Ga'llego, A. Herna'ndez-Solana, M. Canales, A. Valdovinos, (University of Zaragoza, Spain); L. Campelli, M. Cesana, A. Capone, E Borgonovo (Politecnico di Milano, Italy)

Connectivity Aware Routing in Ad-hoc Networks J. Leguuy, 7: Friedman, S. Fdida, (University Pierre et Marie Curie, France); K Conan, A. Cotton, (Thales Communications, France)

A New Approach for TDMA Scheduling in Ad-hoc Networks D.D. Vergados, D. J. Vergados, (University of the Aegean, Greece): C. Douligeris, (Univeristy of Piraeus, Greece)

A Self Organizing Algorithm for Ad-Hoc Networks N. Kettaj A. Abouaissa, F! Lorenz, (University of Haute Alsace, France); H. Guyennet, (University of Franche-Comti, France)

HDSR, Hierarchiacal Dynamic Source Routing for Heterogeneous Wireless Mobile Ad Hoc Networks

M. Naserian, K.E. Tepe, M. Tarique, (University of Windsol; Canada)

Ad Hoc (11)

Analyzing the Effect of Cooperation Approaches M. Frank, M. Holschbach, P: Martini, M. Plaggemeiel; (University of Bonn, Germany)

67

77

87

99

107

115

125

135

Mobility Management in Multihops Wireless Access Networks E Theoleyre, E Valois (Inria Ares, France)

Location Update Protection for Geographic Ad Hoc Routing Z. Zhou, K.C. Yow (Nanyang Technological University, Singapore)

PEDCF: Predictive Enhanced Service Differentiation for IEEE 802.11 Wireless Ad-Hoc Networks based on AutoRegressive-Moving Average Processes

N. Tabbane, S. Tabbane (SUP’COM, Tunisia); A. Mehaoua, (University of Versailles, France)

IEEE 802.11

A Carrier-Sense based Transmission Power Control Protocol for 802.11 Networks

J. Rao, S. Biswas (Michigan State University, USA)

Chaos Shift Keying and IEEE 802.1 1 a G. Plitsis (Aachen University of Technology, Germany)

Effect of Time-Correlated Errors on Power-Saving Mechanisms for IEEE 802.11 Infrastructure Networks

G.A. Safdar, WG. Scanlon (The Queens University of Bevast, UK)

COMPASS: Decentralized Management and Access Control for WLANs A. Hecker (Wavestorm, France); E.O. Blass (University of Karlsruhe, Germany); H. Labiod (ENST, France)

xi

146

154

162

173

181

189

197

QoS Provisioning Mechanisms for IEEE 802.11 W A N : A Performance Evaluation

205

J. Villaldn, R Cuenca, L. Orozco-Barbosa (University De Castilla La Mancha, Spain)

QOS

Impact of Varying the Minimum Value of Contention Window (CWmin) of the IEEE 802.11 MAC Protocol on the QoS Parameters

M. Saraireh, R. Saatchi, R. Strachan, S. Al-Khayatt (Shefleld Hallam University, UK)

219

xii

Statistical QoS Guarantees in Bluetooth under Co-channel Interference J.L. Sevillano, D. Cascado, E Diaz del Rio, S. Vicente, G. Jimknez, A. Civit-Balcells (University de Sevilla, Spain)

Performances Evaluation of the Asynchronous Bluetooth Links in a Real Time Environment

7: Khoutaij E Peyrard (Research Icare Team, France)

System Simulations of DS-TRD and TH-PPM for Ultra-Wide Band (UWB) Wireless Communications

S. Vasana, K. Phillips (University of North Florida, USA)

Global Solution for the Support of QoS by IEEE 802.11 Wireless Local Area Networks

A. Bedoui (ENII: Tunisia); K. Barkaoui (CNAM, France); K. Djouani (University of Paris 12, France)

Traffic

Cross-Layer Design for Dynamic Resource Allocation in Wireless Networks

J.Y Kim, A. Saidi, R.J. h n d r y (The MITRE Corporation, USA)

Coverage Area Analysis of Soft Handoff on Cellular CDMA Systems T.L. Sheu, J.H. Hou (National Sun Yat-Sen University, Taiwan)

LMS vs. RLS for Adaptive MMSE Multiuser Detection over Time Varying Communications Channels

Z.B. Krusevac, PB. Rapajic, R.A. Kennedy, (National ICT Australia and University of New South Wales, Australia)

Performance Analysis of a Preemptive Handoff Scheme for Multi Traffic Wireless Mobile Networks

D.D. Vergados, A. Sgora (University of the Aegean, Greece)

Performance Analysis of DS-CDMA Systems in Multiple-Cell with Correlated Fading Channels

J.I.-Z. Chen, N. Chi-Kuang, EC. Chung (Da Yeh University, Taiwan)

227

235

244

252

263

279

287

295

303

xiii

Cellular Networks

Multimedia Transmission over Third Generation Cellular Networks 317 A. Alexiou, C. Bouras, K Igglesis (Research Academic Computer Technology Institute, Greece)

Distributed Content Sharing in Cellular Networks 325 B. Bakos, L. Farkas, J.K. Nurminen (Nokia Research Centel; Hungary); K. Marossy (Nokia Technology Pla~orms, Hungary)

On UMTS HSDPA Performance 338 I! Matusz, J. Wozniak (Gdansk University of Technology, Poland)

Supporting Flexible Network Operator Policies in EGPRS Trough Admission Control

346

D. Todinca, I. Sora (University Politehnica Timisoara, Romania); f? Perry, J. Murphy (University College Dublin, Ireland)

Performance and Quality of Service Management in GPRS Network 354 0. El Ghandoul; M. Fikry (Helwan University, Egypt); S. El-Ramly (Ain Shams University, Egypt)

Mobile Networks (I)

Enabling Mobile IPv6 in Operational Environments X. Fu (University of Gottingen, Germany); H. Tschofenig, S. Thiruvengadam (Siemens AG, Germany); W Yao, (Brunel University Wenbing, UK)

365

Performance Evaluation of Tunnel-based Fast Handovers for Mobile IPv6 in Wireless LANs

H. Lu, J. Li, I? Hong (University of Science and Technology of China, China)

373

Mobile IPv6-type Route Optimization Scheme for Network Mobility (NEMO) Support

38 1

B. f? Kafle (The Graduate University for Advanced Studies, Japan); E. Kamioka, S. Yamada (National Institute of Informatics, Japan)

xiv

Comparative Analysis of Handoff Delay of MIFA and MIP A. Diab, A. Mitschele-Thiel, R. Boeringel; (Ilmenau University of Technology, Germany)

Integration the Protocols HMIPv6 and Diffserv over M-MPLS in Order to Provide QoS in IP Network Mobility

J. H. Ortiz (University Polytechnic of Madrid, Spain)

Mobile Networks (11)

An Agent-Based Framework for Mobile Multimedia Service Quality Monitoring and Diagnosis

M. Li (Nokia Research Centel; USA)

Neural Network and Self-Learning Based Autonomic Radio Resource Management in Hybrid Wireless Networks

C. Shen, D. Pesch, J. Irvine (Cork Institute of Technology, Ireland)

WebBee: An Architecture for Web Accessibility for Mobile Devices K. Upatkoon, W Wang, S. Jamin (The University of Michigan, USA)

Empowering Wireless UPnP Devices with Webprofiles J.I. Vazquez, D.L. De Ipina (Deusto University, Spain)

UICC Communication in Mobile Devices Using Internet Protocols B.H. Nguyen, H.K. Lu (Axalto, Smart Cards Research, USA)


Modular Proxies for Service Adaptation and Session Continuation over Heterogeneous Networks

I: Seipold, I: Tantidham (RWTH Aachen University of Technology, Germany)

A Channel Preemption Model for Multimedia Traffic in Mobile Wireless Networks

TL. Sheu, YJ. Wu (National Sun Yat-Sen University, Taiwan)

389

397

405

413

421

429

438

449

457

xv

Nonuniform-Detection-Based Fast Mobile IP Handoff for Wireless LANs 465 B. Shen, H. Zhang, K Liu, E Zhao (Beijing Jiaotong University, China)

Analysis of ACD: Autonomous Collaborative Discovery of User and 473 Network Information

T Zhang, S. Madhani (Telcordia Technologies, USA); S. Mohanty (Georgia Institute of Technology, USA)

EasyMN: An Effective IP Mobility Solution for High-Mobility Network 479 L. Wang, M. u! Xu, K. X u (Tsinghua University, China)

Signalization

Proposal of PAPR Reduction Method for OFDM Signal by Using Dummy Sub-can-iers

l? Boonsrimuang, K. Mori, H. Kobayashi (Mie University, Japan)

489

Adaptive Scheduling for Heterogeneous Traffic Flows in Cellular Wireless OFDM-FDMA Systems

497

S. Valentin, H. Karl (University of Paderborn, Germany); J. Gross, A. Wolisz (TU Berlin, Germany)

The Power Spectral Density of the H-Ternary Line Code: A Simulation Model and Comparison

507

A. Glass (Technical Studies Institute, UAE); N. Abdulaziz (University of Wollongong, UAE); E. Bastaki (Dubai Silicon Oasis, UAE)

An Optimized CPFSK-Receiver based on Pattern Classification in Combination with the Viterbi Algorithm

D. Briickmann (University of Wuppertal, Germany)

518

Wireless Sensors

ENERGY-EFFICIENT APPLICATION-AWARE COMMUNICATION FOR WIRELESS SENSOR NETWORKS

RODRIGO M. PASSOS , CLAUDIONOR J. N. COELHO JR ,

Department of Computer Science Federal University of Minas Gerais, Brazil { passos, coelho, loureiro, raquel} Qdcc.ufmg. br

ANTONIO A. F. LOUREIRO AND RAQUEL A. F. MINI

In this work, we propose a new dynamic power management (DPM) technique for wireless sensor networks (WSNs) that selectively shutdowns the sensor node radio and other hardware components based on the application-level information. The DPM technique is modeled using the hybrid automata framework, which is leveraged to represent different duty cycles, according to the application requirements. Both the application-level information and the application requirements determine the sensor node duty cycle and the tradeoff between energy conservation and data delivery rate. We provide a comparison of our DPM technique using the main types of routing protocols to show the performance of our DPM technique in a sensor network modeled for a fire detection application. Simulation results reveal that our DPM technique can change the duty cycle of a node according to the data delivery quality required by the application and the need for communication, saving energy and, thus, extending the WSN lifetime.

1. Introduction

Sensor networks are commonly designed to be used in hostile environments and the re-charge process of typically battery operated sensors can be almost impossible, which makes the energy consumption in sensor nodes crucial for the network lifetime. The sensor nodes energy limitation requires energy-efficiency in all aspects. At design time, much work has been performed to make energy efficient circuits, architecture, communication protocols, routing protocols, algorithms and sensing . At run time, dynamic power management (DPM) techniques have been leveraged to selectively shutdown hardware components to avoid the waste of energy".

In multihop WSNs, communication is the major consumer of energy5 and must be carefully designed and performed to reduce the energy consumption. According to7, the main factors of energy ineffiency can be due to useful energy consumption (transmitting, receiving and routing data)

3

4

and wasteful energy consumption (idle listening, collisions, overhearing, and control packets). The transmission process and the idle listening problem represent the main factors of energy consumption and they must be carefully conducted by the power management (PM) policy, achieving the application requirements in a satisfactory way.

We believe that the best trade-off among energy conservation, packet latency and data delivery rate can be obtained when application-level information is used in the communication process. In4, we have shown the benefits of a system-level DPM technique for single-hop communication. In this work (Section 2), we extend the DPM technique to multi-hop communication, by leveraging the application-level information to selectively turn off the sensor radio, avoiding unnecessary transmissions due to neighbor data similarity, avoiding the idle listening problem and reducing the energy consumption, without modifying the routing and MAC protocols.

As previously introduced in4, the hybrid automata framework is used to model the DPM and the application behavior in a formal way. In order to show the performance of our DPM technique (Section 3), we use the basic routing algorithms for WSNs to show the benefits and impacts of a system-level DPM technique. We also use a fire detection application to show the performance influence of the application-level information.

2. A New Application-Aware DPM Approach

In4, we have previously defined the DPM hybrid automata model in formal way. In this work, we improve the application-driven DPM technique by leveraging the application-level information to analyze the needs for communication and to avoid wasteful energy sources (idle listening). We extend the power state machine, through the hybrid automata framework, representing in a single model control data, neighborhood and environment information, and the sensor node duty cycle.

The DPM hybrid automata represents different control modes. Each control mode represents a duty cycle that performs different rates of sampling, transmitting and receiving operations, according to the required QoS delivery. According to the required quality of service, each location defines the necessary hardware configuration, selectively turning off unnecessary components. The main goal of the DPM hybrid automata is to change duty cycles (locations) according to the QoS requirements of the application. The expected behavior of the variables are used to keep the control mode in a lower power consumption location. Unexpected behaviors may

5

require a higher data delivery rate, forcing the DPM hybrid automata to go to a location that corresponds to the required QoS. The application-level requirements trade off energy conservation and data delivery rate, performing the transitions to the most appropriated location (duty cycle), according to the environment and the neighborhood behavior.

Figure 1 represents the sensor node duty cycle leveraged by the DPM technique. The duty cycles represent the sensor node operation behavior for each location of the DPM hybrid automata. They define the operation rates of sampling (I&), transmitting (Rt) and receiving (I?,.), and the sleeping time (&). The operation rates are related to the maximum execution time of the duty cycle and they axe defined according to the required QoS.

Figure 1. Graphical representation of the DPM duty cycle.

According to each operation rate and to the duty cycle execution time, the main sensor node tasks are performed to guarantee the sensor node basic operation. After a sleeping timeout, if the duty cycle execution time is not over, the sensor node may decide to listen for packets by turning on the radio in receive mode, or it may perform a sensing operation. Otherwise, if the duty cycle execution time is over, the PM policy makes a decision of changing the duty cycle, according to the flow and invariant conditions of the DPM hybrid automata. After a sensing operation, the sensed data may be transmitted, according to the transmission rate. If the sensed data is relevant (it reaches an invariant condition), it may be necessary to change to a new duty cycle that better represents the data delivery quality required for the sensed information. Otherwise, not relevant sensed data may contribute

6

to conserve energy by the leverage of a sleep mode, according to the sleeping rate. In the same duty cycle a change decision is evaluated for the routing process. According to the data received from neighbors of a node, the PM policy verifies if the current sensor node duty cycle is able to attend the routing process in a satisfactory way. Otherwise, a more appropriated duty cycle may be required. The energy-efficiency is obtained by verifying at each sensor node whether it needs to communicate.

If a sensor node is transmitting relevant data, it may need their neighbors (or at least the routing path) to route its data to the sink. This situation may trigger a transition to a location (control mode) that maps a receive-dominant duty cycle with a higher receiving rate and a higher execution time. On the other hand, a transmit-dominant location is useful to report unexpected data behavior detected by the sensing operation of the sensor node. The hybrid automata can also map sleep-dominant and sensing-dominant locations, or any kind of required QoS, in terms of data delivery rate. In fact, the total execution time and operation rates performed in each duty cycle model the power consumption of the hybrid automata locations. Our DPM technique tries to exploit the sleep-dominant locations, according to the application control variables behavior.

In the DPM technique context, the need for communication decision and the communication rate required decision are performed by the leverage of three types of application-level information: (i) environment sensed data, (ii) neighborhood received data and (iii) neighbors current duty cycle.

The environment sensed data is useful to monitor the environment behavior. The environment variables are represented as differential equations (flow conditions) in the hybrid automata. The flow conditions are used to determine the expected behavior of the environment. The sensed data is used to determine if the environment is behaving as expected. An expected behavior may lead a transition to a sleep-dominant location or a transition to a location that implements a low transmitting rate. Otherwise, an unexpected behavior may lead a transition to a transmit-dominant location, according to the application requirements of data delivery quality.

The neighborhood received data is useful to monitor the neighbors environment behavior. An unexpected environment behavior may lead a transition to a receive-dominant location, improving the data delivery rate to the sink. Otherwise, expected data received from the neighbors can be helpful to conserve energy, since there is no need to route a similar data. Data similarity may lead a transition to a sleep-dominant location, specially in high density sensor networks, since there is no need to report the same data

7

to the sink from sensor nodes with similar data. The neighbors current duty cycle is important to adjust the sensor node

duty cycle. Every time a transmission occurs or the node duty cycle is changed, a control packet is broadcasted to all neighbors, informing the current node duty cycle and the duty cycle start-up time. Although the transmissions of control packets are increased, the neighbors duty cycle knowledge may be useful to avoid wasteful communication sources, since the probability of a successful communication operation is also increased.

Transmissions should be performed when there is a higher probability of a receiver in listening mode. A transmit-dominant duty cycle may need the receiver in a receive-dominant duty cycle. However, there is no need for a receive-dominant duty cycle if the neighbors remain in a low transmitting rate. A node aware of its neighbors duty cycles is able to exploit sleep modes, when the traffic is low, and wake up in a more appropriated time.

3. Performance Evaluation

In order to evaluate the energy savings and the communication impacts of the new DPM technique , we simulate it and compare it to the ideal DPM model (represents the best and not realistic DPM model) and a communication model without a DPM scheme. Since the application-driven DPM works on a higher layer than the communication process, we evaluate the communication impacts of the DPM using three kinds of routing approaches3:(i) the classical flooding algorithm used by a sensor node to broadcast data to all its neighbors, (ii) the EF-Dee (Earliest-First Tree) algorithm that builds and keeps a data dissemination tree for the entire network and (iii) the SID (Source-Initiated Dissemination) algorithm which the reactive data dissemination process starts from the data source. The data delivery quality is determined by the environment fire probability, to simulate the behavior of a fire detection application modeled in4.

In terms of energy consumption (Figure 2(a)), the DPM technique indicates a better performance over the EF-Tree (no DPM), independently of the routing approach. The expected environment behaviors, represented by the lower fire probabilities, require lower data delivery rates. In this situation, the DPM hybrid automata is able to take a transition to a sleep- dominant location (duty cycle), conserving energy. Otherwise, unexpected environment behaviors, represented by higher fire probabilities, force the leverage of transmit-dominant duty cycles by the PM policy, increasing the data delivery rate and spending more energy. The need for communication

8

decision is performed by the analysis of the environment sensed information, resulting on the use of the most appropriated duty cycle for each fire probability. Since the EF-Tree (no DPM) is not aware of the application-level information, the energy consumption is not related to the fire probability.

62 ............... .. * a U 4O " " ' . "

10 XI so 40 50 6o 70 60 80 im Fin Pmb.blnt#in)

............. ......... ........................ C'

(a) Energy consumption vs. Fire prob. (b) Data delivery rate vs. Fire prob.

Figure 2. Energy consumption and Data delivery rate, according to the fire probability.

In terms of data delivery quality (QoS), Figure 2(b) shows the expected behavior of the DPM technique that selectively changes the data delivery rate, according to the fire probability and to the required application data delivery quality. The EF-Tree (no DPM) model represents the highest data delivery rate and the ideal model represents the necessary data delivery rate, according to the fire probability. It is clear to see that much energy is wasted when no necessary data delivery rates are performed. Although the DPM technique results on a lower data delivery rate, even for the maximum fire probability, much energy can be saved by trading off data delivery quality and power conservation, according to the environment behavior. Since the fire probability may remain low for hours or days, low data delivery rates are required, saving energy and improving the sensor lifetime.

The analysis of the sensed neighbors information is important to identify data similarity among sensor nodes, avoiding redundant data delivery to the sink and conserving energy. Figure 2(a) shows that the DPM technique conserves much energy due to data similarity, specially for lower fire probabilities, when the environment temperature tends to be more similar.

The leverage of the current neighbors duty cycle is also important to avoid wasteful communication sources. A sensor node that receives its neighbors current duty cycle, is able to evaluate the most appropriated moment to perform transmissions, when there is a higher probability of

9

... .....,.,.,,.,. * ............................................................... .... ................. ......... lM t 1m c 4 m . , . , I , , , m . I . . . I , .

in 20 so 11O 50 60 70 w 82 Im 10 20 JO 110 m 60 70 en so iw nm PmbablllW) nR pmbablri%)

(a) Using environment data. (b) Using envir. and neighborhood data.

Figure 3. Idle listening time, according to the fire probability,

listening neighbors, and it is able to evaluate the most appropriated moment to search for packets, avoiding long periods of idle listening.

We also evaluate the impact of turning off the radio over the data dissemination process, since the DPM technique works independent of the communication protocols. Figure 2 shows the DPM performance when environment and neighborhood data are used on the duty cycle change decision. The SID (DPM) seems to be more adaptive to the network topology changes resulted by the turning off the radio process, although the EF-Tree (DPM) is worse, since the DPM technique may cause network disconnections by turning off an important node in the routing path.

The SID (DPM) and the Flooding (DPM) save more energy, since more information about the network is known, due to the broadcast nature of these approaches. A better knowledge about the neighborhood can be used to keep the sensor node in sleep-dominant duty cycles for longer periods . A not pre-determined routing structure is useful to save energy, without losing much data delivery quality. For the EF-Tree (DPM) approach, although the energy consumption is reduced (see Figure 2(a)), the idle listening time remains high, due to disconnections from turning off the radio process.

In fact, although much energy can be saved by a DPM technique, the data delivery quality can be influenced negatively, according to the data dissemination approach. It seems to be a very difficult task to model a DPM technique unaware of the communication routing process, although on-demand routing approaches, like SID, seems to be an alternative.

Figure 3 shows a comparison in terms of idle listening time, for different levels of application-level information leverage by the DPM technique. Figure 3(a) shows that, although the DPM technique conserves energy by

10

choosing duty cycles according to fire probability, it results in the increase of the idle listening time, due to the respective increase of the average sleeping time that requires more idle listening nodes searching for packets to guarantee the communication process. On the other hand, Figure 3(b) shows that the neighborhood current duty cycle information is crucial to avoid wasteful communication sources, like idle listening, although it does not make much difference for the DPM technique using the EF-Tree routing algorithm, due to the fixed tree routing structure, which requires the parent nodes to listen for packets frequently.

4. Conclusion and Future Work

In this work, we propose a new DPM technique that selectively turns off the sensor node radio to avoid useful and wasteful communication factors. We show that the leverage of application-level information in the power management policy seems to be a good way to trade off data delivery quality and energy conservation. Although the performance evaluation shows that much energy can be saved exploiting sleep-dominant duty cycles, our DPM technique does not perform well for all kinds of data dissemination algorithms, since the process of turning off radio does not consider the routing process. In fact, we show that on-demand routing approaches for DPM techniques work better than other routing approaches, due to their adaptability to network changes. Therefore, as future work, we plan to extend the model to incorporate important parts of the routing process to improve the efficiency of the DPM.

References 1. E . Y . CHUNG, L. BENINI, AND G. D. MICHELI, Dynamic power management

using adaptive learning tree, ICCAD, 1999. 2. A. Z. ET AL, Eficient power management in real-time embedded systems,

ETFA'03, (2003). 3. C. F. ET AL, ,Protocolo adaptatiuo hz3rido para disseminaGCo de dados e m

redes de sensores sem fio auto-organita'veis, SBRC, May 2004. 4. R. M. P. ET AL, Dynamic power management in wireless sensor networks:

An application-driven approach, WONS, January 2005. 5. S. C. ET AL, Lifetime analysis of a sensor network with hybrid automata

modelling, WSNA, 2002. 6. A. SINHA AND A. CHANDRAKASAN, Dynamic power management in wireless

sensor networks, IEEE Design & Test of Computers, 18 (2001), pp. 62-74. 7. W. YE, J. HEIDEMANN, AND G. D. ESTRIN, An energy-efficient mac protocol

for wireless sensor networks, New York, NY, June 2002, IEEE INFOCOM.

SDMA IN CONNECTIONS BETWEEN WIRELESS SENSORS AND WIRED NETWORK

VESA HASU, HEIKKI KOIVO

Control Engineering Laboratory, Helsinki University of Technology P.O.Box 5500, 02015 HUT, Finland

Wireless sensor networks are not just stand-alone applications. Wireless sensors need to communicate to wired networks, where monitoring applications are usually connected to. This paper examines performance of SDMA in connections between wireless sensors and a wired network. SDMA is considered in the form of adaptive beamforming in the wired network base station. Interest is laid especially on connections with and without feedback. Performance is examined through simulations in a line-of-sight and a more realistic multi-path environments.

1. Introduction

The purpose of this study is to examine to the usefulness of the space division multiple access (SDMA) in the connection between wireless sensors and a fixed network. Saving the transmission power and reducing the complexity of wireless sensors is a necessity. Lighter soft- and hardware in sensor nodes leads to easier multiplication of sensor network size and thus opens new application fields.

The communication in wireless sensor networks does not always need to be ad-hoc. For example in monitoring applications, wireless sensor nodes need to transmit to the wired network, in where monitoring takes place. Additionally in very large ad-hoc networks, it might be more useful to utilize a wired network connection than a multi-hop route between two distant nodes. Hence examination of efficient strategies between wireless and wired worlds is needed.

SDMA is considered in the form of the receiving beamforming (BF) ([3], [4]) in the wired end of the connection, Compared to the other multiple access (MA) techniques, SDMA does not require accurate scheduling as time division multiple access (TDMA) or as wide frequency band as frequency division multiple access (FDMA). In addition, it is applicable simultaneously with the other MA techniques for better communicational capacity.

2. Link Between Wireless Sensors And Wired Network

What additional value SDMA and BF antenna arrays can bring to the

11

12

communication between wireless sensors and wired network? In addition to the larger number of supported connections, they give a prospect to use connections without feedback. SDMA is considered for the cellular systems in [5].

Wireless sensor networks, which communicate ad-hoc, are obviously able to communicate with feedback to wired stations. In those cases, communication capacity between nodes and bases can be restricted, and hence the use of SDMA is worthwhile. If the wireless nodes are distant to the base or SINR is small, it is advisable to collect the information to a single node from a larger cluster of nodes and then send to the wired base station [6].

The upsides of connections without feedback are lighter hardware and software, leading to smaller battery usage. The lack of feedback enables wireless sensor node constructions without receivers. The advantage of the receiverless node is the smaller requirement for hardware and complex signal processing, and the smaller power consumption. After all, the power consumption of the analoddigital-conversion of receivers is high (for examples, see [l] and [2]).

The downside of the feedbackless connections is inability to use several multiple access and radio resource management techniques, e.g. TDMA, FDMA, power control (PC) and collision avoidance algorithms, and therefore the communication capacity is reduced. Also, SDMA with BF requires a control signal channel for efficient interference rejection in bursty traffic.

3. SDMA - Beamforming

The receiving BF is an antenna array technique, in which the incoming radio waves are spatially filtered based on channel measurements. Antenna array is a set of co-operating antennas. The spatial filtering corresponds to amplifying or attenuating waves based on their directions of arrival to the antenna array. In practice, BF is done by weighting signals of antennas by a suitable combination of complex-valued weights.

An example of a BF gain lobe is drawn in Fig. 1. One characteristic of BF systems are so called null gain hections, in which directions the incoming radio waves are faded out completely. If number of antenna elements in the array is nu, each antenna array weighting has nu - 1 null gain directions [3]. This means that nu simultaneous connections can be supported through BF, if transmitters are in different directions in respect to the receiver and BF is done appropriately. The null gain steering is the essence in SDMA. A successful steering cancels interference and MA is provided for.

Several BF algorithms are presented in the literature for antenna weights determination [3], [4]. Some of them rely on channel measurements, e.g. signal covariance matrices or steering vectors, and some of them are based on training sequences, i.e. known bits attached to transmissions to study channel conditions.

13

The main algorithm used in this work is maximum signal-to-inte$erence- und-noise (SINR) beamforming algorithm [4], which relies on the channel measurements by signal covariance matrices. The key advantage of maximum- SINR beamforming (MSBF) is, by definition, ability to maximize SINR-levels.

While the interference reduction is vital in wireless communications, it is especially important in feedbackless systems. For SDMA systems, the reduction correspond to an accurate steering of null gains towards interferers. In practice, BF techniques relying on the training sequences are more practical to implement than the SINR-beamformer [3], especially in the bursty traffic. In addition, a control signal is also needed for perfect null gain orientation.

null gain, directed towards interferer I

beamforming lobe, Le. the gain level in dfiertnt spatial directions

0 = transmitter

0

Fig 1 : An example of beamforming gain.

4. Simulation Study Of Line-Of-Sight Connections

While connections without feedback and using SDMA are considered, one essential question is: how often unacceptable signal conditions are expected to occur? The following simulation results examined the effect of transmission power level and adaptive antenna specifications in a BF system.

4.1 Simulation Specifications

Simulations included transmitting wireless sensors and a BF base station. The sensor nodes were considered to be in randomly spread fixed locations in a space with radius r. The base station antenna array had two configurations: four elements in tetrahedron or six elements in triangle with half-wavelength spacing. The base station was set outside the node cluster to distance Y. Geometrical specifications simulate, for example, industrial processes, in which system measurements are sent to a monitoring system. It would be more advantageous for SDMA system )o have more spatial distribution, i.e. the base would be in the middle of the sensor nodes. On the hand, the near/far effect could become significant in those cases. The system used MSBF. This meant that there must be a control channel fiom wireless nodes to the base station in addition to the communication channel.

Signal path losses were modeled with the exponent two. The relative noise

14

power in the system was -80 dB throughout simulations. The SINR target level for connections was 7 dB, which corresponds to the bit-error-rate (BER) of 10" using coherent binary phase-shift-keying-modulation [9].

Since the usefulness of SDMA in feedbackless communication with random transmission times and combined to TDMA was examined, SINR levels of all combinations with 1-8 simultaneously transmitting nodes were simulated, and probabilities of supported connections were determined. As a definition, a connection was supported, if SINR was larger than the SINR target. These results are good for comparing support probabilities in feedbackless connections and in SDMA/TDMA with multiple simultaneous transmitters.

4.2 Results

Fig. 2a shows support probabilities in systems with 15 randomly spread nodes and triangular antenna array with six elements. Fig. 2b presents support probabilities of a system with a tetrahedrical four element antenna array. In order to guarantee spatial diversity, results were averaged over 100 simulations.

Figs. 2 show that the BF theory holds quite well in cases of high transmitting power: n, simultaneous connections are supported, where n, is the number of antenna array elements. On the other hand, n, connections are not always supported. If transmitting nodes are in the same direction, the spatial resolution of the antenna array is not accurate enough and signal can not be separated from the interfering signal. In Figs. 2, the near/far effect of wireless communications can be seen as lower support percentages in cases with lower transmission powers. If Fig. 2a and Fig 2b are compared, the transmission power level is much more significant factor in the case with fewer antenna elements.

O L " " ' 1 J 1 2 3 4 5 6 7 8

LOO

90

80

70

60

50

40

30

20

10

0 1 2 3 4 5 6 7 8

Number of simultaneous transmitters Number of simuhanwus transmitters a) b)

Fig. 2: Support probabilities with different transmission power levels. System includes 15 randomly spread nodes and triangular antenna may with a) six elements, b) four elements.

15

Support probabilities varied a lot over 100 simulations. As an example, confidence limits for probabilities of the four antenna element array case are given in Table 1. It shows that in addition to average performance reduction, low transmission power adds dependency on the spatial distribution.

These results inhcate that SDMA enables the feedbackless communication between wireless sensor nodes and wired base stations, if the control channel can be utilized well enough, e.g. BF can direct nulls. Performance limitations are related to the number of elements in the array nu, as the BF theory suggests. If nodes transmit with random intervals and the average number of simultaneous connections is lower than nu, SDMA approach guarantees a good throughput, Table 1: 95% confidence intervals of the percentages of the supported connections in cases with different transmission powers and simultaneous transmitters.

-20 dB -30 dB -40dB -50 dB [99.8%, lOO?’”,] [99.7%, loo%] [98.9%, loo%] [94.3%, loo%] [77.1%, 92.0%] [99.6%, lOO?’o] [99.1%, 100%] [95.8%, lOO%] [81.9%, 99.9%] [44.8%, 70.1%] [98.9%, lOO%] [95.90/,, 100%] [83.4%, 99.5%] [55.8%, 83.3%] [18.0%, 35.1%] [51.0%, 56.0%] [50.8%, 55.6%] [47.4%, 54.0%] [29.5%,46.3%] [4.0%, 16.4%] [18.9%,31.2%] [18.8%,31.0%] [17.9%,30.0%] [10.6%,24.8%] [0%,8.0%]

4.3 Extensions To SDMA/TDMA Systems With Power Control

SDMA can be applied in wireless sensor networks with and without feedback. In the system point of view, the basic difference between the two versions is that feedback enables TDMA and PC. If there are several transmitters and BF is used with PC and some other MA technique, BF is more like a radio resource control technique. This is since in those cases BF is more for the interference rejection than MA. The closed-loop PC performance in S D W T D M A systems is examined in [7]. Even though results in [7] are on cellular systems, results on the blocking rate are applicable in the node-base station case: the blocking rate of the calls is reduced by a decade, while using the closed-loop PC.

Simulations with varying distance between base and nodes with PC were done. The base station included the tetrahedrical antenna array, and the distance from the nodes to the base station was varied. Decentralized PC was used with five iterations. Admission control was not applied. While more transmitters are on, PC problem becomes easily unfeasible without the admission control, and none of the connections can be supported in the same timeslot. Successful admission control would guarantee at least nu supported connections all the time.

Simulations showed that almost all connections could be supported in all cases with at most nu, i.e. four, nodes were transmitting. This was expected for SDWTDMA: a successful combination of PC enables supporting nu transmissions in each TDMA time slot. In addition, transmission power levels

16

were reduced radically, see Table 2. Further, PC adds robustness against the radio channel variations, which was not examined here. Table 2: Average transmission powers in dBs of different distances between the base and nodes, and different numbers of simultaneous transmitters.

Transmitters I r 2r 5r 1 Or 25r 50r 1 OOr 2 I -52 -48 -41 -3 6 -27 -2 1 -15 3 -47 4 2 -32 -29 -19 -13 -10 4 I -39 -33 -24 -20 -12 -2.8 2.1

5. A Case Simulation Study Of Wireless Nodes In An Industrial Hall

Additional simulations in an industrial hall environment were made. These give insight about the efficiency of SDMA, while lines-of-sight (LOS) are not guaranteed, and reflected non-line-of-sight (NLOS) signals are present.

5.1 Simulation Specifications

The simulations were similar to Section 4 with the following exceptions. The communication is talung place in a hall with four reflected components as shown in Fig. 3. The rest of reflections were assumed to be negligible due to shadowing by machinery etc. Reflections were set to cause -7 dB attenuation and a random shadowing caused -10 dI3 attenuation. The shadowing was added to LOS and NLOS paths with probability of 0.1 and 0.4, respectively.

Simulations included two base station configurations (four or six element arrays) and three different PC-BF-versions: MSBF with constant relative transmission power of -10 dB, MSBF with PC, and maximum gain beamforming (MGBF) with PC. MGBF weight vector is the principal eigenvector of the signal covariance matrix as. MGBF was applied to obtain information about the SDMA efficiency, if gain nulls are not directed towards interferers.

In order to guarantee as large support rate as possible in cases with PC, a simple admission control was also implemented. The admission control denied transmissions from nodes experiencing too poor channel and interference conditions. If the admission control was not implemented, transmission powers were high and support percentages were still small.

Fig. 3: A sketch of the multipath reflections in the industrial hall environment.

17

2 3 4 5

5.2 Results

-10 -17.1 -9.8 -10 -16.8 -12.2 -10 -8.7 -13.8 -10 -10.1 -14.6 -10 -10.3 -18.2 -10 -7.2 -19.4 -10 -1 1.8 -17.6 -10 -7.6 -18.1

Support percentages of different four and six element antenna arrays and PC-BF -configurations are shown in Fig. 4. Average transmission powers of these cases are collected in Table 3.

Figure 5 shows that MGBF is not clearly good enough for MA technique. MGBF does not direct its nulls as does MSBF. Additionally, as expected, PC and more antenna elements clearly increase the percentages of the supported connections. If support probabilities are compared to the ones of LOS connections in Fig. 2, the decreased performance due to reflected signals and shadowing is evident. Multi-path components and shadowing lead to more scattered radio environment, and BF becomes much harder task. The cure for severe multipath conditions is to increase the number of antenna elements n,.

Average transmission powers in Table 3 indicate that PC combined with admission control can save transmission power. Although, Table 3 reveals that occasionally the situation is vice versa. Increases in transmission powers of some cases in Table 3 are consequence of the admission control, and they should be interpreted in respect to support percentages in Fig. 4. Even though the system with MGBF and PC uses more power than the system without the PC in some cases, it has clear benefits in percentages of supported connections.

AGBF+PC I

100

90

80

70

60

x 50

40

30

20

10

- 6-ekm.. MSBF+PC

1 2 3 4 5 6 1 Number of simuheous transmitters

Fig. 4: Support probabilities over 100 cases with base station configurations and power control. System included 15 randomly spread nodes.

Table 3: Transmission powers in dBs of configurations, and different numbers of simultaneous transmitters.

~ ~ ~ ~ ~ _ _ _ _ _ _ _

4 elem., 4 elem., 4 elem., 6 elem., 6 elem., 6 elem., I MSBF MSBF+PC MGBF+PC MSBF MSBF+PC MGBF+PC Transmitters

18

6. Conclusions

All in all, SDMA is a promising technique for node access to the wired network. The greatest advantage can be received by a S D W T D M A combination. The possibility of supporting more connections to the wired network with feedback channel is obvious based on the literature [7], and it is also confirmed by the simulations in both line-of-sight and multipath environments.

The feedbackless communication between wireless and wired networks seems to be possible using SDMA. This would make possible to make smaller, cheaper and simpler hardware for wireless sensor nodes. However, even though SDMA systems have a lot of promise in lightening the configuration of sensor nodes, it must be taken into account that the effective use of BF and null steering is necessary. Moreover, the effect of many multipath signals to the performance must be considered properly in the system design. A well-planned application of control channel, non-moving nodes and fairly stable radio channel are required for transmissions without feedback.

References

1.

2.

3.

4. 5.

6.

7.

8.

9.

J. Sevenhans, Z.-Y. Chang, “ A D and DIA conversion for telecommunication”, IEEE Circuits Devic., 14(1): 32 - 42, 1998. A. Anttonen, T. Rautio, “Performance and complexity analysis for adaptive sample rate converten in GSM/UMTS/HIF’ERLAN2 mobile transceiver”, IEEE Int. Symp. on Circuits and Systems, ISCAS, vo1.4, pp. IV-489 - IV-492,2002. J.C. Liberti, T.S. Rappaport, “Smart Antennas for Wireless Communica- tion”, Prentice Hall NJ, 1995. R.A. Monzingo, T.W. Miller, Introduction to Adaptive Arrays, Wiley, 1980. W.J. Huang, J.F. Doherty, “An Evaluation of Blocking Probability for Three- fold SDMA”, Proc. Military Comm. ConJ, vol. 2, pp. 1248-1252,2001. D. Nicholson, C.M. Lloyd, S.J. Julier, J.K. Uhlmann, “Scalable Distributed Data Fusion”, Proc. F$h Int. Con$ Inform. Fusion, ISIF, vol. 1, pp. 630-635,2002. J. Yunjian, S . Hara, Y. Hara, ”Impact of Closed-loop Power Control on SDMNTDMA System Performance”, Proc. 56th Vehicular Tech. Con$ Fall-2002, vol. 3, pp. 1825-1829,2002. J. Yunjian, S. Hara, Y. Hara, ”Performance of T D D - S D W M A System with Multi-Slot Assignment m Asymmetric Tmffic Wireless Network”, Proc. 13th IEEE Symp. Personal, Indoor, andMobile Radio Comm., vol. 5 , pp. 2317-2321,2002. S. Haykin, Digital Communications, Wiley, 200 1.

10. F. Rashid-Farrokhi, L. Tassiulas, K.J.R. Liu, “Joint Optimal Power Control and Beamforming in Wireless Networks Using Antenna Arrays,” IEEE T. Comm., 46(10): 1313-1324,1998.

A RELIABLE AND ENERGY-EFFICIENT ROUTING PROTOCOL FOR WIRELESS SENSOR NETWORKS

PETER, KOK KEONG, LOH School of Computer Engineering, Nanyang Technological University, Nanyang Avenue,

Singapore 639798

SAY W A N , LONG School of Computer Engineering, Nanyang Technological University, Nanyang Avenue,

Singapore 639798

YI, PAN Department of Computer Science, Georgia State University, 34 Peachtree Street, Suite

I450 Atlanta, GA 30302-4110, USA

In wireless sensor networks, nodes generate important data that are of interest to the network users. These data have to be delivered reliably to the hub in the network which will then be disseminated to the network users. Routing protocols for such networks have to provide reliable delivery of data using minimal energy as nodes are energy- constrained. We propose EAR; a novel routing protocol that achieves reliability and energy efficiency in data delivery.

1. Introduction

With the introduction of wireless sensor networks (WSNs) [l-21, many applications have been conceptualized to leverage on this emerging technology. Some examples include battlefield surveillance, large-scale industrial and environmental monitoring. All these applications require data collected by the nodes to be delivered to some sink nodes also known as hubs which will then disseminate the data to the network users. A routing protocol that can provide reliable delivery of data is essential.

Achieving reliability is not an easy task in WSNs because of high node failure rate and high packet loss. The nodes are battery-operated and therefore have limited energy resource. Routing protocols designed for WSNs have to achieve reliability while expending minimum amount of energy to maximize the network lifetime.

19

20

With this in mind, we propose a novel routing protocol called EAR that provides robust data delivery from the nodes to the hubs and also achieve energy-efficiency in WSNs.

The remainder of this paper is organized as follows. Section 2 describes our proposed routing protocol in details. Simulation results are presented and discussed in Section 3. Finally, Section 4 concludes this paper.

2. Protocol Design Details

2.1. Overview

In EAR, nodes generate Report (RPT) packet that contains information of interest to the network users. The RPT packet is then routed to the hub in the network.

EAR requires a Medium Access Control (MAC) protocol that provides reliable link-to-link transmission. One example is IEEE 802.1 1 [3] MAC protocol that provides reliable link-to-link transmission by using three-way handshaking mechanism. Using such MAC protocol is necessary in a wireless environment because RF links are unreliable and loss of messages is frequent due to message collision. Explicit control messages are therefore needed to detect lost messages.

2.2. Setup Phase

The hub will broadcast an Advertisement (ADV) packet indicating that it is ready to receive RPT packet generated by nodes. When neighboring nodes around the hub receive this ADV packet, it will store the route contained in the ADV packet in their routing table. Nodes will not propagate the ADV packet received.

Every node will back off for a random amount of time before beginning an initialization process. A node begins the initialization process by broadcasting a Route Request (RREQ) packet asking for a route to the hub. When a hub receives a RREQ packet, it will broadcast a Route Reply (RREP) packet containing the route information. Similarly, when a node receives a RREQ packet, it will broadcast a RREP packet containing the route information if it has a route to the hub. Otherwise, it will ignore the RREQ packet. Nodes do not propagate RREQ packet.

When a node receives a RREP packet, it will store the route in its routing table. When it has at least a route to the hub, it skips the initialization process. Therefore, by introducing random delay for each node to begin initialization

21

process, a portion of nodes will have received a RREP packet before they have begun their initialization process. This enables fast propagation of routes and also saves on the amount of control packets generated in the setup phase.

A node will store more than one route to the hub. A route in the routing table is indexed using the next hop node’s ID that is a neighbor of this node. We define Node 1 to be a neighbor of Node 2 and vice versa when they are one hop away from each other. In other words, they are within communication range of each other. A node will only keep one route entry for a neighbor that has a route to the hub. That neighbor could have multiple routes to the hub but it is of no concern to the node because all it needs to know is that this neighbor has a route to the hub so it can forward RPT packet to this neighbor. In the routing table, every entry is uniquely identified by the neighbor’s ID.

2.3. Route Management

In WSNs, nodes have limited memory and therefore the size of the routing table has to be restricted. This leads to the question of how to select the best routes and to only keep the best routes in the routing table at all times. In EAR, two metrics are used to decide the admittance of a route into the routing table.

2.3.1. RouteLength Metric

The primary metric is the number of hops a route needs to reach the hub, which is called RouteLength. The reason for choosing this metric is because the best route is always the shortest route (route with least RouteLength value). It will have the lowest packet latency and expends the least energy.

However, the RF link between a node and each of its neighbors will not be the same because of the difference in physical distance and the type of terrain between them (e.g. two nodes might be obstructed by a tree that attenuates RF signals).

In this situation, the best route is not the shortest route because trying to forward a packet to a neighbor with a shorter path but bad RF link quality will expend more energy in retransmission and also increases packet latency than forwarding to a neighbor with a longer path but with good RF link.

In EAR, the above problem is eliminated by using a concept known as “Route Blacklisting”. Initially, routes are admitted into the routing table using RouteLength as admission criteria to ensure that only the shortest routes are chosen. As RPT packets start to flow through these routes, less desirable ones will start to exhibit high packet error rate and will eventually be blacklisted and omitted from the routing table. Routes that are omitted from the routing table

22

will not be admitted again until after a period of time. Some routes are only affected by temporary external disturbances and so should be given the chance to be re-admitted into the routing table after a period of time.

2.3.2. Routescore Metric

The second metric uses Routescore that is calculated using the formula as shown,

Routescore = (PE x WE + PL X Wd PE - energy level of the next hop node (0.0 to 100.0) WE - assigned weight for PE (0.0 to 1.0) PL - link quality to next hop node (0.0 to 100.0). Uses a sliding window of

WL - assigned weight for PL (0.0 to 1.0) size N gets the average ratio of successes on using this route.

RouteScore takes on a value from 0 to 100 and a higher value indicates a better route. The sum of the weights (WE and WL) must be equal to 1.

2.3.3. Route Replacement

When a RREP packet is received, a check is done on the route carried on the RREP packet. If the route is a blacklisted route, it is ignored. If it is an existing route, then the route information is updated. If it is a new route and the routing table is not fill, then it is stored in the routing table. If it is a new route and the routing table is fill, a route replacement strategy is carried out.

In the replacement algorithm, the first step is to search for the worst route with the largest RouteLength in the routing table. In the event of a tie, the route with the lowest Routescore is chosen. In the second step, the worst route is then compared against the incoming route and the path with lower RouteLength is admitted into the routing table. If there is a tie, then the route with the higher Routescore is admitted.

To calculate the Routescore for the incoming route, an arbitrary value has to be assigned to PL as the link quality is unknown. Assuming the PE of the worst route and the incoming route are the same. The factor in deciding if the incoming route is to be admitted will depend on the value assigned to PL for the incoming route.

2.4. Data Dissemination

After the setup phase, every node in the network will have at least one route to the hub. Depending on the application, nodes will either start generating RF'T

23

packets at periodic intervals or go into idle mode waiting for some event to happen before generating RF'T packets.

When a RPT packet is generated at a source node, it cames two fields that are used by EAR in its header; ExpPathLen and NumHopTraversed. The first field defines the expected number of hops this packet will have to traverse before it reaches the hub. It is initialized according to the formula,

ExpPathLen = NH x a NH is the number of hops from this node to the hub for the route selected to

forward this packet. a is some weight assigned from 0.0 to 1.0. NumHopTraversed records the number of hops a packet has traversed so far and is initialized to 0.

The packet is then queued in the output buffer before being forwarded to the next hop in the route. When the next node receives the packet, it will increment NumHopTraversed by one and then compare it with ExpPathLen that is never altered after initialization.

If ExpPathLen is larger than NumHopTraversed, the routing mechanism will choose a route with the highest RouteScore. Should there be a tie in the RouteScore; the route with the lowest RouteLength is chosen. By assigning a > 0, a packet can take a longer route with better link quality, assuming RouteScore is determined by link quality alone.

If ExpPathLen is smaller than or equal to NumHopTraversed, a simple route selection mechanism requiring only two comparisons is used. Firstly, select route with the lowest RouteLength. If there is a tie, select the route with the highest RouteScore. The logic is that if the number of hops a packet has traversed exceeds the expected number of hops, there must be some changes in the network topology due to node failure or environmental noise affecting the RF communication. During this transient period of instability where routing table undergo changes, the packet will take the shortest route to the hub. The same routing mechanism is used at each intermediate node until the packet reaches the hub.

2.5. Route Update

Nodes in WSNs may be prone to failure and the unpredictable RF link quality between neighboring nodes changes frequently causing the network topology to change with time. Also, node energy levels will decrease according to the amount of data packets they receive for routing. Hence, nodes need to maintain updated and fresh routes in the routing table at all times. EAR uses a novel solution that provides rapid updating of route at negligible cost.

24

The solution uses the handshaking messages used by IEEE 802.1 1 MAC protocol. When node I wants to send a data packet to node 2, node I sends an RTS packet and when node 2 receives the RTS packet, it will send a CTS packet to node 1. The idea is to piggyback route infonnation on both RTS and CTS packet. This enables the neighbors of both node I and node 2 to obtain the latest route information of the two nodes.

3. Simulation

We implemented EAR on GloMoSim [4] and compared it against GRAB [5]. We used GloMoSim to simulate a WSN where the nodes are referenced after crossbow MICA2 mote [6] which is a popular hardware platform for WSNs. The specification together with other settings for the simulator is shown in Table I .

We use the following metrics to measure the performance of the routing protocols. Packet delivery ratio (PDR): The percentage of data packets successfully routed to the hub. Packet latency: The average time it takes for a packet to be routed to the hub. Total Energy Consumption: The total energy expended by all nodes.

We simulated a scenario where a noisy wireless medium is compounded with node failure. A noise model where every node in the network except the hub takes on a random noise factor between 10% and 50% was used. The noise factor of a node indicates the probability that packets received by the node are assumed to be corrupted or lost in transmission. A fault model where 50% of the nodes in the network fail at a random time within the simulation duration was used. We placed a source node generating packet at a rate of 1 packet every 10 seconds at one corner of the terrain area and a hub at the opposite comer. To study the scalability of the routing protocols, we conducted the test from 50 nodes to 500 nodes at 50-node interval. As the number of node increased, we increased the terrain size proportionately so to keep the node density constant.

Table 1: Simulator

25

The nodes are uniformly distributed in the terrain. We used a different seed for each run of the test and the results are averaged over 30 runs.

3.1. Results

The results are shown in Figure 1. EAR has higher PDR and lower packet latency than GRAB while expending lesser energy. GRAB uses broadcast to forward packets, the number of redundant packets generated are high. This results in high bandwidth utilization which increases the rate of packet collision and also incurs higher packet latency. The redundant packets also consumed additional energy. Although GRAB attempts to achieve reliability through forwarding redundant packets on a mesh but using broadcast does not allow a node to know if a packet has been successfully delivered to the next hop. If a packet is lost due to corruption after being broadcasted by the originating node, then that packet is lost for good as the originating node doesn't know of the failure and none of the neighbors will receive the packet.

I Packet Delivery Ratio (PDR) I 100

* 95

L 8 0 5

i m ; 75 70

6 865

€0 0 103 m 3M) 403 5x

Nurnhr of mdes

Packet Latency

26

~___

Energy Consumption

5wo- , 5

t 4-- 3 3 m -

1 L. lo00 -

t /

0 100 200 300 400

Numbr of nodas

Figure 1: Simulation results

4. Conclusions

We have proposed EAR; a reliable routing protocol for WSNs that uses novel techniques to achieve high reliability, low packet latency and low energy consumption compared to GRAB which is another routing protocol designed for reliability and robustness. Encouraged by current results, we will continue to improve the performance of EAR.

References

1. Culler D., Estrin D., Srivastava M., “Guest Editors’ Introduction: Overview of Sensor Networks” IEEE Computer Magazine, vol. 37, pp. 41-49, August 2004

2. Akyildiz I. F., Cayirci E., Sankarasubramaniam Y., Su W., “A Survey on Sensor Networks” IEEE Communications Magazine, vol. 40, pp. 102-1 14, August 2002.

3. LAN MAN Standards Committee of the IEEE Computer Society, “Wireless LANmedium access control (MAC) and physical layer (PHY) specification”, IEEE, New York, NY, USA, IEEE Std 802.1 1-1997 edition, 1997.

4. Ahuja R., Bagrodia R., Bajaj L., Gerla M., Takai M., “GloMoSim: A Scalable Network Simulation Environment”, Technical report 990027, UCLA, Computer Science Department,

5. Lu S., Ye F., Zhang L., Zhong G., “GRAdient Broadcast: A Robust Data Delivery Protocol for Large Scale Sensor Networks” ACM Wireless Networks (WINET), Vol. 11, No.2, March 2005.

6 . CrossBow MICA 2 motes specification http://www.xbow.com

ANALYSIS OF COVERAGE AND CONNECTIVITY IN

WIRELESS AD HOC SENSOR NETWORKS*

JIANXIN WANG, LUPENG WANG

School of Information Science and Engineering, Central South University

Changsha, 410083, China

RENYI XIAO

National Nature Science Foundation of China, Beijing, 100083, China

Coverage and connectivity are two fundamental problems in sensor networks and answer

the questions about the quality of service that can be provided by a particular sensor

network. We consider a new wireless sensor networks model that assumes each node can

just connect with the other nodes that locate at the neighbor cells. And then we provide a

mathematic analysis of the fundamental relationship among cells number, nodes number

and the probability of coverage and connectivity. The analysis shows that the coverage is

related to the number of sensor nodes and the size of the sensing area, the connectivity is

also decided by the network topology besides the above two factors. The analysis results

give underlying insights for treating coverage and connectivity in a unified framework.

1. Introduction

Recently, the concept of wireless sensor networks has attracted a great deal of research attention due to its wide-range of potential applications [ 11. A wireless sensor network consists of tiny sensing devices, deployed in a region of interest. Each device has processing and communication capabilities and the base station aggregates and analyzes the messages received [2].

Since sensors may be spread in an arbitrary manner, two issues are quite important in the deployment of wireless sensor networks. They are [3]:

Coverage: One goal of a sensor network is that each location in the physical space of interest should be within the sensing range of at least one of the sensors.

* The work is supported by the National Natural Science Foundation of China under Grant No. 90304010

27

28

Connectivity: To aggregate all the sensor data, the sensors need to organize themselves into a connected ad-hoc network. The location and placement of the sensors determine the connectivity of the sensor network.

The problem of sensing coverage has been investigated extensively. Some papers analyzed the coverage problem and developed corresponding algorithms, but they didn’t consider or simplified the connectivity problem [4-81.

Several other protocols aim to maintain network connectivity, but do not guarantee sensing coverage, such as ASCENT [9], and GAF [lo].

There are also many works that consider both coverage and connectivity problem. Ref. [ 113 presents the design and analysis of protocols to achieve guaranteed degrees of coverage and connectivity. In [12], the authors at first proved a fhdamental result and based on this, they introduced a density control algorithm named OGDC. In [ 131, the authors consider different sensing ranges and give an extended result of [ 121.

In this paper, we at first define a new model for coverage and connectivity. Then we analyze coverage and connectivity problem and investigate the relationship between coverage and connectivity. The analysis results give underlying insights for treating coverage and connectivity in a unified framework and can be used in the design of corresponding protocol for coverage and connectivity in wireless sensor networks.

2. Network Model and Problem Statement

In this paper, we use the following cell partitioned network model. As shown in Fig.1 (a), the network area A is partitioned into C non-overlapping identical hexagonal cells. There are many sensor nodes in area A that are placed randomly and independently. Each node can just communicate directly with the nodes that locate at the same cell and neighbor cells.

(a) Cell partitioned network Fig.] Cell partitioned nc rork

(b) the corresponding graph (a) . and its corresponding graph

We provide a mathematic analysis of the relationship among cells number, nodes number and the probability of coverage and connectivity. The analysis is based on the basic characters of sensor networks and can be extended to analyze

29

other practical sensor networks. We mainly discuss the following problems that are important in the applications of such network model.

(1) How much is the probability PoccupyaN(h? that all the cells are occupied by sensor nodes when there are N nodes in all?

(2) How many sensor nodes Noccupya@th) are required to get the probability PoccUpya/~Noccupya/~pth)) lager than the threshold pth?

(3) Suppose the probability of an independent cell being occupied is P, that is to say, for any one arbitrary given cell in wireless sensor network, the probability that there is at least one sensor node is P, what is the probability Pconnlp(P) that all the sensor nodes in area A are connected?

(4) What is the probability Pconn(N) that N nodes that are distributed in C cells randomly can be connected?

3. The Proposed Solutions

3.1. Coverage: Poccupy4~dN) and Noccupy4~~(Pt~

Lemma 1. Suppose to put n sensor nodes into k different cells ( n 2 k ), the number of the different ways to ensure that each cell has at least one node is given by

Proof: Let S be a set consist of all k" different ways of n sensor nodes placing in k

cells ( n 2 k ). Supposing s E S , s has the property ai if the ith ( 1 I i I k ) cell

does not have a node. Let K ( ~ ~ , ~ , ~ ... a . )be the number of the elements in S, 4%

which have all the properties

number of different ways of n nodes placing in the k-m cells, that is@-m)" .

S that has none of the properties ail , ai2 ,. . . ,aim is

,..., a i m , then K ( ~ , , ~ , ~ . . . ~ ~ ~ ) equals to the

In terms of the inclusion-exclusion principle, the number of the elements in

k k

Fk(n)=IS I + z ( - l ) " x K ( a i , a i , . . .a im)= k" + x ( - l ) " m=l ISi,<i, c.4, Sk m=l

k

= f!(-l)m(L)(k-m)n = x ( - l ) k - m m=l

m=O

30

Theorem 1. Placing N (WC) sensor nodes into C cells of area A randomly, the probability of all cells being occupied is given as following:

The proof of this theorem uses the meaning of Poccupyo~~(N) and Lemmal, and is omitted for brevity.

Fig. 2 Poccupyall(N) with different number of sensor node The number of sensor nodes N o c c u p y o l ~ ~ ~ h ) in problem (2) can be solved

using theorem 1 through numeric algorithm. Fig.2 shows the relationship between the number of nodes N and the

probability Poccupy~N(If), which indicate the coverage in wireless sensor network. If the probability P[h is 0.98, Noc~pyoN(P,~) must be larger than 215 when C=30.

3.2. The probability Q(k) of k-cells Connection

Definition 1. Assume that an area A is partitioned into C non-overlapping hexagonal cells and there is at least one sensor node in any of cells. The probability Qfi) of k-cells connection is all the sensor nodes just in k cells which are selected randomly from the C cells can communicate with each other without the help of the sensor nodes in the other (C-k) cells.

If we denote each cell by a vertex and for any pair of cells x andy, join the corresponding pair of vertices by an edge if the cells are adjacent, the resulting structure is a graph. We denote the graph whose vertex set is V and whose edge set is E by G(V,E). As shown in Fig.1 @), each vertex in Fig.1 (b) is for each cell in Fig. 1 (a) and each link indicates the neighbor cells.

The computation of Qfi) when k is an arbitrary integer is a difficult problem, which is needed to compute in many researching fields.

31

In the following, we introduce a new concept k-connected induced subgraph.

Definition 2. Consider a graph G(V,E), let V’ be a non-empty subset of V and E’ be a non-empty subset of E. Define the graph G‘(V’,E’) as a induced subgraph of G if V’ contains all vertices of edges in El.

If IV’I=k, G’ is also called k-induced subgraph of G. We denote the set of all k-induced subgraphs of G by k-isub(G).

Definition 3. If G’(V’,E’)is connected and also a induced subgraph of G, G’ can be called a connected induced subgraph of G. If I V’I=k, G ’ can also be called k-connected induced subgraph of G. We denote the set of all k-connected induced subgraphs of G by k-cisub(G).

I k - isub(q I# 0 , the probability Q(k) of k-cells connection can be given by

Theorem 2. For a Graph G( V,E), when k 11 V I and

I k - cisub (G) I Q ( k ) = I k - isub (G) I (3)

It is difficult to get a closed formula for Ik-cisub(G)I. However, we can get useful approximate results by using the compute algorithms and also you will see that the definition of Qfi) can simplify our analysis.

There is an internal relationship between subgraph count and k-combination count, and we can say the former is more general than the latter. When G( V,E) is a complete graph, there is no difference between k-combination count, k-induced subgraphs count and k-connected induced sugraphs count.

4.0-

0.6 - as -

Q a4:

0.2 -

0.0 -

Fig.3 Q(k) with different number of vertices

As shown in Fig.3, the probability Qfi) of k-cells connection is increasing with the increase of the number of vertices and the decrease of the number C of cells. But there is an exception when k is very small. The value of Qfi) is related to C, k and the topology of all the cells.

32

3.3. Connectivity and Coverage Pconnk@)

Theorem 3. Suppose the probability of an independent cell being occupied is P, the probability that all sensor nodes in these cells can be connected is

The proof of this theorem uses Bernoulli trials, and is omitted for brevity. Fig.4 shows the relationship between coverage and connectivity where the

cells partition is 5 X 6. Basically, the connectivity is increasing with the increase of coverage. But there is an exception when P is very small. When the probability of coverage is 0.8, the probability of connectivity is 0.84965.

Fig.4 Connectivity and coverage PcOnob(p)

3.4. Connectivity Pconn(N)

Lemma 2. Placing N sensor nodes into Area A randomly, the probability that these N sensor nodes are located in k cells and can be connected with each other is

According to Lemma 1,we can get the proof easily. Theorem 4. If N sensor nodes are distributed into area A randomly and

independently, the probability POn,(n) that any pair of sensor nodes is connected is

33

1.0-

0.8-

0.6-

Z ' *! 0.4 -

0.2-

0.0,

N I C )

N > C )

Theorem 4 can be proved directly by using Lemma 2, so that the proof is omitted for brevity.

1 0 50 1W

N(lk tmdm of gnsm mds)

Fig3 Connectivity with different number of sensor nodes Fig.5 shows the relationship between the probability of connectivity and the

number of sensor nodes. Basically, the probability Pc,,,,,,(N) of connectivity is increasing with the increase of the number of sensor nodes. But there is an exception when the number of nodes is very small. When the number of sensor nodes is 60 and C equals to 30, the probability of connectivity is 0.95327.

4. Conclusion

In this paper, we defined a new network model and analyzed the fundamental relationship among cells number, nodes number and the probability of coverage and connectivity and exact expressions were derived. The analysis shows that the coverage is related to the number of sensor nodes and the size of the sensing area, the connectivity is also decided by the network topology besides the above two factors. This analysis results gives underlying insights for treating coverage and connectivity in a unified framework.

References

1. I. F. Akyildiz, W. Su, Y.Sankarasubramaniam, E.cayirci. Wireless Sensor Networks: A Survey. Computer Networks, March 2002; 393-422

34

2.

3.

4.

5 .

6.

7.

8.

9.

D.Tian, N.D.Georganas. Connectivity Maintenance on Coverage Preservation in Wireless Sensor Networks. In proceedings of CCECE 2004, May 2-5,2004, Niagara Falls, Canada. K. Kar, S . Banerjee, Node Placement for Connected Coverage in Sensor Networks. Proceedings of WiOpt 2003, Sophia-Antipolis, France, March 2003. K. Chakrabarty, S . S . Iyengar, H. Qi, E. Cho. Grid coverage for surveillance and target location in distributed sensor networks, IEEE Transactions on Computers, 51(12): 1448-1453, December 2002. S. Meguerdichian and M. Potkonjak. Low Power 011 Coverage and Scheduling Techniques in Sensor Networks. UCLA Technical Reports 03000 1. January 2003. S . Meguerdichian, F. Koushanfar, M. Potkonjak, and M. Srivastava, Coverage Problems in Wireless Ad-Hoc Sensor Networks. INFOCOMOl , Vol3, pp. 1380-1387, April 2001. D. Tian and N. D. Georganas. A coverage-preserving node scheduling scheme for large wireless sensor networks. In First ACM International Workshop on Wireless Sensor Networks and Applications, Georgia, GA, 2002. C. Huang and Y. Tseng. The coverage Problem in a Wireless Sensor

Network. WSNAO3, September 19,2003, San Diego, CA. 2003. A. Cerpa and D. Estrin, ASCENT: Adaptive Self-Configuring Sensor Networks Topologies, INFOCOM, June 2002.

10. Y. Xu, J. Heidemann, and D. Estrin, Geography-informed Energy Conservation for Ad Hoc Routing, MobiCom 2001, Rome, Italy, July 16-2 1, 2001.

11. X. Wang, G. Xing, Y. Zhang, C. Lu, R. Pless, and C. Gill, Integrated Coverage and Connectivity Configuration in Wireless Sensor networks, SenSys'03, Los Angeles, CA, November 2003. H. Zhang and J. C. Hou. Maintaining scheme coverage and connectivity in

large sensor networks. Technical report, UIUC, 2003. J. Wu and S . Yang, Coverage and Connectivity in Sensor Networks with

Adjustable Ranges, accepted to appear in 2004 International Workshop on Mobile and Wireless Networking (MWN), Aug. 2004.

12.

13.

CROSS-LAYER’S PARADIGM FEATURES IN MANET: BENEFITS AND CHALLENGES

LAMIA ROMDHANI AND CHRISTIAN BONNET Institut Eurecom

2229 route des CrCtes - BP 193, 06904 Sophia Antipolis, fiance

E-maihomdhani, bonnet0eurecom. fr

Nowadays, the cross-layer design approach, is the most relevant concept in mobile ad-hoc networks which is adopted to solve several open issues. It aims to overcome MANET performance problems by allowing protocols belonging to different layers to cooperate and share network status informations while still maintaining separated layers. The central key of related research studies is what information can be shared and how it used in cross-layer architecture to provide QoS enhancement and enable an efficient resource utilization. In this work, we detail the most coupling features of introducing cross-layer models in mobile ad hoe networks. Then, we discuss the risks and the challenges facing this new architecture.

1. Introduction

Ad hoc networks have many characteristics that meet a lot of node heterogeneity. A fundamental issue in such multihop wireless environments is that network performance can degrade rapidly as the number of hops increases. Major problems to transmit data over available radio channels exist in every layer of the protocol stuck. In one hand, adaptive rate selection, adaptive antenna pattern, adjust power control are issues of the physical layer. In the other hand, the link reliability, the admission control, and the access control to the shared channel are some issues of both routing and MAC layers. Moreover, there are several real-time application requirements that have to be respected in order to provide QoS support and achieve service differentiation.

In the past, a lot of research have been conducted to address these issues separately. One new research direction to optimize data transfer in ad hoc networks is the cross-layer design without respecting the original layered design approach in which each layer operates independently. The layered approach is simple, flexible, and scalable as the case in the Internet, but

37

38

it led to poor performance in ad hoc network even with the optimization applied to the evolved protocols because they are no taking into account network and application constraints. For example, each layer have to react in route failures and collisions in its own way and there are no coupling of different layer informations to meet some parameters in order to address a good coordination of the efforts satisfying as well as possible the application requirements.

As conclusion, the co-operation between layers to enable performance enhancement is very important and useful in wireless ad hoc networks. The global objective of such co-operation is to achieve a reliable communication- on-the-move in highly dynamic environments as well as QoS provisioning.

In this paper, we review the parameters that should be provided by each layer to other layers in order to improve the global performance. In some cases, specific processing should be done by intermediate layers to present the parameters to other layers in a comprehensive and understandable way. Lot of works have been presented in the open literature that introduce several coupling ways and solutions between different communication layers 1,2,4

The remainder of this paper is organized as follow: In Section 2, we discuss the problems of accommodating a good service for each layer in the layered approach going from the physical layer to the application layer. We also identify the most important parameters in each layer to be managed in a cross-layer architecture. Then, we review the most works that have been conducted to study the cross-layer design in mobile ad hoc networks in Section 3. In Section 4, we outline our observations that lead to fix the potential risk of cross-layer design in MANET. Then, we give our recommendations on how a cross-layer architecture should be designed in an efficient and scalable manner in Section 5. We conclude the paper in Section 6.

2. Limitations of Layered Approach in MANET

As it is well known, networks are organized as a series of layers, each one built upon the one below it. The goal of this architecture is to split the network into smaller modules with different functionalities and deal with more manageable design and implementation. The purpose of each layer is to offer certain services to the higher layers, shielding these layers from the details of how the services are implemented. So the advantage behind the layered protocol architecture is to reduce complexity by dividing and

39

conquering approach. This simplicity ensure an easy way to standardize, and to deploy new flexible protocols (easy upgradeable). However, wireless networks don’t come with links. The channel quality changes dynamically. The applications require a minimum of QoS that could not be achieved in such very dynamic capacity networks. Hereafter, we analyze the problems related to each layer and we give an overview of the characteristics and QoS metrics of each layer.

2.1. Limitations Related to Physical Layer7$ Characteristics

The wireless channel varies over time and space and has short-term (or small-scale) memory due to multipath. The channel variation meets the amount of contention, time-varying fading, multi-path, variation of the SNR. Indeed, these variations are caused either due to motion of the wireless device, or due to changes in the surrounding physical environment, and lead to detector errors. This causes bursts of errors to occur during which packets cannot be successfully transmitted on the link. Fast channel variations due to fading are such that states of different channels can asynchronously switch from good to bad within a few milliseconds and vice-versa. Furthermore, very strong forward error correction codes (i.e. very low rates) cannot be used to eliminate errors because this technique leads to reduced spectral efficiency. The techniques that may be used to adapt to rapid SNR changes in wireless links and mobility include: power control, multiuser detection, directional antenna, adaptive modulation and software radio. However, sharing these informations with high layers, has a big benefits on performance as shown in 2 y 7 . For example, characteriz- ing the application requirements help to use the adaptive modulation, the knowledge of channel quality help to avoid useless MAC retransmission..etc.

2.2. Limitations Related to MAC layer’s Characteristics

The main issue at MAC layer is to adress an efficient medium access mechanism to resolve contentions. CSMA-based MAC protocols can yield an efficient operation (under proper loading levels) when the carrier sensing operation is spatially effective. Unfortunately, stations may be geographi- cally located in a manner that induced blocking, leading to masked terminal scenarios. In this case, two major problems have been identified: hidden terminal and exposed terminal conditions. Despite of introducing RTS/CTS handshaking scheme, leading to the Multiple Access Collision Avoidance (MACA) protocol, the MAC layer still suffers from the problems of inter-

40

ference resolution, exposed terminal, efficient medium utilization. Indeed, the optimal strategies of resource sharing issue among different classes of users, still the main challenge also for the FDMA, TDMA techniques.

2.3. Limitations Related to Routing Layer’s Characteristics

Routing protocols for ad hoc networks require to consider the reasons for link failure to improve its performance. It sould be adaptive to cope with the time-varying topology and time-varying network resources. For instance, it is possible that a route that was earlier found to meet certain QoS requirements no longer does so due to the dynamic nature of the topology. In such a case, it is important that the network intelligently adapts the session to its new and changed conditions. So, the goal of QoS routing is to optimize the network resource utilization while satisfying application requirements. Indeed, it is not enough to find a shortest path but also with available resources as battery, bandwidth, and buffer. Note that the factors that can change the topology of an ad-hoc network are: the mobility of nodes, change of power, the MAC layer mechanism because different schedule for the contending nodes, results in different topology, the flow dynamics that flows come and go; if a node has nothing to transmit, its links are gone from the topology, and finally the mode of nodes: sleeping or active mode. If a node goes to a sleeping mode, its links are gone from the topology and hence it can’t participate to route establishment and communication.

2.4. Limitations Related to h n s p o r t Layer’s

TCP combines error control (ARQ), flow control that are not over-running the receiver buffer, and congestion control that is not clogging the network, and not overloading the capacity in the routers. Moreover, TCP enjoys simplicity of control and gains widest acceptance. However, this simplicity of control is at the cost of efficiency loss. TCP is not able to distinguish the presence of congestion in wired networks, mobility, collision in wireless links, and bit errors due to poor quality of wireless links. Single bit error could trigger congestion control mode (TCP getting into slow start phase); even fast retransmit/fast recovery is not effective in coping with packet/bit errors. So, TCP needs to handle delay (RTT) and packet loss statistics that are very different from those in wired networks.

Characteristics

41

2.5. Energy conservation

Some scenarios where an ad hoc network could be used are business associates sharing information during a meeting, military personnel relaying tactical and other types of information in a battlefield, and emergency disaster relief personnel coordinating efforts after a natural disaster such as a hurricane, earthquake or flooding. In fact, in such scenarios, maximizing the network lifetime is a very important deft since recharging battery is very difficult (hard) to do in such conditions. Indeed, the network connectivity is strictly related to the possibility of routing between each node in the network. The energy exhaustion problem leads to network disconnection and resource unavailability problems.

2.6. Limitations Related to Application Layer’s Characteristics

There are some application’s requirements that should be considered in order to maintain as good as possible the performance and offer a minimum service delivery according to their constraints. Indeed, there are time-bounded applications that are sensitive to delay and others require high throughput and/or less packet loss rate. For example audio traffics should reach destinations at most up to 400 sec. The corresponding packets are almost short. They could have the highest priority: minimum waiting time in the queue, and so short medium access time ( e g short contention window size). Moreover, they require short and less congested routes to reach destinations within a short delay. Throughput-constrained applications require less congested routes and available queue to enqueue packets. Hence, successful transmission should be assured and they are less sensitive to delay comparing to above described class. TCP traffcs are very sensitive to both packet loss and delay. Background traffic should not be starved and so a minimum service has to be guaranteed.

The key question is how to adapt physical layer parameters, distribute fairly the access to the medium and achieve an efficient bandwidth sharing while providing service differentiation and application requirements with the less possible complexity?

In the next Section, we describe how these problems and information, related to each layer, are exchanged over the different protocols in the layered stack in order to address cross optimization and QoS provisioning. Hence, we discuss the most various cross-layer approaches that have been proposed in the literature.

42

3. Review and Discussion of Cross-Layer Proposals

Each layer of a stacked set of modules maintains an independent set of statistics for error conditions and performance metrics. When a problem occurs, it may manifest itself as aberrant statistic values in multiple layers in the system. In classical systems, there is no logic that correlates these aberrant statistic values across different system layers. This lets thinking about alternative solution as cross-layer design. The main feature of the proposed studies in the literature is the determination of what information could be shared and how is it used in a cross-layer architecture to provide QoS enhancement and enable an efficient resource utilization? Hereafter, we describe some examples of cross-layer integration for ad-hoc networks.

3.1. Physical layer + MAC

0 Adaptive power control and MAC

The proposed solution is based on MAC and physical layer cooperation. It estimates the channel using RTS packet and transmit the information using the CTS packet. Then, an adaptive power control mechanism is described according to the obtained information from MAC layer. Furthermore, a receiving node should be spatially separated from any other transmitter by at least a distance D that leads to spatial separation. So, the D parameter greatly influences the amount of interference suppression: If we have low D, more users are selected in the valid set, but a lot of interference must be managed by the power control. The drawback is that power control may not be feasible. If we have high D, there are less users scheduled and so easier job for the power control. However, the scheme may be too conservative higher delays resulted from scheduling because of only these users can run the power control algorithm. In the second step, the power control mechanism optimizes the power allocation among different users. If we have few scheduled users, the MAC layer does not re-optimize its selection based on information from the physical layer ,the loop is not closing.

3.2. Physical layer + MAC + routing

0 Adaptive modulation + MAC + routing

A cross layer networking system is described in '. The paper proposes a coordination between routing, MAC, and physical layers. When a node

43

receives the RTS packet, it estimates the SNR. Then, the transmission rate is mapped from the estimated SNR, and appended to the CTS packet. So, the sender transmits data at the adapted rate. An M-QAM scheme is used in which the constellation size changed with SNR. The routing decision is made based on three metrics. The first one is the bandwidth that represents the rate of link between node i and j. The second one is the interference duration that is the interval from the when the RTS packet is sent to when the data packet is received. The third one, is the congestion that is the queuing delay in the buffer of transmit node.

Adaptive beamforming + MAC + routing

In 7, the proposed mechanism describes a cooperation between adaptive beamforming, MAC, and routing. The scheme uses the same MAC for directional antennas, but transmits RTS over multiple hops (MMAC protocol). The presented network performance results depend on the simulated network topology. There are several cases that were studied: Manhattan networks with aligned routes, Manhattan networks with random routes, and Random configuration. For all three cases, the numerical parameters chosen are: Antenna beamwidth equal to 45 Omnidirectional, transmission range equal to 250 m, Directional transmission range (DD link) equal to 900 m. The performance measure only average throughput. It shows that in general MMAC, better than DMAC, better than 802.11. However, when the routes are aligned, using MAC and directional antennas degrades the performance, compared to the case with omnidirectional antennas (802.1 1). For Manhattan networks, more directional interference occurs, due to the aligned paths. The gain is more if we can actually exploit the spatial reuse property of the directional antennas. If not, the performance will be worse because of the increased directional interference (higher gain for the directional antennas).

0 Power control + scheduling + routing

The cross-layer approach introduced in 5, is presented at four levels: First, the proposed adaptive MAC protocol is sensitive to contention. Second, influence of network layer FIFO queuing on better bandwidth utilization. Third, importance of transmission scheduling. Fourth, routing and power control interactions. In the proposed Progressive Back Algorithm (PBOA) Protocol, nodes contend during every contention period. Unsuccessful nodes progressively backoff during progression of contention period. Suc- cessful nodes use remaining contention interval to discover minimum power

44

needed to transmit their data. There are two benefits of this approach: the first that energy conservation is enhanced because of tuning transmission range as possible. The second benefit is that both interference and collisions are reduced thanks to the proposed backoff procedure. The performance of these interactions depends on several constraints. Indeed, The cost of packet control overhead and packet lost could be more significant than the performance improvement in an arbitrary mobile node in a particular scenario. The key observation is that the protocol performance looks worse than some optimal choices because these two protocols are distributed and hence require global knowledge to schedule their transmissions which is hard to achieve in a very high mobile and distributed networks.

3.3. MAC + Routing + network layer

0 MAC utilization+Interface queue+reactive routing protocol

In 8, the authors propose a mechanism for detecting network congestion, in order to improve the performance of all types of traffic. Indeed, there are two metrics which are used to measure the congestion level. The first one, is the average MAC layer utilization around each node. Instantaneously, this metric can be equal to 1 or 0. It is equal to 1 if the MAC layer is utilized (there is at least one packet in the transmission queue, during backoff decrease period, inter-frame space, detection of physical carrier). The second metric, is the instantaneous interface queue length. This metric is used to avoid nodes that are congested even there is no contention. The proposed mechanisms aim to influence routing decisions that will follow other route discovery scheme either than the short hops count used traditionally. In- deed, it is unsuitable to establish routes over nodes that are already busy. However, if we avoid busy nodes in route establishment, there are some routes that cannot be established even they exist. The congestion information is also used when the medium utilization is high, to influence the setting of the Explicit Congestion Notification (ECN) bits in the IP header of packets at each node. ECN is used to prevent the loss of packets along that flow. At transport layer, the MAC layer utilization metrics measured around the node allow TCP sender to tune its parameters according to these metrics since they represent a recent level to the wireless medium utilization. At higher layer, these metrics can be used to decide or not data compression. Indeed, when the medium is busy the sender can decide to compress the data. However, the compression should represent a trade off between bandwidth consumption and the CPU time used for compression

45

and decompression.

0 topology information+Enhanced back-off +Proactive routing

In 9, a cross-layering design has been presented in the context of research project called MobileMAN. This project investigates a local interaction among protocols in a manet node. For example, MAC layer exploits the topology information collected by network layer to achieve fair channel scheduling and fix the problem related to hidden and exposed terminals. An enhanced backoff scheme is introduced. At transport layer, the different events occurring at lower layers such as route failure, route changes and congestion, are analyzed in order to minimize the useless data retransmissions. Moreover, MobileMan considers routing according to the cross- layering principal. Indeed, a path per-formability index is computed using congestion, link quality, and other parameters that can influence system performance. Furthermore, the MobileMan transport protocol exploits information reported by the routing and Wi-Fi layers in the Network Status component to avoid useless data retransmission.The authors suppose that a node has a knowledge of the hole network topology and so a proactive routing protocol should be used. Hence, it seems that for some scenarios, it is very hard, costly, and not efficient to address this cross layer architecture regarding the dynamic traffic nature and the high mobile node speed. Any information has been provided to how compute the path per-formability index or other cross layer parameters considered in this project.

Probability of successful transmission+route selection+energy conservation

In 3, tow cross-layer designs based on energy consumption are presented.The proposed schemes, namely, Energy-Constrained Path selection (ECPS) and Energy-Efficient Load Assignment (E2LA), employ probabilistic dynamic programming (PDP) techniques and utilize cross-layer interactions between the network and MAC layers. They aim to enhance the operation of existing power-based multi-path routing protocols via cross- layer designs and optimal load assignments. The Energy-Constrained Path Selection (ECPS) consists of maximizing the probability of successful transmission in at most n retry. That is mean that the total n transmissions don’t exceed a total amount of energy equal to y. Furthermore, the authors developed four distinct reward schemes for which E2LA assigns routing loads accordingly. In ECPS mechanism, the MAC sublayer provides the network layer with information pertaining to successfully receiving a CTS

46

or an ACK frame, or failure to receive one. ECPS, in turn, chooses the route that will minimize the probability of error or, equivalently, maximize the expected reward. The proposed medium access control (MAC)-based performance studies, revealed that battery capacity may not be efficient for achieving energy-based fairness and system longevity for wireless mobile multi-hop ad hoc and sensor networks. However, energy conservation may be attained only if valuable MAC (and PHY) input is passed to the network layer. In addition, illustrative examples of E2LA were presented, and its diverse properties were introduced and validated.

0 power control + topology information

In a study of cross-layer design based on power conservation, and congestion informations in ad hoc network have been presented. The authors describe a power control based cross layer architecture. Indeed, they detail the significant impact of power control on all protocol stack above the physical layer. Furthermore, they summarize several works that have been done to address power saving in the protocol stack and show how the power information could be considered at each layer. Moreover, the work claims that, exchanging the topology information between different layers through their interfaces, is very important to support QoS such as geometric location, channel, link conditions. A proposed mechanism, that uses the number of neighbors around the node to adjust transmission power, has been presented.

3.4. Physical layer + MAC + Application

0 SNR information +MAC retransmission+ adaptive FEC

Real-time applications, such as audio and video streaming over wireless links, suffer from bandwidth variation, packet losses, and heterogeneity of the receivers. In 6, the authors propose to exploit the mechanisms available at the lower layers of the protocol stack in order to address an adaptive cross-layer protection strategies for robust scalable video transmission. This mechanism uses a multipath channel model to simulate the wireless indoor channel. This channel model provides the bit error rate (BER) of the link for the eight different PHY modes of 802.11a under different channel signal- to-noise ratio (SNR) conditions. Then, the authors analytically derive the packet loss ratios and throughput efficiency at various channel conditions, considering a given packet size, a given number of retransmissions at MAC layer, and an application layer FEC. These parameters are dynamically

47

adapted according to an end-to-end distortion model in order to achieve an efficient transmission of video streams. The presented algorithm presents a good performance for video streaming. However, it is only centralized.

In the next section, we discuss the constraints of introducing cross layer architecture and the recommendations to achieve a good and optimized solution.

4. The Implementation Cost of a Cross-Layer Architecture

The advantages of cross layer design is to exploit the interactions between layers in order to improve QoS support and optimize resource utilization. Moreover, this new architecture promotes adaptability at all layers based on the exchanged information and tight their interdependence. However, understanding and exploiting the interactions between different layers is the core of the cross-layer design concept. For example, the cross layer models introduced in and in 9 , require respectively the congestion information and hole topology information to build routes using layer cooperation mechanisms. Hence, if we consider high variable scenario in term of mobility and traffic load, the collected metrics will be inaccurate and so become inefficient and very costly. Indeed, it is hard to characterize the best and efficient interactions between protocols at different layers. Moreover, joint optimization across layers may lead to complex algorithms. Note that complexity consumes more resources for computing and introduce a new problem of scalability. So, we have to answer the following question: is cross-layer design suitable for all types of wireless networks and all types of applications? If yes, that means that we have to throw away the OSI reference model and we don’t need to consider a network architecture anymore? This is clearly impractical and disaster in terms of implementation, debugging, upgrading and standardization.

The solution is to maintain the layered approach, while accounting for interactions between various protocols at different layers.

5. Achieving a good trade-off between complexity and enhancement in cross-layer architectures

While cross-layer model could enhance the performance of the applications and achieve better QoS support, there is a lot of proposed models that have to be compared and optimized. In the most cases, we have to take into account the benefits of each model that provides layer cooperation comparing to its complexity. Indeed, there are some proposals that compute global

48

or local metrics which are used to make decisions for route establishment, scheduling, tuning transmission rate, etc. However, using these metrics in a cross-layer model could be not efficient because they are have sometimes inaccurate values which do not reflect the real situation around a given node. Moreover, since a node moves with an arbitrary speed and toward an arbitrary destination, the computed metrics (according to the participation of the node in communication and the traffic load level around it) could change during the time. So, other nodes that consider the metrics of that node to build routes for example, could have an inaccurate information since this later change according to mobility, traffic, and capacity.

We believe that cross-layer a QoS model is a somewhat “danger”. In one hand, the modification that we have to add in the protocol stack and the complexity in introducing a new parameters and new algorithms to provide a ”good” layer cooperation are usually introduce a high complexity risk. In the other hand, this could be very interesting given that it captures the characteristics of the capacity, the expected behavior of node load to choose the ”best route’’ between sources and destinations in a way to achieve a global load balancing, and in other cases have knowledge about neighbor density and ”quality” to adapt transmission rate and to use scheduling strategies in an efficient manner.

So, if we recapitulate, cross-layer is a promised solution to address QoS support and service differentiation in mobile ad hoc networks, but it is affected by mobility and so the “lifetime“ of the availability of the accurate available informations. We recommend the following requirements to efficiently design a QoS cross-layer model which leads to the architecture shown in Figure :

(1) Choosing the metrics: choosing of a very useful and efficient metrics such as battery level, available bandwidth, and mobility rate.

(2) Computing the metrics: the way of computing these metrics regarding one path (energy, lifetime of nodes, throughput, delay, etc.) have to be decided. The well-known approach is to minimize a cost function for a given link in the path between a source and a destination then consider the different costs computed for all links in the path. Depending on the nature of the metric, the cumulative value could be additive, concave and multiplicative. Other techniques could be also used such are variance and max-min. Computation and complexity costs should always be taken into account.

49

(3) Adapting metrics’ values: an adaptive method should be used to update the measured metrics: They could be updated even more when mobility increases and less in a stable network while taking into account traffic load variation and application requirements.

(4) Deciding to use or not the metrics: As shown in Figure 1, considering the information useful for model selection, the more efficient model has to be chosen according to the two following parameters:

(a) Regarding to the network behavior: in some cases, when the traffic load and its characteristics change rapidly (high mobility), it is very difficult to compute accurate values of the metrics that can be used to address QoS. Hence, the complexity of the cross-layer model becomes too high comparing to the expected performance enhancement and it is recommended in this case to use the legacy layered approach.

(b) Regarding to the user application: each layer of the protocol stack responding to local variations and informations from other layers. We have to evaluate the benefits and the disadvantages of the cross layer model for each specific user application.

Inlormations uselul for

architecture selection

Transpon layer

Physical layer

Layered Archltecture

Cross-Layer

Architecture

Figure 1. New architecture design

As a conclusion, the decision to use a cross-layer model is very coupled with the nature of the user application and the evolution of the network behavior. The very promising cross-layer design model consists in maintaining the layer isolation in the protocol stack while enabling a cross-layer interaction according to network and traffic characteristics.

50

6. Conclusion

Many subsystems of appliance operating systems are implemented as stacked modules. For example, the TCP/IP subsystem consists of the link layer, the network layer (IP), the transport layer (TCP and UDP) and the application layer organized as a protocol stack.

In this paper, we discussed the most important features based on cross- layer exchanged information, introduced for mobile ad hoc networks. De- spite of the performance improvement that this new design can achieve, there are some risks into changing the legacy layered architecture. Indeed, several issues need to be talcked before these interactions can be successfully exploited such as implementation, debugging, upgrading and standardization. We have to specify and explain whether cross-layer paradigm is suitable for all types of wireless networks and applications or not. Even if the answer is yes, it is necessary to maintain the layered approach, while enabling interactions between various protocols at different layers.

References

1. X.Li and Z.Bao-yu, Study on cross-layer design and power conservation in ad hoc network, PDCAT, Pages324 - 328(2003).

2. ElBatt, T.; Ephremides, A.; Joint scheduling and power control for wireless ad hoc networks, Wireless Communications, IEEE Transactions on , Volume: 3 , Issue: 1 , Jan. 2004 Pages:74 - 85

3. A.Safwati,and al. Optimal cross-layer designs for energy-eficient wireless ad hoc and sensor networks, Performance, IEEE Computing, and Communic& tions Conference, 2003, 9-11 April 2003 Pages:123 - 128

4. Wing Ho Yuen; Heung-no Lee; Andersen, T.D.; A simple and eflective cross layer networking system for mobile ad hoc networks, PIMRC 2002.

5. Toumpis, S.; Goldsmith, A.J.Performance, optimization, and cross-layer design of media access protocols for wireless ad hoc networks, ; Communications, 2003. ICC '03. IEEE International Conference on , Volume: 3 , 11-15 May 2003 Pages:2234 - 2240 vo1.3

6. V.Schaax, and al. Adaptive cross-layer protection strategies for robust scalable video transmission over 802.11 WLANs, Selected Areas in Communications, IEEE Journal on , Volume: 21 , Issue: 10 , Dec. 2003 Pages:1752 - 1763

7. R. R. Choudhury, X. Yang, R. Ramanatham, N. Vaidya, Using directional antennas for medium access control in ad hoc networks, ACM, Mobicom atlanta, September 2002.

8. Y . Hu and D. B. Johnson, Exploiting Congestion Information in Network and Higher Layer Protocols in Multihop Wireless Ad HOG Networks, ICDCS March 24-26 2004 Hachioji, Tokyo, Japan.

9. M.Conti, and alcross-Layering in Mobile Ad Hoc Network Design, Published by the IEEE Computer Society, February 2004.

TWO BANDWIDTH-VIOLATION PROBLEMS AND BANDWIDTH-SATISFIED MULTICAST TREES IN MANETS

CHIA-CHENG 'Department of Computer Science and Information Engineering

National Taiwan University, Taipei, Taiwan

ERIC HSIAO-KUANG WU Department of Computer Science and Information Engineering

National Central Universiiy, Chung-Li, Taiwan

GEN-HUEY CHEI?

In the existing mobile ad hoc network (MANET) Quality-of-Service (QoS) routing/multicasting protocols, the methods of bandwidth calculation and allocation were proposed to determine routes with bandwidth guaranteed for QoS applications. As our observations, two bandwidth-violation problems will be incurred in the above protocols. In this paper, a heuristic algorithm is proposed to determine a feasible bandwidth-satisfied multicast tree without the two bandwidth- violation problems.

1. Introduction

A multicast group contains a special host (server) that is responsible for transmitting data packets to the other hosts (clients) in the same group. A mobile ad hoc network (MANET) is formed by a group of mobile hosts that can communicate with one another without the aid of any centralized point or existing infrastructure. Due to the recent provision of high-speed wireless Internet services, the QoS requirement applications will be crucial for the integrated new-generation wireless multimedia communication Systems. In the existing QoS routing/multicasting protocols [ 1-51, some admission methods were proposed to determine bandwidth-satisfied routes for bandwidth- requirement requests. However as our observations, two bandwidth-violation problems in the above existing MANET QoS routing/multicasting protocols would be incurred.

First, this problem arises because a host only considers its local bandwidth (one-hop) transmission information while determining whether a new route can

51

52

pass it or not. This problem is called as Hidden Route Problem (HRP) in this paper. Second, another bandwidth-violation problem, called as Hidden Multicast Route Problem (HMRP), would mislead the bandwidth reservation for QoS multicast applications in the existing QoS multicasting protocols. In Section 2, the bandwidth-reservation methods in the existing QoS routing/ multicasting protocols are overviewed and the more details for HRP and HRMP will be illustrated.

In this paper, we aim to determine a feasible bandwidth-satisfied multicast tree without HRP and HRMP in MANETs. We minimize the number of forwarders for reducing the number of hosts participating in packet forwarding so as to lower bandwidth and power consumption. To attain the purpose, a heuristic algorithm to obtain a feasible bandwidth-satisfied multicast tree, denoted as FBST, is proposed in Section 3. In Section 4, studies are carried out to evaluate the performance of the proposed algorithm. In Section 5 , this paper concludes with some remarks and future works.

2. Bandwidth-reservation and related problems

In wired networks, since a dedicated point-to-point link liJ between host vi and vj is used, the maximum bandwidth b-maxjj of 1, is only consumed while vi or vj transmits packets to each other. The neighbors of vj and 5 will not consume b-max, when vks transmit packets to vi or vj so that the remaining availability bandwidth b r i j for l j j can be computed by v; and vj. Based on the above, the judgment is achieved easily for the requested flow with bandwidth requirement b-req. If brjJ2b-req, vi has enough available bandwidth to forward the flow to vj. A bandwidth-satisfied route Ra,fv,+vb+vc+vs,. . .+ve+vf is determined as the equation: min{b-r,b, b-rb,,, b-r,d ,..., b-re,f} 2b-req.

The routing protocols [ l , 21 and the multicasting protocols (MCEDAR [4] and M-CAMP [ 5 ] ) utilize the concepts by using point-to-point link from the wired networks for MANETs. Since a host shares the radio channel with all its neighbors in MANETs, the bandwidth will be consumed not only by itself but also by its neighbors. Refer to R,)h the remaining bandwidth in three hosts (va, Vb

and v,), in which v, and v, are located at the transmission range of vb, will be consumed if vb transmits the packets to v,. That is the consumption will occur at the links l,s, in which V ; E {v,, vb, v,} and vjs are the neighbors of v,, vb and v,. Further, other forwarders along Raf make the similar bandwidth consumption to their neighbors.

AQOR [3] proposed a different scheme for computing the bandwidth. A host is modeled as a resource unit in contrast to a link. Suppose the maximal

53

bandwidth and the remaining availability bandwidth of host vi are denoted as b-maxi and b-ri, respectively, to replace b-maxij and b-rij. Referring to the bandwidth-satisfied route R, the total consumed bandwidth b c i to the host vi by the requested flow with bandwidth requirement b-req will be computed as following equations: b-c,2b-req, b_cb=3b_req, hcC=3b-req, b-cp3bpreq,. . ., b-ce=3 b-req and b-crb-req.

However, no matter link resource oriented scheme or host resource oriented scheme in the above QoS routing/multicasting protocols, two bandwidth- violation problems HRP and HMRP would likely occur.

HRP An illustrative example is shown in Figure l(a), where two routes v p v g and v,,, --v,, are established and one route v a j - + v e is being constructed. For convenience, we use + to denote a route of one hop and -- to denote a route of one or more hops.

(a) @) Figure 1. The examples of HRP and HMRP

Suppose that the capacity of each host is constant, say 11 units, and the bandwidth requirements of va--ve, v p v g and v,,--v, are 3 , 2 and 7 units, respectively. When the route determination proceeds to host v,, the available capacity of v, has to be computed in order to determine if v, can be one forwarder in v,--ve. Now that the bandwidth requirement of va--ve is 3 units. If v, can be constructed as one forwarder of v,--v, , the bandwidth requirement for each of the three hops (Vb-v,, V,-Vd and v p v , ) is 3 units, i.e., the total necessary remaining capacity for v, is 9 units. There is an ongoing transmission from vf to vg in the radio coverage of v, whose bandwidth requirement is 2 units. Hence the capacity remaining for v, is 1 1 -2=9 units and v, can be constructed as one forwarder of va--ve. However, the establishment

54

of forwarder v, increases the bandwidth consumption in the radio coverage of v~ to 12 units, which violates the capacity of vy. 0 HMRP Another bandwidth-constraint violation problem, HMRP, will likely happen when the multiple routes from a server to all clients are determined concurrently. The illustrative example for HMRP is shown in Figure l(b), where one multicast route from server v, to clients v, and vi (including va--v, and va++ v;) is being constructed. The problem is encountered when a multicast route is to be constructed from the server by broadcasting the routing request to the multiple clients. Suppose that the bandwidth requirement for each of v,++ve and va++vi is 3 units.

When the route determination proceeds to host v,, the available capacity of v, has to be computed in order to determine if v, can be one forwarder in va+- v,. Similar to the above, the total necessary remaining capacity for v, is 9 units and v, can be constructed as one forwarder of v,+-ve. Because the bandwidth reservation will be activated for the flow only when the real data flows at the host, the remaining capacity for vg is still 11 units such that vg can be constructed as one forwarder of va-+vi at the same time. However, the establishments of v,+-ve and v,++v; cause the bandwidth consumption in the radio coverage of v, to 12 units, in which there are four forwarders (vb, v,, vd and vg) and each forwarder consumes 3 units of bandwidth. That violates the capacity of v,. The same violation also happens in the radio coverage of vg.

3. Heuristic algorithm

In FBST, two procedures are designed. The first one denoted as SBSP is proposed to determine the shortest bandwidth-satisfied path without HRP for a bandwidth-requirement source-destination pair. The second denoted as SBST can determine a feasible bandwidth-satisfied multicast tree without HRP and HMRP for a multicast group. By SBSP and SBST, FBST determines a feasible tree selecting with minimum number of forwarders from a set of feasible trees which are derived from an iterative operation. The information to take as the input parameters of FBST is listed as follows: V: a set with n hosts, denoted by vl, v2, ..., v,. djj: a binary integer djJ=l (did =0) to denote that v; is (is not) a neighbor of vi. b-mmi: maximum transmission bandwidth for vi. bongoing;: the sum of bandwidth transmitted from host vi for the ongoing flows. Dongoing: the destination set of the ongoing flows. v,: the sever of the request multicast group. D-reg: the client set of the request multicast group. b-req: the bandwidth requirement for the request multicast group.

55

PROCEDURE DIAGNOSE (F, D, 6 - req) PROCEDURE SBSP (v , , vd , b - req, D , T ) { for each vi E F do ( x = Y - (vd};

( b-aJieri =b-ongoingi + b - r e q ; foreachv,EY a n d i # d ( if DIAGNOSE( ( vi } u T , D, 6 - req ) == tiue

the n hi = 1 and Pi = ( v i } ; else hi = m and P, = (};

if b - aJieri > b - mar j return false; I * vi without enough remaining bandwidth

I * to diagnose the neighbors of vi for avoiding

and d,,d == 1

}

violation * I for each vi E F do { for each neighbors v of vi ,

i s . , di,, = 1 and i # j

{ if b - ongoing > 0 or v j E (D u D - ongoing) if ( Cb-aJier ,+)>b-mar , return false;

d , . d

I } return true; I * vi can be a candidate forwarder * I

}

Figure 2. High-level description of DIAGNOSE

while X is not empty do ( determine v x E X so that h, = min( hi I vi E X } ;

delete vx from X ; if h, f m

( for each vi E X and di,x = 1 do ( if DIAGNOSE(( v i } u P, u T ,

D , b - r e q ) = = true then hi = h, + 1 and Pi = ( v i } + P,;

}

} return H and P ; I * H and P are the sets containing hi and Pi * I

1 Figure 3. High-level description of SBSP

In SBSP, a sub-procedure DIAGNOSE in Figure 2 is proposed to judge whether HRP will happen for a request, in which the set of candidate forwarders, the set of destinations and bandwidth requirement are defined as F, D and b-reg, respectively. SBSP as shown in Figure 3 is proposed to determine the shortest bandwidth-satisfied path without HRP from the destination v d to all the other hosts, in which the path from v d to the sender v, is also included. Since SBSP will be utilized iteratively by SBST to determine a feasible tree, the input parameters D and T are used to record the relative information of the previous iterations. When SBST executes SBSP in the first iteration, the first client is contained in D and T is empty. In the second iteration, the first two clients and the forwarders determined in the first iteration are contained in T and D, respectively. The rest iterations are similar to the first two iterations.

The second algorithm SBST as shown in Figure 4 is proposed to determine a feasible bandwidth-satisfied tree T. SBST executes ID-seqI times of SBSP for determining T. A template set D that contains some clients is used to make sure that each client in D should have enough bandwidth to receive the flow. Initially, T and D are set to empty. In the first iteration, the first client of D-seq is added to D and the shortest feasible path P, from v, to the first client is obtained from SBSP. Obviously there is no feasible tree for the group if P or P, is empty. If not, we set T=P,. In the second iteration, the second client of D-seq is also added to

56

D such that D contains the first two clients of D-seq. In order to avoid HMRP, the first client and the forwarders belonged to T should be taken into consideration while the feasible path for connecting the second client is being determined. In the way, the feasible path to connect the second client and obtained from SBSP will not incur HRF' to the existing flows and HMRP to the first path derived in the first iteration. To connect the second client and T, the forwarders belonged to Tare the candidate points in SBST. We select a feasible path P,with the minimal number of forwarders, where VFT, so as to minimize the number of forwarders in the final feasible solution. Then, a new T is derived from merging T and Pf If P or Ps is empty, SBST stops the execution and returns possible-no-solution. In the rest ID-seql-2 iterations, the executions are the same as the second iteration.

PROCEDURE SBST (v, , D - ( let T and D are empty;

for i = 1 to I D - seq I do ( let v d be the ith element

let H and P be the sets obtained

i f i = 1

in D - seq ; add V d to D ;

from SBSP( v s , v d , b - req , D , T ) ;

( if P == empty or P, = empty return no - solution ;

else T = P,;

else if P == empty return possible-n o - solution ;

else ( determine PJ E P so that

h , = min( h i I v i E T ); if PJ == empty

else T = T u P J ; return possible-n o - solulion ;

} I return T ;

Figure 4. High-level description of SBST }

"I .

I , , , . . , . . 0 4 1 1 4 I e 7 8 9

Figure 5 . Average admission rate

NMberofgrolps

'--

Nmbarofgmlpl

Figure 6 . Average receiving rate

FBST utilize SBST to determine a bandwidth-satisfied multicast tree without HRP and HMRP. In SBST, a feasible tree is constructed to connect the clients one by one based on the sequence of the clients. However, SBST may fail in finding a feasible tree whenever one exists. To remedy the disadvantage,

57

SBST will be executed several times by an iterative algorithm. A distinct permutation sequence of the clients is given to SBST for computing a feasible solution in the each iteration, and several feasible trees may be derived. In FBST, the final solution tree is selected from these trees. We know that there are ID-reql! permutations, in which D-req is the set of the clients, such that the number of the iterations to execute SBST is ID-reqI!. Suppose that Ti is the feasible tree derived from the ith iteration, where l<illD-reql!. Then, We select the tree as the final solution with minimum number of forwarders from the set {T,I Iltl)D_req)!} determined by the first step.

4. Performance analysis

We note that each problem instance generated may or may not have a feasible solution. Infeasible problem instances should not contribute to the frequency with which FBST fails in finding feasible solutions. Hence, to accurately evaluate the effectiveness of FBST in finding feasible solutions whenever one exists, we have to be able to detect whether a generated instance is feasible or infeasible. In order to do so, each generated instance is first formulated as a 0/1 integer linear programming (ILP) [6], which was proposed by the authors. Then, solving the formulated ILP using branch-and-bound techniques can derive the optimal solution, denoted as OBST. We integrate FBST and OBST with ODMRF' [7] for supporting bandwidth-requirement multicast services.

The simulation environment for modeling a MANET has 50 hosts placed randomly within a 1 OOOmx 1000m. The raw data transmission capability (maximum available bandwidth) of each node was assumed 800 Kbps. Two performance measures are collected in our studies: Receiving rate (the number of data packets received divided by the number of data packets delivered from the servers with admitted multicast trees) and admission rate (the rate to admit the bandwidth-requirement multicast groups).

In Figure 5 and Figure 6, the comparisons of average admission rate and average receiving rate among FBST, OBST and MCEDAR are executed while the hosts are static, and some remarks are concluded as followings.

When there are few groups, both FBST and OBST obtain higher admission rates than MCEDAR because that FBST and OBST select the tree with less number of forwarders so as to reduce the bandwidth consumption and to support more requests. When there are many groups, FBST obtain lower admission rates than MCEDAR. The lower admission rates and high receiving rates of FBST validate that the violation incurred by MCEDAR is avoided by FBST.

58

0 To accurately evaluate the effectiveness of FBST in finding feasible solutions, Figure 5 shows that the admission rate of FBST is close to OBST.

5. Conclusions

In this paper, we introduce two bandwidth- reservation problems HRP and HMRP happened in the existing MANET QoS routing and multicasting protocols. To avoid the two problems, a heuristic algorithm FBST is proposed to determining a feasible bandwidth-satisfied multicast tree. In FBST, the feasible multicast tree with small number of forwarders is selected in order to improve bandwidth and power efficiencies. In comparison FBST with OBST and MCEDAR, simulations validate the following results: 1) To benefit from fewer number of forwarders in the determined multicast tree, FBST has higher admission rate than MCEDAR while the network traffic is not saturated. 2) To correct the misleading bandwidth allocation, FBST can avoid the network performance declination while heavy network traffic load. 3) The admission rate of FBST is close to the optimal solution of the problem.

References

1. R. Sivakumar, P. Sinha, and V. Bharghavan, ”CEDAR: A Core-Extraction Distributed Ad Hoc Routing Algothrim,” IEEE Journal on Selected Areas in Communications, vol. 17, pp. 1454-1465, August 1999. S. Chen and K. Nahrstedt, ”Distributed quality-of-service routing in ad hoc networks,” IEEE Journal on Selected Areas in Communications, vol. 41, pp. 120-124, June 1999. Q. Xue and A. Ganz, ”Ad hoc QoS on-demand routing (AQOR) in mobile ad hoc networks,” ACDEMIC PRESS: Journal of Parallel and Distributed Computing, vol. 41, pp. 120-124, June 2003. P. Sinha, R. Sivakumar, and V. Bhanghavan, “MCEDAR: multicast core- extraction distributed ad-hoc routing,” Proceedings of the IEEE Wireless Communications and Networking Conference, 1999, pp. 13 13- 13 17. E. Pagani, G.P. Rossi, “A framework for the admission control of QoS multicast traffic in mobile ad hoc networks,” Proceedings of the ACM International Workshop on Wireless Mobile Multimedia, 2001, pp. 3-12. C. C. Hu, E. H. K. Wu and G. H. Chen, “Feasible Bandwidth-Satisfied Multicast Tree Determination in MANETs,” will appear in IEEE International Conference on Wireless and Mobile Computing, Networking and Communications 2005. S. J. Lee and M. Gerla, “On-demand multicast routing protocol in multihop wireless mobile networks,” ACM/Kluwer Mobile Networks and Applications, vol. 7 , pp. 441-453,2002.

2.

3.

4.

5 .

6.

7.

LAP-RRP: A RELIABLE ROUTING PROTOCOL WITH LINK

AVAILABILITY PREDICTION IN MANET*

JIANXINWANG, JUEHE

School of Information Science and Engineering, Central South University

Changsha, 41 0083, China

XICHENG LU

School of Computer, National University of Defense Technology

Changsha, 410073, China

In this paper, wc proposc a new method to estimatc thc probability that link cxists at to

can be kept until to+At, where dr is decided by the actual application. Based on the link

availability prediction, we can choose a reliable path for data transmission at to, which

can also be used at to + At. At last we test the performance of the protocol by simulation.

The results show LAP-RRF' can reduce broken times of the transmission path and the

number of lost packets. LAP-RRP protocol improves network performance by reducing

the influence on routing because of the mobility in MANET, especially when the

number of nodes is large and the velocity of nodes is high.

1. Introduction

Mobile ad hoc network is a self-organization multi-hop network, which consists of a group of nodes with wireless receive-and-dispatch devices. Each node in ad hoc network can move rapidly and consequently results in frequent breaks of wireless links, then rerouting comes into being which can degrade the performance of mobile ad hoc networks. So it's very important and meaningful to select the most stable and optimal path in ad hoc networks.

The path stability depends on the stability of all links that make up it. And link prediction includes two stages:

1. 2. There are four proposed strategies of link prediction: Prediction-based

Link Availability Estimation [I], Link Expiration Time Prediction Model [Z],

The acquisition of node movements' status information. Computation of link status using link prediction.

' The work is supported by the National Natural Science Foundation of China under Grant No. 90304010

59

60

node’s Location Prediction Model [3] and Path Availability Model [4], they just give the prediction of LET value or the probability which one link can also be used after LET. But in some environment, we need predict the probability which one link can also be used after random time At, so we propose a new method to predict the link availability p(At) in this paper.

2. Network Model

In the paper, we propose a reliable routing protocol with link availability prediction (LAP-RRP), which bases on the network models as follows.

Assume wireless communication of each node has no direction and communication links between any nodes are bi-directional, and we treat MANET as an undirected graph G(N,L), where N refers to a set of all nodes in the network, L refers to a set of all links that make nodes connect directly.

Definition I s(Li, t ) denotes whether there exists link Li at t . If link Liexists then s(Li,t)= 1, otherwise s(L,t)=O.

Definition 2 A L , to, At) denotes the availability of link Li, which is the probability that link Listill exists at time to+& in the case of Liexisting at time to. That is:

fTLi, to, At)=P{ S(Li , t,+At)=l I s(Lir to)=l } Definition 3 F(P, to, At) denotes the availability of path P which is the

probability that path P is available at to+At while it is established at time to. Assume a path P between A and B consists of LIJ2, ...& k links (P= {LIJ2 , .... Lk}). That is:

F(P, to, At)=f(L,, to, dt)*f(L,, to, At)* ...... *fTLk, to, At) Definition 4 F(Pma(A,B, to, At)) is the availability of the most reliable

path between A and B, which refers to the largest stability path in all paths between A and B. Supposing there are m paths between A and B in the network, PATH={PI,P2, ... ... P,,,} refers to the set of all paths and the availability of each path are F(P,,to,At), F(P2, to, At), . . . . . .,F(P,,,, to, At) respectively. That is:

F(Pm,(A,B,to, dt))=rnax{ F(Pi ,to, At) i=1,2,3.. . . . . ,m}

3. LAP-RRP Protocol

3.1. Link Availability Prediction

The basic idea is to let a node first predict a continuous time period (LET) from time to, then estimate the probability that the link will last to+LET, p(LET), Finally, we estimate the probability that the link can be used at to+At, p(dt), where At is a variable which is decided by the application or the users.

Path stability prediction mechanism contains three steps: 1. Prediction of the expiration time of links (LET)

61

Regarding LET prediction, a measurement-based scheme has been proposed in [5 ] , in which, a node can predict LET for an active link with another node by measuring relative distances between them without knowing the velocities of their movements. In [2], a similar scheme has been also proposed to predict LET, in which, the velocity of a node’s movement is supposed to be known by using Global Position Systems (GPS).

2. Estimation the probability p(LET) Assume that the epoch lengths are ID exponentially distributed with mean

A?? that is E(x)q(Epoch length Ix}=l-e-hx, and the movement of each node is independent. In a simple model, we can get the estimation by subtract the probability that the link may be broken during the epoch lengths, that is

-2 .M - p(LET ) = 1 - e

3. We need predict the link availability of any link during the interval At. The

descended speed of link availability degrades significantly as the increase of the intervaldt. p(At) degrades quickly when At is less than LET and it degrades slowly when At is larger than LET. So we propose the following formula to calculate the availability of link that can be used during the interval At.

Prediction of the link availability p(dt)

1 - p(LET) *At (0 I At I LET) LET

(At >LET) p(LET) * LET At + log(At - LET + 1) * logZ(epoch ) * log V, * log Vz

P(At) =

where J’,, V, are the speed of nodes. According to definition 2, AL,,t0,At) equals to p(At) which is calculated for

link L, at to.

1 I _ (

0 20 do 60 80 1W 120

At( lee59.43)

Fig. 1 The prediction value of the link availabilityp(dt)

to p(LET). When At is larger than LET, p(At) degrades slowly. As shown in Fig. 1, as At varies fiom 0 to LET, p(dt)degrades linearly

62

3.2.

LAP-RFU' routing protocol finds a reliable path from all paths that may be used in the interval At. LAP-RFU' adopts an on-demand method and contains two phases that are route build and route maintenance.

Route build can be divided into two steps: 1. Processing and forwarding RREQ packet After the source node broadcasts a RREQ packet, each node receives the

RREQ packet will process it as follows: Firstly, the node calculates the link availabilityJLi,to, At) that is from the

previous node to the current node according to the formula in section 3.1. The path availability F(P,t,,At) from the source node to the node is also calculated.

If it is not a duplicate and it does not reach the destination, it updates the path availability F(P,to,At), broadcasts the information to its neighbors.

If it is a duplicate and it does not reach the destination, it compares the F(P,to,dt) with the old F(P,to,At) in the routing table. If new F(P,to,dt) is larger, then the protocol replaces the path availability with it and broadcasts this information. Otherwise it dmards the request packet and does nothing.

If it reaches the destination node and is not a duplicate, the destination node replies RREP. If it is a duplicate, the destination compares the F(P,to,dt) with the old F(P,to,At) . If it is larger, the destination node replies RREP and updates its old information, otherwise it will be discarded and does nothing.

A Reliable Routing Protocol LAP-RRP

2. If an intermediate node receives a RREP packet, it updates its routing table,

actives the path and forwards the RREP to its upstream node on the path. If the source receives RREP, it updates its routing table, and then begin transmitting.

We utilize MAC protocol to monitor links, if there is a acknowledged break, it updates routing tables, sends RERR (Route Error) and establishes a new path. In order to get the more accurate route information, we limit the maximum living time of each path. The path will be deleted from routing table after the maximum living time and the source node have to find a more stable path.

Processing and forwarding RREiP packet

4. Simulation Environment and Results

4.1. Simulation environment and merits

The simulation tool is Glomosim 2.03 [6 ] , which is developed by the parallel computing Lab of University California in Los Angeles (UCLA). In the simulation environment, the signal of each node is 15dbm in a area of 1500m*1500m. Simulation time is 1200 seconds. The results in the following

63

figures are the average value of many times simulations in different scenarios. We adopt the random walk-based mobility model in the simulation. To show the performance of LAP-RRP, we compare AODV [7] with it by using the following metrics: 1. Number of broken links: The number of links that are broken during data

transmission. 2. Number of losing packets: The number of all losing packets caused by

broken links. 3. Packet delivery ratio: The number of data packets actually received over

the number of packets sent by all nodes in network layer. 4. Number of control packets: The total control packets transmitted, including

RREQ(Route Request), RREP(Route Reply) and RERR(Route Error).

4.2. The analysis of simulation results

(1) Network performance with different epochs

I --.-W-W --r-AODV

I . . . . . . , . . . . . 4 . . . . . . . . . . . . epoch(second) epoch(smnd)

$0 ,I m n 30 3s 10 (5 20 23 YI 15

(a) Number of losing packets @)Number of broken links

.;::I =I

.- - om 8

epoch(second)

-1

epoch(second)

(c) Packet delivery ratio (d) Number of control packets Fig. 2 Network performance with different mobility epochs

64

In the simulation, the number of nodes is 30 and the average speed of each node is 5 d s . The epoch length varies from 10 seconds to 25 seconds. As shown in Fig.2, the mobility epoch length does not influence the network performance significantly. Fig2 (b) shows that because AODV utilizes path which has probably broken that has been saved by the source and intermediate nodes, whereas LAP-RRP utilizes the prediction of path availability and data are transmitted in a reliable path, the number of broken links of LAP-RRP is less than that of AODV and therefore the number of losing packets is also less than AODV as Fig.2 (a) shows. As shown in Fig.2 (c), LAP-RRP protocol owns higher packet delivery ratio due to its smaller number of losing packets, Fig.2 (d) shows that LAP-RRP has more control packets than AODV. AODV checks whether there is a path or not in routing table when it sends a new data packet, while LAP-RRP need rebuild a new path every time in order to get more reliable transmission path, thus it has more control packets.

(2) Network performance with different mobility speeds

mobility speed(m/s)

(a) Number of losing packets

I l I 4 l l

mobility speed(m/s)

(c) Packet delivery ratio

/ /’

/

mobility speed(m/s)

@) Number of broken links

(d) Number of control packets . .

Fig.3 Network performance with different mobility speeds

65

In the simulation, the number of nodes is 30 and the mobility epoch length is 10s. The average mobility speed varies from I d s to 6 d s . As shown in Fig.3, network performance changes significantly with the increase of mobility speed. Fig.3 (b) shows that the number of broken links increases with the increase of mobility speed. The number of broken links is less significantly than that of AODV and the difference is more and more obvious with the increase of mobility speed. As the increase of mobility speed, topology changes quickly and links break frequently. As shown in Fig.3 (a) and Fig.3 (c), the number of losing packets increases and the packet delivery ratio decreases with the increase of mobility speed. LAP-RRP protocol gets better performance than AODV in the two metrics. Fig.3 (d) shows that the number of control packets increases correspondingly with the increase of mobility speed.

(3) Network performance with the number of nodes The average mobility speed is 5 m/s and the mobility epoch is 20s. The

number of node varies from 20 to 45. As shown in Fig.4, the number of losing packets, broken links and the number of control packets increase with the increase of the number of nodes, while the packet delivery ratio degrades. LAP-RRP gets better performance than AODV in the four metrics.

' . /./---------. 0. , . , . , , , . , . , ,

20 zs M 31 40 45

the number of nodes

(a) Number of losing packets

the number of nodes

(c) Packet delivery ratio

tmo-

/ a k c.

8 , g - : - - - - = .

0 4 , . , . , . I . , . , , 20 21 30 31 40 45

the number of nodes

tmo-

/ a k c.

8 , g - : - - - - = .

0 4 , . , . , . I . , . , , 20 21 30 31 40 45

the number of nodes

@)Number of broken links

the number of nodes

(d) Number of control packets Fig. 4 Network performance with the number of nodes

66

5. Conclusion

In this paper, we propose a reliable routing protocol with link availability prediction, which utilizes the prediction of link stability to find a most reliable route. Simulation results indicate that the protocol can greatly improve stability of the transmission path and reduce the number of broken links, so more data packets are delivered to destinations than in AODV, but the number of control packets increases since it finds a reliable path in routing process. However, LAP-RRP performs better than AODV protocol in mobile ad hoc network, especially when the number of nodes is large and the mobility speed is high.

References

1. Shengming Jiang, D.J. He and J.Q. Rao, A prediction-based link availability estimation for mobile Ad Hoc networks, IEEE INFOCOM 200 1 W. Su, Sung-Ju Lee and Mario Gerla, Mobility prediction and routing in ad hoc wireless networks. International Journal of Network Management,

2.

2001, 11:3 - 30. 3. H. Shah, and K. Nahrstedt. Predictive Location-Based QoS Routing in

Mobile Ad Hoc Networks. In Proceedings of IEEE International Conference on Communications (ICC 2002), New York, NY, April 2002. A.B. McDonald and T.Znati. A mobility based framework for adaptive clustering in wireless ad-hoc networks. IEEE Journal on Selected Areas in Communications, Special Issue on Ad-Hoc Networks, Aug. 1999. D.J. He, S.M. Jiang and J.Q. Rao, "Link availability prediction model for wireless ad hoc networks", Proc. 2000 International Conference on Distributed Computing System Workshop, Taipei, Taiwan, (April 10- 13,

4.

5 .

2000), pp.D7-D 1 1. 6 . http://pcl.cs.ucla.edu/projects/glomosim 7. C.Perkins and E.Royer, and S.Das. Ad hoc on-demand distance vector

routing. Internet Draft, draft-ietf-manet- aodv-lO.txt, 2002

AN EFFICIENT LOAD-BALANCING ALGORITHM FOR SUPPORTING QOS IN MANET

MOHAMED BRAHMA University of Haute Alsace, 34 Rue du Grillenbreit, 68000 Colmar France

KWAN-WOONG KIM University of Haute Alsace, 34 Rue du Grillenbreit, 68000 Colmar France

ABDELHAFID ABOUAISSA University of Haute Alsace, 34 Rue du Grillenbreit, 68000 Colmar France

PASCAL LORENZ University of Haute Alsace, 34 Rue du Grillenbreit, 68000 Colmar France

MIKE MYUNG-OK LEE University of Dongshin, 2.52 Daeho-Dong, Naju, Chonnam, 520-714, South Korea

A Mobile Ad hoc Network can be considered as an autonomous distributed system that consists of a set of identical mobile nodes that move arbitrarily and use wireless links to communicate with other nodes that reside within its transmission range. In our work, we analyze the performance of the original IEEE 802.1 1 b wireless local area networks and we present an efficient load-balancing scheme in MAC for support QoS over MANET. Simulation results aimed to highlight the capability of our algorithm to reduce packet lose rate and to increase throughput in the MAC layer are provided.

1. Introduction

A Mobile Ad hoc Network (MANET) [1][2] is an autonomous distributed system that consists of a set of identical nodes that move arbitrarily and use wireless links to communicate with other nodes that reside within its transmission range. Because of limited radio propagation range, mostly routes are multi-hop.

Ad hoc networks are deployed in applications such as sensor networks, disaster recovery, rescue and automated battlefields, and without any base station or centralized administration. Nodes are free to move randomly and organize

67

68

themselves arbitrarily; thus, the networks topology may change rapidly and unpredictably.

Various constraints are introduced by the ad hoc networks:

- Dynamic topology which evolves very quickly because each node can move arbitrarily and disappear randomly without any notification. From where need for routing mechanisms which adapts with the nodes connectivity at a given moment.

- Radio channel of communication: indeed the connections are with variable rates and limited bandwidth.

- Nodes function with batteries: a reduced autonomy in term of energy. Moreover each node serves as a host as well as a router and uses consequently its own energy to route flows intended for other nodes of the network.

- Limited security: since ad hoc networks are more vulnerable to physical security threats, provisions for security must be made.

The ability to provide an adaptive quality of service (QoS) in such a mobile environment is a key to the success of next generation wireless communications systems. Recently there has been a considerable amount of QoS research. However, the main part of this research has been in the context of framework components, and much less progress has been made in addressing the issue of a group management to provide QoS within an ad hoc network.

In the rest of paper, we first briefly provide an overview of present related work and then describe our load-balancing algorithm in section 3. In section 4, we evaluate performance of proposed algorithm through simulation experiments and compare it to the original MAC IEEE 802.1 1 b [9] [ 131. Finally, we describe future work, before we conclude.

2. Previous and related Works

Due to nodes mobility, the topology of an ad hoc network may change rapidly and unpredictably over time [3]. The design of network protocols for MANETs is a complex issue; these networks need efficient distributed algorithms to determine network organization (connectivity), link scheduling and routing.

Most routing protocols for mobile ad hoc networks, such as: Ad Hoc On Demand Distance Vector Protocol (AODV) [4], Dynamic Source Routing (DSR) Protocol [5], Temporally-Ordered Routing Algorithm (TOM) Protocol [6], Dynamic Destination-Sequenced Distance-Vector Routing protocol (DSDV) [7], are designed without explicitly considering quality of service of the generated

69

route. These routing protocols provide the capability for establishing minimum- hop paths between nodes on a best effort basis [4][1] regardless of node status such as available buffer space. However some applications, such as multimedia and real-time, need not only the capability to establish communications between nodes, but also require some quality of service guarantees on bandwidth, delay and bit error rate. Some researchers are currently working on service differentiation in the MAC layer as it’s the case in [ 101 [ 1 11 or [ 121.

The buffer management is usually one of the most essential and challenging in such a dynamic environment. Therefore, there is a need for load-balancing to allow all nodes the opportunity to enhance their use of network resources. In this paper, we propose a load-balancing algorithm that allows nodes to:

distribute and efficiently use network resources (buffer space), reduce network congestion by change route, increase overall performance (throughput).

3. Efficient load-Balancing algorithm in IEEE 802.1 1

In general, the problem of congestion can occur in a network when offered traffic load exceeds available capacity at any point in the network. In ad hoc networks, congestion causes overall channel quality to degrade and loss rates to be increased, leads to buffer drops and increased delays, and tends to be grossly unfair toward nodes whose data has to traverse a larger number of radio hops. The goal of this work is to propose an efficient load-balancing scheme in the MAC IEEE 802.1 1 layer to reduce the network congestion by changing the route when a node is congested and to distribute and efficiently use network resources, especially the buffer space. The purpose of load-balancing is to:

increase and enhance network resource utilization (buffer, radio distribute excessive load of a node to its neighbors,

channel), reduce collision by load distribution, reduce the number of packets lost by buffer overflow improve overall performance in MANET.

3.1. Scenario

When a node receives DATA packet, if this node is congested, it broadcasts HELP message to its neighbors with sender address ‘B’ as shown in figure 1 (a)

70

Definition: Congested node is which the queue occupancy is more than a certain level. For example: i f the queue occupancy is more than 80% of total queue size, a node is considered as congested.

When a node receive a HELP message, if a node has available buffer space and sender address 'B' is belong to my neighbors than send an OK packet to the congested node as shown in figure 1 (b)

The congested node choose the best node address 'D' and send a NOTIFY message with address 'D' to sender 'B' (figure 1 (c))

After receiving a NOTIFY message, the node 'B' change the route to the node 'D' from the congested node as like as figure 1 (d).

Q >...no 8 n

(a) The c o n p a d node broadcast HELP lo one+ep neighbors I4 The wnossted node send NOTIFY te node B

-a

(dl The node B RdlW Path horn the conw1.d node to tk node 0 (b) some nolghbor node. rend OK to the congesled node &r re0iV.d OK, the Congested nod. ohms. nods D 10 inlermediale nod.

Figure 1. Load-balancing scenario

3.2. Message format and procedure of load-balancing

For providing load-balancing ability in IEEE 802.1 1 MAC, we have designed new messages as depicted in figure 2: HELP, OK and NOTIFY. Where 'queue- status' is the available buffer space of the transmitting node. The size of queue- status field is 2 bit and each value means that; 0 is empty, 1 is low, 2 is medium and 3 is high. IA field is the address of an intermediate node.

Determine queue status In load-balancing algorithm, we determine 'queue-status' with queue occupancy for decision of congestion.

71

Frame Control

If queue occupancy < low threshold than queue-status = 0; Else if low threshold < queue occupancy < medium threshold than

Else if medium threshold < queue occupanq < high threshold than queue-status = 1;

oueue~stalus RA TA IA FCS (2bils)

queue-status = 2; Else queue-status = 3 ; Where queue occupancy is

queue over total queue space.

Frame Con,rOl

the ratio of the number of waiting packets at the

HELP format

FCS Queue-alatus Receiving Transmitang (2bns) Address Addregs

F,Bme cDnlml

OK format

Queue-stalua Recelvlng TmnsmWng Intermediate (2bnU) Address Address nDdeAddress Fcs

NOTIFY format

RA : Receiving Address

TA : Transmitting Address IA : Intermediate Address

Figure 2. Packets’ format

Decision to send HELP message If ‘queue-status’ is greater than 2 when a node receives DATA packet, a node become congested node and sets IA to transmitting address of DATA packet and then broadcasts a HELP message to ask neighbors for help. Decision to send OK message When a node receives HELP message, if the ‘queue-status’ is less than the ‘queue-status’ of HELP packet and IA of HELP message is belong to my neighbors then a node sends OK packet to the congested node. Send NOTIFY packet to the intermediate node When congested node receives OK message, congested node sets IA of NOTIFY message to transmitting address of OK message and sends NOTIFY message to the intermediate node. Receive NOTIFY and reroute DATA packet When a node receives NOTIFY message, a node records NOTIFY to the database. After node receive new DATA packet, if RA of DATA packet is the same as RA in the database, a node replaces RA of DATA packet to IA that coupled with RA in the database

72

4. Performance evaluation

We performed simulation to evaluate the validity of the proposed load-balancing scheme and compared its performances with IEEE 802.1 1 MAC by using NS2 network simulator [8] which is a freely available discrete-event object-oriented network simulator, which provides a framework for building a network model, specifying data input, and analyzing data output and presenting results.

4.1. Simulation configuration

Two network models were used for simulation. In the first one, 20 mobile nodes were configured to create an ad hoc network as shown in figure 3 . The second network is that consists with 50 mobile nodes and node’s position is randomly generated. In all simulation, we used CBR (Constant bit rate) traffic source, each CBR source generates packet every 0.05 seconds. The packet size is set to 1K bytes. For each simulation, we run simulation 10 times to avoid the bias of random number generation. The simulation time of the network model 1 is set to 70 seconds and that of the network model 2 is 100 seconds. In ad-hoc network model 1, we configured 5 CBR connections as described in figure 3 :

4.2.

To evaluate performance of the load-balancing scheme, we executed simulations while varying queue size from 10 to 100 packets. Figure 4 plots an average number of received packets.

Numerical results of 20 nodes model

73

2600-

2000 -

,600

,000

600

0 ,

We observe that our load-balancing scheme increases throughput up to 7 % as depicted in figure 4. Figure 5 and 6 depict average number of lost packets by collision and average number of lost packets by buffer over-flow. Load- balancing scheme provides better performance in lost packets by collision than IEEE 802.1 1 MAC protocol. The reason is that load-balancing scheme provides ability to change the path fkom congested node to neighbor node.

As shown in figure 5, the number of packets lost by collision is more decreased as much as increase queue size of nodes. It is because that the larger queue could contain more packets. The purpose of the load-balancing scheme is that distribute excessive load to neighbor nodes fkom a congested node. So load- balance scheme is able to reduce packet loss by collision.

Avemsa receIyed packets NYm8Cr of paclet.

-______

..................................................................................... ................................. ,

--..-

............................

....................................................................................................

~

............................................................................................................................. ..I I

J2W ....

........................................................................ 2900

-C MAC-80211 2800 ...................................................................................................................

2,110 .........................................................................................................

2- ........................................................................................

25w 10 10 JO 40 so 60 10 ao w 100

QY.W S W

Figure 4. Average received packet with variable queue size

74

----- ,0000

4.3. Numerical results of a random topology model

For evaluating performance of the load-balancing scheme in ad hoc networks, we use the second network model. This model is a random topology that consists of 50 mobile nodes and node’s position is randomly generated. Each node moves randomly by using the “setdest” command of ns2 simulator. Figure 6 plots average number of received packets with number of flows variation.

We observe that overall throughput increase as the number of flows increase. In all cases, IEEE 802.11 MAC with load-balancing protocol provides better throughput than IEEE 802.11 MAC. The load-balancing scheme is able to increase throughput up to 14%.

Averact9 -wad packet. N u m b s d padres

I It- ..... -. .. ./’ ................................................................................ ;..-c ................................... 1

.,-

Figure 6. Average received packets in a random topology Loat by solffiion

“umber d p.a.3

75

Figure 7, 8 depict the average number of lost packet by collision and by buffer over-flow. It shows that the load-balancing is effective in ad-hoc network which change topology by arbitrarily movement of mobile nodes. In general the results were quite positive in the sense that IEEE 802.11 MAC protocol with load- balancing outperformed than IEEE 802.1 1 MAC without load-balancing in most cases. Using our technique, load-balancing scheme may improve performance when the traffic load is heavy and should distribute excessive load of a congested node efficiently.

ll + (MO - U

1000 - /

800

BOO

400

200

0 ( 0 16 20 26 30 56 40

Number otllowm

Figure 8. Average ost packets by buffer overflow in a random topology

5. Conclusions

In this paper, we have presented an efficient load-balancing algorithm in IEEE 802.1 1 MAC for supporting QoS in mobile ad hoc networks. Load-balancing ability of our proposed algorithm is able to distribute excessive load of a node to its neighbors. Our algorithm takes advantage of distributing and efficiently using network resources (buffer space), reducing network congestion and increasing overall performance (throughput).

For evaluate validity of load-balancing algorithm, we compare performance with IEEE 802.1 1 MAC by simulation. Simulation results show that load- balancing algorithm distributes excessive load efficiently. And also it shows that load-balancing algorithm improves throughput and reduces packet loss.

It would also be interesting to relax some of the assumptions in our system model to investigate wider applicability of self-stabilization.

The study of QoS provision in MANETs can be an important area of future research. Load balancing algorithm would result in more efficient QoS provision in MANETs by cooperating prioritized queuing disciplines.

76

References

1. C. E. Perkins. Ad Hoc Networking. Addison-Wesley, Upper Saddle River, NJ, USA, Jan 200 1.

2. I . Chlamtac, M. Conti, J.J. Liu, “Ad Hoc Networks” vol. 1, no. 1, pp. 13 - 6 4 , Jan 2003

3. M.W. Subbarao, J. S. Pegon “Simulation Framework for a Mobile Ad-Hoc Network”, Proceedings of OPNETWORK 1999, Washington DC. Sept. 1999.

4. Elizabeth M. Royer, Charles E. Perkins. “An Implementation Study of the AODV Routing Protocol“ vol. 3, pp. 1003-1008, Proceedings of the IEEE Wireless Communications and Networking Conference, Chicago IL, Sept. 2000.

5. David B. Johnson, David A. Maltz, and Josh Broch. “DSR The Dynamic Source Routing Protocol for Multihop Wireless Ad Hoc Networks. In Ad Hoc Networking”, edited by Charles E. Perkins, chapter 5, pp. 139-172. Addison-Wesley, 200 1.

6. Vincent D. Park, M. Scott Corson. “A Highly Adaptive Distributed Routing Algorithm for Mobile Wireless Networks”. http://www.cs.odu.edu/ skovvuri/tora.pdf cited 10.04.2004

7. Charles E. Perkins, Pravin Bhagwat. “Highly Dynamic Destination- Sequenced Distance-Vector Routing (DSDV) for Mobile Computers“. In Proceedings of the SIGCOMM ’94, http://people.nokia.net/charliep/ txt/sigcomm94/paper.ps cited 0 1.03.2004

8. NS-2 homepage: http://www. isi. edu/nsnam/ns 9. Lahti Marja-Leena, “IEEE 802.11 Wireless LAN“, http://www.tml.hut.fi/

OpinnotITik- 10.55 1 /2000/papers/IEEE~8O2/wlan.html, 2000. 10. I. Aad and C. Castelluccia “Differentiation mechanisms for IEEE 802.1 1”

Proc. of IEEE INFOCOM 200 1, pp. 209-2 18. 11. M. Barry, A. T. Campbell, A. Veres, “Distributed Control Algorithms for

Service Differentiation in Wireless Packet Networks”, Proc. of IEEE

12. G. Bianchi, I . Tinnirello, “Analysis of Priority Mechanisms based on Differentiated Inter-Frame Spaces in CSMA/CA.”, in Proc. IEEE VTC 2003, vol. 3, pp. 1401 - 1405, Orlando (FL), Oct 2003.

13. IEEE 802.11 WG, IEEE Std 802.11e/D8.0, Draft Supplement to IEEE standard for Telecommunications and Information exchange between systems. Local and metropolitan area networks. Specific Requirements Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications, Feb 2004.

INFOCOM 200 1, pp. 582-590

A BANDWIDTH-EFFICIENT CROSS LAYER PROBABILITY ROUTING FOR MANETS *

XIAOMEI WANG HONGYI W CHONGSEN RAN XIA ZHANG WEINING QI Dept. of Communication Engineering, Information Science and Engineering Institute

Zhengzhou, 450002, P.R. China

A cross layer routing protocol is proposed for MANETs in this paper, which is called Bandwidth-Efficient Cross Layer Probability Routing (BECLPR). BECLPR differentiates between the send bandwidth and the receive bandwidth of the node. Based on the available send/receive bandwidth range provided by the MAC layer, a probability model is defined in BECLPR which is called Available-Bandwidth-Aware Probability Model of the Node (ABAPMN). By that each node processes the route request message with a given probability on its ABAPMN in the route discovery procedure, BECLPR probabilistically distributes the loads among the nodes according to the available sendlreceive bandwidth resources in MANETs. Also the output queue length of the node on the path is considered in searching for a path. Simulation results show that BECLPR outperforms the Ad hoc On-demand Distance Vector routing (AODV) in terms of the packet delivery ratio and routing overhead. BECLPR has lower end-to-end delay than AODV when the network load is not too high.

1. Introduction

Mobile Ad hoc Networks (MANETs) [ 11 are multi-hop wireless networks where all mobile nodes cooperatively maintain network connectivity without communication infrastructures for routing. Many different routing protocols have been proposed in the literature [2-51. However these algorithms consider the shortest path with the minimum hop count as the route selection criteria. This may lower the network utilization and increase the end-to-end delay, since some particular heavily loaded mobile nodes have little bandwidth resources to support the packet-relaying function. Several load-aware approaches have been proposed, which consider the effect of the loads of a node on the neighbors [6, 71. But they do not differentiate between the incoming loads and the outgoing loads of the node. In fact, the total amount of a node’s available send (receive) bandwidth is affected only by the incoming (outgoing) loads of neighbors and

* This work is supported by the National Natural Science Foundation of China (NSFC) (Approved No. 60472064) and the National High Technology Research and Development Program of China (863 Program) (Approved No.2003AA123340).

77

78

that of itself. Furthermore, as long as the available send (receive) bandwidth is adequate, a node can send (receive) data successfully even if the available receive (send) bandwidth is little.

A Bandwidth-Efficient Cross Layer Probability Routing (BECLPR) is proposed for MANETs in this paper. BECLPR differentiates between the send bandwidth and the receive bandwidth. A probability model is introduced in the route discovery procedure of BECLPR, which is made based on the node’s available sendreceive bandwidth range and is called the Available Bandwidth- Aware Probability Model of the Node (ABAPMN). By that each node processes the route request message (RREQ) with a certain probability based on its ABAF’MN in searching for a path, BECLPR is able to probabilistically distribute the loads among the nodes in the network and get the path with relative adequate available bandwidth resources. The output queue length is also considered in BECLPR. Simulation results show that BECLPR outperforms the Ad hoc On-demand Distance Vector routing (AODV) [4] in terms of the packet delivery ratio and routing overhead. BECLPR also has better end-to-end delay performance than AODV when the network load is not too high.

This paper is organized as follows. Section 2 presents BECLPR in detail. Performance evaluations of BECLPR are introduced in Section 3. Finally, concluding remarks are found in Section 4.

2. BECLPR

Due to its efficiency, AODV is adopted as the baseline routing algorithm throughout this paper.

BECLPR differentiates between the send bandwidth and the receive bandwidth which are used to send and receive data respectively. When a source node S has a flow for the destination node D, the routing tries to find a path with adequate bandwidth resources. Namely, S should have enough available send bandwidth; D has adequate receive bandwidth; the intermediate node(s) on the path should make both adequate. However it is very difficult to guarantee this due to the shared channel and the dynamics in MANETs. To handle this, three things are done in BECLPR. First, the total amount of the available receivekend bandwidth of the node is computed by the MAC protocol but not by the routing protocol, since the former is more aware of the availability of the bandwidth resources. Second, ABAPMN is adopted in the route discovery procedure to probabilistically achieve a path with relative adequate sendreceive bandwidth resources. Third, the output queue length of the node(s) is considered in searching for a path.

79

2.1. Computation of the available senrt/receive bandwidth of the node

The node in MANETs shares the wireless bandwidth resources with its neighbors. By analyzing the exposed and hidden terminal problem in detail, we realize that the total amount of a node’s available send (receive) bandwidth is affected only by the incoming (outgoing) loads of the neighbors and that of itself. In Figure 1 (a), the maximum data transmission capacity of the node is C. There are three flows. The corresponding used bandwidth is r l , r2 and r3. Now node X’s current available send bandwidth is C-rl-D, where r l is the used receive bandwidth of Y 1 for receiving the incoming flow and r3 corresponds to the used receive bandwidth of Y3. Similarly, X’s available receive bandwidth is C-r2-r3, where r3 is X’s used send bandwidth and r2 is that of Y2. From this example, we make two conclusions. First, the amount of the available send bandwidth of the node and that of the available receive bandwidth may be different. Evidently, it is good for routing to differentiate them. Second, the amount of the available send (receive) bandwidth of a node equals the value of subtracting the total amount of the used receive (send) bandwidth of all the neighbors and that of itself from the maximum data transmission capacity.

Figure 1. The illustration of node X sharing the wireless bandwidth resources with its neighbors.

Now let’s take a close look at the total amount of the used receive bandwidth of all the neighbors. Shown in Figure 1 (b), Y 1 has a data flow from Z1 and its used receive bandwidth is r l . Y2 is receiving from 22. Its used receive bandwidth is r2. Y1 is out of range of 22 and Y2 out of range of Z1. The two transmissions do not interfere with each other and may be transmitted simultaneously. So the total amount of the used receive bandwidth of all the neighbors (Y1 and Y2 in this example) of node X is the bigger one between r l and r2. When the nodes move about, another topology may form as shown in Figure 1 (c), where Y2 is in the range of Z1. The two transmissions interfere with each other and can’t be transmitted simultaneously. So the total amount of

80

the used receive bandwidth of all the neighbors of X is the sum of r l and r2. It is similar for computing the total amount of the used send bandwidth of the neighbors. As a result, the available sendreceive bandwidth range of node X is achieved shown in Eqs. (1) and (2) .

Where BZilab,e and Bkilable stand for the available send and receive bandwidth of node X, C is the maximum data transmission capacity, Nb(X) is the neighbor collection of node X, B2ed and BU2, are the used receive and send bandwidth of node i. Let M and N be the number of the incoming flows and that of the

outgoing flows of node i respectively. We obtain B2ed = c,=, B,

and B2ed = c,=, Bk , where Bk is the bandwidth required for transmitting flow

k. It pays to say that a flow forwarded by a node means that this node has an incoming flow and an outgoing flow.

(if

2" < 0 , let BZn = 0 ), we get BTLi,able E [B2n, B2m] . Also we may

get BZilable E [ B k , B L ] , where BZx = C - BTx - m a x e e d and

~2 = C - B ~ X used - CB~Y used (ifBzn < 0 , let Bzn = 0).

M

N

Let B2m = C- Bu2d - max Bu2d and ~ 2 " = c- Y S N b ( x )

Y G N b ( x )

Y E N b (x) used

YeNb(x)

2.2. ABAPMN

The ABAF'MN of node X is defined as follows:

P,' , X is the source node.

P," , X is the destination node. Px = P," . P,' , X is the intermediate node. (3) 1

Where PX is the probability for node X to process the RREQ message (Px~[O,l]), P,' andP' are the send probability and the receive probability as

81

presented in the following equations, in which B, is bandwidth requirement carried in the RREQ message.

2.3. Modifications made for BECLPR based on AODV

In original AODV, nodes don't consider their available bandwidth resources but forward the first received RREQ message definitely, while in BECLPR nodes process the RREQ message probabilistically according to their available receiveisend bandwidth range. Each node uses the probability computed according to Eq. (3) to process the RREQ message in the route discovery procedure based on AODV. Table 1 describes the basic operation behavior.

Table 1. The basic operation behavior in the route discovery procedure of BECLPR.

Node 1 Steps I Compute the probability Ps according to Eq. (3), do the following with the

Source node I probability Ps:

,

(node S) I Generate a RREQ message; Piggyback the output queue length of S into the fields: NqI, and Pql,; Broadcast this message.

If the output queue length is bigger than the value of the Nqlen field,

Add the value of the output queue length to Pqlen field;

Compute the probability Px, do the following with the probability Px:

replace the latter with the value of the output queue length of X; Intermediatenode

(node X)

........... ~ Establish the reverse route and rebroadcast ". this first received RREQ. j If it has not sent any RREP for the request with the same source-destination

Destination node ' pair, do the following with Po: (node D) If Nqlen I a*Qlmit (a is a constant coefficient), establish the reverse route ' and send RREP back to the source node S;

j Otherwise, add a record in the route list maintained in node D to record i the corresponding route information. i After a given interval of time, if no reply has been sent, the record with / least Pqlsn is selected from the route list and used to set up the reverse route I and send RREP to the source node.

The output queue length of the node(s) on a path is also considered in BECLPR. Two additional fields, called Nqlen and Pqlen, are added into the RREQ message. Nqlen is used to carry the maximum value of output queue length of the nodes on the path. Pqlen is to piggyback the sum of the output queue length of

82

each node on the path. Qlimit stands for the limit of the output queue length. The output-queue related operations are shown in Table 1. If the destination node receives a RREQ message with Nqlen no bigger than a*Qamit, it doesn’t need to wait for the arrival of other RREQ messages but replies to the source immediately. Otherwise, the Pqlen field is used to help BECLPR to get a better path, which usually occurs in the heavy network load condition. Even if the intermediate node has a route to the destination, it doesn’t reply to the source.


We used ns-2 [8] to simulate and evaluate the performance of BECLPR. We are interested in the packet delivery ratio, end-to-end delay and routing overhead.

The evaluations are based on the simulation of 40 mobile nodes moving about in an area of 800 X 800 sq. meters for 100 seconds of simulated time. The nodes move randomly according to the “random waypoint” model, in which the maximum moving speed of the node is 1 d s . The average pause time of nodes between two movements is 20s. The channel capacity is 2Mbits/s. Constant bit rate (CBR) is adopted as the data traffic. Packet size is 512 bytes. The sending rate of each CBR source is randomly set between 5pkts/s and 20pkts/s. The coefficient a in Table 1 is 0.8. A slight modification of IEEE 802.1 1 MAC DCF is adopted as the MAC protocol. Due to the MAC packets, routing packets and other overhead, the net channel capacity for transmitting data is far smaller than 2Mbits/s. Through simulations we found the maximum capacity for transmitting data is approximately 550kbitds and used this value for C in Eqs. (1) and (2).

Figure 2 compares BECLPR with AODV under different load conditions in terms of the packet delivery ratio. BECLPR has a much higher packet delivery ratio than AODV. As the number of CBR sources increases, the difference between them becomes apparent. Since original AODV does not,consider the available bandwidth of the node, the selected nodes may be congested while other nodes are very light loaded, thus many packets are droRped by the heavily loaded nodes. BECLPR probabilistically distributes the data flows to the nodes according their available sendreceive bandwidth, so more packets are delivered. The considering of the output queue length also avails to improve the performance of packet delivery ratio. Although BECLPR delivers more packets than AODV, it still outperforms AODV in terms of end-to-end delay when the network load is not too high (here the number of sources is smaller than 14.), shown in Figure 3. When the network load is so high that the load on each discovered path is also very heavy, the delay of BECLPR becomes higher than that of AODV.

83

a 0.5

0.4 2 4 6 8 10 12 14 16 18

Number of sources

Figure 2. The comparison between BECLPR and AODV in terms of the packet delivery ratio.

0.03 h

23 0.025 4 2 B 0.02

3 0.015

w" 0.01

a

.? a

0.005' ' ' ' ' ' ' ' ' 2 4 6 8 10 12 14 16 18

Number of sources

Figure 3. The comparison between BECLPR and AODV in terms of the average end-to-end delay

3oo00

5000

0 2 4 6 8 10 12 14 16 18 2 4 6 8 10 12 14 16 18

Number of sources Number of sources

(a) Normalized routing load (b) The comparison between the number of RREQ and RREP messages of BECLPR and that of AODV

Figure 4. The comparison between BECLPR and AODV in terms of the routing overhead.

Figure 4 (a) shows the routing overhead in normalized routing load. Normalized routing load is the ratio of the number of routing messages propagated by every node in the network and the number of data packets received by the destination nodes. As the number of the sources increases, the routing overhead of BECLPR retains low, while that of AODV is increasing.

84

This is due to two main causes. First, the number of the RRFiQ and RREP messages propagated in BECLPR is far less that of AODV, shown in Figure 4 (b). As the network load increases, the available sendreceive bandwidth of the node decreases, which causes some RREQ messages dropped by the nodes running BECLPR with a given probability. The number of the RREP messages of BECLPR also decreases, since only the destination responds to the source node. Second, the destination nodes under BECLPR receive more packets than those running AODV, since BECLPR has a higher delivery ratio than AODV.

4. Conclusions

A Bandwidth-Efficient Cross Layer Probability Routing for MANETs (BECLPR) is proposed in this paper, BECLPR differentiates between the send and receive bandwidth, which are computed by the MAC protocol. Based on the proposed probability model ABAPMN, each node processes the RREQ message with a given probability. BECLPR is able to probabilistically distribute the loads among the nodes in the network according to the available sendreceive bandwidth. Also the node’s output queue length is considered in the route discovery process. Simulation results show that BECLPR outperforms AODV in terms of the packet delivery ratio, end-to-end delay and routing overhead.

References

1 .

2.

http://www,ietf.org/html,charters/manet-charter.html. IETF mobile ad hoc networks (MANET) working group charter. David B. Johnson, David A. Maltz, and Yh-Chun Hu. The Dynamic Source Routing Protocol for Mobile Ad Hoc Networks (DSR). JNTERNET-DRAFT <draft-ietf-manet-dsr-lO.txt>, Internet Engineering Task Force, July 2004. C.E. Perkins and P. Bhagwat. Highly Dynamic Destination-Sequenced Distance-Vector Routing (DSDV) for Mobile Computers. Comp. Comm. Rev, Oct. 1994, pp.234-244. C. Perkins and E. Belding-Royer. Ad hoc On-Demand Distance Vector (AODV) Routing. RFC 3561, Internet Engineering Task Force, July 2003. T. Clausen and P. Jacquet. Optimized Link State Routing Protocol (OLSR). RFC3626, Internet Engineering Task Force, October 2003. H. Hassanein and A. Zhou. Routing with Load Balancing in Wireless Ad Hoc Networks. Proc. ACM MSWIM, Rome, Itary, July 2001. K. Wu and J. Harms. Load-Sensitive Routing for Mobile Ad Hoc Networks. Proc. IEEE ICCCN’Ol, Scottsdale, AZ, Oct. 2001.

3.

4.

5.

6.

7.

8. http://www,isi.edu/nsnam/ns/doc/ns-doc.pdf. December 13,2003.

Ad Hoc(I)

EFFICIENT BANDWIDTH ALLOCATION FOR BASIC

ADHOC MAC PROTOCOL " BROADCAST AND POINT-TO-POINT SERVICES IN THE

J.R. GALLEGO', L. CAMP ELL^, M. CESANA~, A. CAP ONE^, F. BORGONOVO~, A. HERNhJDELSOLANA', M. CANALES', A. VALDOVINOS

'Departamento de Ingenieria Electrbnica y Comunicaciones, University of Zaragoza, Maria de Luna I , 50018 Zaragoza, Spain,

' Dipartimento Elettronica e Informazione, Politecnico di Milano, Piazza L. Da Vinci 32, 20133 Milano, Italy

An effective Medium Access Control for communications in wireless Ad hoc networks should be able to support both broadcast and point-to-point communications paradigms. The ADHOC MAC protocol, recently proposed within the European Commission funded CarTALK2000 project, seems to match these requirements. As a matter of fact, it allows the exchange of connectivity information among wireless terminals which can be usefully exploited to devise both broadcast and point-to-point services. In this paper we evaluate through simulation the efficiency of the protocol in a mixed traffic scenario where broadcast and point-to-point communications coexist. An adaptive bandwidth allocation strategy is proposed to share the resources between both services in a dynamic situation. The capability of the protocol to establish parallel point-to-point data communications and the corresponding improvement in the point-to-point efficiency is also evaluated

1. Introduction

The transmission media in wireless environment has to be shared by definition. Further on, the radio resources are often limited in comparison with the number of users which access them, thus the capacity of any wireless network is highly determined by the capability of the medium access control mechanism to handle the access process and to achieve high resource reuse [ 11.

ADHOC MAC [2] is a medium access control protocol recently introduced within the European Commission funded CarTalk2000 project [3] for providing connectivity in ad hoc inter-vehicles networks [4]. ADHOC MAC works on a slot synchronous physical layer and implements a completely distributed access technique capable of dynamically establishing a reliable single-hop Basic

This work is supported by projects fiom CICYT and FEDER, TEC2004-04529/rCM.

87

88

broadcast CHannel (BCH) for each active terminal, i.e., each transmission within a BCH is correctly received by all the terminals within the transmission range of the transmitter. Each BCH carries signalling information that provides a prompt and reliable distribution of layer-two connectivity information to all the terminals. This information provides a valuable basis for the efficient implementation of point-to-point data services, exploiting parallel transmissions, and also supplies a prompt means to manage different QoS requirements for these services, through the use of priorities.

In [5] and [6] we have studied the performance of ADHOC MAC broadcast services in a static scenario and with users’ mobility respectively. In this paper we evaluate through simulation the efficiency of the protocol in a mixed traffic scenario where broadcast and point-to-point communications coexist. An adaptive bandwidth allocation strategy is proposed to share the resources between both services in a dynamic situation. The goal of the proposal is to guarantee access requirements for BCH whereas capacity for extra data communications is optimized. The capability of the protocol to establish parallel point-to-point data communications and the corresponding improvement in the point-to-point efficiency is also evaluated. The remaining paper is organized as follows. In Section 2 we briefly summarize the basis of the ADHOC MAC protocol and the proposed bandwidth allocation strategies for basic broadcast and point-to-point services. In Section 3, both the resource sharing strategies and the point-to-point service efficiency are evaluated through simulation. Finally, in Section 4 some conclusions are provided.

2. The ADHOC MAC Protocol

2.1. Basic Operation Mode for BCH and point-to-point Communications

ADHOC MAC operates with a time slotted structure, where slots are grouped into virtual frames (VF) of length N, and no frame alignment is needed. In the BCH, each terminal broadcasts information on the status of the channel it perceives. The BCH contains a control field, namely, Frame Information (FI) field, which is an N-elements vector specifying the status of the N slots preceding the transmission of the terminal itself. The slot status can be either BUSY or FREE: it is BUSY if a packet has been correctly received or transmitted by the terminal, otherwise it is FREE. In the case of a BUSY slot the FI also contains the identity of the transmitting terminal.

89

Figure 1. Example of the FI information propagated by the terminals 1-7 in the one-hop clusters A, B, and C represented by ellipses.

Based on received FIs, each terminal marks a slot, say slot k, either as RESERVED, if slot k-N is coded as BUSY in one FIs received in the slots from k-N to k-1 at least, or as AVAILABLE, otherwise. If a slot is AVAILABLE, it can be used for new access attempts. Upon accessing an AVAILABLE slot, terminal j will recognize after N slots (a frame) its transmission either successful, if the slot is coded as "BUSY by terminal j" in all the received FIs, or failed, otherwise. In figure 1, an example of FIs transmitted by a set of terminals is given. The union of all one-hop (OH) clusters with a common subset is denoted as two-hop (TH) cluster. The terminals belonging to the same OH-cluster see the same status (AVAILABLE or RESERVED) for all the slots; terminals belonging to different OH-clusters of the same TH-cluster mark as RESERVED all the slots used in the TH-cluster, whereas terminals belonging to disjoint OH-clusters usually see a different channel status. As a result, slots can be reused in disjoint OH-clusters, but can not be reused in the same TH- cluster and, therefore, the hidden-terminal problem can not occur [4]. The BCH provides a reliable single hop broadcast channel which can be used both for signaling and for data traffic purposes. Upon this basis, point-to-point data communications among terminals can be effectively established by exploiting the distributed signaling provided by the FIs. To this end, each entry of the FI encloses a PointToPoint (PlT) flag, which is handled as follows:

A terminal sets the PTP flag of a given slot in the FI, if the packet received in the slot is a broadcast one or if it is destined to the terminal itself.

90

Figure 2. Examples of parallel transmissions. Transmission fiom terminal I is established first. Allowed transmissions by terminal 2 are indicated by solid arrows.

In order to set up a point-to-point communication, all the AVAILABLE slots can be used. Further on, even some RESERVED slots can be used according to the following rule:

A RESERVED slot can be accessed if: 1. The PTP flag is signaled off in all the received FIs and 2. The FI received from the destination terminal signals the slot as FREE.

The conditions above allow point-to-point transmissions to share the same slot when there is no collision at the receivers. This can be seen referring to the four cases shown in figure 2. The cases a and b in the figure consider two transmitting terminals, say 1 and 2, belonging to different not disjoint clusters. Assuming that terminal 1 has already activated a PTP channel with 3, terminal 2 can transmit using the same slot if these two conditions above are satisfied. In case a, terminal 2 can use the same slot as terminal 1 even if it is signalled as RESERVED. In fact, the only PTP flag ON is that in the FI transmitted by terminal 3 and not received by terminal 2 (satisfylng condition (l)), and the FI generated by terminal 4 marks the slot as FREE (satisfjmg condition (2)). In case b the FI, generated by terminal 3 and received by terminal 2, prevents terminal 2 from transmitting (not satisfylng condition (1)). In this case parallel transmission would, in fact, interfere at terminal 3. In cases c and d the two transmitting terminals belong to the same cluster. In case d terminal 3 can use a RESERVED slot since both conditions (1) and (2) are satisfied (in fact, this is the exposed-terminal case) whereas in case c condition (2) is not satisfied and a collision would occur at terminal 4. If several access attempts occur concurrently, collisions can still occur. Then, the transmitting terminal has to perform a further check according to the following rule:

91

The point-to-point transmission is successful if the slot is coded as BUSY in the FI of the destination terminal; otherwise the transmission is failed.

2.2. Bandwidth allocation strafegies

Once a terminal has acquired its BCH channel, it can establish additional broadcast data communications if the data payload in the BCH is not Iong enough. In the same way, different PTP data communications with all its neighbours can be established. In this paper, only extra PTP communications are considered, so in the remaining of the paper these additional communications are referred as PTP. However, the proposed strategy can be generalized since dimensioning is made only according to BCH demands.

In its basic operation mode, every slot in the frame can be used for both PTP and BCH transmissions. In this situation, as the number of PTP communications grows, the number of AVAILABLE slots for new terminals accessing the system decreases, leading to a reduction in the number of terminals that can access the system for a given number of slots. As the acquisition of a basic broadcast channel is mandatory to access the system, an appropriate dimensioning of the network must guarantee certain resources for BCH transmissions. As a metric for the BCH performance, we take into account the outage probability. A terminal is declared in outage if it does not acquire a BCH within a period of a given number of frames after birth. According to this situation, it must be guaranteed a trade-off between ensuring an acceptable outage probability for BCH channels while providing the maximum throughput for PTP data communications. In order to guarantee an outage probability for new terminals accessing the system, we propose a frame subdivision into two subframes, where the performance of BCH is not limited by the amount of PTP traffic in the network: A frame with N slots is divided into NBCH and Npr~ slots for BCH and PTP communications.

(1) For this assumption, it is required a slot and frame time synchronization of

each terminal in the network, that can be obtained with the Global Position System (GPS) or other solutions [7], [S]. With this subdivision, the probability of access the system is higher. When a terminal tries to access the system, it looks for an AVAILABLE slot. The existence of an AVAILABLE slot for a new terminal can only be statistically guaranteed: if the neighbours have enough FREE slots, it is probable that there is a common FREE slot for all of them. The frame subdivision brings together the FREE BCH slots making more probable for a terminal to find an AVAILABLE slot.

N = NBCH + NPTP

92

If a static frame subdivision is considered, NBCH limits the maximumdensity of terminals supported by the system. A lower density of terminalsimplies that resources are wasted, since extra FTP communications could beallocated in the free BCH slots. On the other hand, when the density grows overexpected, terminals declared in outage could access the system using slots ofthe FTP subframe. To overcome these limitations, an adaptive subdivisionstrategy that moves the border between the slots dedicated to each type of trafficwithin the frame according to the channel dynamics is proposed and evaluated:A set of W possible values for NBCH [Ni < N2 <...< NW} is defined. Terminal /chooses the value NBCH,i within this set according to the density of neighbours,Pi, it observes. We have considered two possibilities to measure this density:

pt = NB, (2)

where NBt is the set of neighbours for the terminal i, and its dimension |MJ,-| isequal to the number of BCH channels received by this terminal.

Pi =1

m\+i\ jeNB,(3)

Equation (3) represents the mean number of neighbours in thesurroundings of terminal ('. This can be obtained through the FI informationtransmitted by each neighbour of terminal / and its own observed neighbours.According to this density, each terminal updates the NBCH.I value every frameand includes it in the FI transmitted to all its neighbours.

N BCH.i

if />,</*,if /A, < PI < th2

(4)

w-\ <thw-\w if A

where tfy represents the maximum density of terminals tolerated for a numberof slots NJ of the BCH subframe.

In the same way that a terminal i sends this value, NBCH,I, it receives thecorresponding NBCH values of all its neighbours. Since it must be guaranteedthat FTP communications can not be established in any of the BCH subframesof the neighbours, the number of slots where terminal / can establish FTPcommunications, as a transmitter, is given by

NFTP-JXJ =N- maaNgcaj) (5)

93

whereas the subframe where it can receive PTP communications is limited just by its own NBcH,i

NPTP-Ry,i = - NBCH,i (6)

P , < P 2 < P 3 ~ N B C X , i N B C I I . I i N B C H , 3

IFWconn-

Figure 3. Example of how PTP transmissions can be established according to the subhme division

An example of how PTP transmissions can be allocated according to the resource sharing algorithm is given in figure 3. In that situation, terminal 2 cannot establish a PTP communication as a transmitter with terminal 1 in that position, even if the slot is AVAILABLE, because this slot belongs to the BCH subframe of terminal 3. However, terminal 1 can transmit to terminal 2 in the same slot, since this transmission does not affect to terminal 3. This slot belongs to the set of NpTp-Ry,2, but not to Nm-m,2.


In order to evaluate the proposed resource sharing strategies and the performance of the point-to-point services provided by ADHOC MAC, we have built up an event driven simulator which implements all the functionalities of the medium access control protocol. Since we mainly focus on performance evaluation of the medium access control protocol, as first step of analysis, we simplifj the physical layer assuming neither fading nor shadowing in the calculation of the received power. The connectivity among terminals is simply determined by their respective distances and no power control procedures are implemented. As consequence, a transmission either broadcast or point-to-point can be erred due to collisions only.

3.1. Bandwidth allocation evaluation

The bandwidth allocation strategies are evaluated in a dynamic situation, where terminals are generated within the network according to a Poisson process with average rate Y [new terminals/s]. Each active terminal has a lifetime random variable exponentially distributed with mean L=500 [frames], thus the

94

parameters Y and L define the offered traffic of the basic broadcast service. Terminals are randomly positioned within a square area with edge equal to 1Km. Under these conditions, PTP communications are generated according to a Poisson process with intensity X [PTPconnections/s]. The source of each point-to-point communication is randomly chosen among the users with an active BCH, while the destination is randomly chosen among the source’s neighbours. The duration of each point-to-point communication is exponentially distributed with mean D [frames]. The parameters X and D define the point-to-point offered traffic. We define a common framework of simulation by setting the length of a frame F = 100 ms, the number of slots within a frame N = 30, the coverage radius R = 100 m and the point-to-point communications mean duration D = 50 frames. The modification of the simulation parameters only impacts on the absolute values of the performance figures, whereas the comparative results obtained in this paper still hold.

. . 0 0.1 0.2 0.3 R4

O u o d B c H & . G b ~ a . N . & ! / I , , , l l k W A ! ! ! ! . . , ,

1 2 3 4 5 6 7 8 9 10

M,.m ..l*rd..khL.m

( 4 (b) Figure 4. Outage probability for BCH transmissions versus the BCH offered traffic when varying the intensity of PTP traffic (a) and versus mean number of neighburs with static subdivision (b). Standard Setting of simulation parameters and basic operation.

In the basic operation of ADHOC MAC, each slot can be used for both BCH and PTP communications. Under these conditions, figure 4a reports the outage probability for the broadcast channels versus the offered broadcast traffic, defined as the density of BCH channels per slot, when varying the point-to-point traffic intensity. The number of frames a terminal tries to acquire a BCH is set to 10. As it was expected, the outage probability of broadcast traffic increases as the point-to-point traffic grows, since few slots are available for accessing a BCH, and systematic collisions happen.

Capacity in the system can be considered as BCH capacity, referred to the capability of accepting users in the system and PTP capacity, as the extra bandwidth for data communications. Results from figure 4 confirm that a more efficient resource management is necessary to share both capacities. The static frame subdivision in NBcH and Nm slots tries to guarantee at least the access to the system for new terminals. Figure 4b shows the outage probability for BCH

95

channels with this static subdivision for several values of NBCH (10, 15, 20, 25 and 30). It is represented versus the mean number of neighbours in the network, which is directly related to the offered broadcast traf3ic. If a specific value, for example 0.01, is considered as an acceptable limit for outage probability, figure 4b shows, that according to the number of terminals in the network, the minimum NBcH that guarantees this requirement changes. Since in a real situation the density of users in the network is not supposed to be known, the use of a dynamic frame subdivision tries to optimize the network dimensioning making the allocation density of terminals.

locally, according to the geographical

a

7.5

I ? " 15

5

0

Figure 5 . Outage probability for BCH transmissions (a) and mean number of slots allocated for BCH and PTP (b) versus the BCH offered traffic with dynamic subdivision using the number and the spatial mean number of neighbours. Standard Setting of simulation parameters.

I Eal

I il

a g 1En2

I Eal

I il g 1En2

5 1

I E a 3

8 1 U

am 0 1 015 0 2 an 03 a35

ofbd m d r G b r d d t l a l am ai au az au 03 m OmredBCHtrulb G(wtnhn.lilol)

(a) (b) Figure 6. Outage probability for BCH transmissions (a) and mean number of slots allocated for BCH and PTP (b) versus the BCH offered traffic wth dynamic suMvision (mean number of neighbours) for different sets of thresholds. Standard Setting of simulation parameters.

4 5

I E a 3

8 1 U

am 0 1 015 0 2 an 03 a35

ofbd m d r G b r d d t l a l am ai au az au 03 m OmredBCHtrulb G(wtnhn.lilol)

(a) (b) Figure 6. Outage probability for BCH transmissions (a) and mean number of slots allocated for BCH and PTP (b) versus the BCH offered traffic wth dynamic suMvision (mean number of neighbours) for different sets of thresholds. Standard Setting of simulation parameters.

In order to make this dimensioning, results from figure 4b has been used as a reference to establish the values thj with 1 < j < W, which determine NBCH. The set of NBCH values chosen for this dynamic subdivision is {N1=lO, N2=15, N3=20, N4=25, N~=30} and, according to figure 4b, the chosen set of th is {thl=3, th2=5, th3=7, th4=9}. The decisions are taken according to (4). Figure 5 shows the performance of the adaptive algorithm. Using the mean number of

96

neighbours according to (3) clearly outperforms the use of only the own number of neighbours (2). With (3), outage probability is lower and more stable for different densities of users, as it is shown in figure 5a. This is confirmed in figure 5b, where the mean number of slots allocated for BCH is higher with (3). Moreover, the number of slots allocated for PTP transmissions is also higher, i.e., the differences between Nm-m and Nm-Ry are reduced. The use of the mean number of neighbours allows a terminal to adapt its subframe borders to the variability of the density of neighbours in its surroundings. For example, the loss of a single neighbour, which can be a result of multiple factors (movement, fading, loss of battery, temporal switching off.. .) has a lower effect over the measured density than if just the own number of neighbours is used, making the frame subdivision more stable. Through the variation of the values of the set of th, it is possible to adjust the balance between BCH and PTP communications, according to the BCH requirements. Figure 6 shows similar results for three different sets of thresholds, (3, 5 , 7 , 9 } , (3.5, 5.5 ,7 .5 , 9.5) and (4, 6, 8, lo}. In order to guarantee an outage probability for BCH around 0.01, the set (4, 6, 8, lo} could be enough, whereas it is the option that provides a higher bandwidth for PTP transmissions.

3.2. Analysis of the point-to-point efficiency

Once the outage probability for new accesses is guaranteed by means of the proposed bandwidth allocation strategy, the remaining PTP capacity must be efficiently managed. As a first step, a maximum theoretical value for this capacity has been obtained through simulation. These results are valid as a boundary for the maximum capacity provided by the protocol, but they will not be reachable in a dynamic situation where the BCH requirements limit the actual PTP capacity.

In order to analyze the maximum point-to-point capacity without interacting with the BCH tra€lic, the simulation has been carried out considering a stationary broadcast situation where all the terminals have an active BCH. For this purpose, a number of terminals that will be active through all the simulation is generated at the beginning of the simulation. This number defines the offered traf€ic of the basic broadcast service. Upon generation, each terminal tries to acquire a BCH, thus after a certain transient time, all of them have acquired their BCH and a stationary scenario is arranged. On the other hand, simulations in a dynamic situation have been also carried out. For these simulations, the adaptive algorithm with the set of thresholds (4,6, 8, lo} and the spatial mean number of neighbours have been used, since according to

figure 6 they provide the higher available bandwidth acceptable outage for BCH.

97

for PTP with an

Figure 7. Maximum PTP throughput versus the BCH offered traffic with and without the PTP flag in the FIs. Standard Setting of simulation parameters. Stationary and Dynamic BCH traffic.

Figure 7 reports, for both scenarios, the maximum throughput of point-to- point communications, defined as the maximum density of successful PTP transmissions per slot, versus the broadcast offered traffic, using the standard setting of simulation parameters. For all cases, the amount of PTP offered traffic is high enough to fully occupy the available resources. Two curves are reported for each situation: the dotted one depicts ADHOC MAC workmg mode with the use of PTP flag in the FIs, while the other one refers to the simplified case where the PTP is not used, i.e., only the AVAILABLE slots can be accessed by point-to-point traffic.

The ADHOC MAC provides higher reuse due to parallel point-to-point communications on the same slot when using the PTP flag, as it solves the exposed terminal problem. As the broadcast offered traffic increases, the use of this flag in the FIs provides an increasing gain with respect to the case where it is not used. As a matter of fact, high broadcast offered traffic means high terminals density, and consequently high probability of exposed terminal situations. Further on, in a stationary scenario, the maximum point-to-point throughput decreases if the broadcast offered traffic keeps increasing above the value 0.2 (terminals/area/slot), since there are less AVAILABLE slots for point-to-point. Below that value the point-to-point throughput increases with the broadcast offered traflic, since the number of point-to-point communications which can be set up is limited by the low number of terminals within the network. In a dynamic situation, since some slots must be kept FREE within the BCH subframe in order to guarantee the access for new terminals, the maximum achievable throughput is lower than in a static scenario, where only the set of BCH channels already acquired by the current terminals is not AVAILABLE for PTP connections.

98

4. Conclusions

In this paper, resource sharing strategies for basic broadcast channels and point-to-point data communications have been proposed and evaluated through simulation. Basic broadcast tr&c and point-to-point communications can efficiently share the total resources with a frame subdivision that allocates separately both services. Moreover, in a dynamic situation, the proposed adaptive strategy, that makes this allocation according to local densities of terminals provides a trade-off between both services guaranteeing the access requirements (outage probability). Regarding point-to-point communications, the performance of the protocol can be improved by means of the slot reuse provided by the parallel transmissions that can be allocated using the PTP flag. Upon these results, the management of point-to-point services with different QoS requirements through the use of priorities will be evaluated in coming works.

References

1.

2.

3. 4.

5.

6.

7.

8.

T. Rappaport, Wireless Communications: Principles and Practice, Prentice Hall, New Jersey (1996). F. Borgonovo, A. Capone, M. Cesana, L. Fratta ADHOC MAC: a new MAC Architecture for ad hoc Networks Providing Efficient and Reliable Point-to-Point and Broadcast Services. Wireless Networks. 10 (2004) 3 59- 366 CarTALK2000 Home page: www.cartalk2000.net M. Aoki, Inter-vehicle communication: technical issues on vehicle control applications. IEEE Communication Magazine. 34. (1996) 90-93. F. Borgonovo, L. Campelli, M. Cesana, L. Coletti, MAC for ad-hoc inter- vehicle network: service andperformance. In: Proc. IEEE VTC fall 2003, Orlando, USA. 5. (2003) 2789-2793 F. Borgonovo, L. Campelli, M. Cesana, L. Fratta, Impact of User Mobility on the Broadcast Service Eflciency in ADHOC MAC Protocol. In: Proc. IEEE VTC 2005 Spring, Stockholm Sweden (2005), to appear. A. Ebner, H. Rohling, R Halfmann, M. Lott, Synchronization in ad hoc networks bused on UTRA-TDD. In: Proc. IEEE PIMRC 2002, Lisboa, Portugal (2002) A. Ebner, H. Rohling, M. Lott, and R. Halfmann, Decentralized slot synchronization in highly dynamic ad hoc networks. In: Proc. WPMC 2002, Hawaii, USA (2002)

CONNECTIVITY AWARE ROUTING IN AD-HOC NETWORKS

JCREMIE LEGUAY, TIMUR FRIEDMAN, SERGE FDIDA Universite' Pierre et Marie Curie, Laboratoire LIP6 - CNRS

8 rue du Capitaine Scott, 75015 Paris - France Tel: -1-33 I 44 27 71 34, Fax: +33 I 44 27 53 53

Email: Cfirstname.name) @lip6.fr

VANIA CONAN, ANDRE COTTON Thales Communications

160 bd de Valmy - BP 82,92704 Colombes cedex - France Tel: +33 1 46 13 22 16, Fax: +33 I 46 13 26 68 Email: Cfirstname.name) @frthalesgroup.com *

Ad-hoc networks that use the IEEE 802.1 1 MAC layer suffer from severe performance issues because nodes compete to access the wireless channel. In such a context, the network topology has a great influence on the overall network performance. In this paper, we present a connectivity aware QoS routing framework that takes routing decisions with regards to the local characteristics of the network topology. The advantages of the approach is to rely solely on observations made by each node locally and applies with existing MAC layers.

1. Introduction

Ad-hoc networks allow the spontaneous set up of communication systems when deploying an infrastructure is a non-trivial task or may take too much time. An ad-hoc network is composed of several mobile nodes sharing a wireless channel without centralized control or an established structure. Furthermore, all the nodes communicate only with the ones within their transmission range. As a consequence, nodes need routing capabilities to allow multi-hop communication and the topology is expected to change frequently.

This work is situated in the context of an ad-hoc network using the popular IEEE 802.1 1 MAC layer. In such a network, all nodes compete to access the same wireless channel. Network topology thus has a strong impact on performance.

*this work is funded by Thales Communications and the European Community through the European project WIDENS (http://www. widens .org) .

99

100

Indeed, the geographic positions of nodes greatly influence the ambient level of interference and the level of competition between nodes. We present in this paper a QoS routing framework that we use to investigate the benefit of using connectivity as the metric for routing.

The rest of the paper is structured as follows. Sec. 2 provides some simulation results that motivate this study. Sec. 3 describes the QoS routing framework that we use. Sec. 4 gives an overview of existing work on connectivity metrics and describes the ones we have chosen for this study. Sec. 5 presents the simulation results and Sec. 6 a discussion around this work. Sec. 7 concludes the paper.

2. Motivations

As a preliminary study, we ran simulations to determine the impact of the local level of connectivity on the local network performance. We analyzed the performance of an ad-hoc network composed of fixed nodes placed on a 100 square meter playground having a radio range of 250m. This means that for this simulations any node is at a one hop distance from any other node. They all compete for the same channel. A number of Constant Bit Rate (CBR) connections at the rate of 4 packets per second are established between pairs of nodes. The packet size is 512 bits and the routing protocol is AODV'. We used the network simulator 1x2' with nodes having the 802.1 1 MAC layer at 2 Mbits with the RTSKTS mechanism. In order to evaluate the overall performance, we measured, for data packets, the average delay and the packet delivery ratio, i.e the ratio between the number of received packets and the number of sent packets.

Fig. 1 shows average results for 30 instances of the experiment. Fig. l(a) shows that the delay increases both with the number of nodes and the amount of traffic. Despite the fact that the amount of traffic seems to have a more severe impact on the delay than the number of nodes, the effect of the number of nodes is not negligible. Note that even if a node does not send traffic, it runs the routing protocol, which needs to periodically send packets toward its neighbors. The effect of higher connectivity is also visible for the packet delivery ratio (see Fig. l(b)). For a constant amount of traffic, the performance decreases with an increase in the number of nodes involved in the scenario. Thus, when the connectivity is high, competition increases dramatically, meaning that the delay becomes higher due to contentions, which results in a poor use of network resources.

3. The QoS Routing Framework

Since connectivity has a strong impact on network performance, our goal here is to use connectivity as a constraint for routing to improve network utilisation. To

101

..S

3 5

2.5

1 5

0.5

(a) Evolution of the delay (b) Evolution of the delivery ratio

Figure 1 . Influence of the connectivity level on network performance.

our knowledge, connectivity has never been the focus of a QoS routing study. We decided to focus our work on the integration of QoS routing with the proac-

tive link-state protocol OLSR 3. In OLSR, all the nodes are aware of a subset of all links and use Multi-Point Relays (MPR) to minimize the amount of control traffic. The heuristic for the MPR set selection can be changed to reserve link advertizement to those having certain properties. Munaretoo et al. and Ge et al. change the heuristic to make the OLSR route computation algorithm able to find routes having a good level of available bandwidth and delay. In our case, we simply use the Dijskstra algorithm 7, with links weighted according to our chosen connectivity metric. The metrics are combined using the additive operator. We plan to study multiplicative combination in future work.

In regular constraint based routing protocols, there is a need to propagate and to maintain QoS metric values. In our case, we rely on OLSR to maintain an image of the topology of the network. Thus, no additional network overhead is generated since no QoS information needs to be exchanged between nodes. Furthermore, we do not incur the measurement costs of classical metrics such as delay or bandwidth, which can be non-trivial 516.

4. Connectivity Metrics

This section presents an overview of connectivity metrics used in the literature and the ones we used for this study.

Several studies related to ad-hoc networks have dealt with connectivity metrics as a parameter to vary for 'simulation scenarios. There have been the k- connectivity ', the number of shortest pathsg, the node densityg, and the directed

102

connectivityg. Lets use the following notation. The computation of the connectivity metric

v(a,b) attached to link ab is usually done on the undirected graph of its k-hop neighborhood. A node j is in the k-hop neighborhood of link ab if there exists shortest paths from j to a and from j to b that have a length lower or equal than k hops. Also, a link cd is in the k-hop neighborhood of link ab if c or d is in the k-hop neighborhood of link ab.

In this study, because they can be computed efficiently, and because they appear relevant, we choose to use the following metrics:

The k-hop node density: It represents the number of nodes in the k-hop neighborhood. The k-hop link density: It represents the number of links in the k-hop neighborhood. The clustering coefi cient': It represents the probability that two neighbors of a node are connected. The clustering coefficient of a node u is defined by C(u) = k u ~ ~ ~ - l , . E, is the number of existing links between the neighbors of u and k, is the number of neighbors of u. The k-hop beta index": ,6 = $ with E the number of edges and V the number of nodes in the k-hop neighborhood.

We have conducted experiments with k equal to 1 and 2. Using a value of k greater would not have made sense in this study because, for computational reasons, we did not perform our simulations on networks having a very large diameter.

5. Evaluation

To evaluate the connectivity aware QoS routing scheme described above, we ran graph oriented simulations to understand the metrics behavior and network oriented simulations to measure their benefit in term of performance.

5.1. Metric performances

To evaluate the connectivity metrics we have implemented a stand alone simulator that measures several properties:

Path length injution: the difference in terms of number of hops between paths and minimum hop count paths. It should remain small since inflation in ad-hoc networks impacts performance.

103

Routing discrimination level: the difference between the average metric value along paths found by the routing algorithm and along minimum hop count paths. This measures how well the new routing scheme performs in finding different routes than the minimum hop count.

0 Path stability: the number of path changes that occur in a certain amount of time.

We consider a network composed of 200 nodes having a radio range of 250m on a square playground 20OOm large. We analyze the behavior of the metrics on three types of network graph, one with a low node degree variance, one with a medium node degree variance and one with a high variance. We have studied these three cases to artificially create three connectivity patterns.

This study aims to discover relevant metrics that imply a low frequency of routing changes, and that give a good level of discrimination without providing a high level of inflation.

(a) Average path length inflation (b) Average routing discrimination levei

Figure 2. Metric performances

Fig. 2(a) shows the average path length inflation. We can see that beta-lhop does not produce any inflation. The other metrics engender an inflation that increases with the node degree variance of the graph, which is normal. Only density-nodes2hops does not obey this law. We can also see that density-links-2hops and clustering introduce significant inflation compared to the other metrics.

104

Fig. 2(b) shows the average routing discrimination level. All the metric values have been normalized in order to be compared. We can see that they all have a quite high power of discrimination except for beta-lhop. More generally, the discrimination increases with the node degree variance of the graph.

Fig. 3 shows the number of routing changes that occurred for each metric. We considered that nodes move according to the random waypoint mobility model with a maximum speed of 1 0 d s and a maximum pause time of 5s during 300 seconds. The routing tables are dumped every second. We observe that a11 the metrics involve more route changes than standard OLSR. We note in particular that clustering engenders much a larger number of route changes than the others.

From the present results and studies we can conclude that clustering and densityJinks-2hops suffer from undesirable properties. The others seem to give better results.

Metric I Number of route changes

hop count I 82.99 densi tynodes-1 hop densi tynodeszhops densi ty-links-lhop densi tyJinks2hops clustering betalhop beta2hoDs

109.047 109.849 110.366 109.223 140.723 128.592 123.094

Figure 3. Route stability.

5.2. Network simulations

Fig. 4 shows simulation results for 30 nodes obtained with the help of the click router for a network where nodes use the IEEE 802.11 MAC layer at the bitrate of 2 Mb/s. We measured the average delay and delivery ratio for data packets in networks having different connectivity patterns (low, medium, high as in Sec. 5.1) and we varied the number of CBR connections (same as in Sec. 2) chosen at random between nodes from 10 to 90.

We can see that when traffic is low, all the metrics perform the same. Whereas when traffic is high, the average delay is lower and the delivery ratio higher when conectivity metrics are used. Futhermore, we observe that one of the simplest connectivity metrics, density-links-lhop, clearly out-performs the others when traffic is high and especially when the connectivity level is high. For instance, when the connecitivy pattern is medium and when the number of CRR connections is 90, densityJinks-lhop leads to a diminution of 8.77% for the average delay and a gain of 3.65% for the delivery ratio.

linked to the network simulator ns2

105

0.01

h m v

2 8 - 0.m

0.006

.

metric

(a) Average delay (10 CBR)

metric

@) Average delivery ratio (10 CBR)

metric metric

(c) Average delay (90 CBR) (d) Average delivery ratio (90 CBR)

Figure 4. Average metric performances on 5 instances of the experiment @lack bars are for low connectivity patterns, grey for medium and white for high ones).

6. Discussion

We found that using connectivity as a metric for routing is only interesting when the amount of traffic is high and becomes even more relevant when the network connectivity is disparate. This may introduce the need for a hybrid system working as follows. One can extend the QoS framework to handle multiple constraints: the number of hops, the connectivity level, etc., and can imagine that nodes are observing the variance of the node degree to take the decision whether or not to integrate the connectivity level in the routing decision.

Such a routing scheme may introduce a lack of diversity in the routing decisions. This can be solved by adding some randomness in the choice of the routes. For instance, for a given destination, a node can choose a subset of possible next

106

hops and distribute the traffic among them. Regular QoS routing protocols suffer from two phenomena: the overhead in-

duced by the additional exchanges of QoS information between nodes and the self-interference caused by the fact that the routing decisions have an impact on network resource availability, which can lead to flapping (as with bandwidth for instance). However the connectivity metrics used for the route computation in this work are calculated on the graph already maintained by OLSR, thus no real- time network performance information needs to be measured and no additionnal information has to be exchanged between nodes.

7. Conclusion and future work

This paper has highlighted some interesting properties of ad-hoc networks related to the network connectivity. It has presented a QoS routing scheme that benefits from the network topology, using connectivity metrics. We have shown that it improves the network utilization when the amount of traffic is high and even more when connectivity level is not constant through the network.

References 1. E. Belding-Royer C. Perkins and S. Das, “RFC 3561 : Ad hoc on-demand distance

vector (AODV) routing, July 2003,”. 2. “The network simulator ns-2. http://www.isi.edu/nsnam/ns/,” . 3. T. Clausen and P. Jacquet, “RFC 3626 : Optimized link state routing protocol (OLSR),

October 2003,” . 4. Amir Qayyum, Laurent Viennot, and Anis Laouiti, “Multipoint relaying: An effi cient

technique for fboding in mobile wireless networks,” Tech. Rep. Research Report RR- 3898, INRIA, February 2000.

5. K. A1 Agha H Badis, “An efficient QOLSR extension protocol for QoS in ad hoc networks,” in IEEE VTC’O4-Fal1, September 2004.

6. L. Lamont Y. Ge, T. Kunz, “Quality of service routing in ad-hoc networks using OLSR,” The 36th Hawaii International Conference on System Sciences (HICSS-36), 2003.

7. E. W Dijkstra, A note on two problems in connexion with graphs, Numerische Mathe- matik, 1959.

8. Christian Bettstetter, “On the minimum node degree and connectivity of a wireless multihop network,” in Proceedings of the 3rd ACM international symposium on Mo- bile ad hoc networking & computing. 2002, pp. 80-91, ACM Press.

9. Andree Jacobson, “Metrics in ad hoc networks,” M.S. thesis, Lulel University of Tech- nology, LuleB, Sweden, May 2000.

10. Charles E. Perkins and P. Bhagwat, A Handbook of Graph Theory, 2003. 11. “The click modular router. www.pdos.lcs.mit.edu/clicW,” .

A NEW APPROACH FOR TDMA SCHEDULING IN AD-HOC NETWORKS*

DIMITRIOS D. VERGADOS University of the Aegean

Department of Information and Communication Systems Engineering GR-832 00, Karlovassi, Samos, Greece

DIMITRIOS J. VERGADOS university of the Aegean

Department of lnformation and Communication Systems Engineering GR-832 00, Karlovassi, Samos, Greece

CHRISTOS DOULIGERIS Univeristy of Piraeus

Department of Informatics Karaoli & Dimitriou St,. GR- 185 34, Piraeus, Greece

Medium Access Control (MAC) for Ad-hoc networks has been a major concern of both academia and industry for many years. Even though the dominant MAC for ad-hoc networks is CSMNCA, it leads to a number of disadvantages like high overhead, increased access delay, high jitter and limited QoS capabilities. TDMA appears as a promising solution for overcoming these issues, but the application of TDMA in ad-hoc networks leads to the known NP complete Broadcast Scheduling Problem (BSP). This paper presents a new polynomial algorithm that can provide a suboptimal solution the BSP. The efficiency of the new algorithm is evaluated in terms of complexity, delay, throughput and fairness, and is compared with other solutions

1. Introduction

Ad-hoc networks are becoming more and more popular as a networking solution for scenarios where infrastructure is not available or desired, like battle field, disaster, etc. Furthermore they are an attracting solution in cases where a network must be deployed rapidly, or at low cost.

* This Research work is funded by the Ministry of Education and Religious Affairs and co-funded by E.U. (75%) and National Resources (25%) under the Grant “Pythagoras - Research Group Support of the University of the Aegean”.

107

108

Despite the vast research activity in the field of ad-hoc networks, there are still numerous open issues, including routing, security, medium access control, quality of service, clustering, addressing, scalability etc. The most popular medium access control scheme for ad-hoc networks is the IEEE 802.1 1 DCF [ 11, which uses the CSMA algorithm and has numerous disadvantages like high overhead, increased access delay, high jitter and limited QoS capabilities. Furthermore, 802.1 1 can not overcome the Exposed Terminal problem. Even though TDMA can overcome all these issues, a solution for the Broadcast Scheduling Problem (BSP) [4] is needed for using TDMA in an ad-hoc environment.

2. Network Model

In an ad-hoc network there are a number of nodes who communicate with each other, either directly, or through a number of relay nodes, who forward incoming packets towards their destination. Every transmission is broadcasted over the wireless channel, and all nodes located close to the transmitting node can receive the transmission, where as far away nodes cannot receive the transmission. Nodes that can transmit with each other are called neighboring nodes. Furthermore, the receiving nodes can only receive one transmission at a time without errors, and nodes cannot transmit and receive packets at the same time.

We assume that multiple access in the wireless channel is achieved by TDMA. All nodes in the ad-hoc network must have at least one transmission opportunity within each TDMA frame. The TDMA frame consists of a number of TDMA slots. The number of slots in each TDMA frame is called the frame length. More than one ad-hoc nodes may transmit in every TDMA slot without collision, if they do not have any common neighbors. The purpose of TDMA scheduling is to determine the slots used by every ad-hoc node for transmitting its packets, in a way that ensures collision avoidance and at the same time minimizes the delay each node experiences, and maximized the total network capacity in a fair manner.

If we represent all nodes in the ad-hoc network as vertices of a graph with edges between neighboring nodes then there are no collisions in the network, if the distance of all transmitting nodes in the graph is at least two hops. So the broadcast scheduling problem is to determine how to schedule every node in the network into the appropriate slot, so that maximum capacity is achieved in the shortest possible frame length. As shown in [4] broadcast scheduling is an NP- complete problem.

109

3. The proposed algorithm for Ad-hoc Networks

Many different algorithms have been proposed for solving this problem [3] - [lo], most of which try to minimize the TDMA frame length and increase the network capacity, using techniques derived by graph theory and neural networks.

In this paper we present a new algorithm that provides a near optimal solution for the Broadcast Scheduling Problem in polynomial time, in order to enable the operation of TDMA in the ad-hoc scenario. The algorithm provides low delay, increased capacity and takes fairness into consideration.

The TDMA frame is a set of TDMA slots. Each TDMA slot is a set of nodes. If node i can transmit during timeslot k, then i E S, . So, the algorithm decides which node can transmit in each timeslot. When the algorithm ends, every node is assigned to (at least) one timeslot, in a way that ensures collision avoidance.

Two stations can be in the same slot if they have no common neighbors. Consider a one-hop neighboring table A, where

1 if node i and j are neighbors, or i = j otherwise (1) AI., = 0 {

Obviously A is a symmetric table. Also, consider vector F , where

If node k E S,and there exists a node 1~ S, where k and 1 have z as a > 0 . On the common neighbor, then Az,l=Az,k=l. Therefore F, 2 1 and 2, .

contrary, if

2, . F = O

there is no 1 E S, that has a common neighbor with node k, so node k can be added to the slot.

In order to produce the desired schedule, all nodes in the network are tested in a specific order. If equation (3) is true for the tested node (k), then the node is added to the slot; otherwise it is not added and the next node is tested. When all nodes in the network are tested, then the first time slot is produced. The nodes are re-ordered, and the nodes are re-tested to produce the second timeslot. This procedure is repeated until every node in the network is in at least one timeslot.

110

Obviously, the set of nodes in every slot is determined by the order by which the nodes are tested. Since the objectives of the algorithm are fairness and low delay, we created a vectors W and Q , where

W, = Ai,j .je N

(4)

and Qi is the number of slots where node i has transmitted. When creating

bi < b, if Qi < Q,. Nodes that have not transmitted in previous slots should transmit with a higher priority. if Q, = Q,, then bi < b, if Wi > W, . Nodes with many neighbors should be checked first, because nodes with few neighbors have a grater chance of transmitting (This happens because there are more possible slots, where a node with few neighbors can transmit). After each slot is created, the values in Q are updated and the nodes are re-

ordered. Then the next slots are created until all nodes have transmitted (Qh,, > 0).

each slot, if bk is the order of node k, then the nodes are ordered as follows: 1.

2.

4. Performance Analysis

4.1. Complexity Analysis

In order to produce the TDMA transmission schedule according to the proposed algorithm, all nodes have to be tested once (in the worst case) for every TDMA slot. During each test, there is a vector multiplication that requires N multiplications (if there are N nodes in the network). Also the maximum number of slots required for every node to transmit at least once is N. Therefore the worst case complexity is at N3.

The complexity in most cases is expected to be significantly smaller. If F, > O,Vi , then no other node can be added to the slot, so the remaining nodes don’t have to be tested. Also, in most cases the produced frame length is significantly smaller than the number of nodes in the network.

4.2. Frame Length

The frame length produced by the algorithm is the number of slots required for all nodes to transmit at least once. The frame length is desired to be as small as possible, because the access delay increases proportionally to the frame length. The frame length can be reduced by placing as many nodes as possible in every slot. This is taken into account during node ordering

111

Nodes 15

30 40

4.3. Capacity

The capacity of the TDMA scheduling algorithm is the number of transmissions that take place at each frame, or the sum of nodes in each slot. Th capacity usually increases as the frame length increases.

MFA HNN-GA NEW FL C Th FI FL C Th FI FL C Th FI 8 18 2.25 0.90 8 20 2.50 0.78 8. 19 2.38 0.89 12 39 3.25 0.83 10 35 3.50 0.87 10 34 3.40 0.92 9 71 7.89 0.65 8 67 8.38 0.68 8 62 7.75 0.80

4.4. Throughput

The throughput of the TDMA scheduling algorithm is equal to the capacity of the network, divided by the frame length. Thus, it is equal to the rate by wich information is transmitted.

4.5. Fairness Index

A scheduling algorithm is expected to share the available resources fairly among the nodes. This is taken into account in the node ordering procedure. We evaluated the fairness of the algorithm using Jain’s Fairness Index, that is

5. Simulation Results

5.1. Comparison with previous algorithms

In order to evalure the efficiency of the new algorithm, we tested it on the network topologies found in [2] and compared its performance with the HNN- GA [2] and MFA [S] algorithms. For these networks we measured the frame length, capacity, throughput and fairness. The results are in table 1

The new algorithm is found to be better than the MFA algorithm and equivalent to the HNN-GA algorithm in terms of frame length, capacity and throughput. However the new algorithm is superior that the other ones in terms of fairness.

Table 1. Comparison of the new algorithms with GNN-GA and MFA

112

5.2. Simulation of random networks

We generated numerous random networks and simulated the efficiency of the algorithm on these networks. For the first series of simulations a constant neighboring probability is considered. Every node could be a neighbor with any other with probability p = 0.2. We studied the resulting frame length for networks with different number of nodes (Figure 1)

p = 0.2

20

g 15 c - E 2 10

y = 0,531~'

1 I

I

I I

7 I

0 5 10 15 20 25 30 35 40 45 Number of Nodes

Figure 1. Frame length versus Number of nodes. p = 0.2

nodes = 40

0 0.1 0.2 0.3 0.4

probability 0.5 0.6

Figure 2. Frame length versus neighboring probability. n = 40 and 100

113

A second series of simulations was also carried out to illustrate the affect of the neighboring probability. These simulations where carried out for networks with 40 nodes and for networks with 100 nodes. The results are in Figure 2.

6. Conclusions

TDMA appears as a promising medium access control scheme for overcoming various issues in ad-hoc networks. This paper introduced a new algorithm for overcoming the NP-complete Broadcast Scheduling Problem that appears in TDMA ad-hoc networks. After comparing the new algorithm to some existing algorithms on three different networks, we found it superior than the MFA algorithm [2] and equivalent to [8] for most of the tested network topologies in terms of delay and throughput, whereas the new algorithm is far superior in terms of fairness. Therefore, the low worst-case complexity of the algorithm, in condition to its good performance makes it ideal for real-time operation for TDMA scheduling in ad-hoc networks.

References

1. IEEE std. 802.11, "Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications", 1999.

2. Gangsheng Wang, Ninvan Ansari, "Optimal Broadcast Scheduling in Packet Radio Networks Using Mean Field Annealing", IEEE Journal on Selected Areas in Communications, vol. 15, no. 2, February 1997 pp. 250-260

3. J. Ju, V. Li, "TDMA Scheduling Design of Multihop Packet Radio Networks Based on Latin Squares", IEEE Journal on Selected Areas in Communications, vol. 19, is. 8, August 1999 pp. 1345-1352.

4. Anthony Ephremides and Thuan V. Truong, "Scheduling Broadcasts in Multihop Radio Networks", IEEE Transactions on Communications, vol. 38, no. 4, April 1990, pp. 456 - 460

5. H. Fattah, C. Leung, "An overview of scheduling algorithms in wireless multimedia networks", IEEE Wireless Communications, vol. 9, is. 5, October 2002, pp. 76-83.

6. R. L. Cruz, Arvind V. Santhanam, "Optimal routing, link scheduling and power control in multi-hop wireless networks", IEEE INFOCOM 2003 - The Conference on Computer Communications, vol. 22, no. 1, Mar 2003

7. Chiu Y. Ngo, Victor 0. K. Li, "Centralized broadcast scheduling in packet radio networks via genetic-fix algorithms", IEEE Transactions on Communications, vol. 51, no. 9, Sep 2003 pp. 1439-1441

8. Sancho Salcedo-Sanz, Carlos Bousofio-Calz6n and Anibal R. Figueiras- Vidal, "A mixed neural-genetic algorithm for the broadcast scheduling

pp. 702-71 1

114

problem", IEEE Transactions on Wireless Communications, vol. 2, no. 2, Mar 2003, pp. 277 - 283

9. Goutam Chakraborty, "Genetic algorithm to solve optimum TDMA transmission schedule in broadcast packet radio networks", IEEE Transactions on Communications, vol. 52, no. 5, May 2004 pp. 765-777

10. Joseph L. Hammond, Harlan B. Russell, "Properties of a transmission assignment algorithm for multiple-hop packet radio networks", IEEE Transactions on Wireless Communications, vol. 3, no. 4, Jul2004 pp. 1048- 1052

A SELF ORGANIZING ALGORITHM FOR AD HOC NETWORKS

NOUREDDINE KETTAF University of Haute Alsace, 34 Rue du Grillenbreit, 68000 Colmar France

ABDELHAFID ABOUAISSA University of Haute Alsace, 34 Rue du Grillenbreit, 68000 Colmar France

PASCAL LORENZ University of Haute Alsace, 34 Rue du Grillenbreit, 68000 Colmar France

HERVE GUYENNET University of Franche-Comtt!, 16 route de Gray 25030 Besanqon France

In this paper, we present a new leader election algorithm for mobile ad hoc networks, in order to handle arbitrary topology changes and group self-organizing or self-stabilising. Our proposed leader election algorithm ensures that eventually nodes of a communication system network can form a group. Within this group start nodes trigger the leader election, in order to organize nodes in one hierarchical level with leader election. We also simulate the algorithm in mobile ad hoc network settings. Through our simulation study, we point out several important issues that can significantly influence performance of such a protocol for ad hoc networks, such as choice of signalling, broadcast nature of wireless medium, etc. The simulation study shows that, furthermore the proposed algorithm auto-organize mobile nodes; it brings the group to support QoS, in several operating conditions.

1. Introduction

Leader election algorithm in wired and wireless systems is a paramount control problem, especially when failures can occur. For example, if a node failure causes the token to be lost in mutual exclusion algorithm, then the other nodes can elect a new leader to hold a replacement token [15]. However, for wireless ad hoc networks, designing and developing an asynchronous distributed algorithm is a very challenging task, since the communication topology may change when the hosts move.

Leader election problem requires that a unique leader can be elected from a given set of nodes. In mobile ad hoc networks [9], much of the work has focused

115

116

on routing and medium access control protocols, and there is little work done on distributed systems. The self organizing concept was introduced by Dijkstra in [3], in which a system is defined as self stabilizing if starting from an arbitrary configuration; it is guaranteed to converge to an acceptable configuration in finite time. The execution starting from that legitimate configuration shows the desired behaviour [7]. However, there is much less work on self-stabilization for mobile ad hoc networks with leader election supporting QoS in asynchronous system [2] [5] [8].

Our algorithm uses the concept of message exchange to perform leader election. Briefly, the algorithm operates as follows: in a communication system network, to form a homogeneous group, nodes must check two fixed parameters (signal power and mobility). Within the formed group, arbitrary nodes also called start nodes trigger leader election operation, by transmitting messages to other nodes of the group to inform them about their characteristics. Other nodes receive the message and compare their characteristics with the received one, to elect the most valued node. However, nodes broadcast a confirmation message to inform the chosen node that is the leader and to warn other nodes.

We emphasize a study of performance of the algorithm, regarding bandwidth loss, energy consumption, and the throughput.

The rest of the paper is organized as follows. The next section discusses existing and related work. In section 3 , we describe our model’s assumptions and objectives. In section 4, we focus on the leader election algorithm. Simulation settings and the performance metrics are described in section 5. Section 6 discusses simulation results. In Section 7, we present our conclusions and perspectives.

2. Overview on existing works

Given the wide variety of existing works in leader election in wired networks, we highlight only the major relevant algorithms in wireless mobile ad hoc networks. In [5] the algorithm works by constructing several spanning trees with prospective leader at the root of the spanning tree and recursively reducing the number of spanning trees. However, this algorithm works only if the topology remains static and cannot be used in a mobile setting [2]. Other algorithms presented in [12], [4], and [S] for clustering and hierarchy constructions schemes assume static networks and synchronous systems; therefore they cannot be used in asynchronous systems like ad hoc networks, which are characterized by mobility and arbitrary topology changes.

117

In self stabilizing and organizing systems there has been some work on spanning tree algorithms using a shared memory model that have been proposed in [lo] and [S]. These algorithms assume a shared memory model and are not suitable for a message-passing system such as an ad hoc network [2]. In addition, the stabilisation time within these bounds is proportional.

The aim of this paper is to design a practical asynchronous, leader election algorithm for mobile ad hoc networks, which can adapt to arbitrary changes of topology caused by nodes mobility and get an organized group of nodes, in a finite time, as well as support better QoS. Our work was concentrated on the conservation of energy, because a node remains operational and reachable within the group as long as it emits a powerful signal.

3. System model and assumptions

The system model contains a set of n independent mobile nodes; we can model an ad hoc network as an undirected graph that changes over time as nodes move. The vertices in the graph correspond to mobile nodes and an edge between a pair of nodes represents the fact that two nodes are within each other’s transmission radii and, hence, can directly communicate with one another. The graph would be disconnected if the network were partitioning due to node movement [2] . Assumptions on the mobile nodes network and system architecture are:

Automatically, each node has a property associated with it. The property indicates the node’s performance related attribute like the quantity of energy, the resource (memory), the power of the signal and its mobility. The nodes have unique node identifiers. Communication links are bi-directional, reliable and FIFO. Topology is dynamic and nodes movement is random.

4. Leader election algorithm

The leader election problem requires that only one leader can be elected from a given set of nodes. In the wired networks domain the problem has been widely studied, but in wireless networks, this area is recent. However, the leader election algorithm that we have proposed is based on two phases:

In the first one, we dynamically group nodes in one or several homogeneous groups. The homogeneity here is in terms of power of signal and mobility. We have chosen these two conditions, because communication in ad hoc networks is wireless, and within a group a node is reachable if it still emits signals to other nodes. The second condition is the mobility, we cannot form a stable group from

118

nodes having different speeds. In this case, we set a given threshold to separate the groups formed by high-speed and slow nodes.

In the second phase, start nodes trigger leader election within the formed group.

4.1. Phase one: dynamic nodes grouping

As described before, to define the members of a group, each node pertaining to the communication system carries out the following tests:

4.1.1. Signal power:

In wireless communication the average value of power of the received signal decreases in function of the distance between the transmitter and the receiver [ 1 11. In order to receive the signal correctly, it is necessary that the report signal to noise (signalhoise) should be higher than a Q threshold usually equal to 10. If the condition of the report (signalhoise) > 10 is satisfied, then all the nodes which satisfy to this condition can form a virtual group.

4.1.2. Mobility:

Ad hoc networks are characterized by the mobility of nodes, in which one group can contain several nodes moving with different speeds. Our goal is to form a stable group; therefore we set a given S threshold to differentiate mobile nodes having high speed from others with medium or low speed. We gather nodes moving with high speed in one virtual group, and the same task with other nodes moving with medium or slow speed. In basic ad hoc networks nodes can exchange [RTS, CTS, DATA, and ACK] messages, via a complete virtual graph, in order to guarantee group self-stability realized by the homogeneous mobility of nodes having practically the same power of signal. Furthermore, this method ensures a reliable communication between wireless mobile nodes.

4.2. Phase two: leader election

In this second and last phase, the goal is the election of a node to be the leader of the formed group. Initially, every node possesses appropriate values; these values are in terms of energy, of signal power and mobility; within a defined group, nodes exchange messages to elect a node as the leader. The election algorithm is based on the work presented in [13] for the broadcasting of messages in wireless ad hoc networks, in which nodes do not consume a large quantity of energy during the control messages exchange. Indeed, to introduce

119

the election, start nodes broadcast a request message called Req(Prop) to the other nodes of the formed group, where ‘LPr~p77 represents energy, signal power and mobility. When other nodes receive the Req(Prop) message, they compare their own “Prop” with those received from start nodes.

The aim of these comparisons is to estimate the most powerful node in order to elect a leader within a defined group. Once the leader is designated, every node sends a confirmation message called Conf() to the leader, and another message to inform the other nodes.

G,= {nl, n2, n3... ni. .. nj.. . nk}

Algorithm:

For any node nk from G, :

We consider a group Si :{ in order to establish contacts between nodes in G , }

V ni, c G, :

ni Sends and Receives (RTS/CTS/DATA/ACK]i to and from Si in order to establish contacts

ni Sends Req (Prop)i To Si

when Si nodes receive ni message

Begin

G, : is a virtual group formed by nodes having checked the two conditions.

/* comparison of properties */

if (ni Req (Prop)i nj Req (Prop)j )

then the leader is a node ni has the greatest (Prop)

nj informs the other nodes that ni is the leader

else nj is the leader, ni informs other nodes that a node nj is the leader

all nodes send a ConfO message to the leader

end

/* node nj c Si */

End

5. Simulations

In this section, we report the results of simulations that we have performed to study the performance of the proposed algorithm, for a better comprehension of how to elaborate the leader election algorithm for ad hoc networks. The performance metrics that we consider in our simulations are bandwidth loss, energy consumption, and the throughput.

5.1. Simulation environment

With ns2 [14], the simulated network consisted of 50 nodes uniformly distributed over a zone size of 1000 square meters. For all network sizes, nodes

120

1.20

1 ,l.z

1 .I0 ,,o

move according to the Random Waypoint mobility model [6]. We set the node speed to 1 - lOm/s. Each node was equipped with a radio transceiver, assuming Free-Space Propagation path-loss model. We used the IEEE802.11 like the MAC layer protocol. In our simulations, a leader node transmitted a heartbeat message to all other nodes. We set the value of heartbeat message interval to 30 seconds, and the maximum time of heartbeat message loss to 5 seconds. Nevertheless, a start node initiated a new leader election if it had not received a beacon message from the leader for 90 seconds. It should be noted that the duration of messages were chosen to accommodate the random waypoint mobility model, and can be set according to application requirements.

....... . , . . , I , . . I . . , , . . . , , I I , : : : ; ;

....... ...... :L ...... J ....... :.* .... .:.. ..... ~ ....... ; ....... L ..............

....... : ...... ...................... ; ....... : ....... : ...... 2 ...... .: ......

....... i ....... 1 .............. i ............... : .... ...; ....... , ; ...... ....... . , I . , . . . , I , I . I I , , , I I , , I I I . ,

. I . . % . . . . . , . , . . . . . I , . , . . . . , 0 . . , . . , . ,

* . . . . , . I / * .

, , . - . . . . , I I , . I . . I

6. Results and Discussion

6.1. Bandwidth

In wireless ad hoc communications, periodic transmission of control messages consumes a part of the available bandwidth. To estimate bandwidth loss, we set up a simple scenario, in which several nodes of the group send Req(Prop) and Conf() messages with constant rate, in order to saturate the bandwidth.

The figure 1 represents the loss of bandwidth caused by MAC layer messages added to request and confirm election messages [Req(Prop), Conf()], during communication between a node and group members. Our scenario allows to measure the impact of the size of Req(Prop) and Conf() messages, on the

121

bandwidth consumption. In figure 1, we can observe that the bandwidth degradation percnetage is equal to 0.11% for 100 exchanged messages.

This fall is due to the medium control messages added to our election messages.

Generally, a quantity of bandwidth is saturated by the medium control messages. In our case, the leader election messages exchanges consume a slight part of available bandwidth, which makes our method more advantageous.

6.2. Energy

Among the most important problems of mobile ad hoc networks, there is energy consumption, because the life span of mobile nodes depends on it. More exactly, as each node is autonomous and has a limited energy, it is necessary to avoid excessive energy consumption. It is even necessary to use as little energy as possible during the election, so as to leave the biggest part for the applications that will follow.

Within this frame, we supposed for our simulations a scenario to calculate energy consumption during four independent stages.

90

80

70

60

30

20

10 10 15 20 25 30 35 40 45 50

GROUP Figure 2. Power consumption for different mode of transmission

Figure 2 depicts four curves, the first and the second curves (from the horizontal axis) represent a broadcast and a multicast of Req(Prop) and Conf() messages, between group members. We notice that energy consumption is important in the first, because message exchanges were done by a Broadcast and the consumed energy was equal to 26% of available energy.

122

The second curve represents a multicast of [Req(Prop) and Conf()] messages. We observe that the energy level decreases rapidly which means that the consumption of energy is 39%. The third curve illustrates the consumed energy during the leader election, by exchange of Req(Prop) and Conf() messages using the method presented in [13]. We notice that the consumed energy (20%) is clearly weaker than the energy consumed by broadcast or multicast.

We prove by this, that the operation of the election consumes a little quantity of energy. On the last graph (the highest one), we observe that after leader election operation, energy consumption decreases more (9%), because nodes do not exchange control messages, insofar as the leader manages its group.

6.3. Throughput

The reason for which ad hoc network nodes cannot have a perfect and durable knowledge of network topology, is that any node in such a network can move, appear, or disappear all the time. This constraint makes the conception of a routing or signalling protocol very complex.

Figure 3. Network capacity

Our aim is to set up a feasible method within a mobile ad hoc group, to provide group stability that would be able to support QoS applications. In this context, we proposed a simulation scenario to calculate the rate of UDP sent packets, in which we performed sending parquets of 1 and 10 Mbps from the leader toward group nodes, and observed the network behaviour during packets transfer. Figure

123

3 shows simulation results and that performance degradation in terms of delivered packets is very slight.

In fact, we can explain this degradation by the nodes mobility that entails packets loss and IEEE802.11 MAC layer also generates interferences.

The obtained results prove that group self-organizing by leader election, is able to insure QoS.

7. Conclusion and perspectives

In this paper, we have presented a new asynchronous, distributed leader election algorithm for mobile ad hoc networks, based on new metrics which represent the node’s properties. To obtain homogeneous groups of nodes, the proposed algorithm is based on two phases. In the first phase, we form a group in which nodes must check two conditions: signal power for communications and mobility for group stability. In the second phase start nodes trigger the election to determine the most valued node within the defined group.

Finally, we have simulated the algorithm and have provided useful insight, based on our experiences in designing leader election algorithm. Although in this paper we described our algorithm as an extrema finding one, it can be used in scenarios where a unique or several leaders are desired.

Currently, we are investigating on a new routing protocol using leader election metrics.

References

1. Sudarshan Vasudevan, Jim Kurose, Don Towsley, “Design and Analysis of a Leader Election Algorithm for Mobile Ad Hoc Networks” ICNP Proceedings of the Network Protocols, 12th IEEE International Conference on (ICNP‘04) - Pages: 350-360.2004.

2. W.B. Dunbar, E. Klavins, and S. Waydo. “Feedback controlled software systems”. CDS technical report 2003- 002, California Instit. of Tech, 2003.

3. T.Arici,B.Gedik,Y.Altunbasak, and L. Liu, “PINCO: a Pipelined In- Network Compression Scheme for Data Collection in Wireless Sensor Networks,” in Proceedings of IEEE International Conference on Computer Communications and Networks, 2003.

4. B. Wang and S. K. S. Gupta, “On maximizing lifetime of multicast trees in wireless ad hoc networks “in International Conference On Parallel Processing (ICPP-03), Kaohsiung, Taiwan, China, Oct. 2003.

5. J . Yoon, M. Liu and B. Noble. “Random Waypoint” Considered Harmful In Proc. of IEEE INFOCOM, 2003.

124

6. Yu Chen, Jennifer L. Welch “Self-stabilizing mutual exclusion using tokens in mobile ad hoc networks”. Workshop on Discrete Algorithms and Methods for MOBILE Computing and Communications 2002.

7. Robert E. Schapire“The boosting approach to machine learning: An overview”.In MSRI Workshop On Nonlinear Estimation & Classification, 2002.

8. N. Malpani, N. Vaidya and J. Welch. “Distributed Token Circulation in Mobile Ad Hoc Networks”. In Proc. 9th International Conference on Network Protocols (ICNP), November 200 1.

9. A. Arora, M. Demirbas, and S. S. Kulkarni. “Graybox stabilization”. Proceedings of the International Conference on Dependable Systems and Networks, pages 389-398, July 2001.

10. R E. Lee and Y. Xiong, “System-level types for component-based design” in Lecture Notes in Computer Science, Embedded Software, vol. 221 1 . Heidelberg, Germany, 2001, pp. 237-253.

11. D. Coore, R. Nagpal and R. Weiss. “Paradigms for Structure in an Amorphous Computer. Communications of the ACM archive. 2000.

12. J.E. Wieselthier, G.D. Nguyen, A. Ephremides, “On the Construction of Energy Efficient Broadcast and Multicast Trees in Wireless Networks,” in Proceedings IEEE INFOCOM 2000, pp. 586-594.

13. L. Breslau et.al. “Advances in network simulation” IEEE Computer, 5-2000. 14. A.Arora and M.Gouda.“Distributed Reset”.ln IEEE Transactions on

Computers, 1994. 15. Navneet Malpani, Jennifer L. Welch, Nitin Vaidya “ Leader election

algorithms for mobile ad hoc networks”. Workshop on Discrete Algorithms and Methods for MOBILE Computing and Communications. Proceedings of the 4th international workshop on Discrete algorithms and methods for mobile computing and communications Boston, Massachusetts, United States, 96 - 103. 2000.

HDSR: HIERARCHICAL DYNAMIC SOURCE ROUTING FOR WIRELESS HETEROGENEOUS MOBILE AD HOC

NETWORKS *

MOHAMMAD NASERIAN, KEMAL E. TEPE, AND TARIQUE MOHAMMED

University of Windsor, Department of Electrical and Computer Engineering,

Windsor, Ontario, Canada E-mail: { naseria, ktepe, tarique} Quwindsor. ca

In this paper, hierarchical dynamic source routing (HDSR) protocol is introduced for heterogeneous mobile ad hoc networks architecture. In this network architecture, there are two tiers: Forwarding nodes (FN) and mobile nodes (MN). FNs route the packets and MNs host the applications. Dynamic source routing (DSR) protocol is modified and optimized for mobile multi-tiered ad hoc network architecture. The new routing protocol, HDSR, separates route discovery (route requests and replies), and packet forwarding functionalities. Those two functionalities are distributed into two node types, where only route requests and replies are functional in MNs when they are source or destination, and both route discovery and packet forwarding are functional in FNs. HDSR combined with hierarchical architecture has lower routing overhead and shorter route discovery delays, because in HDSR there are smaller number of route reply and route request messages than in regular DSR. HDSR is implemented by network simulator (Network Simulator-2 of University of California). It was shown via computer simulations that HDSR improves average end-to-end delay per packet, network throughput and packet delivery ratio in wireless ad hoc networks compare to regular DSR.

1. Introduction

Mobile ad hoc networking is becoming increasingly popular as a mean of providing instant networking to groups that may be within the transmission range of one another. These networks are self-initializing, self-configuring and self-maintaining, all of which can be coined with term “self-organizing” . Routing is an essential part of network protocols to provide self-organizing

‘This work is supported by Natural Science and Engineering Research Council of Canada (NSERC).

125

126

capability, and it is the most widely studied element for ad hoc networks. Numerous routing protocols have been developed or adopted from Internet routing protocols for ad hoc networks. Broadly, those protocols can be classified as: (1) Proactive routing protocols, and (2) On demand (Reactive) routing protocols. In proactive routing, routing information is periodically exchanged among network nodes, like Dynamic Sequence Distance Vector (DSDV) Because of those periodic updates and information exchanges, the network consumes large portion of the useful bandwidth for routing control overhead (i.e., route maintenance and update packets). That is why proactive protocols do not scale well. On the other hand, in on demand routing, routes are discovered when they are needed. Such provision eliminates periodic routing updates, hence allows protocols to operate more efficiently (i.e. less routing overhead) than proactive routing protocols. That is why most of the recently proposed routing protocols for ad hoc networks fall under the on demand category, like Ad Hoc On Demand Distance Vector (AODV) and Dynamic Source Routing (DSR) 3. Al- though on demand routing protocols offer lower average routing overhead than proactive routing protocols, there are still two problems that need to be solved. The first one is the large routing overhead during route discovery phase, and the second one is the large end to end packet delay.

There are proposals to improve routing overhead and scalability. One of those proposals is clustering and cluster based routing 5*13114. Clustering (active or passive) can be described as grouping nodes into clusters. A rep- resentative of each group can be named as cluster head and other members are called cluster members. There are proposals to provide efficient formation of clusters, selection of the cluster heads and its member nodes 13. To form clusters and maintain the clusters, network nodes need to cooperate and exchange information with each other, which can increases the control overhead packets. Passive clustering l4 has recently been proposed to exploit ongoing traffic to propagate cluster related information. Although it requires less overhead packets to form and maintain clustering, it still requires some partial information about the neighbors.

Other proposals that provide improvement and scalability to routing protocol suggest systematically reducing the number of messages generated and transmitted during the flooding. Those schemes can be loosely classified as probabilistic schemes 5 9 6 9 7 7 8 9 9 and location based schemes 'O1l2. The major problems of probabilistic schemes is that the probability at which a node should rebroadcast is not universal, but specific to each topology. There is no analytical formula to obtain that probability. Local topology

127

information is used to avoid unnecessary rebroadcasts in location based schemes. In Ref. 11, 12, self-pruning and neighbor-coverage schemes were proposed, in which a node does not rebroadcast if the packet is delivered to all neighbors of this node by a prior broadcast.

Our approach to improve the performance of reactive routing, particularly in DSR, is to introduce hierarchy. The new protocol derived from DSR, called Hierarchical Dynamic Source Routing (HDSR), limits the number of nodes that participate in the route discovery of the protocol, which in turn reduces overhead and delay compare to DSR. That architectural change provides HDSR to reduce the routing overhead significantly because number of nodes that involve in the route discovery is smaller and they can find and return routes faster to the source. That can reduce the end to end delay too. In addition to those, MNs do not need to acquire and maintain any statistical information about the neighbors or do not need to send maintenance messages or location information about the neighbors. Such reductions could significantly save bandwidth of the network, hence improve the throughput of the network. Rest of the paper is organized as follows. In Section 2, we will explain HDSR in details, which will provide its differences from DSR. Later in Sections 3 we will provide simulation model and results. Finally Section 4 summarizes our conclusions.

2. Hierarchical Dynamic Source Routing Protocol

In HDSR protocol, we classify the participating nodes of the network as Mobile Node (MN) and Forwarding Node(FN). We assign different functionalities to those nodes depending on what type of node they are. MNs initiate route discovery. FNs help them to find source route to the destination MN. The destination MN replies back through the FNs to source MN. Once source MN discovers the routes, it starts sending packets to the destination. FNs assist the MN to forward packets to destination MN. Route discovery and route maintenance in HDSR are different from those in DSR.

When a source MN originates packet to a destination MN. If the source cannot find a source route in its route cache, it initiates a route discovery by transmitting a ‘route request packet’ as a local broadcast packet. Only FNs, which are within the range of the source MN receives the broadcast packet. Other MNs, which are also within the range of source MN and which are not the destination of this packet, discard the broadcast message and do not broadcast further. Only the FNs re-broadcast the request to other FNs unless the destination MN receives this route request packet.

128

@ ................ 5 ..,

. , .. .. I I 1 ,.....,, ..’ ,

Figure 1. Limiting Route Request storming in HDSR

Figure 2. Limting Route Reply Storming in HDSR

The destination MN then replies back to the source MN through the FNs. After receiving the route reply, the source MN record the source route in its cache and starts sending packets to the destination MN using the source route it has just discovered.

Route maintenance is performed by FNs only. When a F N detects that the next link from itself to the next MN or FN is broken, it updates its own route caches by marking all the paths which use the broken link as invalid and sends route error message to the source MN and all other F N which uses the broken link for packet transmission. We will now explain how HDSR reduces overhead packet during the route discovery processes and prevent route request and route reply flooding.

Figure 1 shows how a route is discovered in HDSR. In this scenario nodes 1, 2, 3, 5 and 6 are MNs and nodes 4 and 7 are FNs. Route discovery is initiated by MN-1 to find a source route to destination MN-8. MN-1 transmits the route request packet as a local broadcast message. MN-2, MN-3 and FN-4 are within the range of MN-1. MN-2 and MN-3 are restricted not to re-broadcast the route request further. They are not forwarding node and they are not the destination as well. Only FN-4 will rebroadcast the request packet after adding itself in the request packet. FN-7 will only accept the route request packet only because it is the only FN within the range of FN-4. FN-7 rebroadcasts the request packet and the route request packet finally reaches the destination MN-8. MN-8 replies back to source node. Upon receiving the reply packet, source MN-1 record its route cache and starts sending packet through the source route it has just learned from the reply packet. In this case only three broadcast messages are generated. Redundant route request broadcasting by MNs except the source MN have been eliminated in HDSR which saves bandwidth by reducing packet colli-

129

sion. Figure 1 illustrates how route reply flooding is prevented in HDSR. In

this case there is only one FN and all other nodes are MNs. Route discovery was initiated by MN-1 to find a source route to the destination MN-7. MNs 2, 4, 5, 6 and FN-3 are within the range of MN-1. Assume each MN and FN has a source route in its cache. In HDSR protocol, only FN-3 will reply back to MN-1 in contrary to replying procedure used in DSR where all the MNs reply back to MN-1. All other MNs which received the route request message discard it. MN-1 starts sending packet to destination MN using the route 1-3-7. Thus route reply flooding is limited in HDSR when each node replies from its route cache.

3. Simulation Model and Results

We used Network Simulator 2 (NS-2) to implement and test the performance of the new proposed protocol. In NS-2, the effective transmission range of wireless radio is 250 meters and the medium access control (MAC) protocol is based on IEEE 802.11 with 2 Megabits per second raw capacity. Traffic sources are Constant Bit Rate (CBR) with 512 bytes per packet. The mobility model uses the random waypoint model in a rectangular field. In this model each node starts its journey from a random location to a random destination with a randomly chosen speed, which is uniformly distributed between 0-20 mls. When the mobile node reaches its destination it stays at that location for the period of a pause time, p seconds, and then it chooses another random destination and moves toward this new destination with a new randomly chosen speed. We vary the pause time p , which affects the mobility scenarios. Each CBR source starts randomly in the first ten seconds of the beginning of the simulation and simulation runs for 600 seconds. In order to increase reliability of the simulations, each connection scenario is simulated 10 different times with new topologies and different mobility scenarios. The reported results are average of these 10 simulations. In order to obtain these new simulation environment, we modified NS-2 source code to suit our needs, and at the same time implemented HDSR protocol and hierarchical architecture.

Figures 3, 4, and 5 show performances of simulations of 80 MNs scenarios versus the pause time of mobile nodes. The size of the rectangular area that mobile nodes are located is 1000 x 1000 meters. There are 20 CBR sources with data packet rate of 2 packets per seconds. In HDSR, there are 12 FNs in additions to MNs and locations of the FNs are chosen randomly

130

as well. Figure 3 shows the routing overhead in HDSR and DSR. The overhead in HDSR is consistently lower than DSR in all scenarios, and for this scenario it is approximately 50 times lower. We repeated these experiments with differing number of nodes, and we found that overhead improvement in HDSR is higher when the number of nodes in the network grows. The difference between HDSR and DSR overhead increases when the mobility is

' DSR -+- HDSR x

"\

PaUte lme

Figure 3. 80 MN Scenario


I 0 50 100 1% 200

IW ' Pawe The(%)



Figure 7. 50 MN Scenario Figure 8. HDSR, 50 MN Scenario

Different number of FNs in

131

higher (i.e., shorter pause times). Due to the higher number of routing overhead packets, the network with DSR routing protocol has lower bandwidth for data packets, which we think adversely affects performance metrics in DSR compared with HDSR. For example, throughput of the network with HDSR is improved 3 times in high mobility and 20-30 percent in low mobility cases compared with that of DSR (Figure 4). In different scenarios, the throughput is always better with HDSR. The average end-to-end delay is also improved with HDSR. Figure 5 shows average end-to-end delay of scenarios with 80 mobile nodes. In that case, the delay in DSR is 3 times higher than that of HDSR for very high mobility (i.e., pause time less than 50 seconds) and few tens of times in low mobility cases. Delivery ratios of HDSR was better than DSR too. We also repeated our simulations choosing scenarios similar to Ref. 15. Figure 6 shows how HDSR saves overhead in a 50 Node scenario which results better throughput (Figure 7). Number of FNs in the network naturally affects the performance of HDSR. We varied the number of FNs in the 50 node network where the mobile nodes are moving with zero pause time. We observed that increasing the number of FNs in the network improves the throughput up to a certain point. After that point(9-11 FNs), increasing the number of FNs will increase the routing overhead and deteriorate the performance, which is depicted in Figure 8. That is why we think that distribution of FNs in the network is important for optimization of the performance figures, but this study does not focus on the optimization problem.

4. Conclusion

We presented a new routing protocol based on Dynamic Source Routing (DSR) for heterogeneous ad hoc networks. HDSR limits the role of nodes during the routing discovery phase of the protocol, consequently increases the routing efficiency. We have shown via computer simulations that the HDSR improves network performance figures, namely throughput, delay and packet delivery ratio significantly. Now FNs are randomly distributed but we are working on the efficient and adaptive FN selection mechanisms for HDSR in homogeneous networks as well as optimization of number of FNs in a given network scenario.

References 1. C.E. Perkins and P. Bhagwat, “Highly dynamic destination-sequence distance-

vector routing (DSDV) for mobile computers”, Proc. of ACM SIGCOMM 94, pp. 234-244, London, UK, August 1994.

132

2. C.E. Perkins, “Ad Hoc On Demand Distance Vector (AODV) routing”, IETF Internet-Draft, draft-ietf-manet-aodv-OO.txt, November 1997.

3. J . Broch, D. B. Johnson, and D. A. Maltz, “The Dynamic Source Routing Protocol for Mobile Ad Hoc Networks”, IETF Internet-Draft, draft-ietf-manet- dsr-OO.txt, March 1998.

4. K. Fall and K. Varadhan, “Ns Notes and Documentation Technical Report”, University of California Berkeley, LBL, USC/ISI, and Xeron PARC, 2003.

5. S. Ni, Y. Tseng, Y. Chen, and J. Sheu, “The Broadcast Storm Problem in a Mobile Ad Hoc Networks”, in Proc. MOBICOM 1999, pp. 151-162 Seattle, Washington, August 1999.

6. W. Peng and X. Lu, “On the reduction of broadcast redundancy in mobile ad hoc networks”, Proc. of the First Annual Workshop on Mobile Ad Hoc Networking and Computing, pp. 129-130, August 2000.

7. B. Krishnamachari , S.B. Wicker, and R. Bejar, “Phase transition phenomenon in wireless ad-hoc networks”, Proc. of GLOBECOM 2001, Vol. 5, pp. 2921- 2925, San Antonio, Texas, November 2001.

8. Y. Sasson, D. Cavin and A. Schiper , “Probabilistic Broadcast for flooding in Wireless Mobile Ad Hoc Networks”, Swiss Federal Institute of Technology, Switzerland, Technical Report IC/2002/54.

9. Z.J. Haas, J.Y. Halpern, and L. Li, ‘‘Gossip Based ad hoc routing”, Proc. of INFOCOM 2002, Vol. 3, pp. 1707-1716, New York, June 23-27, 2002.

10. H. Lim and C. Kim, “Multicast tree construction and flooding in wireless ad hoc networks”, Proc. of 3rd ACM International Workshop on Modeling, Anal- ysis and Simulation of Wireless and Mobile Systems,Boston, Massachusetts, August 20, 2000.

11. Y. Tseng , S. Ni, E. Shih , “Adaptive Approaches to Relieving Redundant Storms in a Wireless Multihop Mobile Ad Hoc Networks”, Proc. of IEEE INFOCOM 2001 , April 22-26, Anchorage, Alaska.

12. H. Lim and C. Kim, “Flooding in wireless ad hoc networks”, Proc. ACM MSWIM workshop at MOBICOM, August 2000, Computer Communication J., Vol. 24, No. 3-4, February 2001.

13. C.R. Lin and M. Gerla , “Adaptive Clustering for Mobile Wireless Network”, IEEE Journal on Selected Area in Communication, Vol. 15, No. 7, pp. 1265- 1275 September 1997.

14. Y. Yi, T. Kown and M. Gerla, “Passive Clustering (PC) in Ad Hoc Net- works”, Internet Draft, draft-ietf-yi-manet-pac-OO.txt, November 2001.

15. C. E. Perkins, E. M. Royer, S. R. Das, and M. K. Marina, “Performance comparison of two on-demand routing protocols for ad hoc networks“, IEEE Personal Communications, Vol. 8, Issue 1, pp. 16-28, 2001.

Ad Hoc (11)

ANALYZING THE EFFECT OF COOPERATION APPROACHES

M. FRANK, M. HOLSCHBACH, P. MARTINI, M. PLAGGEMEIER INSTITUTE O F COMPUTER SCIENCE IV

UNIVERSITY O F BONN, ROMERSTR. 164, 53117 BONN, GERMANY {MATTHEW, HOLSCHBA, MARTINI, MP}QCS.UNI-BONN.DE

Mobile Ad Hoc Networks totally depend on cooperation. Each node is involved in forwarding data. Unfortunately, it is not obvious why a node should cooperate: forwarding data costs battery power and bandwidth. Research in this area is focused on approaches making stations behave more cooperatively. In contrast to this, the issue whether those approaches really improve the network performance is not well understood. In this work we present a general model for analyzing the effect of cooperation approaches in mobile ad hoc networks. To characterize the level of cooperation we use several classes of nodes, which forward data with class-specific probabilities. We show that cooperation approaches can improve the network performance in certain circumstances.

1. Introduction

Mobile ad hoc networks are becoming more and more popular. First, routing protocols like AODV or DSR were developed to get ad hoc networks started. Security and trust were no big issues. However, this has changed. There are a lot of new threats in ad hoc networks. One new point of interest is cooperation. There is no doubt that selfish or malicious nodes may cause significant problems. Several approaches (e.g. 19233,497*8), which try to solve this problem were proposed. The question whether those approaches really help the network is rarely asked. To our best knowledge, the first work presenting a simple analysis was 5 . In our work, we assumed that each node decides whether to forward a data packet or not. The probability for each node forwarding the packet was assumed to be identical. We have shown that cooperation approaches can be useful for the network under certain circumstances, but: The more selfish nodes are part of the network, the less significant is the improvement by a cooperation approach. The model used in has one drawback: It allows some essential statements on cooperation approaches but does not distinguish between reliable and malicious nodes.

135

136

Reliable nodes forward each data packet without delay whereas malicious nodes try to harm the network and forward packets with a low probability. The most relevant question is: Which kind of nodes form the majority in an ad hoc network? Is the network basically good or a collection of malicious stations. We study several kinds of networks and answer the question how a cooperation approach effects a network.

In this paper we present a general approach to analyze cooperation approaches. Our model neither depends on an underlying routing protocol nor on a specific cooperation approach. It is not important whether the cooperation approach is motivation based or detection based.

This paper is structured as follows: In section 2, we give an overview of known cooperation approaches. In the next section, we describe how we model an ad hoc network to allow for an analysis. Thereafter, we present our general approach to analyze the effect of cooperation approaches. Sec- tion 6 shows some simulation results. Finally, we present conclusions and discuss further work.

2. Related Work

Current cooperation approaches may be classified either as motivation based or detection based.

Motivation based Approaches: Most of the motivation based approaches use virtual currency (also called virtual nuggets or beans) to motivate the participants. One approach which can be used with a real currency has been introduced by Lamparter, Paul, and Westhoff 4 . It uses human minor motives and intensifies correct routing behavior in ad hoc networks by promising personal monetary benefits. An AAA-Service is used to secure transactions. To contact such an AAA-Service, the ad hoc network has to be connected to the Internet. In contrast to 4, B u t t y h and Hubaux

use virtual nuggets to motivate the participants to forward packets. They use tamper resistant hardware to protect transactions. A game-theoretic approach was proposed by Zhong, Chen, and Yang '. Each node gets credit when forwarding packets, the sender is charged when sending packets. The authors provide a formal proof to show that their system is secure.

Detection based Approaches: One of the first detection based approaches was the Pathrather/Watchdog model proposed by Marti, Giuli, Lai, and Baker '. It uses the promiscuous mode to detect whether a node forwards a packet or not. It tries to mitigate the effect of selfish nodes by using routes without those nodes. Buchegger and Boudec proposed a

137

similar approach '. CONFIDANT uses a reputation system to rate selfish- ness of nodes. Selfish nodes can be excluded from routing. In 7, Michiardi and Molva present another reputation based system to increase the level of cooperation. It is based on a collaborative monitoring technique. The first group based approach was presented in (CineMA - Cooperation En- hancement in MANETs). A group of nodes observe the neighborhood and punish malicious nodes by reducing their throughput. CineMA uses the level of cooperation: The more selfish a node, the worse the punishment.

3. Modelling Ad Hoc Networks

In 5 , we only considered one kind of nodes: Each node had the same forwarding behavior in terms of the percentage of packets discarded. This assumption provided us with some essential results. However, this assumption must be expected to be far away from real life behavior.

Modelling ad hoc networks is a big challenge. To analyze the effect of cooperation approaches we have to consider numerous classes of nodes. In our model we use classes of dropping behaviors. The more classes we consider, the more accurate the analysis. On the other hand: The more classes we have, the more complex the mathematical analysis. The model presented in this paper is not limited to a certain number of classes.

Let us now describe the model in detail. Equation 1 may be used to calculate the probability that a single packet reaches its destination over a pre-established path with h intermediate nodes in absence of a cooperation stimulating approach:

n

n is the number of different classes, ui is the percentage of the total nodes belonging to class i, lci (with Ici E [0,1]) is the percentage of packets forwarded by a node belonging to class i. In presence of a cooperation approach - which improves the forwarding behavior of a node - the probability rises to:

n

i=l

Ak, denotes the gain resulting from a cooperation approach. To quantify the benefit of a cooperation approach in terms of additional stations, a threshold must be used that reflects the percentage of packets which have

138

to reach the destination. We call this threshold T. The absolute hop gain Ah, quantifies the number of additional intermediate stations that can be used without falling below this threshold. To get Ah,, equation 1 and equation 2 are equated to T , solved to h and finally substracted. We get:

4. The Accuracy of the Model

As mentioned above, the model presented in this paper is not limited to a certain number of classes. However, to simplify the illustration of the effects, we use three classes, only. In this paper, we refer to the classes as class 1, class 2 and class 3 to make clear that the classes do not reflect a certain kind of nodes. Only the dropping behavior of the class - the percentage of packets forwarded to the next station and the percentage of packets dropped - is important. Each class has a characteristic dropping behavior called Cooperation level. Class 1 forwards 90% of the packets, class 2 forwards 60% of the packets and class 3 only 20% of the packets. We used these values because they reflect the behavior from "forwarding most of the packets" to "dropping most of the packets". In this paper, the cooperation level of the network is denoted as F = (Class l/Class 2/Class 3).

Equation 1 is used to calculate the probability that a single packet reaches its destination. This equation is based on an urn model with replacing. So, for each path of the ad hoc network, h - 1 balls of the urn are drawn and replaced. It could be argued that this model is inaccurate because an ad hoc network is modeled by an urn model without replacing - a node cannot be used twice on a route. But with the help of this model, it is much more easier to study the effects of a cooperation approach. This section discusses the accuracy of the model. We will show that the error by our model is very small for most of the networks.

Figure 1 shows the comparison of our model and the urn model without replacement. The classification of the classes used in this example is C = (60/10/30) that is there are 60 nodes of class 1, 10 nodes of class 2 and 30 nodes of class 3. All possible probabilities are plotted in this figure, each mark represents the probability of a ball drawn. The y-axis shows the probability of our urn model with replacement, the x-axis shows the probability of the urn model without replacement. The nearer a point to the dashed line the smaller the differences between the two models.

Obviously, the difference increases with the number of balls drawn. We

139

Y 1

- P 001 -

Draw wilhoul replace [PI 1

Figure 1. Comparison of the urn Figure 2. Prob. of Successfully Re- model ceived Packets

also see that the deviation of the two models is negligible. It can be shown that the expected maximum deviation of our model is negligible as well. Assuming that the length of a route in an ad hoc network is between three and five hops, then the expected maximum deviation is between 0.0005 and 0.002 per draw. The more nodes are part of the network the smaller is the deviation, 100 nodes are sufficient to have such a small difference.

5 . Analyzing the Effects

We analyze two different effects of a cooperation approach. First, we analyze the effect of improving the forwarding behavior of a single class or all classes. Then, we analyze the effect of increasing the percentage of nodes of a single class only.

5.1. Networks without any Cooperation Approach

Figure 2 shows the probability of a packet reaching its destination over five hops. Due to the fact that the sum of all classes must be 1 - the equation x:=l ai = 1 must hold - the fraction of one class depends on the fraction of the other two classes. Thus, the figure shown is a surface with three edges. The corners of the surface represent the maximum values. The lower left corner represents the case where all nodes belong to class 3, the upper corner represents the case where all nodes belong to class 2 and in the lower right corner all nodes belong to class 1. Moving in the direction of one axis means to increase the relevance of the corresponding class. For example, moving upwards the y-axis, leads to a higher percentage of class 2. The z-axis represents the probability of reaching the destination. The darker the surface, the lower is the probability of reaching the destination.

1 Della k

Figure 3. Prob. of rev. Packets. Figure 4. Longer Routes are possible

As it is shown in this figure, the probability of a single packet reaching its destination is close to zero in networks consisting basically of nodes belonging to class 3. The more nodes belong to class 1, the higher is the probability of reaching the destination. Figure 2 shows that the surface is curved. That is, the influence of nodes belonging to class 1 is higher than the influence of the other two classes. This results in the effect that small changes of the amount of class 1 can cause significant increase or decrease of the probability of reaching the destination.

For most applications, it is important that many packets reach their destination. Assuming that this threshold is 60% - we believe that this is a pessimistic threshold: normally this threshold should be higher - the number of stations that can be used without falling under this threshold is small. For example, in networks with a classification of C = (90/7/3) the maximum number of hops that can be used without falling under the threshold of 60% is 4.33. In networks consisting basically only of nodes belonging to class 3 (e.g. C = (20/10/70)), only one hop (1.52) - that is a direct connection between source and destination - can be used. These values illustrate that a cooperation stimulating approach may be necessary to improve the cooperation.

5.2. Improving the Forwarding Behavior of one Class

Which effect can be achieved when a cooperation approach improves the forwarding behavior of one class of nodes only? Figure 3 shows the effect when increasing the cooperation level of class 3 from 20% to 60%. The cooperation levels of our classes are now F = (0.9/0.6/0.6).

When compared to figure 2 the probability of a packet reaching its destination increases - we see that the surface is raised. The benefit of

140

141

networks where cooperative nodes form the majority is rather small. As- suming the classification of a network is C = (90/7/3). In this case, the probability raises from 0.54 to 0.57. Networks consisting of selfish nodes have a much higher benefit. For example, the probability of a network with a classification of C = (20/10/80) raises from 0.03 to 0.27. However, the overall throughput of good networks is much higher than the throughput of networks consisting of "selfish" stations.

The surface is not as curved as the surface of figure 2. This means that the influence of class 1 is smaller, the impact of the other two classes is more significant than before. The area where 50% of the packets are discarded on their way to the destination is now wider and is moved to the left. That is, compared to figure 2, networks with more selfish or malicious nodes have a higher probability to deliver successfully a packet to its destination.

Increasing the cooperation level of one class also leads to longer usable routes. Figure 4 illustrates this fact using a threshold of 60%. This means, at least 60% of the packets sent to the destination are successfully delivered. We see that increasing the cooperation level of class 3 leads to a small increase of the route in networks where nodes belonging to class 1 and class 2 form the majority. The more nodes belong to class 3, the higher is the benefit. However, the maximum number of hops that can be used without falling under the threshold of 60% is higher in networks consisting basically of cooperative nodes. Only high values of Ak lead to longer paths in networks where nodes of class 3 form the majority. Conclusion: Improving the forwarding behavior of one class which contains a lot of nodes is worthwhile. This means, improving the behavior of class 1 in good networks is worthwhile, whereas improving class 1 in networks consisting basically of selfish nodes has a small benefit, only. On the other hand: Improving the forwarding behavior of selfish nodes is worthwhile in "selfish" networks. However, to be as good as good networks, Ak must be very high.

5.3. Improving the Forwarding Behavior of all Classes

Figure 5 and figure 6 show the effect of increasing the forwarding behavior of all classes. In this example, the forwarding behavior of our three classes is F = (0.95/0.8/0.3) that is class 1 forwards 95% of the packets, class 2 forwards 80% and class 3 forwards 30% of the packets. Compared to figure 2, the surface is raised, which means, the probability of receiving a packet over five hops is higher than before. Even for networks only consisting of cooperative nodes (class l), the probability is now higher than

142

Y 1

0 0 1 0 2 0 3 P I 0 5 0 6 0 7 0 8 0 8 1 Delta k

Figure 5 . Prob. of rev. Packets. Figure 6. Longer Routes are possible

80%. Without any cooperation approach, the probability was 65%, only. Networks where selfish nodes form the majority have an even higher benefit. The probability of a network with a classification C = (20/10/80) raises from 0.03 to 0.32.

Studying figure 6 , it is remarkable that the curves consist of three parts: Values between Ak = 0 and Ak = 0.1, values between Ak = 0.1 and Ak = 0.4 and values larger than Ak = 0.4. This effect is caused by the different classes and the corresponding forwarding behaviors. Whenever the probability with which a node forwards a packet reaches the maximum of k = 1.0, the probability cannot be increased anymore: no additional benefit is possible. From figure 6, we observe that for small Ak the hop gain of networks consisting of cooperative nodes is much higher than of the other networks. This is due to the fact that in such networks the cooperating nodes form the majority. Ak = 0.1 leads to a maximum number of more than 8 hops without falling below the threshold of 60%. In this case, 90% of the nodes forward all packets. Conclusion: Improving the forwarding behavior of all classes increases the probability of receiving packets. The benefit is quit high for all kinds of networks. If Ak is small, the improvement of networks is much better for networks consisting of cooperative nodes than for networks consisting mainly of selfish nodes.

5.4. Increasing the Percentage of one class

Another effect of cooperation approaches may be to increase the percentage of one class. One reason might be the insight of a node to cooperate due to the fact that the punishment by the cooperation stimulating approach is worse. In figures 3 and 5 we observe that increasing the percentage of one class leads to a higher probability of receiving a packet. The more the

143

Figure 7. Inc. Class 1, decr. Class 3 Figure 8. Throughput of a Network

surface is curved, the higher is the benefit of networks consisting basically of nodes belonging to class 1. Even small changes of the classes can lead to a high benefit.

Figure 7 shows an example where the number of nodes belonging to class 3 is decreased and the number of nodes of class 1 is increased, denoted as Aa3,I. We see that networks consisting of cooperative nodes have a high benefit even for small changes. To get the same maximum number of hops for networks where selfish nodes form the majority, much more nodes have to change from class 3 to class 1.

The same statement holds in the case where nodes of class 3 and class 2 are decreased and the number of nodes of class 1 is increased. The curves are similar to the curves in figure 7. The benefit of the network is somewhat higher and leads to longer paths. Conclusion: Increasing the percentage of class 1 is worthwhile. The relative gain for good and bad networks is all about the same. However, the absolute gain is higher in cooperative networks. To achieve the same effect in selfish networks, Aai must be rather high.

6. Simulation Results

This section presents our simulation results validating our analysis. Fur- thermore, we show that the simulation yields the same conclusions as the analysis. The parameters of the simulation are as follows: NS2 was used as simulator, the area used was 1000m~1000m, 100 nodes equipped with IEEE 820.11b and a transmission range of 200m were used. 6 concurrent flows with 256 Bytes each 0.2 seconds were sent from a randomly choosen source to a randomly chosen destination. In addition, a Random Direction Model was used to ensure that a uniform distribution of the nodes in the area was maintained. Simulations using static scenarios were also performed. The

144

Classes. Cooplevel Hops Th.put Classes. (lOO/O/O) (l.O/-/-) 4.45 100% (80/10/10) 1100/0/0~ 10.9/-/-) 3.92 70.46% (80/10/10)

results are similar to those presented here, therefore they are not presented. First we show that with the help of our analysis we are able to compute

the throughput of an ad hoc network. In figure 8 we see that the throughput measured in the simulation was close to our analysig using equation 1. The difference between the two curves is negligible.

Table 1 summarizes our results. Increasing the percentage of a class which contains a lot of nodes is worthwhile. For example, increasing the percentage of class 1 from C = (80/10/10) to C = (90/10/0) leads to a benefit of 11.16%. Increasing class 1 in networks consisting of a few selfish nodes leads to a benefit of 3.49% only. The more nodes belong to a class the higher the benefit.

Increasing the cooperation level of class 3 yields a better benefit in selfish networks. From Table 1 we see that the throughput of the network C = (80/10/10) increases from 53.76% to 59.93%. This is a gain of 6.17%. Increasing the cooperation level of selfish nodes improves the network much more. In our example, the throughput increases from 19% to 34.80%, that is a gain of 15.8%. However, the throughput of selfish networks is much smaller than the throughput of cooperative networks.

The analysis has shown that in a network with the classification C = (20/10/70) the cooperation levels of F = (0.9/0.6/0.6) is similar to a network with the classification C' = (60/10/30) and the cooperation levels of F' = (0.9/0.6/0.2). In fact, this result was confirmed by the simulations. The first network has a throughput of 37.47%, whereas the throughput of the second network is 34.80%. That is a difference of 2.67%, only.

Cooplevel Hops Th.put (0.9/0.6/0.2/) 3.46 53.76% (0.9/0.6/0.6/) 3.65 59.93%

Table 1. Throughput and mean Number of Hops

I i90/i/3)' j i0.9/0.8/0.2/) j 3.72 j 62.44% j j i80/10/1oj j i0.95/0.8/0.6/) j 3.95 j 72.63% 1 (90/7/3\ I f0.9/0.6/0.6/\ 1 3.7R 1 64.92% 11 (90/10/01 1 f0.9/0.6/0.2/) 1 3.78 64.92% \~ I - ~ , I , ~ , ~ - , , ~~

.. \ - - , I -, (90/7/3) i i0.95/0.8/0.6j) i 4.09 i 78.95% / / (20/10/70) j i0.9/0.6/0.2/) i 1.82 j 19.00% (93/7/0) I (0.9/0.6/0.2/) I 3.83 I 66.80% 11 (20/10/70) I (0.9/0.6/0.6/) I 2.76 I 34.80%

7. Conclusion and Next Steps

In this paper, we have analyzed the effect of cooperation approaches. First, we have motivated our model and shown that the error of the model is negligible. Thereafter, we have analyzed two different effects: Improving the forwarding behavior of nodes and changing the percentage of a class.

145

The answer to the question whether cooperation stimulating approaches have a significant effect on mobile ad hoc networks was shown to be: "It depends on the network.

In case of networks where most of the participants are cooperative, a cooperation approach has a strong effect and yields longer possible paths whenever the cooperation approach increments the probability of the cooperative nodes. Improving the forwarding behavior of selfish nodes does not lead to a significant benefit. On the other hand, improving the forwarding behavior of selfish nodes yields longer pathes in networks where selfish nodes are in the majority.

Increasing the percentage of cooperative nodes improves the network as well. In case of basically cooperative networks, the effect is much stronger than for networks with a lot of selfish nodes. In networks with many reliable nodes, even a small increment yields significantly longer paths. Is the network a collection of malicious stations, the increment must be very high.

A mathematical analysis of complex networks with hundreds of stations and many different classes of nodes is very difficult. The impact of the MAC-Layers (e.g. wireless LAN), routing protocols, and interface queues have to be discussed as well. These issues are subjected to further research.

References 1. S. Buchegger, J. -Y. L. Boudec, Performance Analysis of the CONFIDANT

Protoco, Proc. of IEEE/ACM Symposium On Mobile Ad Hoc Networking and Computing, 2002

2. L. ButtyAn, J. P. Hubaux, Stimulating Cooperation in Self-organizing Mobile Ad Hoc Networks, ACM Journal For Mobile Networks and Applications, Spec. Iss. On Mobile Ad Hoc Networks, 2002

3. M. Fkank, P. Martini, M. Plaggemeier, CineMA: Cooperation Enhancement in MANETs, Conference on Local Computer Networks, 2004

4. B. Lamparter, K. Paul and D. Westhoff, Charging Support For Ad Hoc Stub Networks, Elsevier Journal of Computer Communication, 2003

5. B. Lamparter, M. Plaggemeier and D. Westhoff, Estimating the Value of Co- operation Approaches for Multihop Ad Hoc Networks, Elsevier Ad Hoc Net- works Journal, 2005

6. S. Marti, T.J. Giuli, K. Lai, and Mary Baker, Mitigating Routing Misbe- haviour in Mobile Ad Hoc Networks, 6th International Conference On Mobile Computing and Networking, 2000

7. P. Michiardi and R. Molva, CORE: A Collaborative Reputation Mechanism to Enforce Node Cooperation in Mobile Ad Hoc Networks, Sixth IFIP Conference On Security Communications, and Multimedia, 2002

8. S. Zhong, S., J. Chen and Y.R. Yang, Sprite: A Simple, Cheat-Proof, Credit- Based System For Mobile Ad-Hoc Networks, Proc. of IEEE INFOCOM '03

MOBILITY MANAGEMENT IN MULTIHOPS WIRELESS ACCESS NETWORKS

FABRICE THEOLEYRE AND FABRICE VALOIS CITI - INRIA ARES, INSA Lyon

21, Avenue Jean Capelle, 69621 Villeurbanne Cedex, fiance Email: { fabrice. theoleyre, fabrice. valois) @insa-lyon. fr

Tel: ( f 3 3 ) 4 72 43 63 27 and (+33) 4 72 43 64 18 Fax : ( t 3 3 ) 4 72 43 62 27

Ad hoc networks are spontaneous networks of mobile nodes which use wireless communications. The nodes must collaborate to route and forward data packets from a source to a destination. We consider multihops wireless access networks which are ad hoc networks connected to the Internet via Access Points (AP). A terminal can send and receive data packets to and from the Internet. To achieve this goal, we propose here a micro-mobility management solution. The solution uses a virtual backbone to centralize information and to limit the overhead. The proposition mixes the reactive and the proactive approaches to propose a trade-off between the delay and the overhead. In upload, each node has a default proactive route toward the AP introducing no latency. In download, the AP initiates a reactive localization to find a route to the destination inside the ad hoc area, reducing the overhead. A solution of paging with several AP and a solution of power-energy saving are also proposed.

1. Introduction

Mobile Ad hoc networks (MANET) could be defined as spontaneous networks: a collection of terminals organizes itself to exchange packets with each other via wireless communications. A source is not always a neighbor of the destination, so a route must be provided. Moreover, neither wired nor wireless router exists to manage the network. Thus, some terminals must collaborate to forward the data packets from the source to the destination. Ad hoc networks do not distinguish the routers and the clients: a terminal plays both roles. In a MANET, all terminals are independent and can move freely. MANET could be interconnected to the Internet via special gateways: the Access Points (AP). Such networks are often called Hybrid Networks or Multihops Wireless Access Networks.

Routing is one of the major issues in the MANET: a packet must be forwarded from a source to a destination without loss and with reduced

146

147

delay and overhead. Many protocols were proposed and can be classified in 2 major categories: the reactive protocols discover routes on-demand, flooding the network to find a route, whereas in proactive protocols, each node floods the network to create proactively routes toward it. Because of the particular constraints of MANET, we think that a self-organization is required. Some nodes must be chosen to manage the network, helping weakest nodes. These managers will stabilize the logical network view and will allow to reduce the control traffic. A self-organization6 was already proposed: strongest nodes form a virtual backbone, acting as routers and managers. This backbone can optimize floodings and reduce the control traffic for other nodes. We propose to use this virtual backbone to provide a new solution for mobility management. The data packets are forwarded proactively from the source toward the AP, constituting a default route to Internet. Inversely, routes from the AP toward a node could be discovered reactively. A solution of paging and power-energy saving are also proposed.

Next, we expose related work about routing in hybrid networks. Section 3 presents briefly the virtual backbone and the routing solution. Paging and power-energy saving solutions are also described. Section 4 presents some simulation results, before the conclusion and perspectives.

2. Related Work

MIPMANET3 proposes to integrate a reactive routing protocol (AODV) and Mobile IP. AODV is used for internal communications whereas external communications use Mobile IP. The Access Point acts as Mobile IP gateway (Foreign Agent): it must periodically flood Agent advertisements in the network, giving some Mobile IP parameters. Floodings present many problems in MANET4 : redundancy of transmissions exists and collisions occur creating a lack of reliability. The advertisements could be discovered reactively by clients5, but inducing a delay. A proactive and reactive combination is also possible, but the trade-off is complex to set. Moreover, a node must Aood several times the network before deciding that the destination is outside the MANET area.

To the best of our knowledge, only MEWLANA' proposes a solution optimized only for hybrid networks. The authors present 2 approaches. MEWLANA-TD allows both internal and external communications and is inspired from MIPMANET, but using a proactive routing protocol (DSDV). Each node knows instantaneously a route toward each node in the MANET area. However, the overhead for Agent Advertisements re-

148

mains unchanged. Moreover, the ratio of internal communications must be sufficient to justify the important control traffic amount required by DSDV. The second approach, MEWLANA-RD, is specifically designed for hybrid networks. The AP sends periodically Agent Advertisements. Each node registers the forwarder as the default route to the AP and forwards the Agent Advertisement. In parallel, each node answers in sending a Registration to the AP, on the new default route. However, the periodical Agent Advertisements and Registrations overload the radio medium, and many collisions occur, inducing many packet losses. Moreover, the systematic periodical reconstruction of the tree is suboptimal.

3. Proposition

Our solution mainly focus on the problems of both routing and mobility management in hybrid networks. The AP being the gateway to the Internet, it may constitute the default route, routing all the traffic, and acting as a Mobile IP Foreign Agent. If the AP has packets to deliver to one of its nodes with no associated route, one is discovered reactively.

We use a virtual backbone6. The backbone nodes, the dominators, are selected according to a stability weight representing their aptitude to act as network managers. Some nodes are elected dominators to form a connected structure where each normal node (or dominatee) is at most Ic hops far from its dominator. The backbone constitutes a tree of dominators where the leaves are the dominatees. Each dominator maintains the identity of its parent (except the AP) and the identity of the dominators for which it is a parent: they constitute its children. MANet are volatile environments. Hence, we have proposed a maintenance protocol to maintain the efficiency of the virtual backbone. To maintain the backbone connectivity, ap-hello are periodically sent by the AP, but forwarded only by dominators, limiting the overhead. Procedures for backbone reconnections are also proposed.

3.1. Mobility Management

3.1 .l. Upload

The AP can represent a suited default router. When a node wants to send a packet, it delivers it to the AP. Then, the AP acts as a proxy to find a route in the Internet, to do Network Address Translation if required,. . . The backbone is a tree rooted at the AP where a parent represents the next hop through the backbone to reach the AP. Each nodes maintains proactively

149

the identity of its k-neighbors, with classical hel lo packets. A dominatee knows the identity of its dominator, its distance, and a next hop toward it. This next hop can appears to be the default route. A dominator maintains the identity of its parent in the tree, from which it receives the ap-hellos. This parent represents its default route. We assume that communications will mainly be initiated by the nodes creating an efficient proactive feature: the route knowledge requires no latency and no additional overhead. An inverse route can be learned gratuitously, as described further.

3.1.2. Download

When a data packet is received from the node N , the node registers N as the next hop toward the source before forwarding the packet. Hence, an inverse route, i.e. in download, can be gratuitously learned when a node sends a data packet. Each data packet refreshes this proactive route, in triggering the associated timer in the cache of intermediary nodes.

However, if the node did not send a data packet, or if the timer of the route expired, the AP must implement a localization process. Because we assume this case seldom, we propose a reactive solution. When the AP receives a packet, and no route to the destination D is known, it buffers the data packet, the memory of the AP being supposed to be important. Then, the AP sends in multicast to its children in the backbone a Route Request. Each dominator N which receives this packet forwards the request if D is unknown. Else, N acts as proxy and sends a Route Reply to the AP, the source being seen as D.

The Route Request is forwarded along the backbone until D is found. If D is a dominator, one of its dominator neighbors or in the worst case the node itself will send a Route Reply. In the same way, if D is a dominatee, its dominator is a t most k hops far. Hence, D is in the neighborhood table of its dominator: a Route Reply will be generated. To limit the impact of the backbone disconnections, any dominator is allowed to act as proxy for Route requests. This limits the Route Request failure, at the cost of a negligible overhead. The Route Reply is sent on the default route, to the AP. A route is hop by hop created in the cache of each intermediary dominator, like with data packets. A distributive route cache is created, a timer being associated to this route. To adapt the solution to topology changes, a dominator which loses one of its children sends a Route Delete. This packet, forwarded to the AP, deletes all outdated routing entries.

150

3.1.3. Ad HOC Routing

As an extension, ad hoc connectivity can be proposed. A source 5’ sends a data packet to the destination D via its default route toward the AP. If a node receiving it knows a route to the destination, it sends directly the packet along this route. Else, the AP will finally receive the data packet. The AP will know if D is in the Internet or in the ad hoc area (with the address prefix, its own paging cache.. .). If D is in its covering area, it adds the data packet to the data buffer, and sends a Route Request. When the Route Reply arrives, the AP sends the buffered packets. Other data packets to D will be forwarded until it reached the first common ancestor of D and the source in the backbone. The route length and the delay are in conclusion not optimal.

3.2. Paging and Power-energy saving Schemes

Paging is used in cellular networks to limit the overhead of registrations. The node registers itself less frequently in its paging zone than in its AP. Paging consists in finding the AP serving a destination. A Paging Master (PM) is connected to all the AP constituting the paging zone. The Paging Master adds the mobile in its Membership Cache with a long timeout. When a packet arrives, the PM verifies that the destination is present in its Membership Cache. Then, it searches an associated entry in the Paging Routing Cache associating an AP to a node. If an entry is found, the data packet is directly sent to the corresponding AP. Else, the PM buffers the data packet and sends a Paging Request to all AP of its paging zone. These AP will send a Route Request. When an AP receives a Route Reply, it sends a Paging Ack to the PM. The PM adds an entry in its Paging Routing Cache with a short timeout and sends the buffered packet.

Terminals have limited energy reserves. However, turn off its radio is the only way to economize its energy2. Such a node does not participate in the network life, it sleeps. The backbone is particularly suited for such a feature. Elections are based on a weight depending on the energy reserve. Thus, a node with a too low energy reserve will not be elected as backbone member. Moreover, dominatees have a role of clients, they can spare their energy. The degree must be sufficient so that eventual backbone reconnections could be forwarded. Thus, a dominatee is allowed to sleep if the number of not-sleeping 1-neighbors is superior or equal to 6. Finally, a dominatee is allowed to sleep according to the probability l /n , n being the number of 1-neighbors with a lower weight.

151


We simulate our solution wit OPNET Modeler 8.1, with the WIFI standard model (300m radio range), and the random waypoint mobility model. The default parameters are a speed of 5m.s-l, 40 nodes and a degree of 9. Data flows of 8 data packets interspaced by 0.25s are sent according to the exponential distribution with a mean value of 2s. Data flows are sent in the same way from the AP to one random destination, and from a random source to the AP. The maximum distance from one node to the backbone is 2 hops. To evaluate the solution, we investigate the behavior according to the mobility, the load, the number of nodes, with the paging and the power-energy saving activated or not. We compare MEWLANA-RD and the proposed solution, labeled cdcl.

I D I I0 ,I I Z 1 x

o m

6P.M L-d l

Figure 1. Impact of the mobility on performances

Our solution is robust according to the mobility: the delay increases slowly, even with a very high speed of 30m.s-l. The delay of MEWLANA- RD follows the same tendencies. Our solution mixing the reactive and the proactive approaches doesn't suffer from the delay compared to the full- proactive approach. The delivery ratio is almost symmetrical in upload and in download. Oppositely, the delivery ratio in download is in MEWLANA- RD 10% lower than in upload. MEWLANA-RD reconstructing periodically the whole backbone, collisions occur. Finally, the delivery ratio in both directions is higher for cdcl thant for MEWLANA-RD.

Figure 2(a) presents the horizontal scalability of both solutions, i.e. the performances when the number of nodes increases. The delivery ratio of cdcl is higher than that of MEWLANA-RD. Moreover, the symmetrical property of the delivery ratio remains identical for cdcl. Both solutions are simulated with a constant degree when the number of nodes increases. Hence, the average route length growth when the number of nodes increases, increasing the delay too. However, this increasing remains acceptable.

152

ow

, I

-%%TE z= I I 1D I a €a 40 a 10 s IYJ 0 I I D , 6 . . 3 0 ~ 4 0 ~

* n . . d l l d l h l H m w.dBn)U*ol-sN*

(a) impact of the number of nodes (b) impact of load Figure 2. Scalability

Figure 2(b) presents the performances of both solutions according to the load of the network, i.e. the number of simultaneous communications. We can observe that both approaches are very scalable according to the load for the delivery ratio, although MEWLANA-RD keeps on presenting more packet losses. The delay increases lightly when the load becomes very important, when every node is in communication. However, such an augmentation remains below 50ms. Finally, we can note that the overhead of MEWLANA-RD (1.4 packet per second) is higher than for cdcl (0.8 pps).

We implement our solution of paging, creating a Paging Master. We place 2 AP on the surface, on the top left and top right corners. This network can deal with a more important load, the data packets being on average distributed among the AP, which is the requested property (fig.3(a)). However, the delivery ratio suffers more from the mobility. The delay of convergence of all routing caches and paging caches could be important, some data packets are dropped.

(a) Paging (b) Power-energy saving Figure 3. Impact of the Paging and the Power-energy-saving solutions

The power-energy solution was simulated. Third of the nodes don't generate or receive data packets, they can sleep if they are dominatees.

153

This power-energy saving solution has almost no impact on the delay, but the delivery ratio decreases slightly. If a collision occurs for an hello, the neighbors can choose as next hop a sleeping node, creating data packets losses. A non-communicating node is sleeping on average 25% of the time.

5. Conclusion

We propose here a routing and localization solution optimized for hybrid networks. This solution uses a virtual backbone to structure routing caches. In upload, information to maintain the backbone is used to find a route to the AP, constituting a gratuitous default route. In download, the AP implements a localization procedure to discover a route reactively. To minimize the frequency of the localization process, an inverse route is created on the fly when a node sends a Data Packet toward Internet. The s e lution presents a very high delivery ratio, and a limited delay although a reactive solution is used besides the proactive part. A solution of paging and power-energy saving are also proposed, taking into account the natural heterogeneity of the hybrid network. The backbone hides many topology changes, offering a stable view of the topology. Next step of this study will be the implementation of new functions like the handover with the choice of the optimal AP, the load balancing among the AP according to the number of nodes and a multicast routing solution.

References 1. Mustafa Ergen and Anuj Puri. Mewlana-mobile ip enriched wireless local area

network architecture. In 56th Vehicular Technology Conference, Vancouver, Canada, September 2002. IEEE.

2. L. Feeney and M. Nilson. Investigating the energy consumption of a wireless network interface in an ad hoc networking environment. In INFOCOM, Anchorage, USA, April 2001. IEEE.

3. Ulf Jonsson, Fredrik Alriksson, Tony Larsson, Per Johansson, and Gerald Q. Maguire. Mipmanet - mobile ip for mobile ad hoc networks. In Proceedings of the first ACM international symposium on Mobile and ad hoc networking and computing, pages 75-85, Boston, USA, May 2000. ACM, IEEE Press.

4. S.Y. Ni, Y.C. Tseng, Y.S. Chen, and J.P. Sheu. The broadcast storm problem in a mobile ad hoc network. In MobiCom, Seattle, USA, August 1999. ACM.

5. Yuan Sun, Elizabeth M. Belding-Royer, and Charles E. Perkins. Internet connectivity for ad hoc mobile networks. International Journal of Wireless Infor- mation Networks, 9(2), April 2002.

6. Fabrice Theoleyre and Fabrice Valois. A virtual structure for mobility management in hybrid networks. In Wireless Communications and Networking Conference (WCNC), pages 1035-1040, Atlanta, USA, March 2004. IEEE.

LOCATION UPDATE PROTECTION FOR GEOGRAPHIC AD HOC ROUTING

Z. ZHOU AND K. C. YOW School of Computer Engineering

Nanyang Technological University Singapore 639798

E-mail: (pg04570215,askcyow) Ontu. edu.sg

Security research in mobile ad hoc networks(MANETs) has been receiving increasing attention due to its evident importance in practical deployments. The main progress achieved so far is the protection of the critical Route Discovery component of on-demand routing protocols. In this paper, we identify that Location Update is the key component for geographic routing. To focus on securing the basic local location update, we propose three schemes in this paper with a detailed discussion.

1. Introduction

One of the most active areas of security research in MANETs is securing ad hoc routing. Considerable work 1 9 2 1 3 1 4 has been done to deal with security issues of some prominent on-demand routing protocols 5 9 6 . The main effort of their work is to design efficient security mechanisms to protect the Route Discovery procedure from malicious attacks. However, we are aware of a severe missing of proper study in geographic routing security. With the increasingly easier acquisition of location information, geographic routing schemes 1 3 9 1 4 are attracting more attention. It utilizes location information to make localized forwarding decisions, and is considered to be more scalable and resilient to node mobility compared with on-demand protocols. Our work focuses on securing geographic routing. We found that the coun- terpart of route discovery in geographic routing is location update, which is extremely vulnerable to a variety of attacks that compromise the security and performance brought by geographic routing.

Our paper is organized as follows. In Section 2, we identify the importance of Location Update and its vulnerabilities. To protect against malicious attacks, we propose three schemes for securing location update in Section 3. Section 4 discusses the security and gives the conclusion.

154

155

2. Threats of Geographic Routing

We consider Location Update as the major target for attacks in geographic routing, since it essentially drives the routing functionality. Basically, two underlying mechanisms make a typical geographic routing work: neighboring exchange and location service 8,9. We further generalize the tasks of geographic routing into one issue, i.e. location update. Neighboring exchange is essentially a constant update process to the node’s neighborhood, while location service could be generally viewed as a constant update to some remote nodes who behave as proxies of the node, or need the node’s location directly. The main difference of the two update processes is the “distance” between who updates and who is updated. Therefore, in another way, we define neighboring exchange as local location update (LLU), and location service as remote location update (RLU).

Obviously, location update including LLU and RLU will be the main target of attacks that threat the correct functionality of geographic routing. One of the attack instances is the Black hole attack 2, which is quite common in on-demand protocols, and it becomes much easier to launch in geographic routing. Basically, black hole attack causes the traffic to go through the attacker or certain point of the network, where most possibly the traffic is suppressed. To achieve this goal, the attacker may simply send a local update message with such a forged location that makes it compelling in the next hop selection of a certain traffic flow. In this case, the future traffic will be redirected to the attacker node.

Another typical attack that may happen is routing loop. A simple attack instance is illustrated in Figure 1. The notation of X ( z , y) means that the coordinate of node X is (z, y). M keeps on refreshing the spoofed location of E and A, causing the traffic destined for D to a routing loop A, E , B , A.

M F M? E r (1,3.5) i i A (1,4)

*A (1,2) C(1,3.4) 0 0-0 0 0

Figure 1. out of B to spoof as E and moves out of A but close to B to spoof as A.

Routing loop: A-E-B-A. The malicious node M moves in the range of A but

156

Identity spoofing is the main means for routing loop attacker to succeed. If spoofing of network identities is exploited excessively, it may even cause denial of service to a certain victim. The attacker can simply fabricate enough nodes around the edge of the victim’s radio range, leading to all the traffic from that node to the forged nodes controlled by the attacker. Suppression of all packets, in fact, denies the access of the node.

Threats of location update is not limited to those mentioned above. For more comprehensive discussion, please refer to our prior work lo. Since LLU and RLU share most of security problems and required properties such that solutions to LLU could be potentially extended to RLU scenarios. So, in this paper, we will focus on securing local location update.

3. Location Update Protection

In this section, we propose three secure LLUs to prevent malicious attacks and provide desirable security goals such as authentication, integrity and confidentiality. Schemes are denoted as LLU-2.

3.1. LL u-0

In LLU-0, each node periodically broadcasts an update message with its current location to its neighbors. The update message should be digitally signed by the sender with its private key, and its certificate is also attached with the message. The main fields of LLU message include message identifier, identity of the sender, current location of the sender, timestamp of the message about to send, a digital signature computed over all fields ahead, and the sender’s certificate. We denote the LLU message like <{LLU, id, loc, ts}sign, cert>. Each recipient of the message first verifies the attached certificate by using the public key of the trusted third party. If it is valid, it means the public key inside the certificate is genuine and truly associated with the claimed sender, and could be used to verify the signature of the message. F‘urthermore, the timestamp included can be used to determine if it’s a message replay or not. The location update will be finally accepted to update the recipient’s neighbor table if there is nothing wrong in all verifications. In this case, the node without a valid certificate is denied by the network since it will not be able to sign the update message properly. And any fabrication or spoofing of identities is not possible either.

LLU-0 applies the digital signature scheme to make the solution quite straightforward. It provides most of the expected security properties, such as the authenticity of the sender, integrity of the location, and even sender

157

non-repudiation, which is potentially useful as a deterrent of misbehavior since only the node with the corresponding private key can generate the signature. However, performing public-key cryptographic operations might be computationally expensive to low-end mobile devices. A node has to periodically sign its update messages, and has to verify all its received messages. And potentially it is more vulnerable to resource consumption attack, where a malicious node injects a large number of fake LLU messages that need processing. The computing resource or energy of innocent recipients may be exhausted in a short time. In fact, this problem exists in almost all cryptographic protocols that perform usually heavy cryptographic operations. Our defense solution is to mount an incoming traffic monitor, and each node could set a proper threshold value for incoming traffic volume that it can bear. Once the packet rate exceeds the threshold, it indicates that either a network exception or an attempt of resource consumption attack is taking place, and the node could respond accordingly to the event.

3.2. LLU-1

LLU-1 targets on reducing the computation cost and accordingly the risk of potential consumption attack when a pure public key scheme is applied. LLU-1 utilizes symmetric cryptography for authentication of update messages while it avoids the impracticality of using pairwise shared key. It’s basically an authenticated key exchange approach based on key transport. The goal of LLU-1 is that every node generates a “group decryption key” within its own domain, i.e. its neighborhood. The group decryption key will be used for neighbors to authenticate and decrypt its update messages. We require a periodic beacon message to detect new neighbors and to remove “dead” ones. Each beacon message includes the sender’s certificate. The event of new neighbor coming into node A’s range triggers a direct group key transport of A. Thus, the group decryption key is delivered to the new neighbor so that it will be able to decrypt messages from A.

A neighbor table will be maintained at each node with each entry at least including neighbor id, status of key distribution, group key of this neighbor, location, timestamp of location. An entry is added when a new neighbor is detected by receiving the update message, and is deleted when no group key is timely distributed or when timeout happens on this neighbor due to the node failures or a broken link. Note that every node is able to receive messages to maintain its neighbor table, but will not be able to verify the

158

update message and the source until it receives the corresponding group key from the source.

LLU-1 combines periodic beaconing and location update into a single update message, which maximally reduces the required messages exchanged to save power and bandwidth. A basic instance of this scheme is illustrated in Table 1. A and B are neighbors. Ek denotes the encryption operation with key k. gkx denotes node X’s group key. As a regular base, A broadcasts its update message to its neighbors. B found A is not in its neighbor table and started to transport its group key to A. To securely deliver the key, B extracts A’s public key from its certificate and encrypts the message after signing. On receiving B’s key transport, A was aware that B is also a new neighbor, and delivered its group key to B in a similar way. By exchanging group keys, they are able to authenticate each other and be updated with fresh and reliable location of neighbors. Although LLU-1 also involves public key operations, yet they are not performed periodically but reactively as needed.

Another feature of LLU-1 is that group key update for improving security becomes easier. It is simply another round of key transport. However, in order to avoid potential incompatibility of different versions of group key, sequence number could be introduced to identify the specific group key applied or transported.

3.3. LL u-2

Traditional symmetric authentication is not able to provide broadcast authentication property as LLU-0 can, i.e. a broadcast message can be directly verified by all recipients. If this property is available, multiple unicasts of key distribution in LLU-1 could be avoided. We explore, in LLU-2, the way of applying TESLA to achieve efficient broadcast authentication property based on symmetric cryptography. Each node using LLU-2 generates a chain of keys in advance by hashing an initial random value consecutively. These keys are used in a reverse order as generated, and will be disclosed in a few time slots after used. An update message includes these fields,

159

< “LLU”, id, Etki (id, loc, ts) , i , tk+l >, where E, means the encryption operation with key x, and tki is the ith TESLA key. Each node in the network is supposed to send LLU messages periodically.

In order to correctly determine the sender’s key disclosure, a recipient needs to know a few parameters that constructs a key disclosure schedule. It includes To, Tint, and d, where To is the starting time of the first time slot, Tint is the duration of a time slot, and d is the number of slots to pass before the key is disclosed. The very first commitment of the key chain is the last value of hashing. Assuming that any node has reliable key disclosure schedule information and the commitment of key chains of other nodes, the authentication at each receiver has the following two steps:

(1) In order to authenticate the message, the receiver has to know the sender’s key disclosure schedule, where it checks if the TESLA key tki used in the received message is disclosed or not. If the key is disclosed, the update will not be accepted. Otherwise, the message is buffered for future verification.

(2) Authenticating the disclosed key tki-1 requires the commitment of the sender’s key chain, and certain early-verified keys of the sender’s key chain. With the commitment, TESLA keys are completely self- authenticated, because hashing any disclosed key a certain number of times is supposed to be equal to the commitment (recall that TESLA keys are used in the reverse order of the chain generated by consecutive hashing). The authenticated key could be used to verify the old update message buffered. Furthermore, the disclosed key is qualified as a new commitment to replace the old one.

By using TESLA, LLU-2 allows nodes to efficiently generate and verify MACs over the periodic location updates. But LLU-2 presents another problem due to the temporal effect of TESLA for broadcast authentication, i.e. the node will always have a list of neighbors with old location information authenticated. How old is the location is tightly dependent on the frequency of location update as well as the key disclosure schedule. In a low mobility environment, it may have little impact. Our solution to this issue is that the sender could anticipate its future position based on its current speed and direction, and send the update message with the anticipated future location after key disclosure time. We consider that movement prediction should not be a difficult job, especially when only predicting into a several-seconds future. In this case, every node will have a reasonably accurate and authenticated position estimate of its neighbors.

160

As readers may have noticed, the LLU-2 scheme we discussed above has an assumption that the key disclosure schedule and the key chain commitment are the prior knowledge of every verifier. That means TESLA requires a prior receiver setup, i.e. these critical information has to be securely distributed to all expected receivers. It is not a good idea that key schedules for all nodes are fixed throughout the network lifetime and are pre-distributed to all expected receivers, since the storage requirement is similar to pairwise shared key which is linear to the network size. There- fore, LLU-2 also employs digital signature to secure the distribution of key disclosure schedule. But potentially dynamic neighborhood membership also requires this distribution to be periodic. However, LLU-2 does not require the node signing messages periodically. The signature generation for key disclosure schedule and key chain commitment could be an offline operation.

4. Conclusion

The schemes we proposed can thwart all attacks we mentioned in Section 3. Without valid authentication data, attackers are not able to masquerade as other nodes and to disseminate false location information. In addition to the authentication and data integrity that LLU-0, LLU-1 and LLU-2 can provide, in fact, some applications require the prevention from unauthorized eavesdropping, especially when location information becomes sensitive for individuals or the whole task. In this case, the confidentiality of LLU is required. LLU-0 and LLU-1 have the inherent ability to provide confidentiality while LLU-2 does not. Therefore, in terms of efficient confidentiality the tendency of using group key to encrypt messages becomes obvious.

As we can see, all three proposed schemes assume that each node has a prior certificate which is typically issued by an external CA. We consider this inevitable because most of the practical security proposals more or less rely on a public key infrastructure or certificate scheme. Providing PKI support for MANETs is also another important research track for MANET security. There are several distributed PKI schemes l1 ,12 proposed, which could be applied in our context besides having an external CA. The main difference of applying different prior security setup system is that they have different underlying trust relations in nature, which may fit into different applications or scenarios.

161

References 1. M. G. Zapata and N. Asokan, “Securing Ad Hoc Routing Protocol,” in

Proceedings of the 2002 ACM Workshop on Wireless Security (Wise 2002), September 2002, pp. 1-10,

2. Y.-C. Hu, A. Perrig, and D. B. Johnson, “Ariadne: A Secure On-Demand Routing Protocol for Ad Hoc Networks,” in Proceedings of the Eighth Annual International Conference on Mobile Computing and Networking (MobiCom 2002), Sept. 2002.

3. K. Sanzgiri, B. Dahill, B. N. Levine, C. Shields, and E. M. Belding-Royer, “A Secure Routing Protocol for Ad Hoc Networks,” in Proceedings of IEEE International Conference on Network Protocols(ICNP), November 2002.

4. P. Papadimitratos and Z. J. Haas, “Secure Routing for Mobile Ad Hoc Net- works,” in Proceedings of the SCS Communication Networks and Distributed Systems Modeling and Simulation Conference (CNDS 2002), 2002.

5. D. B. Johnson and D. A. Maltz, “Dynamic source routing in ad hoc wireless networks,” in Mobile Computing, Imielinski and Korth, Eds. Kluwer Academic Publishers, 1996, vol. 353.

6. C. Perkins and E. Royer, “Ad-hoc on-demand distance vector routing,” in Proceedings of the 2 nd IEEE Workshop on Mobile Computing Systems and Applications, Feb 1999.

7. A. Perrig, R. Canetti, D. Tygar, and D. Song, “The TESLA Broadcast Au- thentication Protocol,” R S A CryptoBytes, vol. 5, no. 2, pp. 2-13, 2002.

8. J. Li, J. Jannotti, D. De Couto, D. Karger, and R. Morris, “A scalable location service for geographic ad-hoc routing,” in Proceedings of the 6th ACM Inter- national Conference on Mobile Computing and Networking (MobiCom ’00),

9. Y. Xue, B. Li, and K. Nahrstedt, “A scalable location management scheme in mobile ad-hoc networks,” in 26th Annual IEEE Conference on Local Computer Networks (LCN’Ol), 2001.

10. Z. Zhou and K. C. Yow, “Geographic ad hoc routing security: attacks and countermeasures,” to appear in Ad HOC €4 Sensor Wireless Networks, Old City Publishing, 2005.

11. S. Capkun, L. Buttyan, and J. Hubaux, “Self-organized public-key management for mobile ad hoc networks,” in ACM International Workshop on Wireless Security( W i s e ’02), 2002.

12. H. Luo, J. Kong, P. Zerfos, S. Lu, and L. Zhang, “Self-securing ad hoc wireless networks,” in IEEE Symposium on Computers and Communications), 2002.

13. B. Karp and H. T. Kung, “GPSR: greedy perimeter stateless routing for wireless networks,” in Proceedings of Mobile Computing and Networking, pp.

14. P. Bose, P. Morin, I. Stojmenovic, and J. Urrutia, “Routing with Guaranteed Delivery in Ad Hoc Wireless Networks,” Wireless Networks, vol. 7 , no. 6, pp.

120-130 2000.

243-254, 2000.

609-616. 2001.

PEDCF: PREDICTIVE ENHANCED SERVICE DIFFERENTIATION FOR IEEE 802.1 1 WIRELESS AD-HOC

NETWORKS BASED ON AUTOREGRESSIVE-MOVING AVERAGE PROCESSES

NABIL TABBANE Mediatron, SUPCOM, Route de Raoud Cite' El Ghazala, Ariana 2083, Tunisia

SAM1 TABBANE Mediatron, SUP'COM, Route de Raoud Cite' El Ghazala, Ariana 2083, Tunisia

AHMED MEHAOUA Prism, University of Versailles St-Quentin-en- Yvelines, 45, av. des Etats- Unis

Versailles, 78035, France

In this paper, we tray to reduce the degree of QoS degradation in IEEE 802.1 1 wireless ad-hoc network while enhancing at the same time the estimation of the quality of service of the networks. We shall present PEDCF: Predictive Enhanced Service Differentiation methods for forecasting resources to meet the QoS requirements for real-time service support based on AutoRegressive Moving Average processes: ARMA. These processes provide a range of models, stationary, that adequately represent the time Contention Window (CW) variations. The results obtained (in terms of throughputs and end-to-end delays) show that the PEDCF protocol, based on ARMA processes, performs better than conventional EDCF.

1. Introduction

We present methods for forecasting resources to meet the QoS requirements in IEEE 802.1 1 wireless ad-hoc network [ 11. Our approach, called Predictive Enhanced Distributed Coordination Function (PEDCF), is derived from the EDCF introduced in the IEEE 802.1 l e standard. It is combined with Mixed AutoRegressive-Moving Average (ARMA) processes to give forecasting about the behavior of contention window.

We have implemented PEDCF in the NS-2 network simulator. Several simulation scenarios have been used to evaluate its performance and to determine the optimal value of certain parameters. We analyze through

162

163

simulations the efficiency of the slow and predictive decrease of the CW after each successful transmission and compare it with the classic scheme proposed in the standard. Results show that PEDCF outperforms the basic EDCF, especially at high load conditions. Indeed, our scheme increases the medium utilization ratio and reduces considerably the collision rate. While achieving delay differentiation, low jitter is also maintained, and the overall goodput obtained is up to 30% higher than EDCF. Moreover, the complexity of PEDCF remains similar to the EDCF scheme, enabling the design of cheap implementations.

2. The Predictive EDCF (PEDCF) scheme

Let n the number of active stations, and i the priority class. The flows sent by each station may belong to different classes of service with various priority levels. In each station and for each class i, the following parameters are defined:

Note that i varies from 0 (the highest priority class) to 7 (the lowest priority class).

In order to efficiently support time-bounded multimedia applications, we use a dynamic procedure to change the contention window value after each successful transmission and after each collision. We believe that this adaptation will increase the total goodput of the traffic which becomes limited when using the basic EDCF, mainly for high traffic load.

After each successful transmission, the EDCF mechanism resets the contention window of the corresponding class i to CW,,,[I'] regardless the network conditions. Motivated by the fact that when a collision occurs, a new one is likely to occur in the near future, [ 6 ] propose to update the contention window more slowly (not reset to CW,,,,,[I']) after successful transmission to avoid bursty collisions. The simplest scheme to update CW[I'] is to decrease it by a multiplicative factor such as O.5*CWold. This approach is denoted the Slow Decrease (SD) scheme. However, a static factor cannot be optimal in all network conditions. In our scheme, we propose that every class updates its CWparameter in a predictive way taking into account the previous CWs in each station. Indeed, the past values of C W can give an indication about network states. The predictive C W is calculated during a constant period (i.e. a fixed number of slot times). This prediction is based on ARMA: mixed AutoRegressive Moving Average model.

CW[I']: the current contention window, CW,,,,[I']: the minimum contention window, and CW,,[iJ: the maximum contention window.

164

2.1. ARMA: mixed AutoRegressive-Moving Average model

A stochastic model which can be extremely useful in the representation of certain practically occurring time CW is the so-called autoregressive model [7]. In this model, the current value of the process is expressed as a finite, linear aggregate ofprevious values of the process and an error term.

Let us denote the values of a process at equally spaced times t, t-1, t-2, ... by z,, zt-l, zt-2, ... ; Also let it, &, .. . be deviations from p (the mean about which the process varies), it = Z, - p. it represents the estimated contention window at time t.

To achieve greater flexibility in fitting of actual time contention window, it is sometimes advantageous to include both autoregressive and moving average terms in the model [3]. This leads to the mixed autoregressive-moving average model [4]

(1) which employs p+q+2 unknown parameters p; cDl, ..., cDp; el, ..., 8,; 6; (the variance of the error term at), that are estimated from the data.

In practice, occurring stationary time CW can be obtained with autoregressive, moving average, or mixed models, in which p and q are not greater than 2 (to obtain a parsimonious model). Eq. (1) will define a stationary process, provided that the characteristic equation has all its roots lying outside the unit circle.

For the ARMA(p,q) process, there will be q autocorrelations p,, pq+ . . ., pI whose values depend directly, on the choice of the q moving average parameters 8, as well as on the p autoregressive parameters cD. Also, the p values p,, pql, . . . , pq-p+l provide the necessary starting values which then entirely determines the autocorrelations at higher lags.

By substituting estimates r, for pi, initial estimates for the parameters cDi and 8i can be obtained [S].

In the following sections, based on simulation, we analyse the performances of the ARMA models combined with EDCF protocol in congested environment.

iIt = cD18t-l + ... + (Dp8t-p + a, - 81a,-l - ... - 8,a,.,

2.2. Setting CW after each Update Period TepdO,,

The predictive CW is computed dynamically in each period TuPdure expressed in time-slots. This period called Update Period should not be too long in order to get good estimation and should not be too short in order to limit the complexity. For each update period, the predictive CW is computed according to Eq. (1).

165

We have fixed these parameters according to a large number of simulations done with different scenarios. In PEDCF and after each successful transmission of packet of class i, CW[g is then updated as follows:

Eq. (2) guarantees that CW[g is always greater than or equal to CWm,,[g, so the priority access to the wireless medium is always maintained.

In the current version of EDCF, after each unsuccessful transmission of packet of class i, the CW,,,,,[g is then doubled, while remaining less than the maximum contention window CWmux[iJ:

We propose in PEDCF, after each unsuccessful transmission of packet of class i, the new CW of this class:

CWnav[g = mm(CWmJi/, CWpmi[g) (2)

C Wnav[g = min(C Wmm[g, 2 * C Wc,,d[iJ (3)

Cwne,v[g = min(CWmux[U, CWppedil) (4)

2.3. Complexity of PEDCF

The complexity of PEDCF is similar to the complexity of EDCF, only a few more resources are required. Some registers are necessary to buffer the parameters defined above: p, cDl , cD2 , , O2 , 6; in ARMA(2,2) parsimonious model.

The calculation of each CWpmd[g requires four additions, four multiplications and one comparison for each active class. Finally, during the update period, two counters are needed to increment collisions and data sent.

3. Simulation Methodology and Results

We have implemented the PEDCF scheme in the ns-2 simulator [ 5 ] . In this section, we investigate and analyze the performance of PEDCF under several scenarios. As mentioned before, the PEDCF adapts the contention window values according to the previous CW each T,,pdure time slots (Tupdare value equal to 5000 time-slots [6]). We have n stations indexed from 1 to n. Each station generates the same traffic consisting of three data streams labeled according to their priorities with high, medium and low. Station n sends packets to station 1. Station i (i < n) sends to station i + 1 three flows of different classes: Audio (high priority), Video (medium priority), and Background Traffic (denoted by BT of low priority). In the following simulations, we assume that each wireless station operates at IEEE 802.11a PHY mode-6 [2].

166

3.1. Impact of the trafflc load

To evaluate the performance of PEDCF, we investigate in this section the effect of the traffic load and compare it with EDCF and SD schemes.

Our simulations use different types of traffics to evaluate service differentiation. Three queues are used in each station. The highest priority queue in each station generates packets with packet size equal to 160 bytes and inter- packet interval of 20 ms, which corresponds to 64 Kbit/s audio flow. The medium traffic queue generates packets of size equal to 1280 bytes each 10 ms which corresponds to an overall sending rate of 1024 KBit/s. The low priority queue in each station generates packets with sending rate equal to 260 KBit/s, using a 200 bytes packet size. To modify the load of the network, we have used a different number of stations, which gradually increases during the simulations. Every station is able to detect a transmission from any other station, they are not moving during the simulations. We start simulations with two wireless stations, then we increase the load rate by incrementing the number of stations by one every eight seconds. Figures 5 - 9 show the averages of delay, goodput gain, medium utilization and collision rate over 5 simulations. The number of stations is increased from 2 to 44 which correspond to load rates from 7.5% to 160%.

The relationship between the load rate and the number of stations is shown in Table 1. To evaluate the performance of the different schemes, the following metrics are used: Gain of goodput, Mean delay, Latency distribution, Medium utilization and Collision rate.

Table 1. Correspondence between number of stations and load rate

Number of stations Load rate 2 7.5 Yo 5 10 15 20 25 30 35 40 42

19 % 37 % 56 % 75 % 94 % 110% 131 % 150 % 160 %

44 170 %

Figure 1 shows the average delay for the audio flow corresponding to the high priority class. The PEDCF scheme is able to keep the delay low even when

167

the traffic load is very high, i.e., with a large number of stations. We can observe that the average delay of audio for PEDCF is 50% smaller than that for the basic EDCF when the load rate is up to 100% (26 stations). Moreover, when the number of stations is more than 17, the delay obtained by the SD increases faster than PEDCF and EDCF schemes, while PEDCF always keeps a lower mean access delay less than 8 ms. We can also note that PEDCF offers an average delay 45% less than the SD scheme and 40% less than the EDCF scheme when the load rate reaches 170%.

In Fig. 6, we plot the gain on goodput as a function of the traffic load of PEDCF and SD schemes. We observe that the goodput gain of PEDCF increases when the traffic load increases. It reaches about 23% when the load rate is about 130% (i.e. for 35 stations). Moreover, the goodput of PEDCF is 15% higher than the SD scheme when the load rate is 170%. Indeed, PEDCF is much more efficient during high load rate.

+ PEDCF 20 8 2

: 2

2 0 5 10 15 20 25 30 35 40 45

= 8 c 10

c 5 E :

.- .-

5 1 0 1 5 2 0 2 5 3 0 3 5 4 0 d o

Number of stations

Figure 1 : Audio average delay

Number of stations

Figure 2: Gain on goodput Figure 3 shows the medium utilization as a function of the traffic load. For

the three schemes the medium utilization decreases when the traffic load increases. However, PEDCF achieves the highest medium utilization whatever the number of stations. Indeed PEDCF offers 15% of resources moreover than SD and 30% moreover than EDCF.

Number of stations Number of stations

Figure 3: Medium utilization Figure 4: Collision rate The corresponding collision rate is shown in Fig. 8. Collision rates achieved

by the three schemes are similar when the traffic load is low, i.e. the number of

168

stations is less than 8. However, when the traffic load increases, PEDCF is able to maintain a lower collision rate than EDCF and SD schemes. We can explain this behavior by the fact that PEDCF uses an predictive technique (which approaches reality) to change the contention Windows according to the collision rate. The reduction of collision rate of PEDCF leads to significant goodput improvement and reduces the delay.

We have used a different simulation to study performance on delay and jitter. This new experiment has the same topology than before, but the number of stations is increased from 2 (4 sec) to 25 (100 sec) and the simulation stops at t = 1 15 sec. The delay variations of both EDCF and PEDCF schemes are plotted in Fig. 9. PEDCF meets target by maintaining the delay lower than EDCF and stable during the live of audio sessions. However, we can note that both delay and jitter for EDCF are three times higher than PEDCF, which degrade the quality of audio flows.

0 100 0 25 1W 25 SirnulatioiYtime (in seTj

(b) Audio delay for PEDCF SirnulatiorYtirne (in seEj

(a) Audio delay for EDCF Figure 5: The audio class delay variation

From the simulations, we can conclude that both PEDCF and SD schemes outperform EDCF. Using a predictive algorithm, PEDCF get much higher goodput than the SD scheme. Moreover, the PEDCF scheme can improve the goodput and delay performance of all types of traffics.

4. Conclusion

This paper has described an efficient means for predicting Quality of Service for IEEE 802.1 1 ad-hoc WLANs, to support real-time services. Our approach, called Predictive Enhanced Distributed Coordination Function (PEDCF) combined with forecasting contention window Method based on Mixed AutoRegressive-Moving Average ( A M ) processes improves drastically the performance of the real-time service support.

Results show a clear advantage of QoS mechanism based on PEDCF over EDCF or SD mechanisms. The PEDCF scheme is able to keep the delay low even when the traffic load is very high. We observe that the average delay of audio for PEDCF is 50% smaller than that for the basic EDCF when the load rate is up to 100%.

169

PEDCF is much more efficient during high load rate. Indeed, it reaches about 23% when the load rate is about 130%. Moreover, the goodput of PEDCF is 30% higher than the EDCF when the load rate is 170%. The protocol PEDCF allows more resources than EDCF or SD. Indeed PEDCF offers 15% of resources moreover than SD and 30% moreover than EDCF.

We note that when the traffic load increases, PEDCF is able to maintain a lower collision rate than EDCF and SD schemes. The reduction of collision rate of PEDCF leads to significant goodput improvement and reduces the delay. we can note that both delay and jitter for PEDCF are three times lower than EDCF, which ameliorate the quality of audio flows.

In a future work, we will study EDCF protocol combined with Seasonal Processes which provide a non stationary models, that can represent the optimal forecasting procedure of QoS to support real-time services in IEEE 802.1 1 ad- hoc WLANs.

References

1.

2.

3.

4.

5.

6.

7.

8.

IEEE WG, Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, IEEE 802. I I Standard, (1 999). IEEE 802.1 la, Part 11 : Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications: High-speed Physical Layer in the 5 Ghz Band, Supplement to IEEE 802. I I Standard, (Sep. 1999). N. Tabbane, S. Tabbane, A. Mehaoua, "Stationary Stochastic Models for Forecasting QoS in Ad Hoc Networks for Real Time Service Support", WSEAS Transaction (2004). N. Tabbane, S. Tabbane, A. Mehaoua, "Autoregressive, Moving Average and Mixed autoregressive-moving average processes for Forecasting QoS in Ad Hoc Networks for Real Time Service support", in Proc. of VTC Spring, Milan, Italy, (May 2004). K. Fall, K Varadhan, "NS Notes and Documentation," A Collaboration between researchers at UC Berkeley, LBL, USCLSI, and Xerox PARC, (February 25,2000). L. Romdhani, Q. Ni, T. Turletti, AEDCF: Enhanced Service Differentiation for IEEE 802.1 1 Wireless Ad-Hoc Networks, INRIA, Rapport de recherche No 4544, (September 2002). G. E. P. Box and G. M. Jenkins, "Discrete models for feedback and feedforward control," The Future of Statistics, ed. D. G. Watts, 201, Academic Press, New York, (1968). G. E. P. Box and G. M. Jenkins, "Discrete models for forecasting and control," Encyclopedia of Linguistics, Information and Control, 162, Pergamon Press, (1 969).

IEEE 802.11

A CARRIER-SENSE BASED TRANSMISSION POWER CONTROL PROTOCOL FOR 802.11 NETWORKS

JAYANTHI RAO, SUBIR BISWAS Department of Electrical and Computer Engineering, Michigan State University

@yanihi,sbiswas}@egr.msu. edu

This paper presents a new transmission power control protocol, CSNE-PC that addresses the energy inefficiency of basic power control protocols (BPC) in which RTS and CTS packets are exchanged at full power while data and acknowledgement transmissions are carried out at low power. This is accomplished by reducing hidden data collisions which is responsible for energy inefficiency in BPC. CSNE-PC implements a mechanism for inference of low-power data transmissions in the presence of power control. A node measures the duration of its carrier-sense activities and analyses it for detecting 802.11 CTS packets. Upon detecting a CTS packet, the node sets or extends its NAV and enters into a silence mode for a pre-determined packet duration. This enables transmission power control without incurring additional data collisions as described above. Evaluation of CSNE-PC through simulation demonstrates that under sustainable loading conditions, reduction in combined transmit and receive energy can be as high as 38%, and that is while retaining the throughput and delay characteristics of the regular 802.1 1 protocol.

1. Introduction

A number of Transmission Power Control (TPC) schemes have been proposed with the intention of minimizing the energy spent on transmission and potentially increasing spatial channel reuse at the same time [2][3][6][7]. The idea behind TPC is to use an optimal transmission power to reach an intended receiver. In addition to reducing the transmit energy consumption, TPC implies that the reception and the carrier sense zones for packets transmitted at lower power are also reduced. Hence, the number of nodes that overhear [S] and carrier-sense the packets are fewer. Thus, as a second order effect, TPC can lead to a reduction of reception and carrier sensing energy expenditure as well.

Problems with Basic Power Control (BPC): In this scheme [4][5][6], transmitter and receiver transmit RTS and CTS at the maximum power level supported by the wireless interface. But data and acknowledgement are sent at the minimum required power level. It can be shown [l] that power control for DATA and ACK packets in BPC can significantly increase packet collisions. As reported in [ 11 and Section 4, such collisions can severely impact the energy performance of BPC. In the example presented in Figure 1, node A transmits an RTS at maximum power and node B responds with a CTS. The circular regions around

173

174

the nodes indicate the reception and carrier sensing zones for maximum power transmissions. Nodes in the carrier-sensing (CS) zone of the transmitter and the receiver such as nodes C, E and D sense the RTS/CTS and defer their own transmission as needed. Upon receiving the CTS from B, the transmitter A transmits the DATA at the minimum required power level. In this example, the hexagonal regions around the nodes A and B represent the reception and carrier- sensing zones for the low-power DATA and ACK transmissions. DATA Collisions: These collisions can be caused by nodes in the carrier-sense zone of the receiver and they are not completely prevented even in the plain 802.1 1 protocol without power control [ 13. However, in BPC, DATA is more prone to collisions due to the following reasons.

CS zone for max-

power transmission

Figure 1. Collision Scenano for Basic Power Control

First, consider the example in Figure 1 executing plain 802.11 without power control. When node A transmits DATA to node B, node C and node E sense the transmission and defer their own transmissions for EIFS duration. So, node C and node E do not cause a DATA collision at node B. However, node D is not in the carrier sense range of node A and

hence is not aware of the DATA packet being transmitted from A to B. Therefore, only node D can cause a DATA collision at B.

Now, consider the example in Figure 1, when BPC is used. Nodes C and D are outside the low-power carrier-sensing zones of A and B. Therefore, they are not able to sense the low power data transaction between A and B. If either node D or node E has a packet to transmit, it will transmit an RTS at maximum power which may collide with the ongoing data transmission from node A to B. As a result, in BPC, node E can also cause a DATA collision in addition to node D. ACK Collisions: In 802.1 1 without power control, collisions of ACK packets at the transmitter by the nodes in its carrier sense zone are prevented [ 11. In Figure 1, even though node C is not in the carrier sense zone of node B, it cannot cause an ACK collision at node A. This is because node C had set its NAV for EIFS duration when it sensed the DATA and EIFS is long enough to ensure that the ACK packet is successfully received at the source. Now, consider the case when BPC is used. Node D, E and C sense neither the DATA from A to B, nor the ACK from B to A. Therefore, nodes E and node C can start transmitting

175

RTS at maximum power and can cause an ACK collision at A. Both DATA and ACK collisions, as described above, can affect the energy performance of BPC. 2. Related Work

Transmission power control has been incorporated at the MAC layer in two classes of protocols. In the first class, unlike BPC, data, acknowledgement and control packets (CTSRTS) all are transmitted at the minimum required power level [7]. A node uses distinct power levels for transmissions to each of its neighbors. This scheme is prone to collisions due to a special hidden node problem, which is occurs because the reserved floor does not include all the nodes that can potentially interfere with the transmission [2]. Such collisions can reduce the effectiveness of this class of power control protocols.

In the second set of protocols, which are designed around the Basic Power Control (BPC) scheme, RTS/CTS control packets are transmitted at the maximum power level while DATA/ACK packets are transmitted at the minimum required power level [4][5][6]. By transmitting RTS/CTS at maximum power, these protocols reserve the same floor for a transmission as in the plain IEEE 802.1 1 without power control. For every packet transmission, RTS/CTS packets are used to determine the minimum required power level for the DATA/ACK transmissions. Protocols that use the above scheme suffer from the collision problem described in Section 1.

PCM protocol [l] attempts to avoid these collisions using the following approach. It requires that irrespective of the minimum required power level, the data be transmitted periodically at the maximum power just for a duration that is long enough for the nodes in the carrier sensing region of the transmitter to lock on it. This approach aims to avoid the ACK collisions at the transmitter that are introduced by BPC. In the process, protection of DATA packets also becomes similar to plain 802.11. Note that since the reserved floor is same as that of IEEE 802.1 1, PCM delivers throughput comparable to plain 802.1 1.

3. CARRIER-SENSE BASED NAV EXTENSION FOR EOWER - CONTROL (CSNE-PC) In this paper we propose an approach that is different from PCM but it

attempts to accomplish similar goals. 3.1. Protocol Description

The basic goal of CSNE-PC is to ensure that all nodes in the carrier-sensing (CS) zone of the receiver are able to infer the presence of any low power data transmissions between that transmitter-receiver pair. To achieve this, in CSNE- PC, a node measures its surrounding channel activities by monitoring the received control and data packets as well the carrier-sense signals that it cannot

176

decipher. Based on the durations of the carrier-sense signals, the node attempts to determine if the carrier-sense matches the duration of an 802.1 1 CTS packet. If the node decides that it has sensed a CTS, it sets or extends its NAV.

Consider the scenario depicted in Figure 2, in which node Q sends a max- power RTS to node P. Node C does not receive or sense the RTS as it is out of the carrier-sensing range of Q. Node P responds to the RTS by transmitting a CTS at maximum power. Note that node C is outside the reception zone but within the CS zone for P’s transmission. Therefore, Node C senses the CTS signal from Node P, but it cannot decipher it. As a result, in the plain 802.11 protocol, no action would have been taken by node C in this situation.

I I In CSNE-PC.

ACK

however, based on the carrier-sense duration, C concludes that the signal is a CTS, and then it sets its NAV for a pre-defined packet duration. Now, Q starts transmitting the data to P at the minimum required power. Even though C is not able to sense or receive DATA from Q, it does not

transmit for the DATA duration because of the NAV set based on the carrier- sense of CTS. Thus, CSNE-PC avoids the DATA collisions that exist in BPC.

Now consider the following scenario. As the data is being transmitted from Q to P, node A sends a max-power RTS to B. Node B sends a CTS at maximum power which is not received, but carrier-sensed by node C. Based on the duration of the carrier-sense signal, node C determines that the sensed signal is a CTS, and then it extends the already set NAV for another pre-defined packet duration. This ensures that node C defers its transmission till both Q and A complete their transmissions. This NAV extension in CSNE-PC ensures that node C does not prematurely initiate a transmission and cause collisions with A’s transmission at node B or Q’s transmission at P.

Carrier-sense based NAV extension could be based on detection of RTS packets, CTS packets or both. In CSNE-PC, a CTS based NAV extension (and hence DATA collision avoidance) is chosen for the following reasons. First, in 802.11, an RTS transmission does not necessarily result in a data. But, by the time a CTS is transmitted, the first hurdle namely RTS collision has been overcome, and therefore it is logical to assume that a CTS is more likely to be

Figure 2. Timing-diagram of CSNE-PC

177

followed by data than an RTS. Second, due to their large size difference, the likelihood of a DATA to experience collision at a receiver is much larger than that of an ACK to experience collision at the transmitter. Hence, by performing a CTS based NAV extension we expect to reduce DATA collisions and hence energy consumption. During simulation experiments, NAV extension based only on RTS packets were performed. We found that NAV extension based on RTS impacts throughput adversely as nodes can needlessly defer their transmissions.

3.2. Measurement based CTS Detection

Detection of CTS packets based on carrier-sensing is the key for the CSNE- PC protocol. In practice however, it is not always possible to detect all CTS packets based on carrier-sense duration. The reason for this is that often the carrier-sense signals from multiple nodes can overlap at a receiver, and that makes it particularly hard to isolate the leading and trailing edges of individual signals representing CTS packets. Keeping this in mind, we implemented a pattern extraction algorithm that detects a CTS signature when a carrier-sense signal corresponds to a pure CTS without any overlapping signal, or a CTS is the most recent among a set of overlapping signals.

Other than transmitting, at any given time an 802.11 wireless interface can be in one of PHY-Idle (Idle), PHY-CS (Carrier-Sense) or PHY-RX (Receive). In order to detect CTS packets from carrier-sense (CS) signals, the wireless interface keeps track of the times at which all state transitions occur. Transition times TidleTocs, K d , e T o , T ~ s T ~ , and TcsToIdle are used for the detection of CTS signature. Note that a CS to CS transition happens when the wireless interface is in CS state and a new signal is sensed. The sensing of a new signal can be inferred based on an increase in the strength of the resultant carrier-sense signal. When a state transition occurs from carrier-sensing to idle, the wireless interface

equals the CTS duration (TCTS + E ) where E is a small value. If the outcome of the condition check is true then a CTS signature is assumed to be detected. Note that duration-based CTS signature detection, as explained above, requires that the CTS packet size be distinct from all other packets in the network. 4. Performance Analysis

The performance of CSNE-PC has been evaluated and compared with plain 802.1 1 and Basic Power Control (BPC) using Qualnet 3.6 network simulator.

Simulation Setup: The network topology used for our experiments consisted of 25 nodes randomly placed in an area of 1000x1000 meters. The traffic model is similar to that in [I], where each node generates a 5 12 byte packet destined to

checks if (TCSToIdle - TidleToCS) or (TCSToIdle - TidleToRX) or (TCSToldle - TCST~CS)

178

its nearest neighbor according to a Poisson process. A maximum transmit power of 15db has been used. Receive threshold and carrier-sensing threshold are set to -84db and -94db respectively. The rate at which data is transmitted is 2Mbps. The minimum required transmission power is computed using the procedure described in [ 1,3]. Results: Figure 3 shows the transmission and total communication energy consumption per packet as a function of varying network load. Transmission energy indicates the average energy consumed for delivering a packet successfully. Communication-related energy includes not only transmission energy, but also energy spent on receiving, overhearing, and carrier-sensing. Performance data shown in Figure 3 are obtained by gradually increasing the load till the MAC drop rate becomes unacceptable (exceeds 5%). This cut-off point was found to be 65 Kbps per flow.

At low loads, BPC results in modest energy savings that is expected as a result of TPC. However, as the load increases, the energy savings of BPC are overtaken by additional energy expended due to collisions caused by nodes in the CS zone (see Section 1). As a result, for moderate and high load situations, BPC ends up consuming more communication related energy than plain 802.1 1. Note that the energy consumed per packet drops as load increases for 802.1 1.

Experimental data indicate that energy that is spent on overhearing and carrier- sensing go down as load increases while energy spent on transmission and on receiving packets goes up. Since nodes have more data to transmit as load increases, they tend to

(a) Transmit Energy (b) Total cammication Energy spend relatively more energy on transmitting and

receiving data than carrier-sensing/overhearing remote transmissions. Hence, more packets are delivered, and therefore total energy per packet goes down.

CSNE-PC on the other hand consistently consumes up to 24% less transmit energy and up to 38% less total communication energy and that is irrespective of the loading conditions. This savings can be attributed to the reduction of the collisions from nodes in the CS zone. As described in Section 3, in CSNE-PC, nodes in the CS zone detect CTS packets, infer low-power remote transmissions and extend their NAV for the duration of those transmissions. This results in lower number of collisions and hence retransmission costs.

Figure 3. Energy consumption as a function of load

179

0 50 100 Load (Lbps per Bow)

Figure 4 (a) shows throughput of the three protocols under consideration, and 4 (b) shows the MAC layer drop rate. As the load increases, network throughput for 802.11 increases linearly up to a certain point (about 65kbps per flow in our experiments) and then it starts to saturate. The saturation is a result of increased collisions at MAC layer and packet overflow at the link layer buffer.

0 20 40 60 Load &bps pa Bow)

At higher loads, the ability of BPC to successfully deliver packets erodes as the network spends considerable amount of time on back-offs and packet retransmissions resulting from collisions. Packet drops due to failed retransmissions and overflow of the link layer buffers escalate and lead to lower throughput. The MAC

layer drop rate shown in Figure 4 (b) further corroborates this reasoning.

802.11 at all loading conditions. transmission is same for CSNE-PC and 802.1 1.

Throughput of CSNE-PC, on the other hand, is comparable to that of This is because the floor reserved for a

Packet delivery delays for all three protocols are depicted in Figure 5. Llke the other metrics, delay for BPC is much larger than both 802.11 and CSNE-PC due to higher collisions for the former. Higher collision rates in BPC cause higher delay due to back-offs and retransmissions. In CSNE-PC,

preserve the delay performance of 802.11 and even achieve lower delays at some loads in the vicinity of the point of throughput saturation. For example, at the load of 6Okbps per node, 802.11 has an average delay of 257 ms, whereas CSNE-PC has delay of only 89 ms. 5. Conclusions

Figure 5 . Delivery delay as a function of load limiting the co~lis,on rate helps it to

In this paper we have proposed and evaluated the performance of a new transmission power control protocol, CSNE-PC. This protocol addresses the energy inefficiency of basic power control protocols (BPC) in which RTS and

180

CTS packets are exchanged at full power while data and acknowledgement transmissions are carried out at low power, This is accomplished by reducing data collisions which are responsible for energy inefficiency in BPC. In CSNE- PC, DATA collisions are reduced by analyzing carrier-sense signals detected at a node to infer the presence of remote low-power data transmissions, and by using that information to avoid DATA collisions at the receiver. Simulation experiments demonstrate that such a mechanism can result in up to 38% savings in combined transmit and receive energy under a wide range of loading conditions. Also, this energy saving is achieved while maintaining the delay and throughput performance of regular 802.1 1 MAC protocol. Results also indicate that CSNE-PC can be used in networks with different packet sizes.

References 1. Yo E. Jung and N. Vaidya, “A power control MAC protocol for Ad hoc

networks.” ACM International Conference on Mobile Computing and Networking (MobiCom), September 2002.

2. A. Muqattash and M. Krunz, “A distributed transmission power control protocol for mobile ad hoc networks.” IEEE Transactions on Mobile Computing, 2003.

3. J. Monks, V. Bhargavan, and W.M.Hwu, “A power controlled multiple access protocol for wireless packet networks.” In Proceedings of the IEEE INFOCOM Conference, volume 1, pages 2 19-228,2001.

4. M. B. Pursley, H. B. Russell, and S. Wysocarski, “Energy-efficient transmission and routing protocols for wireless multiple-hop networks and spread spectrum radios.” In Proceedings of EUROCOMM Conference, pages 1-5,2000.

5. D. Qiao, S . Choi, A. Jain, and K. G. Shin, “Adaptive Transmit Power Control in IEEE 802.11a Wireless LANs.” In Proceedings of lEEE VTC 2003-Spring, Jeju, Korea, April 22-25,2003.

6. S.-J.Park, R.Sivakumar. “Load sensitive transmission power control in wireless ad hoc networks.” IEEE Global Communications Conference (GLOBECOM), Taipei, Taiwan, November 2002.

7. S . Agarwal, R.H. Katz, S.V. Krishnamurthy, and S . K. Dao. “Distributed power control in ad-hoc wireless networks.” In IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, volume 2, pages 59-66, Oct. 2001.

8. S . Datta and S. Biswas. “Limiting carrier-sense energy consumption by low-power interface idling in 802.1 1 protocols.” In proceedings of IEEE ICWN, Phoenix, Arizona, April, 2004.

CHAOS SHIFT KEYING AND IEEE 802.11a

G. PLITSIS Aachen University of Technologv,

Germany

Chaos engineering for mobile communications is an emerging field of research. Chaos- based modulation schemes such as Chaos Shift Keying (CSK) and Differential Chaos Shift Keying (DCSK) have been proposed as alternatives to conventional digital communication systems. This paper introduces a way to apply CSK in broadband wireless access systems such as IEEE 802.1 la. As well as this, it provides in brief the advantages and disadvantages of such a system. Results concerning spectrum are provided with the use of a link-level simulator with basic functions.

1. Introduction

IEEE standardized 802.1 la [I] with an aim to provide a combination of high bit rates and mobility. Nomadic users can have with the use of IEEE 802.1 l a wireless access to the core network while in business, at work, or at home. Chaos-based mobile communications is an emerging field of research that also tries to integrate chaos-based modulation schemes in conventional ones. As well as this, investigations are being made to take advantage of chaotic oscillations for spread spectrum communications. The idea of using chaotic signals for digital communications has been triggered and inspired by L.M. Pecora (Pecora and Carroll 1990) that observed that two identical chaotic electronic circuits that start with different initial conditions can synchronize.

Different modulation schemes based on chaos have been proposed during the last years. Some of them such as the Differential Chaos Shift Keying (DCSK) do not use the property of synchronization and transmit reference as well as modulated signals. Others such as Chaos Shift Keying rely on the property of synchronization in order to transmit information but are susceptible to channel noise.

2. IEEE wireless LAN overview

Wireless Local Area Networks (WLANs) is an existent technology in many sectors of our everyday life. While in the office, the production line, in hospital, warehouse etc., WLAN technology provides mobile users with high-bit rate data. The IEEE 802.1 1 family embraces many different protocols such as the

181

182

Filter Chaos

802.11a [l]. It is optimized for a coverage area of about lOOm with a typical transmitter power of 100mW. It was first designed for quasi-static network deployment and is basically centralized.

It typically uses Orthogonal Frequency Division Multiplexing (OFDM), Direct Sequence Spread Spectrum (DS-SS), Carrier Sense Multiple Access Collision Avoidance ( C S W C A ) Medium Access Control (MAC) layer scheme, wireless Ethernet, and star topology with an Access Point (AP). By CSMNCA, collisions can be avoided by sending a short “ready to send” message that tells the other nodes the destination and duration of the message that is about to be sent.

HiperLAN/2 is the ETSI proposal for broadband wireless access systems, which has some advantages over the 802.1 la such as Quality of Service (QoS) guarantee, greater spectrum efficiency as well as lower interference but came late in the market and thus did not have the expected success.

Chaotic modulation schemes aim to spread the transmitted signals in order to reduce the power spectral density and as result to minimize interference with other radio communications operating in the same frequency as well as overcome the problem of multipath propagation.

. f

3. Chaotic modulation schemes

A Chaos Shift Keying (CSK) communications system typically contains chaos generators for the data sequences and some filters and equalizers for the modulation and demodulation process. The complexity of the hardware [6] needed for the implementation of such a device is much less than that required by conventional systems. Nevertheless, no synchronization methods have been yet found.

Demodulator Filter 4

Figure 1. CSK communications system.

183

06

04

0 2

0

0 2

0 4

0 6

I I I I I I 0 50 100 150 m 250 300

Figure 2. CSK using cubic map.

Multiple-access CSK can be created by either slightly changing the initial conditions (perturbation-based) or by using a different chaotic map for each user (e.g. logistic, quadratic, exponent, bernoulli, sine circle, tent, henon, lozi, baker, etc.) A theoretical example of perturbation is the amplitude-phase perturbation for the pendulum equation. Considering,

(1) X" + x = E ( Y C 0 S Z - Mc' -@ + x ' )

X(&,Z) = (ro + rn, )COS(Z + a() + M,) - a; c o s 3 ( ~ +a , ) + O ( E 2 )

It can be shown that the solution is equal to

(2)

where p is the spreading factor, N the number of users, and E, = 2pPS . As far as the DCSK is concerned, for the lth symbol period s ( t ) is given

by 1

c(t), (1 - l)Tb I t I (I - X ) T b

(1 - xyb I r I IT, c(t - 'd), s(t) = (3)

184

ld

L 1

t --.-.. 10-'

!$

Id

t I I 1 I 5 10 15

(a) 4

Figure 3. Comparative results of BPSK, FSK, DCSK.

4. The simulation model using chaos generators

The simplified simulation model consists of a Base Station (BS) as well as Mobile Terminals (MTs) that have chaos generators. Data is broadcast to MTs and once the MTs are synchronized, they can communicate with each other and with the BS. For synchronization purposes, VCOs and PLLs are used. The channel is modeled as AWGN [3].

The chaos generator can use different maps for different users such as cubic map, skew tent map, logistic map, and bernoulli-shift map given by,

(4) Xk+, =4x: - 3 x ,

OIx, 10.6

i4 0 . 6 5 ~ ~ 51 Xk+l -

2 X,+, =1-2x,

185

1.2x, +1 x, < o 1.2Xk -1 Xk > o xk+l = (7)

The model is simplified as no special scramblers and interleavers were used. On the other hand, the 802.1 l a link level simulator is created according to the IEEE standard [ 1 J with all functionalities.

I

Figure 4. The simulation model with chaos generators.

5. Spectrum results

The results that appear in the following graphs present the spectrum provided by link level simulations. As far as the advantages and disadvantages of a system based on the link level simulator with chaos generators is concerned, its hardware implementation is easy as well as cheap and it provides high encryption but the synchronization of CSK is still an open question and the Bit Error Ratio (BER) of such a system is inferior to that of the conventional. On

186

10

8

6

4

0 2 .3

I f rn -2

-4

-6

-8

-I05

the other hand, DCSK provides easy synchronization but with a lower data bit- rate when compared to CSK.

-4 -3 -2 -1 0 1 2 3 4 5

Figure 5. Rx power spectrum of 802.1 la PHY (simulation results).

10

0

-1 0

-20

-30

Figure 6. Equalized power spectrum of 802.1 la PHY (simulation results).

187

15

10

0 i 5 m

nl

-10

-5 -4 -3 -2 -1 0 1 2 3 4

Figure 8. Spectrum when skew tent map and AWGN channel is applied (simulation results). Frequency (kHz)

I I -5 -4 -3 -2 -1 0 1 2 3 4 5

Frequency (kHz)

Figure 9. Spectrum when benoulli-shift map and AWGN channel is applied (simulation results),

6. Conclusions

Chaotic modulation schemes have been described and have been proposed ways to be used in contemporary systems. A brief overview of IEEE 802.1 l a has been provided as well as a comparison with chaos-based communication systems. Possible suggestions for W h e r research could be the invention of a synchronization mechanism for CSK and its hardware implementation as well as the circuit design of the various VCOs and PLLs.

188

References

1. IEEE Std 802.11a-1999, “Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications: High-speed Physical Layer in the 5 GHz Band,” ISO/IEC 8802-11:1999/Amd 1:2OOO(E). G. Burel and X. Sammut, “Digital Transmissions with Chaotic Signals: Fast Receiver Synchronization Using Duplicated Chaotic Oscillators,” ISPACS 2000, Honolulu, USA. M. Hasler and T. Schimming, “Chaos communication over noisy channels,” Int. J. Bifurc. Chaos, vol. 10, pp. 719-735,2000. IHP, “Protocol Stack for the IEEE 802.1 la MAC Protocol”, Datasheet, Nov. 2002. A.S. Dmitriev, M. Hasler, A.I. Panas, and K.V. Zakharchenko, “Basic Principles of Direct Chaotic Communications,” Nonlinear Phenomena in Complex Systems, 4: 1 (2002) 1 - 2.

6 . G. Kolumban, J. Schweizer, J . Ennitis, H. Dedieu, B. Vizvari, “Performance evaluation and comparison of chaos communication schemes,” Proc. NDES‘96, pp. 105-1 10, Sevilla, Spain, June 1996.

7. A.S. Dmitriev, B.Ye. Kyarginsky, A.I. Panas, and S.O. Starkov, “Direct Chaotic Communication Schemes in Microwave Band,” Radiotehnika I Elektronika. 46, no. 2,2001 224-233. A. Dmitriev, B. Kyarginsky, A. Panas, and S. Starkov, “Direct Chaotic Communication System Experiments,” Proc. of NDES’OI, Netherlands, 2001 157-160. F. C. M. Lau, M. M. Yip, C. K. Tse, and S. F. Hau, “A multiple access technique for differential chaos shift keying,” IEEE Trans. Circuits Syst. I, vol. 49, no. 1, pp. 96-104, Jan. 2002.

10. M. Hasler, G. Mazzini, M. Ogorzalek, R. Rovatti, and G. Setti (Eds.), “Special issue on applications of nonlinear dynamics to electronics and information engineering,” Proc. IEEE, vol. 90, no. 5, May 2002.

11. M. Sushchik, L. S. Tsimring, and A. R. Volkovskii, “Performance Analysis of Correlation-Based Communication Schemes Utilizing Chaos,” IEEE Transactions on Circuits and Systems-I: Fundamental Theory andApplications, Vol. 47, No. 12, Dec. 2000.

12. N. F. Rulkov, M. M. Sushchik, L. S. Tsimring, and A. R. Volkovskii, “Digital Communication Using Chaotic-Pulse-Position Modulation,” IEEE Transactions on Circuits and Systems-I: Fundamental Theory and Applications, Vol. 48, No. 12, Dec. 200 1.

2.

3.

4.

5.

8.

9.

EFFECT OF TIME-CORRELATED ERRORS ON POWER SAVING MECHANISMS FOR IEEE 802.11 INFRASTRUCTURE

NETWORKS

G. A. SAFDAR & W. G. SCANLON School of Electrical and Electronic Engineering, The Queens University of Belfast,

Ashby Building, Stranmillis Road, Belfast, Northern Ireland, BT9 5AH, UK

The performance of energy-saving protocols for IEEE 802.1 1 wireless local area networks was analysed under time-correlated error conditions. Using OPNET, simulations were performed to compare the performance of the infrastructure power saving mode of IEEE 802.1 1 (PCF-PS) with our proprietary protocol, pointer controlled slot allocation and resynchronisation protocol (PCSAR). The results demonstrate a significant improvement in energy efficiency without significant reduction in performance when using PCSAR. For a wireless network consisting of an access point and 8 power saving stations, energy consumption was up to 28 % lower with PCSAR compared to PCF-PS. The results also show that PCSAR offers significantly reduced uplink access delay over PCF-PS, while modestly improving uplink throughput.

1. Introduction

Energy consumption is a major performance metric for wireless communication networks that support portable devices. Mobile nodes may be in doze, transmit or receive states, and this determines their communication-related energy consumption [ 11. Likewise, system architecture and philosophy also have a role to play in energy efficiency [2]. Considerable research work has been published concerning energy conservation techniques at various layers of the protocol stack, including work on MAC [3], routing [4] and transport control protocols [5]. Regardless of the target standard (e.g., IEEE 802.1 l), a number of general principles have been established. For example, collisions should be eliminated as much as possible within the MAC layer [6] since they result in frame retransmissions and unnecessary power consumption. Likewise, transmission power management is known to improve energy saving in wireless networks and also helps to avoid collisions [7]. The removal of additional overheads [8] also results in improved energy saving by reducing the data transmission time, while for centralised networks higher energy saving may be achieved by avoiding address listening using reservation-based channel access techniques ~91.

189

190

In this paper, the performance of a centralised access protocol, Pointer Controlled Slot Allocation and Resynchronisation (PCSAR), is compared under time correlated channel error conditions with the standard point coordination function (PCF) power saving mode (PCF-PS) of IEEE 802.11. PCF is a centralised MAC protocol for infrastructure networks, while PCSAR is a combination of the pointer-based channel access mechanism of [9] and PCF-PS. PCSAR was developed by the authors and has been described in detail elsewhere [lo]. PCSAR achieves a further reduction in energy consumption over PCF-PS by increasing the opportunities for mobile stations to power down their transceiver components. This paper is organised as follows; the main features of PCF-PS and PCSAR are summarised in Section 2. Section 3 describes the simulation setup with the results presented and discussed in section 4. Section 5 concludes the paper and gives suggestions for hture work.

2. PCF-PS and PCSAR Protocols

For IEEE 802.1 1 wireless networks [ 1 13, infrastructure modes such as PCF-PS provide the greatest opportunity for energy saving. In PCF-PS downlink (DL) traffic destined for power saving (PS) mobile stations is buffered at the access point (AP) allowing stations to save energy by remaining in sleep mode. Frame transfer between the Ap and PS stations under channel error conditions is shown in Figure l(a).

PCSAR is a contention-free protocol where each station is assigned at least one DL transmission opportunity within the contention-free period (CFP), normally followed immediately by a corresponding UL opportunity. Each PCSAR beacon and every DL frame includes a next slot pointer (NSP) and resynchronisation slot pointer (RSP) used for slot allocation and association with the picocell respectively, Figure l(b) shows frame by frame scheduling in PCSAR by making use of NSP. In PCSAR, any incoming station wishing to join the wireless picocell extracts the RSP from any DL frame and contends during the CP period for association. Although the implementation considered in this paper is symmetric (with one UL frame following each DL frame), PCSAR can support adaptive asymmetrical operation in the DL direction. More details about these pointers and the architecture of the protocol and can be found in [lo].

191

(a) LTN = Channel Listening, SLP = Sleep, err = Error Packet

BEACON

using NSP

---------------

I 4 L---+ siotsby NSP uslnp NSP

Figure 1: Timing Diagram for 4 PS stations with channel errors (a) PCF-PS, (b) PCSAR

192

SIFS Time

3. Simulation Details

Using OPNET*, PCSAR was compared with PCF-PS for an IEEE 802.11b network of 8 PS stations and an AP with symmetrical DL and UL traffic conditions. A time-correlated channel error model [ 121 was used to simulate non-ideal transmission conditions. The success or failure of each frame was determined by using the standard OPNET error allocation algorithm.

~~~

10 ps

Simulation models were developed for both PCF-PS and PCSAR and the key parameters are given in Table I .

PIFS

Table 1: Main simulation parameters.

Slot Time + SIFS Time

I Parameter I Value I

Layer Three Mean Packet Size Layer Two Payload MAC Frame Time

CP Beacon Time

I Slot Time I 20 us I

11000 bits 7008 bits

36 * Slot Time MAC Frame Time + SIFS Time

3 * Slot Time CF-END Time PS-POLL Time

Centre Frequency Bandwidth

I Data Rate I 11 Mbos I

3 * Slot Time 2 * Slot Time

2.45 GHz 22 MHz

Max Doppler Frequency

Transmit Power Station Velocity

30 Hz (0.5 - 1 .O) m/s

-10 dBm

In this work, the AP carried all DL traffic and each station had an independent UL traffic source. Network-layer data was simulated by using bursty traffic sources with an exponential inter-arrival rate and a uniformly distributed packet size. To investigate the performance of the network under increasingly poor channel conditions, the number of stations experiencing channel errors was increased from 1 to 8, with all other stations operating error-free. The traffic load on each flow (8 UL and 8 DL) was maintained at 60 % for all of the simulations, the load being specified relative to the useful data capacity of the

* Opnet Technologies, Inc., Bethesda, MD.

193

network evenly allocated across the flows. The useful data capacity wasdetermined by considering the possible network layer data transfer between thetransmission of two consecutive beacon frames. Segmentation and reassemblywas used in both UL and DL due to the bursty nature of the traffic generated bythe traffic sources. In all of the simulations in this paper, the length of the CFPwas variable for PCF-PS as it depended on the status of the TIM (i.e., DLtraffic) at the time of the beacon transmission, where as CFP was fixed at 128slots (187.77 ms) for PCSAR.

4. Results and Discussion

4.1. DL and UL Throughput

DL throughput, which includes frame retransmissions, was found to be almostconstant at 485 kbps for both PCF-PS and PCSAR. Figure 2 shows how thenumber of stations experiencing channel errors affected the average number ofretransmissions, on time, off time in PCF-PS and poll frames in PCSARrespectively. UL throughput for PCSAR and PCF-PS also remained fairlyconstant (at 472 Kbps for PCF-PS and 500 Kbps for PCSAR, an improvementof 6.2 %). This was due to the bandwidth that was taken by NULL frames beingreleased and reallocated to UL retransmissions for both PCSAR and PCF-PScases (Figure 3).

1200

01800

400

4 5

Error Stations

— PCF-PS OnJTime PCF-PS Off_Time— PCSAR_Poll_Frames PCSAR_DL_Re - Tx Frames- PCF-PS DL Re - Tx Frames

Figure 2: Average DL retransmissions, on / off Time and Poll frames for PCF-PS & PCSAR

194

600 E E

400

200

0

1 2 3 4 5 6 I 8

Error Stations

- - PCSAR-NULL-Frames - PCF-PS-NULL-Frames - - PCSAR-UL-Re - Tx-Frames - PCF-PS-UL-Re - Tx-Frames

Figure 3: Average UL retransmissions and NULL frames.

4.2. Access Delay

Figure 4 shows the mean access delay calculated across all 8 stations under channel error conditions. DL access delay with PCSAR was significantly higher (60 % averaged for all stations) than PCF-PS due to the fixed CFP duration of PCSAR. However, PCF-PS DL access delay increased at a much higher rate than PCSAR with the increased number of error stations in the picocell. Furthermore, even under channel error conditions, PCSAR mean UL access delay results were on average 91 % lower than with PCF-PS.

10 . .... -.

1

0.01 J - - _ _ - - _ I 1 2 3 4 5 6 7 8

Error Stations

L-PCSAR-DL - -PCSAR-UL - - 'PCF-PS-DL -PCF-PS-UL 1

Figure 4: DL and UL Access Delay

195

4.3. Energy Consumption and Transmission Efficiency

The energy consumption was calculated for both protocols (Figure 5), assuming 290 mA for transmitting, 205 mA for receiving and 62 mA for sleep or doze mode (Prism 2.5, 3.3 volts IEEE 802.1 lb network interface card). Regardless of the number of error stations in the picocell, average PCSAR station energy consumption was significantly lower than for PCF-PS. Averaged over all scenarios, PCSAR requires 26 % less energy than PCF-PS. With an increased number of error stations in the picocell, the difference in energy consumption for PCF-PS and PCSAR also increases, e g , averaged over all the 8 stations having channel errors, PCF-PS required 28 % more energy than PCSAR.

I 2 100 w

0 1 I

1 2 3 4 5 6 7 8

Error Stations

I - PCSAR - -PCF-PSI

Figure 5: PCSAR and PCF-PS station energy consumption.

5. Conclusion

PCSAR makes use of pointers to increase the degree of control exercised by the AP and reduces energy consumption by removing the need for power saving stations to remain awake and listen to the channel until their transaction with the AP is completed. The results presented demonstrate that PCSAR has consistently improved performance over PCF-PS in terms of energy consumption and transmission efficiency. Uplink throughput and delay performance was much better with PCSAR, regardless of traffic or channel conditions. However, downlink performance of PCSAR was relatively poor, particularly under light traffic and reduced number of stations with errors. PCSAR has the potential to be an extremely effective low power protocol for

196

wireless data networks. However, it remains that further work should address the downlink performance issues, asymmetric operation and improved quality of service performance through intelligent frame scheduling

Acknowledgement

G. A. Safdar would like to acknowledge and thank the Association of Commonwealth Universities for providing funding for this work.

References

1. Christine EJ, Sivalingam K, Agrawal P, Chen J. A survey of energy efficient protocols for wireless networks. Wireless Networks 200 1; 7: 343- 358.

2. Hadjiyiannis G , Chandrakasan A, Devdas S. A low power low bandwidth protocol for remote wireless Terminals. Wireless Networks 1998; 4: 3-15.

3. Chockalingarn A, Zorzi M. Energy Efficiency of Media Access protocols for mobile data networks. IEEE Trans. Communications 1998; 46: 1418- 1421.

4. Xu Y, Heidemann J, Estrin D. Geography informed energy conservation for ad hoc routing. Proc. ACM Intl. Conf: mobile computing and networking 2001: 70-84.

5 . Karvets R, Krishnan P. Application driven power management for mobile communication. Wireless Networks 2000; 6: 263-277.

6. Sivalingam K, Chen J, Agrawal P, Srivastava M. Design and Analysis of low-power access protocols for wireless and mobile ATM networks. Wireless Networks 2000; 6: 73-87.

7. Ebert JP, Stremmel B, Wiederholt E, Wolisz A. An Energy efficient power control approach for wireless LANs. Jnl. of Communication and Networks

8. Jung ES, Vaidya N. A Power Saving MAC protocol for Wireless Networks. Technical Report, Texas A & M University, July 2002.

9. Chui TY, Scanlon WG. A Novel MAC protocol for power efficient short range wireless Networking. IEEE Intl. Con$ on Wireless LANs and Home Networks 2001: 187-196.

10. Safdar GA, Scanlon WG. Pointer Controlled power saving medium access control protocol for IEEE 802.1 1 Infrastructure Networks. Proc. IEEE PIMRC 2004: Barcelona.

11. Gast M, Gast MS. 802.11 Wireless Networks: The Definitive Guide. O’Reilly Networking: 2002.

12. Punnoose R, Nikitin P, Stancil D. Efficient Simulation of Ricean Fading within the Packet Simulator. IEEE VTC. 2000: 764-767.

2000; 2: 197-206.

COMPASS: DECENTRALIZED MANAGEMENT AND ACCESS CONTROL FOR WLANS

ARTUR HECKER heckerawave-storm.com,

Wavestorm SARL 37-39 rue Dareau. 75014 Paris, France

ERIK-OLIVER BLASS erik@erik-blass. de

Karlsruhe, Germany

HOUDA LABIOD labiod@enstjir,

GET-Tklkcom Paris, LTCI-UMR 5141 CNRS, ENST 46 rue Barrault, 7501 3 Paris, France

In this paper, we propose COMPASS, a new decentralized access control architecture for modem WLANs. As traditional centralized access control systems like AAA do not scale well, we propose the use of P2P technologies for the distribution of management data directly between the deployed WLAN access points. Our system COMPASS does not require any additional equipment or central entities. Using auto-organization and fault recovery mechanisms of modem P2P systems, it is robust and easier to maintain. Standard 802.IX mechanisms on the user link guarantee compatibility to the existing user equipment.

1. Introduction

Wireless local area networks as defined by the IEEE 802.1 1 standard [ l ] experience a tremendous popularity. However, some deployment hesitation has been observed in the industry which is believed to be due to the security and management issues with such installations.

Modern 802.1 1 security is based on the IEEE 802.1X standard [2]. 802.1X defines an extensible authentication framework using IETF’s Authentication, Authorization and Accounting (AAA) protocols [3][4]. Thus, 802.1X is typically used with a central AAA server. 802.1X is the chosen base technology for the expected IEEE 802.1 l i standard [ 5 ] .

197

198

The scalability, robustness, and cost of centralized solutions like AAA become an issue in the broad installations of WLANs ranging from small private networks to big, multi-site WLANs. Indeed, the AAA approach results in a centralized server farm which is either over or under-dimensioned for the most network sizes. In that sense, it does not support an natural evolutionary growth of the installed network. Also, from the robustness perspective, it introduces a single point of failure (SPF). These two points can be partly resolved by introducing redundant servers. This however results in a more complicated, costly and difficult to manage AAA infrastructure. Besides, given the low access point (AP) prices in the 802.11 segment, the cost of the central server and its maintenance can hardly be amortized in a typical WLAN (up to 30 APs).

However, 802.1X does not oblige the use of AAA. In this paper, we propose to replace the central AAA server through a distributed P2P based access control architecture in the wired network built directly by the access points. We thus present our Configuration Management P2P-based Access Security System (COMPASS). It integrates P2P technology with the 802.1X access control and does not require any central entity*. It relies on modem P2P technologies which provide highly scalable, efficient and fault-tolerant distributed data retrieval mechanisms [6] [7] [8].

The rest of the paper is organized as follows. In the next section, we present our distributed access control architecture and discuss several important points like self-organization and user management. Then, we give a qualitative comparison and an outlook to the future work.

2. COMPASS: Our New Proposed Architecture

2.1. Main Idea

To reduce costs; we propose to use the access points directly for the storage of data relevant to access control and network management. However, since the resources of access points are limited, the idea is to distribute the administrative load over all access points. Thus, in our proposal every AP holds only a part of the whole management database.

The difficult part is to provide a scalable and robust mechanism for distributed data retrieval. Herein, the problem is not the data transfer but locating the data [ 6 ] . Distributed Hash Tables (DHTs) [6][7][8] have been designed to overcome these difficulties.

While a radius connects a point on the circumference to the center, a compass directly interconnects the circle points.

199

2d

Secure Service Directory

#Elements in routing table

Used in

A DHT is a hash table divided into multiple parts (called zones) and distributed over several nodes. Famous DHT examples are file-sharinglP2P networks such as EDonkey2000 or Kazaa. We have to choose a P2P system tailored to the restricted resources of an ordinary AP. We namely compare CAN [6 ] , Chord [7] and Pastry [8]. Without going into structural or design details of these networks, here we present a short overview of properties that seem crucial for our work. Table 1 shows properties of common DHTs for a network with n nodes. ‘‘#Hops for lookup/store” is the expected number of nodes a request for a lookup or store has to pass. The second criterion is the number of elements each node has to store in its neighbor or routing table. Both properties are expected values for a well-balanced DHT network.

In n B*ln n

CFS-file-system Oceanstore, Scribe

O(n”d) O(ln n ) #Hops for lookup/store

Chord and Pastry seem to perform better regarding the average path length. This means less communication overhead and shorter latencies. On the other hand, CAN’s advantage is a constant O(1) memory consumption which is known at node setup time, prior to network use. For that reason, we choose CAN.

2.2. Basic System Architecture

Based on standard TCPIIP networking in the core network, our P2P management network is formed of the deployed 802.1 1 access points. This is illustrated in Figure 1. Every access point acts as a P2P node building a logical overlay network over the physical core network. This overlay stores different logical databases, primarily user and management databases (DB). The user DB stores AAA-like user profiles. The management DB helps the administrator manage all connected APs and stores AP settings expressed in the respective syntax (e.g. 802.1 1 MIB variables, proprietary manufacturer settings, etc).

On user demand, the solicited node retrieves the correspondent profile by using overlay’s lookup method. Using the retrieved profile the serving AP follows the usual 802.1X procedure acting as Authenticator with a local Authentication Server [2].

Table 1:

200

2.3.

Access Point Access Point

Figure 1 Main entities in COMPASS

Preliminary AP Configuration

Each AP needs a minimum configuration prior to its deployment in the network. This is necessary for a secure management access to this AP, the overlay discovery and the classical 802.1 1 settings. The trust relationship with the AP is expressed by the installation of a signed certificate on every AP. In addition to the usual 802.1 1 parameters, the administrator supplies the bootstrap-address of the overlay network and deploys the AP in the desired location.

2.4. Bootstrapping (AP join)

The original CAN proposal [6] makes use of a bootstrap method based on a well-known DNS address. In CAN this method guarantees a uniform partitioning of the index space. However, it also means that a physical neighborhood does not result in CAN neighborhood.

We want to be able to tie the overlay to the network topology. This can potentially shorten the handoff-delay since the new AP can rapidly retrieve the profile data from the old AP using the overlay (since the old AP is also the overlay neighbor). The other reason for reflecting the physical topology in the overlay is a transparent load balancing: we suppose that the administrator installs an additional AP in the neighborhood of every AP which suffers a heavy traffic. If the APs are CAN neighbors, they also share the administrative load. Otherwise

201

they only share the 802.11 traffic load. CAN’s landmark-ordering method to reflect the physical network topology to the overlay explicitly targets the IP layer topology. We define a 802.1 1 -adapted landmark-ordering mechanism:

1. Booting up, a preconfigured AP searches the 802.1 1 environment for 802.1 1 neighbor APs. The necessary mechanisms are defined in the 802.1 1 standard and include an active and passive discovery of neighboring APs of the same SSID [l]. The joining AP compiles a list of (wireless) MAC addresses of all neighboring APs configured with the same SSID.

2. The joining AP now sends a discovery request to the predefined DNS address of the overlay. The request contains the MAC address list (step 1). If it is not empty, the solicited bootstrapping node chooses from it the node whose zone is the biggest. It is essential to provide a mechanism resolving a wireless link MAC address into the management IP address without global network knowledge. We achieve this by storing these data into the overlay itself, but other mechanisms could be applicable as well. Thus, the solicited node executes a lookup in the overlay for every listed MAC address. The returned value is a pair ( IP-address, zonesize). The solicited node chooses the pair with the biggest zone size. If the received wireless MAC address list is empty, the bootstrapping node proceeds as in CAN randomly choosing a node. In either case it replies with the IP address of the chosen node.

3. The join procedure itself is like in CAN. After the joining procedure, the new AP executes a store command in the overlay, posting its own wireless MAC address, the management IP address and the zone size. Each AP is responsible for the validity of that entry.

Following this scheme, the new installed AP automatically becomes an overlay neighbor of one of its 802.11 neighbors. The advantages of this scheme are an equal pre-configuration of all APs and the requested binding of the overlay to the physical topology. Our method does not affect the scalability since the preconfigured overlay address can correspond to several APs. This can be achieved with a round robin DNS or with a multicast address (e.g. “all overlay APs”). Moreover, the join events are expected to be much rarer than the operational procedures like user access. During an initial deployment (or after a complete system failure), no node is available under the bootstrap address. This case is rarer than join events and needs a special treatment. It can be resolved by weakening the equality of the AP configuration (choose some nodes for join requests) or by dynamically updating the round robin DNS on new node joins and departures. Such mechanisms are however out of scope of this document.

202

2.5. APLeave

AP leave events can occur because of an AP failure (e.g. power down) or because of a admin shutdown. If an AP is shut down correctly, it proceeds like a CAN node. In a case of a power down, the zone databases held by this AP are lost. CAN redundancy mechanisms are used in that case.

2.6. User A d d User Delete

To add a new user record to the system, the network administrator executes the commands store (username, profile) or delete (username) on one of the nodes. Herein, the profile is a list of authorizations. Principally, such profile could be in an arbitrary suitable format (e.g. attribute value pairs). The profile defines the authentication method, the restrictions and session parameters A typical profile hardly ever exceeds 1 Om.

2.7. User Network Access

When a user accesses a COMPASS network, user’s mobile station (MS) and the solicited AP start the typical 802.1 X authentication process. Within this process, at some point of time MS sends an EAP Response/Identity message containing the identity string. The solicited AP retrieves the corresponding user profile from the overlay by invoking an overlay lookup for this string as a key. On the receipt of the profile, the AP can continue the EAP conversation as defined in the profile acting as a 802.1X authenticator with a local 802.1X authentication server (AS). The identity used by the AP is an abstract identity of the whole overlay which acts as one logical entity.

2.8. Failure Management and Optimizations

We distinguish two major possible failures: the path failure (i.e. some of the nodes in the lookuplstore path fail) and the end node failure (i.e. the zone database is not available). CAN provides mechanisms to counter the impacts of such failures [ 6 ] . In our particular application the path failure can be countered by increasing the number of CAN’s dimensions d. That increases the number of tried paths. If the overlay stores only one DB copy, the failures of nodes holding a zone database result in associated data not being available. CAN provides data replication methods. By using multiple realities [6] or multiple different hash functions, the same zone database can exist on multiple nodes. We encourage the use of such mechanisms for user DB.

203

We demonstrate this at an example of CAN with k realities [6]. Let f of n nodes fail. Assuming that at least one overlay path still works, the probability that a stored pair can be retrieved is:

Example: in a CAN with 4 realities, if % of all nodes fail without path failures, there is still 99.6% probability that an authorized user can use the service.

3. Discussion

CAN technology has been recently used in sensor networks [9]. Compared to sensors, the WLAN APs are powerful machines with about 16-32MB RAM and a CPU of about 1 5OMHz. Recent APs with an embedded Linux 0s show that the available resource safety margin is sufficiently large for additional tasks.

with onc DB

As mentioned, CAN has constant memory requirements. The management database stores settings valid for every AP. Its size is thus independent of the number of APs. Given a typical profile size of about 1-lOkB, 30-100 user profiles per AP can principally be stored without any impact on AP performance. User profile size can be further reduced by using group management. Redundancy mechanisms can not be reasonably used when the overall number of APs is very low (say for up to 5-10 APs). When using redundancy mechanisms, the overall database size has to be multiplied by the redundancy factor k. For instance, in a network with 10 APs, 100 user profiles and the redundancy factor

tABLE 2 cOMPARATIVE CHART OF USER MANAGEMENT METHODS IN MODERN wlanS

204

k 2 , the zone database on each AP is about 5kB*100*2 f 10 = 1MB. These values seem to be realistic requirements for modem APs.

In Table 2, we compare different access control architectures in terms of user mobility, administration effort, network extensibility and robustness (expressed by the worst case impact of a partial system failure).

4. Conclusion

In this paper, we propose to integrate P2P technology with access control to provide a system supporting a natural scaling of the management infrastructure. As a P2P architecture, COMPASS inherits the scalability and fault tolerance of the existing P2P technologies. COMPASS is backwards compatible by using standard access methods on the user link and AAA-like user management in the core. It features an easy network extensibility by defining automatic node join. By storing the necessary management settings in the overlay, COMPASS can also be used as network management infrastructure.

References

[I] IEEE Standard 802.1 1, “Wireless LAN medium access control (MAC) and

[2] IEEE Standard 802.1X, “Port-based network access control,” June 2001. [ 3 ] C. Rigney, S. Willens, A. Rubens, W. Simpson, “Remote Authentication

Dial-In User Service (RADIUS),” RFC 2865, IETF, June 2000. [4] P. Calhoun, J. Loughney, E. Guttman, G. Zon, J. Arkko, “Diameter Base

Protocol”, RFC 3588, IETF, September 2003. [5] IEEE Draft 802.1 li, “Draft supplement to IEEE Std 802.1 1. part 11:

specifications for enhanced security”, work in progress. [6] S. Ratnasamy, P. Francis, M. Handley, R. Karp, S. Shenker, “A Scalable

Content-Addressable Network , Proceedings of the 200 1 Conference on Applications, Technologies, Architectures, and Protocols for Computer Communications.

[7] I. Stoica et al., “Chord: A Scalable Peer-to-peer Lookup Service for Internet Applications”, Proceedings of the 2001 Conference on Applications, Technologies, Architectures, and Protocols for Computer Communications.

[ S ] A. Rowstron, P. Druschel, “Pastry: Scalable, distributed object location and routing for large-scale peer-to-peer systems”, Lecture Notes in Computer Science, 200 1.

[9] H.-J. Hof, E.-0. Bla13, T. Fuhrmann and M. Zitterbart, “Design of a Secure Distributed Service Directory for Wireless Sensornetworks”, EWSN 2004, Berlin.

physical layer (PHY) specifications,” 1999.

QOS PROVISIONING MECHANISMS FOR IEEE 802.11 WLANS: A PERFORMANCE EVALUATION*

JOSE VILLALON Instituto de Investigacidn en Informbtica, Universidad de Castilla La Mancha, Albacete,

02071, Spain

PEDRO CUENCA Instituto de Investigacibn en Informbtica, Universidad de Castilla La Mancha. Albacete,

02071, Spain

LUIS OROZCO-BARBOSA Instituto de Investigacidn en Informcitica, Universidad de Castilla La Mancha, Albacete,

02071, Spain

The IEEE 802.1 1 wireless LAN (WLAN) is the most widely used standard nowadays for wireless LAN technology. However, the current standard does not provide QoS support for multimedia communications. Thus, a large number of enhancements to the standard are being proposed. This paper introduces some of the most relevant schemes: the Priority-Based Distributed Mechanisms. We also present the upcoming IEEE 802.1 le (EDCA) standard being introduced. The IEEE 802.11e standard is a proposal defining the mechanisms for wireless LANs aiming to provide QoS support to multimedia communications. In this paper, we carry out a comparative study of Priority-Based Distributed Mechanisms and the 802.1 l e (EDCA) upcoming standard, when supporting different services, such as voice, video, best-effort and background traffic.

1. Introduction

WLANs are gaining popularity at an unprecedented rate, at home, at work, and in public hot spot locations mainly due to their low cost, their ease of deployment and, above all, by allowing the end users to freely move around withm the area they cover. Another influential factor is the appearance in 1997 of the standard IEEE 802.11, with its subsequent revision in 1999 [l], and its subsequent amendments that nowadays enable transmission speeds of up to 54 Mbps, allowing the use of multimedia applications.

~ ~~

This work was supported by the Ministry of Science and Technology of Spain under CICYT project TIC2003-08154-CO6-02 and the Council of Science and Technology of Castilla-La Mancha under project PBC-03-001.

205

206

The basic access function in IEEE 802.11 is the Distributed Coordination Function (DCF) where before transmitting, a station, the source station, must determine the state of the channel. If during an interval of time, called Distributed InterFrame Space (DIFS), the channel is sensed free, the station can initiate its transmission. If the channel is sensed busy, once the transmission in progress is finished and to avoid the collision with other stations in the same situation, a backoff algorithm is initiated. This algorithm consists in choosing an interval of time (the backoff time) at random during which the station delays the transmission of its frames. Once having transmitted the source station, it will wait to get back a reply from the destination station. If after waiting for a time interval, denominated Short InterFrame Space (SIFS<DIFS), the source station does not get the Immediate Positive Acknowledgement (ACK) from the destination station, it simply assumes that there has been a collision. The source station can then attempt to retransmit a finite number of times using a longer backoff time after each attempt. This access function is easy to implement and suitable for most applications. However, it does not provide Quality of Service (QoS) support. Guaranteeing those QoS requirements in IEEE 802.11 is a very challenging task due to the QoS-unaware functions of its MAC layer. Thus providing QoS to IEEE 802.1 1 has become an area of active research giving rise to several new service differentiation schemes. In this paper, we carry out a comparative study of Priority-Based Distributed Mechanisms and the 802.1 l e (EDCA) upcoming standard. Our main aim has been to compare schemes based on the same principle of operation than the upcoming IEEE 802.1 1 e standard.

This paper is organized as follows. Section 2 provides an overview of the Priority-Based Distributed Mechanisms. In Section 3, we also describe the upcoming IEEE 802.1 l e QoS enhancement standard. In Section 4, we carry out a comparative performance evaluation of Priority-Based Distributed Mechanisms and the 802.1 l e (EDCA) upcoming standard, when supporting different services, such as voice, video, best-effort and background traffic. Finally, Section 5 concludes the paper.

2. Priority-Based Distributed Mechanisms

2.1. Deng et al. Mechanism

The proposed mechanism [2] uses two properties of the IEEE 802.11 standard for providing various priority services: the Interframe Space (IFS) used between data frames and the backoff mechanism. By making use of two different time length for IFS, the stations implementing the shortest IFS will be assigned a

207

higher priority than a station with a longer IFS. By using these two different interframe spaces (PIFS and DIFS), traffic can be differentiated and classified into two classes. To further extend the number of available classes, the backoff mechanism could be used to differentiate between stations. This can be simply achieved by defining several time lengths for the backoff interval used in the backoff algorithm depending on the priority level assigned to the station.

Table 1. Deng Priority Classes

Priority IFS Backoff Algorithm 3 PIFS LRand() x 2*+'/2J 2 PIFS 2'+'/2 + LRand() x 22"/21 1 DIFS LRand() x 22f'/2J 0 DIFS 2'+'/2 + LRand() x 2*"/21

Table 1 shows four different priority classes (0-3) defined using two different interframe spaces, and two different backoff algorithms [2]. The backoff algorithm guarantees that the low priority stations always generate longer backoff intervals than the stations with higher priority.

2.2. Aad et al. Mechanisms

Aad et al. present in [3] three differentiation mechanisms.

2.2.1. Different backoff increase (referred from now on as Aad-Backoffscheme)

Each priority level has a different backoff increment function. A shorter contention window is assigned to the higher priority stations ensuring that in most (although not all) cases, high-priority stations are more likely to access the channel than the low-priority ones. This method modifies the contention window of the priority level j after i transmission attempts as follows: cw,,,, = P;'' X CW,, , where Pi is a factor used to achieve service differentiation with the highest value for lowest priority stations. In each retry, a higher increment factor implies a higher waiting time and therefore a smaller throughput.

2.2.2. Different CWmin (referredfrom now on as Aad-CWmin scheme)

The main motivation is that, with a small number of stations contending for the access to the channel, most of the time CW remains at its minimum value (CWmin). Therefore, an Aad-backoff scheme will not help to promptly resolve

208

access conflicts as the CWs will rarely increase. This led to the authors to use different values for CWmin. This scheme consists on defining a different CWmin, value per priority level. Assigning the shortest CWmin to the highest priority level ensures that in most (although not all) cases, the highest-priority stations are more likely to first gain access to the channel than the lower-priority stations.

2.2.3. Different DIFS referred from now on as Aad-DIFS scheme)

Each station uses a different DIFS corresponding to its priority level. Under this approach, each priority level uses a different DIFS, for example, DIFSj+I < DIFSj. Before transmitting a packet, the stations having priority j + I will wait for an idle period of length DIFSj+I slot times. To avoid collision between frames within a same priority, the backoff mechanism is maintained in a way that the maximum contention window size added to DIFSj is DIFSj.I-DIFSj. This ensures that no station of priority j has queued frames before station of priority j-I starts transmitting. This may lead to starvation if a station with a high priority always has frames to transmit, since none of the stations with lower priority will ever be able to access the channel.

2.3. TCMA (Tiered Contention Multiple Access) Mechanism

This mechanism by Benveniste [4] defines the Arbitration-time InterFrame Space (AIFS) as the mandatory waiting time before a transmission can be initiated. It also suggests the use in the backoff algorithm of the Backoff-Counter Update Time (BCUT) regulating the decrement rate of the backoff counter (aSZotTzme in DCF). The priority classes are set by making use of several parameters. First, different AIFS and BCUT values are assigned to each class (a smaller AIFS allows the station to promptly access the channel, while a smaller BCUT enables a faster decrement of the backoff counter). Second, as in the Aad- Backoff mechanism, different Persistence Factors (PF) are assigned. In the case of stations with delay sensitive traffic, the backoff time diminishes instead of increasing exponentially like in DCF. Finally, a deadline is set for the time that a frame stays in the MAC layer waiting to be transmitted, and therefore a frame surpassing this time limit is discarded. Since PCF accesses the channel after an idle period of length PIFS, assigning the top priority class an IFS=PIFS and a minimum backoff delay is equal to 1 will avoid conflict with the centralized protocol.

209

j

3. The Upcoming IEEE 802.11e Standard

The IEEE 802.11e draft [ 5 ] is a proposal defining the QoS mechanisms for wireless LANs for to supporting time-sensitive applications such as voice and video communications. IEEE 802.1 l e incorporates two new access mechanisms: the contention-based Enhanced Distributed Channel Access (EDCA), known in the previous drafts as the Enhanced DCF (EDCF) and the HCF Controlled Channel Access (HCCA).

One main feature of IEEE 802.11e is the definition of four Access Categories (AC) queues and eight User Priorities (UP) queues at MAC layer. Another main feature is the concept of Transmission Opportunity (TXOP), which defines the transmission holding time for each station.

EDCA has been designed to be used with the contention-based prioritized QoS support mechanisms. In EDCA, two main methods are introduced to support service differentiation. The first one is to use different IFS values for different ACs. The second method consists in allocating different CW sizes to the different ACs. Each AC forms an EDCA independent entity with its own queue and its own access mechanism based on DCF with its own Arbitration Inter-Frame Space ( AIFS[AC] = SIFS + AIFSr\rlAC] x SlotTime) and its own CW[AC] (CWmin[AC’ I CW[AC] 5 CWmax[AC])(see figure 1). If an internal collision arises among the queues within the same QSTA, the one having higher priority obtains the right to transmit. It is said that the queue that is able to gain access to the channel obtains a transmission opportunity. Each TXOP has a limited duration (TXOPLimit) during which an AC can send all the frames it wants.

scheduler (resolves virtual Mllisbns by granting TXOP lo highest priority

802.11e: up to 8 User Priorities (UPS) per QSTA

n 8 UPS mapping to 4 Access Categories (A&)

i n n n n

, - - - -

Figure 1. Enhanced Distributed Channel Access (EDCA)

210


In this section we undertake a comprehensive evaluation of various QoS provisioning mechanisms of the IEEE 802.1 1 WLAN. As starting point, we evaluate the DCF operating mode. This first part will serve as a basis for having a clear picture of the performance exhibited by this mechanism in the presence of multiple types of traffic flows. In the second part of our study, we focus on assessing the behavior of the upcoming IEEE 802.1 l e (EDCA) standard. Finally, we carry out a comparative study of five different priority-based distributed mechanisms (namely, Deng, Aad-Backoff, Aad-CWmin, Aad-DIFS, and TCMA mechanisms) and the IEEE 802.11e (EDCA). The QoS provided by all these schemes have been evaluated in terms of five different metrics: throughput, collision rate, access delay, delay distribution and packet loss rate. For all our studies, we have made use of the OPNET Modeler tool 10.0 [6].

In our simulations, we model an 11 Mbit/s IEEE 802.11b wireless LAN supporting four types of services: Voice(Vo), Video(Vi), Best-effort(BE) and Background(BK). We assume the use of a wireless LAN consisting of several wireless stations and an access point connected to a wired node that serves as sink for the flows from the wireless domain. Each wireless station transmits a single traffic type (Vo, Vi, BE or BK) to the access point. We assume the use of constant bit-rate voice sources encoded at a rate of 16 kbits/s according to the G.728 standard. For the video applications, we have made use of the traces generated from a variable bit-rate H.264 video encoder[7]. The average video transmission rate is around 480 kbitsls. The best-effort and background traffics have been created using a Pareto distribution traffic model with average sending rate of 128 kbit/s and 256 kbit/s, respectively.

For all the scenarios, we have assumed that one fourth of the stations support one of the four kinds of services: voice, video, BE and BK applications. We start by simulating a WLAN consisting of four wireless stations (each one supporting a different type of traffic). We then gradually increase the Total Offered Load of the wireless LAN by increasing the number of stations by four. We increase the number of stations 4 by 4 starting from 4 and up to 36. In this way, the normalized offered load is increased from 0.12 up to 1.12. We have preferred to evaluate a normalized offered load, rather than the absolute value. The normalized offered load is determined with respect to the theoretical maximum capacity of the 11 Mbit/s IEEE 802.1 l b mode, i.e. 7.1 Mbit/s. In order to be able to make a fair comparison, we have used the parameter setting of the priority-based distributed mechanisms (see table 2) as close as possible to the settings used by Chesson [8].

21 1

Table 2. Parameter Settings

AC IFS CW.,,” CW,., PF Vo 2 x Slot-time + SlFS 31 1023 2

!=. Vi 2 x Slot-time + SlFS 31 1023 2 8 Be 2 x Slot-time + SlFS 31 1023 2

Bk 2 x Slot time + SlFS 31 I023 2 Vo 2 x Slot-time + SlFS 31 1023 2

4 8 Be 4xSlot time+SlFS 31 1023 2 0 v1 Vi 2 x Slot-time + SlFS 31 I023 2

Bk 7 x Slot-time + SlFS 31 1023 2 Vo 2 x Slot-time + SIFS 31 1023 I

Be 2 x Slot-time + SIFS 31 I023 8 Bk 2 x Slot time + SlFS 31 I023 16

n 5 Vi 2 x Slot-time + SlFS 31 1023 2

Vo 2 x Slot-time + SlFS 1 1023 2 n 1 vi 2 x Slot-time+ SIFS 15 1023 2 8 Be 2 x Slot-time + SlFS 31 1023 2

Bk 2 x Slot time + SlFS 31 1023 2

f vi 2 x Slot-time + SIFS 31 1023 2 Be 4 x Slot-timc + SlFS 31 1023 8 Bk 8 x Slot time + SIFS 31 1023 16

5 vi 2 x Slot-time + SIFS 15 31 2 a Be 4 x Slot-time + SlFS 31 1023 2 Bk 7 x Slot time + SlFS 31 1023 2 AC IFS BackoB Algorithm v o Slot-time + SlFS [Rand() x 2*”/2J Vi Slot-time + SlFS 22”/2 +LRand() x 2*”/2J Be 2 x Slot-time + SlFS LRand() x 2*”/21 Bk 2’*’/2 + LRand() x 2’”/21

v o Slot-time + SlFS 31 1023 I

Vo 2 x Slot-time + SlFS 7 15 2

2 x Slot time + SlFS -__

In the first part of our performance evaluation study, we have focused on showing the legacy DCF mode of operation of the IEEE 802.1 1. Figure 2 shows the simulation results for throughput and delay. When the load exceeds 0.5, the throughput of the four types of traffic dramatically decreased. The decreasing rate of each type of traffic is dictated by their respective bit rates. This decrease results from the increasing number of collisions experienced by the four types of traffic. This phenomenon also explains the access delay performance. When the offered load is higher than 0.5, the means delays of the audio and video traffic are above the tolerable limits delays (1Oms and looms for Vo and Vi, respectively). These results clearly shows that the DCF mode of operation does not provide at all any throughput or delay differentiation between the different types of traffic, i.e., all types of traffic access to the medium under the same conditions. In the second part our performance evaluation, we have focused on assessing the behavior of the upcoming IEEE 802.1 l e (EDCA) standard. Figure 2 shows the simulation results for the throughput and delay. EDCA can support service differentiation for different types of traffic according to their priority. With EDCA the throughput of each type of traffic decreases according to its relative priority. EDCA maintains the throughput of the voice and video traffic at

212

the expense of the BE and BK traffics. On the other hand, when the offered load is less than 0.8, the means delays of the voice and video traffics are kept below the delay bounds.

DCF

EDCA EDCA

Figure 2. Performance Evaluation of DCF and EDCA

In the third part our performance evaluation, we have carried out a comparative study of the different priority-based distributed mechanisms and the upcoming IEEE 802.1 l e (EDCA) standard, when supporting the four different types of services under consideration.

Figure 3 shows the normalized throughput obtained for the four types of services when making use of each one of the seven mechanisms being considered. Figure 3 shows that except for the mechanism proposed by Deng, all schemes prove to be effective in maintaining the normalized throughput of the voice service close to 1. As it will be explained later, the deviation from this maximum is mainly due to the increasing number of collisions. Regarding the video traffic, some of the schemes start to exhibit a drop on the video throughput at loads as low as 0.5. However, most of the schemes, namely EDCA, TCMA, Aad-Backoff and Aad-DIFS are able maintain the normalized throughput to 1 for loads of up to 0.8. Beyond this point, only the TCMA is able to maintain the normalized to its maximum value. The decrease on the throughput is mainly due

213

to the repeated number of packet retransmissions resulting hom the increasing number of collisions. This will be further explained when analyzing the results on packet retransmissions. In the case of the BE and BK traffics, these are severely affected as the network load is increased. l lus is on line with the specifications of the various QoS-aware mechanisms by guaranteeing the QoS to the voice and video services.

Voice traffic Video traffic

Total Offered Load Total Offered Load

to the repeated number of packet retransmissions resulting hom the increasing number of collisions. This will be further explained when analyzing the results on packet retransmissions. In the case of the BE and BK traffics, these are severely affected as the network load is increased. l lus is on line with the specifications of the various QoS-aware mechanisms by guaranteeing the QoS to the voice and video services.

Voice traffic Video traffic


Best Effort traffic Background traffic


Figure 3. Average Throughput


Figure 4. Average number of Attempts per packet

Figure 4 shows the mean number of packet retransmissions for voice and video services. In the case of the EDCA mechanism, it is clear that the decrease on the voice normalized throughput as the offered load reaches 0.75 is due to the

214

repeated number of packet retransmissions. The same can be said for the video traffic. The TCMA scheme exhibits much better performance since the retransmission rate is two to three times lower than the one for the EDCA scheme. This also has a direct impact on the mean access delay.

Figure 5 shows the mean access delay per service class. Even though the EDCA scheme shows the lowest mean delay access for the voice service, its dropping rate due to the repeated transmission attempts is higher than the one obtained for other scheme. This is mainly due to the use of a short IFS and very short CWmin and CWmax by limiting the number of backoff slots to 15. For the video service, the results depicted in Fig. 5 provide further insight into the performance of the EDCA scheme. Below an offered load of 0.75, the EDCA scheme proves effective in meeting the QoS of the high-priority classes. However as the load further increases, the performance of EDCA falls dramatically, due to the high number of collisions.

Voice traffic

’$0 0.2 0.4 0.6 0.8 1 Total Offered Load

Best Effort traffic

Video traffic

Total Offered Load

Background traffic

Figure 5 . Average Access Delay

Figure 6 shows the cumulative distribution function of the access delay for all mechanism operating at a load close to 0.75. Figure 6 shows that the EDCA is able to guarantee the timely transmission of all the voice packets, while the TCMA scheme gets very close to it. In turn, for video packets, Fig. 6 shows that the TCMA outperforms all the other schemes under study and that EDCA scheme exhibits the second best result. It is clear from Fig. 6 that under the

215

scenario being considered, all the other schemes are unable to meet the QoS requirements of the real-time (voice and video) applications (1 Oms and 1OOms for Vo and Vi, respectively).

U 0 0

0.4 - A& sackoff - - - AadDlFS - Aad CWmln - - - Den@

O0 Medkkccess del$%eg) 0.3

.. 1

0.8

0.6

0.4

0.2

U

0 a

'0 20 40 60 80 100 Media Access delav (seal

Figure 6 . Cumulative Distribution (CDF) of the Access Delays Voice traffic Video traffic


Figure 7. Packet Loss Rate due to Maximum Number of Attempts

Figure 7 shows the packet loss rate due to maximum number of attempts (seven). The results shows in this figure are similar to the ones depicted in Fig. 4. The mechanism by Deng provides the worst results, a large number of packet losses are due to the parameter settings used in the backoff mechanism. Furthermore, since the backoff interval for the voice and best-effort services will be set to 0-3 and for the video and background services to 4-7 as the number of stations increases, it is much likely that two or more stations attempt to transmit

216

at the same time. The EDCA scheme shows the second highest loss rates under extreme load conditions. Similar to the Deng scheme, this is due to the value being used for CWmax (15 for voice and 3 1 for video).

5. Conclusions

In the first part of this paper, we have overviewed some of the most relevant works in the area of QoS provisioning for IEEE 802.11 wireless LANs. We have paid particular attention to the upcoming IEEE 802.1 l e standard. In the second part of the paper, we have evaluated various QoS provisioning schemes as well as the IEEE 802.1 le. We have considered a scenario consisting of a wireless LANs where the stations support four different types of services: voice, video, best-effort and background. Our results show that by limiting the number of collisions, the network performance and QoS provisioning can be effectively achieved. From our results, we can conclude that the EDCA scheme does not perform well when exposed to heavy loads mainly due to the excessive number of collisions.

References

1. LAN MAN Standards Committee of the IEEE Computer Society, ANSUIEEE Std 802.1 1, “Part 1 1 : Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, 1999 Edition

2. D.J. Deng and R.S. Chang, “A priority scheme for IEEE 802.11 DCF access method”, IEICE Trans. Commun., vol. E82-B, num. 1, pp. 96- 102. 1999

3. I. Aad and C. Castelluccia, “Priorities in WLANs”, Computer networks, vol. 41/4, pp. 505-526, February 2003

4. M. Benveniste, “‘Tiered contention multiple access’ (TCMA), a QoS-based distributed MAC protocol”, The 13th IEEE PIMRC, vol. 2, pp. 598-604, September 2002.

5. IEEE 802 Committee of the IEEE Computer Society, IEEE P802.11e/D8.0 Draft Amendment to IEEE Std 802.11, “Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications: Medium Access Control (MAC) Quality of Service (QoS) Enhancements”, Feb.2004.

6. 0pnet.Technologies.Inc. OPNET Modeler 10.0 01987-2004. 7. ITU-T Recommendation H.264. Advanced Video Coding For Generic

Audiovisual Services. May 2003. 8. G. Chesson, W. Diepstraten, D. Kitchin, H. Teunissen and M. Wentink,

“Baseline D-QoS proposal”. IEEE 802.1 1 working group document 802.11- 00/399 (2000).

IMPACT OF VARYING THE MINIMUM VALUE OF

PROTOCOL ON THE QOS PARAMETERS CONTENTION WINDOW (CW,i,,) OF THE IEEE 802.11 MAC

MOHAMMAD SARAIREH

Faculty of Art, Computing, Engineering and Science, Shefield Hallam University, Howard Street, Shefield, Sl 1 WB, UK [email protected]. ac. uk

REZA SAATCHI , REBECCA STRACHAN , SAMIR AL-KHAYAT'T Faculty of Art, Computing, Engineering and Science, Shefield Hallam University,

Howard Street, Shefield, Sl 1 WB, UK

There is a growing demand for the transmission of multimedia applications over wireless computer networks. The aim of this study was to analyse the effects of minimum Contention Window (CW,,,) on Quality of Service in IEEE 802.1 1 MAC protocol. This study was carried out using ns-2 simulation software. The analysis was based on four Variable Bit Rate (VBR) connections with different packet sizes, distributed randomly in the same Basic Service Set (BSS). It was demonstrated that poorly selected values of CW,,. result in considerable network performance degradations. Moreover, the CW,,, value for achieving optimum QoS proved to be different. Based on this work, an intelligent method for determining the CW,,, size to achieve optimum QoS has been proposed.

1. Introduction

Wireless systems are widely used for transmitting a mix of applications, i.e. voice, video and data. Wireless transmission requires a controller to manage both the way the medium is accessed is fair and controlled. Random transmission may lead to incomprehensible or unpredictable results. Therefore, a controller for accessing and sharing the medium is an essential tool for achieving a successful transmission between the communication devices.

IEEE 802.1 1 is the original IEEE Wireless Local Area Networks (WLAN) protocol standard [ l , 21. The Medium Access Control (MAC) protocol in the wireless network controls access to the shared medium by applying rules and procedures that permit the communication pairs to communicate with each other in an efficient and fair manner.

219

220

The Distributed Coordination Function (DCF) method of medium access control is considered in this study as it allows nodes mobility without the need for control access point. The DCF mechanism behaviour depends on different parameters such as the backoff time, slot time, Inter Frame Space (IFS), retry counts, Contention Window (CW) value and others. The value of CW has a major effect on the performance of the protocol. Several studies have been carried out to examine its impact on the performance of IEEE 802.1 1 DCF protocol [3,4, 5, 6 , 71.

The majority of previous studies have only considered one or two QoS parameters. This study aims to investigate the impact of varying the value of CWmi, on the network performance when multiple QoS parameters, i.e. throughput, delay, jitter, and packet loss, are considered. The packet size in VBR traffic may critically affect the optimum value of CW,i, size, since small packet sizes may require a small CW,i, size and large packet sizes may require a large CWmi, size. A novel method based on Artificial Intelligence (AI) techniques for optimizing the value of CW,i, is currently being developed. Its structure is outlined in this paper.

This paper is organised into six sections. In section 2, the basics of the IEEE 802.1 1 MAC protocol and the DCF are introduced. The experimental procedures for the study are introduced in section 3. The findings and discussions of this study are included in section 4. The proposed approach is outlined in section 5. The conclusion and future work are presented in the last section.

2. IEEE 802.1 1 MAC Protocol

The IEEE 802.1 1 standard defines two channel access mechanisms, called coordination functions [2]. These coordination functions determine when a station is permitted to transmit, and when it must be prepared to receive data. The basic access mechanism is the DCF which adopts the CSMA/CA mechanism to provide services for asynchronous data transmission. The optional mechanism is the Point Coordination Function (PCF) which incorporates a polling coordinator that locates at the Access Point (AP). The main focus of this paper is based on the DCF. A flow chart of the CSMA/CA protocol of the basic access mechanism is shown in Figure 1.

In IEEE 802.11, DCF also identifies another mechanism which is an optional method of transmitting data packets. The optional mechanism is involved in transmission of special short control messages called RTS and CTS frames. These short messages are always transmitted prior to the transmission of the data packet.

22 1

start

Wait for packet to transnit

Listen to the medium

* ,

update cmnter

wai t IFS

Listen to the medium

-b wait ps 4

intew* I Initiatetransmission I

I TransmissionOK I

Figure 1. Flow chart of the basic access mechanism

3. Experimental Procedures

The ns-2 simulation software [8] was used to carry out the experiments. The experiments analysed the impact of varying the minimum value of CWmi, on the QoS parameters when transmitting VBR traffic.

Four VBR connections, with interval that had a mean value equal to 0.89 second and standard deviation equal to 0.56 second, (all within the transmission range of each other) in a 20 fixed stations topology provided the network with its entire load. Each connection was specified as a source - destination pair, station-0 and station-1 5 composed the first connection, station-1 and station-1 6 composed the second connection, station-:! and station-1 7 composed the third connection, and station-3 and station-1 8 composed the fourth connection.

Each source (station), station-1 , station-2 and station-3) was associated with a VBR traffic generator, which transmitted packets at a fixed packet size and variable interval. The simulation was carried out for different packet sizes

222

(500-byte, 900-byte and 1400-byte). The reason of selecting different packet sizes was that the optimum CW,, for small packet sizes should have small CWmin and for large packet size should have larger CWmin. The channel capacity was set to 2 Mbps to serve all the generated VBR traffic.

4. Results and Discussions

In this section the results obtained are discussed.

4.1. Throughput

The relationship between the sum of averaged throughput and the CW,i, value a

is shown in Figure 2. The experiments were carried out with packet sizes 500, 900, 1400 bytes. It can be observed for all packet sizes that small values of CW,in provided small average throughputs. For instance, at CW,, value 15, for packet size 500-byte, the sum of average throughput was 1.19 Mbps. This value was 8% less than the peak value that was achieved at the optimum CW,i, &e. 47). At large values of CW,i, for instance, at CW,k equal to 512 for packet size 500-byte, the sum of average throughput was 30% less than the peak value that was achieved when CWminvalue was 47.

1.7 - 12 g- Y lp, 1.5

0 E 1.4

5 p 1.3 2 ki g 1.2 c

9 1.1 m

1 0 20 40 60 80 100 120 140 160 180 200 220 240 260

Minimum CW

Figure 2. Sum of averaged throughput of four VBR connections for different packet sizes

a The summation of the average throughput for the whole connections that are performed in the simulation and we refer to it as the average throughput in this section.

223

4.2. End-to-End Delay and Jitter

Initially an increase of CWmin value resulted in gradual decrease in the average end-to-end delay as shown in Figure 3. Smallest values of average end-to-end delay were obtained at the following CW,,, values. At packet size 500-byte the optimal value of CWmin was equal to 64 (670.5 ms). At packet size 900-byte the optimal value was equal to 80 (878.1 ms) and at packet size 1400-byte the optimal value of CW,,, was equal to 112 (1.12 second). Any hrther increase of CWmi, value above the optimal values led to an increase in the average end-to- end delay.

The average end-to-end delay of packet size 500-byte was 38% smaller than the value average end-to-end delay that was achieved at CWmi, 512.

Regarding the average jitter, smallest values were achieved at CWmin value 112 for packet sizes 900, 1400 bytes respectively. While the smallest value of average jitter for packet size 500-byte was achieved when CWmin value was 128.

As shown in Figure 4, at CWmin value 15, large values of average jitter were obtained for packet sizes 500, 900, 1400 bytes. For example, at packet size 500-byte, this value was 32% larger than the value of average jitter that obtained at optimal CWmin (i.e. 128).

5 1.2

= 1

4 0.8

2 0.6 Y

4 0.4

iri h - V

V

0

4 E 0.2

0

-Average End-to-End Delay [900 bytes] -Average End-to-End Delay [1400 bytes]

0 20 40 60 80 100 120 140 160 180 200 220 240 260

Minimum CW

Figure 3. Average end-to-end delay of four VBR connections for different packet sizes

224

0.021

0.018

0.015 6

v) I

b 0.012 P .- & 0.009

0.006

0.003

0

0 20 40 60 80 100 120 140 160 180 200 220 240 260

Minimum CW

Figure 4. Average Jitter of four VBR connections for different packet sizes

4.3. Packet Drop

The relationship between the data packet drop due to buffer overflow and CW- is shown in Figure 5. The experiments were repeated with packet sizes 500,900, 1400 bytes. It was observed in all the cases that the packet drop gradually fell to a minimum value as CW- was increased. Thereafter any further increase in CW- caused an increase in the packet drop. The cause the trend followed by the three graphs was that very small values of CW- resulted in more collisions and very large values of CW- resulted in a high waiting time for the stations to transmit.

I

0 20 40 60 80 100 120 140 160 180 200 220 240 26

Minimum CW

Figure 5. Average end-to-end delay of four VBR connections for different packet sizes

225

In summary, the variation of the value of CW,,, has a considerable impact on the network performance. Therefore, in order to achieve a high network performance the value of CW,i, has to be optimised.

5. Fuzzy Logic Based Approach for Optimising CWmin Value

For a system with little complexity (i.e. little uncertainty) mathematical equations provide precise definition of its operation. For more complex systems but for which significant data are available, model-free techniques such as fuzzy logic provide an effective means to reduce the complexity [9]. For the most complex systems where few numerical data exist and only imprecise information is available, artificial intelligence provides an effective way for understanding them [lo].

Realisation of medium access control which caters for QoS is a complex task which involves imprecise information from the measured data (i.e. delays, jitter, and throughput). Furthermore, the dynamics of the channel varies in space and time in a complex manner. Therefore the use of fuzzy logic as part of IEEE 802.1 1 WLAN medium access control will prove valuable.

Following on from these results discussed in this paper an approach based on using fuzzy inference system to assess the QoS and to achieve the optimum value of CWmi, is currently being developed. Simulations have been carried out when transmitting video stream (CBR traffic with packet size equal to 1280 bytes and interval equal to 0.01 second). The values of CW,,, and CW,,, were set to the default value of the simulation software (31 and 1023, respectively) [8]. Each QoS parameter was represented by three labels, low, medium and high. Then, all these were forwarded to the fuzzy inference system to be evaluated with respect to the QoS provided and to suggest an optimum CWmi, value.


The impact of varying the value of CW,i, was investigated by using the ns-2 simulation software. Results indicated that CW,i, value had a significant impact on the network performance, particularly, the QoS parameters. High values of CWmi, caused long defer of data packets, whereas small values caused large number of collisions. Therefore, a sharp increase or decrease of the value of CWmi, led to a large number of collisions and a large number of drops at the buffer particularly when the transmission over the channel was loaded or in a congested state.

226

The variations of the value of CW,i, above or below the optimal value degraded the performance of the network. Small values of CWmin resulted in large number of collisions and large values of CWmin resulted in long waiting time of data packets at the queue.

The study indicated that the optimum values of CWmin for different QoS parameters were different. Therefore, more sophisticated and intelligent techniques are required to determine the optimal value of CWmin for a specific application. This proposed approach will focus on optimising the MAC protocol transmission parameters using A1 techniques such as fuzzy logic, neural networks, and genetic algorithms to deal with optimising CWmi, value.

References

1. IEEE standard 802.1 1, "Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications", (1 997).

2. IEEE standard 802.1 1, "Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications", (1 999).

3. W. Haitao, C. Shiduan, P. Yong, L. Keping and M. Jian IEEE 802.11 distributed coordination function (DCF): analysis and enhancement", Proceedings IEEE International Conference on Communications ICC 2002.

4. A. Veres, A. Campbell, M. Bany and L. Sun "Supporting Service Differentiation in Wireless Packet Networks Using Distributed Control," IEEE J. Selected Areas in Comm., special issue on mobility and resource management in next-generation wireless systems, vol. 19, no. 10, pp: 2094- 2104, (2001).

5. 1. Aad and C. Castelluccia "Differentiation mechanisms for IEEE 802.1 1," Proceedings of IEEE INFOCOM2001, vol. 1, pp: 209-2 18,200 1.

6. Y. Kwon, Y. Fang and H. Latchman I' A novel MAC protocol with fast collision resolution for wireless LANs If : Proceedings - IEEE

Distributed Coordination Function", IEEE journal on selected Area in Communication. vol. 18, no. 3, pp: 535-547, (2000).

8. The network simulator-ns-2. [Online]. Last accessed on 5 September 2004 at URL: http://www. isi .edu/nsnam/ns/.

9. L. Zadeh "Fuzzy sets", Information and Control.vol.8, pp: 338-353, (1 965).

V O ~ . 1, pp: 605-609, (2002).

INFOCOM, V O ~ . 2, pp: 853-862, (2003). 7. G. Bianchi "Performance Analysis of the IEEE 802.1 1

10. T. Ross "Fuzzy logic with engineering applications", McGrow-Hill, (1 995).

STATISTICAL QOS G U W T E E S IN BLUETOOTH UNDER CO- CHANNEL INTERFERENCE*

J.L. SEVILLANO+, D. CASCADO, F. D h Z DEL RiO, S. VICENTE,

ETSII. Universidad de Sevilla. Av. Reina Mercedes, sln. 41012, Sevilla, Spain G. JIMEmZ, A. CIVIT-BALCELLS.

Abstract: Bluetooth is a suitable technology to support soft real-time applications like multimedia streams at the Personal Area Network (PAN) level. In this paper, we evaluate the worst-case deadline failure probability (WCDFF') of Bluetooth packets under co-channel interference as a way to provide statistical QoS guarantees.

Keywords: Bluetooth, co-channel interference, QoS, worst-case deadline failure probability.

1. Introduction

A typical Bluetooth system is composed of a small number of devices that form a wireless network called apiconet. Connections are established ad-hoc by a Bluetooth unit that becomes a muster so that the other units (slaves) synchronize with it. The Bluetooth channel is then divided into slots of length 625p so that time slots are alternatively used by master and slaves (Time Division Duplex). Any unit may function as a master or as a slave (this role is maintained only for the duration of the piconet), but although it may participate as slave in multiple piconets it may only be a master in one piconet.

Two types of connections can be established in Bluetooth [l]: Synchronous Connection-Oriented (SCO) and Asynchronous Connectionless (ACL) links. In ACL links the master of the piconet performs a polling among all the slaves. Slaves can only transmit if the master in the preceding slot has requested them. ACL packets require acknowledgement and they are retransmitted in case of errors using a fast

This work was supported by the spanish Ministry of Science and Technology under contract Heterorred: TIC2001 -I 868-CO3-02.

+ Corresponding author: [email protected].

227

228

ARQ scheme until they are successfully delivered. Several types of ACL packets are defined depending on whether they use FEC (Forward Error Correction) or not. Also, multi-slot packet transmissions are allowed (3 or 5 slots packets), although in this paper we assume that each packet occupies a single slot.

On the other hand, SCO links are based on a fmed and periodic pre-allocation of slots (every two, four or six slots) for guaranteed transmission of continuous (audio- video) streams. SCO packets are always one slot length and they are never retransmitted. If errors occur during transmission (not corrected by FEC) they are ignored and the packet is delivered as it is received. Finally, the master does not have to poll the slaves in SCO links, so a slave may transmit an SCO packet without a previous request.

The use of SCO links to support soft real-time applications like multimedia streams monopolizes bandwidth and leaves very little room for other ACL links. Several authors [2,3] have proposed that ACL links could be used to carry voice and other isochronous traffic. These applications require QoS guarantees in terms of delay, delay variation and loss rate [4], so several attempts have been made to provide these guarantees in ACL links [5,6]. However, when using wireless connections a system with several interfering devices is probabilistic in nature. In the case of ACL links, although the polling performed by the master provides determinism within a piconet, there is always the possibility of a collision with other Bluetooth devices connected to different logical channels, as well as with other interfering devices operating in the ISM (Industrial, Scientific & Medical) band.

To the best of our knowledge, previous attempts to provide QoS guarantees for Bluetooth connections do not take into account co-channel interferences. Maybe the only exception is [7], but they do not take into account the effect of the polling algorithm or the interference between SCO and ACL packets.

In this paper, we propose the use of schedulability tests like those used to provide guarantees on message delays in distributed real-time systems [8]. These analyses use the worst-case transmission times to bound message response times, and to assess the schedulability of the system. However, when packet transmission times cannot be upper-bounded, as is the case with Bluetooth piconets under co-channel interferences, then a probabilistic analysis like SRMS (Statistical Rate Monotonic Scheduling [9]) is needed. A simpler approach is the probabilistic time-demand analysis used for single-processor systems with semi-periodic tasks (tasks released periodically but with random computation times) in [lo]. However, any of these approaches are difficult and computationally expensive. In this paper, we prefer to evaluate the worst-case deadline failure probability (WCDFP) in a similar fashion to

229

that used to take into account transmission errors in CAN based systems in [ 111. In the following section we first compute the worst-case response times ignoring interferences, and then in section 3 retransmissions due to interferences are included.

2. Worst-case Response Time

Consider a piconet composed of a master and up to 7 slaves, with several SCO and ACL links. SCO packets have pre-assigned slots and always pre-empt ACL packets. Furthermore, they are never delayed, so their response times remain constant regardless of possible interferences. Therefore, in this paper we focus only on the computation of worst-case response times of ACL packets, while SCO packets are considered simply as a source of interference.

Therefore, consider N (N17) slaves with active ACL links. Only a single ACL link can exist between the master and every slave [ 121, and the polling is performed only with slaves with ACL links. Outgoing ACL packets are queued whenever they suffer delays fiom three possible sources:

The polling mechanism. An ACL packet has to wait for the poll when it's the other nodes' turn. Preemption fiom SCO packets. Interferences fiom other ISM devices. These interferences may destroy the packets so they introduce additional delays.

We first present an analysis to compute worst-case response times of ACL packets including the delays due to the first two sources of interferences, while in the following section we extend this analysis to include interferences from ISM devices, particularly Bluetooth devices.

In our model, an SCO packet j is characterized by its period (Tscoj) and its constant transmission time (Csc0=2 slots, since we count two slots per transmission due to the Time Division Duplex mechanism). On the other hand, an ACL packet i is characterized by its period (Ti), deadline (Di) and transmission time (Ci). If the traffic in the ACL link is not isochronous, then the period should be interpreted as the minimum time between successive packets. In what follows, we assume that ACL packets transmission times do not depend on the priority level (the Bluetooth MAC layer does not support priorities except for the priority of SCO over ACL packets). As a result, we write Ci=C, for all i. To simplify the model, we also assume that packets must be received before the end of the period of the sending task (i.e. Di<Ti), to avoid packets fiom successive invocations of the sending task to delay each other.

230

Several intra-piconet polling algorithms have been proposed [13]. In this paper, we consider the simple Pure Round-Robin (one packet per visit), used in most current Bluetooth connections [ 5 ] . Although simple, it has been shown that PRR has a good performance at high loads [ 131.

We are interested in the computation of the worst-case response time of an ACL packet that may be delayed by other ACWSCO packets. With the PRR algorithm, when we can distinguish between two cases: . Packets queued at a given node. Since only a single ACL link can exist between

the master and every slave, there is no interference on a given ACL packet rn due to local (within the node) higher priority packets except for SCO packets. Packets queued at other nodes (ACL l i i between the master and the other slaves). While our reference ACL packet m is waiting for the poll, these packets are being transmitted. They can be modeled as a single "high priority packet" with periodicity equal to the polling period TPou [8]. The interference due to packets in other nodes is given by the number of times

this "higher priority packet" with period equal to TpoLL is scheduled for transmission during the worst-case packet response time, that is re, /TpoLL ITpou - C) , where Qm

is the queuing time of our reference ACL packet rn (time from being queued to the time transmission begins). In the PRR case, nodes are only allowed to transmit a single packet per visit. Therefore, Tpou=N*C. Every time this "packet" interfere a given packet rn, we have to wait TPou-C=(N-I) *C.

We also have to add the periodic and fixed interference &om SCO packets, which does not depend on the node they may be generated. The master can support up to three simultaneous SCO links while slaves can support two or three SCO links. Therefore, the interference from these packets can be found by calculating the number of times successive instances of SCO packets could be scheduled for transmission in front of a given ACL packet m:

I

where SCOs is the set of SCO links in the piconet. We assume that the jitter (the maximum time variability between subsequent packets of an SCO link) is null. The queuing time can now be found iteratively [8]:

23 1

The total response time of a given ACL packet m in the worst case is then:

R, =Q, , ,+C . (2)

If interferences from other ISM devices are ignored, then whatever polling scheme is used collisions do not occur within a piconet. Therefore, the packet transmission time is a constant. Since we are considering single-slot packets, then C=2 slots of 6 2 5 ~ s (Time Division Duplex).

3. Considering Co-Channel Interference

In order to reduce the interference from other devices, Bluetooth uses Frequency Hopping, with a pseudo-random hopping sequence. We can then assume that a Bluetooth device transmits using randomly chosen fiequencies. Therefore, there is a possibility that several independent (but interfering) Bluetooth devices coexist in the same area and that they choose the same hop carrier. In that case, a collision occurs and the packet will be received incorrectly. The sender is notified of this error in the slot directly following the unsuccessful transmission using a fast-ARQ scheme [ 11. The packet is then retransmitted at the next opportunity (in alternate slots) until it is successfully received. Therefore, if co-channel interference is considered, the transmission time C cannot be considered a constant anymore. Instead, C and Qm become (discrete) random variables.

We can assume that hops are evenly distributed over 79 different frequency bands [14]. In [15] we show that if Mis the number of interfering piconets and r is the normalized load over every piconet (we assume a homogeneous traffic, so r is the same for all the piconets) then the probability of successhl transmission is

The approximation symbol comes from the fact that in this equation we neglect the term (r/79,f<0.00016. Furthe~more, although a time and frequency coincidence does not always destroy the packet, depending on the strength of the interference signal arriving to the receiver [16], we assume the worst-case: interference ofjust one bit is enough to destroy the whole packet. Finally, the fact that a single slot packet is only of duration R=366/625=0.5856 slots and that independent piconets are not synchronized are taken into account through lbe factor 2R [17]. We will use this model to show how to include co-channel interferences in the computation of the worst-case

232

response time of ACL packets. However, many other interference models could also be used, for instance the multi-slot case discussed in [ 181.

In a piconet with ACL links, the wasted slots due to collisions correspond to a sequence of Bernoulli trials with probability of success Ps. The overhead for every collision is just a packet retransmission C=2 slots. Therefore, the effect of interferences can be included in our model of section 2 by adding a collision overhead E M that is a function of the number of collisions during a time period X. If SCO packets are ignored, E(X=k*C, with k being the number of collisions during any time period X. However, we have to take into account that SCO packets are not retransmitted in case of collisions (in that case, the packet is simply corrupted but delivered "as is"). Therefore, k should be computed as the number of collisions occurring in any time period Xexcluding the transmission of SCOpackets.

Anyway, considering co-channel interferences, eq. (1) becomes

Note that C is again a constant G 2 slots (for single slot packets) because all the effects of collisions are included in function E. The response time of packet m with k collisions can now be computed iteratively. First we iterate eq. (4) with k=O, obtaining a first result for Q,, and through eq. (2) for R,. Then we increment k and repeat the iterative computation to obtain new results for Q, and R,. We repeat this algorithm until R,>D, (or the iteration does not converge). In this way we obtain K,, which we define to be the maximum number of collisions for which R,<D, (since the final value kfmakes R, exceed the packet deadline, we have K,=kfl). Let Quax be the last obtained value of Q,, that is, the queuing time obtained through the following iteration:

and also let R- be the last obtained value of R,, R , = Q- + C (the worst-case response time when the maximum tolerable number of collisions K, occur).

Once K, and R- are obtained, we are now able to compute the worst-case deadline failure probability (WCDFP) simply as the probability that more than K,

233

collisions occur during R-. If we assume that time is slotted, with slot length 625ps, then the probability that a collision occurs in a given slot is Z-Ps. Therefore, we have that the probability that k collisions occur during a time period Xis

In our case, the time period is R- but transmission of SCO packets should be excluded. Therefore,

Finally, the worst-case deadline failure probability (WCDFP) for an ACL packet m, with Xobtained using eq. (7), is given by:

K

WCDFP, = 1 - g( ;)(l- Ps )" P:-k k=O

4. Conclusions

In this paper, the worst-case delay of ACL packets is computed including preemption fiom SCO packets, the polling mechanism and the effect of interferences fiom other Bluetooth devices. Then, the probability that this delay exceed the packet deadline (the so-called worst-case deadline failure probability- WCDFP) is computed. Our model allows us to evaluate this statistical QoS parameter as a function of the number of interfering devices and quantify how this number degrades QoS. This model may be useful to predict how soft real-time applications behave when several active Bluetooth devices are used in a common area, including other nodes in the piconet, active SCOIACL links, independent devices, etc.

References

[ 11 J.C. Haartsen, The Bluetooth Radio System. IEEE Personal Communications

[2] D. Famolari, F. Anjum: Improving Simultaneous Voice and Data Performance in Bluetooth Systems. Proc. Of Globecom 2002, pp. 1810-1814.

7(2000) 28-36.

234

[3] R. Kapoor, L. Chen, Y. Lee and M. Gerla: Bluetooth: Carrying Voice over ACL links. Proc. IEEE MWCN 2002, pp. 379-383. [4] M. Hamdi, F.L. Lee: Providing deterministic packet delays and packet losses in multimedia wireless networks. Wirel. Commun. Mob. Comput. 3(2003)3-22. [5] S. Chawla, H. Saran, M. Singh: QoS Based Scheduling for Incorporating Variable Rate Coded Voice in Bluetooth. Proc. IEEE ICC, pp. 1232-1237,2001. [6] W.P. Chen, J.C. Hou: Provisioning of Temporal QoS in Bluetooth Networks. Proc. IEEE MWC”O2, pp. 389-393. Stockholm, Sweden. 2002. [7] A. El-Hoiydi, J.D. Decotignie, SofbDeadline Bounds for Two-way Transactions in Bluetooth Piconets under co-channel Interference, 8* IEEE Int. Cod. Emerging Technologies and Factory Automation, pp. 143-150. Oct. 2001. [8] K. W. Tindell, A. Burns, and A. J. Wellings: Guaranteeing hard real-time end-to- end communications deadlines. Technical Report RTRG/9 1/107, University of York, UK, December 199 1. [9] A. Atlas, A. Bestavros: Statistical Rate Monotonic Scheduling. IEEE Real-Time System Symp.. Madrid, 1998. [lo] T.S. Tia et al.: Probabilistic Performance Guarantee for Real-Time Tasks with Varying Computation Times. IEEE Real-Time Tech. And App. Symp., pp. 164-173. Chicago (USA), May 1995. [l 11 N. Navet, Y.Q. Song, and F. Simonot: Worst-case deadline failure probability in real-time applications distributed over controller area network. Journal of Systems Architecture 46 (2000) 607-617. [ 121 “Specification Of the Bluetooth System - Core Vol. 1 V1.l”. Bluetooth Special Interest Group. Feb 2001. pp-1084. [13] A. Capone, M. Gerla, R. Kapoor: Efficient polling schemes for Bluetooth picocells, Proc. IEEE ICC, pp. 1990 -1994,2001. [ 141 J.C. Haartsen, The Bluetooth Radio System,. IEEE Personal Communications

[15] D. Cascado et al., Performance analysis of single-slave Bluetooth piconets under cochannel interference. Proc. IEEE PIMRC’2004, pp. 954-958. Barcelona, Spain. [16] D. Cascado et al.: Including propagation effects on packet error rate of Bluetooth packets. Proc. WoWCAS 2004, Vancouver, Canada. May, 2004. [ 171 J.L.Sevillano et al., An analytical model of Inter-channel Interference in Bluetooth-Based Systems. Proc. IEEE MWCN’02, pp. 384-388. Stockholm, 2002. [I81 T.Y. Lin, Y.C. Tseng: Collision Analysis for a Multi-Bluetooth Picocells Environment. IEEE Comm. Letters 7-10 (2003) 475477.

7(2000) 28-36.

PERFORMANCES EVALUATION OF THE ASYNCHRONOUS BLUETOOTH LINKS IN A REAL TIME ENVIRONMENT

TARIK KHOUTAIF, FABRICE PEYRARD Research ICARE Team - EA3050

1, place Georges Brassens BP 60073 31 703 Blagnac Cedex France

Jkhoutaif; peyrard) @iut-blagnac.fr

This article presents the implementation of a real time platform needed for critical wireless communications. The objective of this work is to evaluate real time system temporal behaviors based on RTAI and the Bluetooth (WPAN) real time communications management for control/command applications (control/command of remote mobile robots). We present a necessary aspect for the optimized operations (interruptions, management of the clock) of the asynchronous ports in the real time environment. Then the different asynchronous Bluetooth communication channels are presented and implemented in order to evaluate their behavior and performance.

1. Introduction

Today wireless communications between the control unit and the robot are needed by many robotic applications, while respecting various temporal constraints. The respect of these temporal constraints linked to mobility, reliability and system availability justifies our research activities for the improvement of response time in the loop of mobile robots control/command. Instructions commands and responses will be conveyed through a “Bluetooth” (2) wireless network for short distance transmissions (a few meters). To achieve these goals, we proceed by stages. In the first part of this article, we well present the principal points of RTAI (Real Time Application Interfaces), the selected real time operations systems. Work presented in a precedent article (8) will justify the choice of RTAI (Real Time Application Interfaces) as a real time operating system. The following stage is the general presentation of Bluetooth and particularly various asynchronous links for data transmission. Then we will present all the real time tasks involved in communication architecture, and will evaluate the temporal perfonhances of the principal Bluetooth links in a real time communicating environment.

235

236

2. Real time system and RTAI

Among operating systems, we can recognize two families, time-sharing operating systems and real time operating systems. In a time-sharing operating system, the scheduler shares the processor time between the various tasks on the wait to being performed. Moreover a real time operating system should guarantee the availability of the material resources for defined tasks, within precise temporal limits. In a real time operating system, data processing and acquisition times should be shorter than the occurrence data times (4). We made a state of the art (7) of the various real times and/or embedded operating systems, and according to performance criteria, kernel, size, latency, "open-source" ..., we selected RTAI (1 1). RTAI is a project inspired from RTLinux (5) which improve real time modes and the floating numbers management. It also allows the development of real time tasks in user space with a fine enough temporal granularities thanks to LXRT (LinuX Real Time) extension (1).

2.1. RTAIstructure

RTAI is a real time kernel using the module theory (3). Communication between the various tasks is provided by the following mechanisms: queues (Fifo), semaphores, shared memories and mailboxes. Material interruptions are collected by the real time kernel and transmitted to the Linux kernel, only if they do not correspond to real time tasks. This characteristic is fundamental since we wish to use the asynchronous interface for data reception in interruption mode.

2.2. Serialport interruptions under RTAI

RTAI uses two interfaces to manage the asynchronous ports. The RT-COM interface resulting initially from RTLinux and its evolution for RTAI, called the SPDRV interface (Serial Port Drivers) (9).

Figure 1 - RTAI and interruptions management

237

However, in our real time communication problem, we have two possibilities; either to write our own asynchronous interface with interruptions management, or to use one of the interfaces offered by RTAI. According to the SPDRV specifications and temporal measurements (7), we chose SPDRV. The various entities which contribute to the management of the RTAI interruptions are presented in Figure 1 . We also present the real time primitives which enter in the interruption management, as well as their levels of intervention on the software layer.

2.3. Conclusion

In this part, we have presented the structure of the real time RTAI system, along with these specificities offered for asynchronous communication based on interruptions. We remind you that such a real time environment should be integrated in to an embedded architecture of mobile robots. In this article, we will not present the aspects related to the optimization of this real time system to embark it on mobile entities. This will be one of the objects of our future studies. We will now present the communicating architecture based on Bluetooth network

3. Bluetooth communicating systems

Among all the local wireless or personal area networks, we retained Bluetooth for our experimental environment (6). Indeed this type of network has been the subject of many works in our laboratory (13), and also allowed us to test it in a real time environment. Experimental network architecture led us to use point-to-point topology in order to control Bluetooth behavior in a real time environment.

3.1. Communication channels

Bluetooth offers two types of communication channels: synchronous channels SCO (Synchronous Connection Oriented) mainly dedicated to voice transmission, asynchronous channels ACL (Asynchronous Connection Less) mainly dedicated to data transmission.

Table 1 summarizes all the asynchronous links offered as well as the associated flows. The various flows offered by the ACL links are mainly due to a variable use of the number of temporal intervals (time slot) to transport information (1 2)( 10). Moreover, the use of a correcting FEC code (Forward Error Correction) offers better service quality but reduces of course the useful band-width.

238

Table 1

Name Type

- Asynchronous channels offered by Bluetooth Useful Detecting Correcting size of Theoretical data Code Code temporal maximum

(bytes) occupation rate DMI ACL 18 CRC 213 FEC 1 slot 108 kbits/s

DHI ACL 28 CRC 1 slot 173 kbitsls

DM3 ACL 123 CRC 2/3FEC 3 slots 387 kbitds

DH3 ACL 185 CRC 3 slots 586 kbitds

DM5 ACL 226 CRC 213 FEC 5 slots 478 kbits/s

DH5 ACL 341 CRC 5 slots 723 kbits/s

The flows indicated in table 1 correspond to the theoretical maximal flows mentioned by the Bluetooth standard.

3.2. Simplified protocol architecture

Nowadays, Bluetooth is integrated perfectly in the whole of wired and wireless, essentially in IP protocol basis, however in our experimental real time application, we are at the bottom of the protocol architecture, in order to avoid the additional delays caused by the crossing of the upper protocol layers. In order to avoid the constraints of USB protocols, we first designed our experimental platform on the HCI interface via the RS-232 asynchronous ports.

4. Real time modules to control Bluetooth peripherals

4.1. HCZ interface andpackets

Thanks to specific requests, the HCI interface allows to pilot the Bluetooth module for its own operation and the emissiodreception of user data. These requests (finished series of bytes) correspond to packets classified in to three categories:

The command, event and ACL data packets on HCI are represented in Figure 2.

request packets, used by the host to control the Bluetooth module, event packets, used by the Bluetooth module generally to answer a request, data packets, for the transmission of user specific data to another Bluetooth module.

239

ldcnafiir rmsxion (2 bWs)

Parpmlen daru number(2 b y W (65535 b p s )

The use of these three types of packets require the use of a single code (on 1 byte) before each packet so as to identify it:

code Ox01 for HCI command packets, code 0x02 for HCI ACL data packets, code 0x04 for HCI event packets,

The ACL data packets (presented in Figure 3) thus sent on the HCI interface will undergo a specific modification from of the Bluetooth module so that is built, well formatted for a forwarding on an ACL radio channel.

Action

Aecm code Hesdcr Payload (72 bib) (54 bib) (2744 bils)

. /.- I

mode of Software Reset automatic Activate Rate loading

detection of events to 15 Creation of

module management kbits/s connection of Bluetooth

module

Figure 3 - Data packet structure on an ACL channel

4.2. Piloting the Bluetooth modules in real time

In order to optimize the configuration of a Bluetooth module in our real time environment we have defined an initialization sequence. The sequence of the following command (Table 2) represents the commands necessav and sufficient to initialize the Bluetooth module.

Processing time

Table2 - Initialization sequence of Bluetooth module Activate the

(Inquiry) 1.905595

(s) 8.198023 (ms) 4.063218 (ms) 2.73233 (ms) 8.05421(ms)

This sequence is thus independent of the type of Bluetooth module in the MastedSlave communication, which allows us to define the programmed init-BT.0 module for initialization. Each entity Master or Slave must perform this module and moreover the only Master must achieve the connection with the

240

Slave whose address it knows. Once the connection achieved, the Slave will generate a connection creation event. We have evaluated the processing times of each command of this software module, whose time off connection is approximately 23 ms. Yet we note that the connection creation time it-self represent approximately 2s. It is significant to consider this in the Bluetooth connection creation phase and particularly in a multi-robot (Piconet) environment. Once the connection is created between Master and Slave, we will transfer data between these two entities in real time environment.

5. Communicating real time tasks

From our experimental platform, we have defined two real time tasks of same priority on the system. These tasks perform data emission and reception. The Master task is regarded as the cockpit and the Slave task as the mobile robot. The behavior of the Master emission is aperiodic (by pulling a random time), which requires the use of the reception mechanism on interruption from the Slave. Considering the behavioral symmetry between Master and the Slave in a point-to-point connection, we have defined the module send-data-ac1.o programmed for data packets emission, and the programmed receive-data-ad. o module for the reception. The chronogram in Figure 4 shows the behavior of these two real time tasks.

The various temporal references presented in Figure 4 are used during the Bluetooth ACL links performance evaluation in a real time environment.

6. Evaluations of performances

6.1. Measurements on the wired medium

In order to validate the measurements taken in real time communicating environment with Bluetooth, we constructed a set of preliminary measurements in the following context: a real time system based on RTAI, on which two asynchronous serial interfaces are inter-connected with a cross cable. This

fIGURE 4. cHRONOGRAM OF THE mASTER/sLAVE TASKS

241

platform allows us to make temporal measurements without the clock synchronization constraints in a multi system environment. Figure 5 presents the values measured in a wired environment with a serial port offering 1 15200bitsIs flow.

I Bytes number I Figure 5 - Temporal measurements on the wired medium

We can note that the measured curves are quite higher than the theoretical curves, since they include the transport of all of the bits of the serial interface management (stop bit, parity...). The curves of command and response transmission correspond to the emission of an order from the transmitter followed by the emission of a response from the receiver. This makes it possible to evaluate the future behavior in the robot or cockpit mobile environment. The stability of communicating real time system thus led us to make measures on the wireless Bluetooth environment, with the initial measurements as reference.

6.2. Measurements with Bluetooth

Temporal measurements were taken in an optimum context, i.e. on the same and single real time system. This allows us to realize time captures (Figure 5) from the Master (T-d-cmd and T-f-rep) and Slave tasks (T-f-cmd and T-d-rep). Figure 6 presents the usehl flow measured with Bluetooth in comparison with the wired useful flow as well as the theoretical 1 15200bits/s flow.

Bytes number I Figure 6 - Performance evaluation of Bluetooth in a real time environment

242

We first note a weakening of the useful Bluetooth flow due to the various bytes constituting the frame envelope (heading, CRC ...), on top of the bits of the serial interface management (Figure 6). The user flow offered by Bluetooth on this real time platform does not exceed the 4Okbits/s via the HCI interface. This is in conformity with the flows measured of approximately 30kbitsh on the software layers higher than HCI

Unfortunately the physical characteristics of the asynchronous ports of real time system are limited to 1 15200bits/s, which did not allow us to evaluate explicitly DM3DH3 and DM5/DH5 packets. Indeed, the serial ports appear in this case like a bottleneck for the ACL Bluetooth channels beyond 115kbitds. However, we made a similar measurements as those made with DMlDHl packets. We present in Figure 7 a comparative analysis of DMl and DHl ACL channels whose principal difference is a higher payload for DHl packets, but without error correcting code (Table 1). We note that the useful flow in DMl is lower of about 6kbits/s, which corresponds indeed to the difference between the maximum flow of DM1 (108kbitds) and that of the serial port (1 15kbitsh).

(12).

4 0 0 0 0

35000 30000 25000

2 20000 1 5 0 0 0

10000 5000

0 1 5 8 1 1 5 1 7 2 2 2 9 2 8 6 3 4 3 4 0 0 4 5 7

B y t e s n u m b e r

Figure 7 - Comparative analysis of DMl and DHl ACL channels

7. Conclusion

This work has allowed us to validate interoperability between the real time RTAI system and the personnel local area Bluetooth networks, for asynchronous communications via the serial interface of the system. This experimental model was produced as part of the implementation of a network composed of a cockpit and mobile robots in a real time communicating environment. We have implemented a real time piloting module managed by interruptions to use the Bluetooth modules via the HCI interface. We have performed a whole set of measurements in order to evaluate the behavior of such a system on real time architecture. These measurements proved to be satisfactory in spite of the flow constraints of the serial port not allowing an optimal evaluation of all the Bluetooth asynchronous channels.

243

To avoid the bottleneck of the serial port, we will evaluate the feasibility of the use of USB ports on our real time platform so as to be able to use other packets than DM 1 /DH 1.

References

1. Bianchi E., Dozio L., (( Some Experiences in fast hard real time control in user space with RTAIILXRT n, Real time Linux Workshop, Orlando, 2000.

2. (( Bluetooth Core Specification 1.12 i), http://www.bluetooth.org/spec/ 3. Cloutier P., Mantegazza P., Papacharalambous S., Soanes I., Hughes S.,

Yaghmour K., DIAPAM - RTAI position paper i), workshop (IEEE Real- Time Systems Symposium) RTSS 2000, Orlando, Florida, 27-30 November 2000.

4. Ficheux P., (( Linux embarque )i, Edition Eyrolles, 2000. 5. Real Time Linux (RTLinux), FSMLabs, http:// www.fsmlabs.com/ 6 . Khoutaif T., Peyrard F., Val T., Vers l’utilisation des RdPTS pour

I’evaluation de performances d’un systkme de communication sans fil embarque et temps reel dedie a la robotique i), CNRZUT’O3, Tarbes, France, 2003.

7. Khoutaif T., Peyrard F., Val T., Juanole G. ((Evaluation de performances des communication entre tlches temps reel sur une interface asynchronei), EDSYS’OI (Congrks annuel des doctorants de 1’Ccole doctorale systkmes de Toulouse), Toulouse, France. Mai 2004.

8. Khoutaif T., Peyrard F., (( Evaluation de performances des communications asynchrones d’un systkme temps reel base sur RTAI i i, IEEE SETIT’OS (Sciences of Electronic, Technologies of Information and Telecommunications), 27-3 1 mars 2005, Tunisie.

9. Mantegazza P., Renoldi G., (( Real Time Application Interface documentation of Serial Port Drivers i), RTAI 2002.

10. Peyrard F., Val T., (( Simulation et analyse de performances de liens synchrones Bluetooth avec Opnet D, Colloque Francophone sur l‘lngknierie des Protocoles CFIP’2002, Montreal, 27-30 mai, 2002.

1 1. RTAI Programming Guide1 .O, DIAPM, September 2000. 12. Val T., Juanole G., (( Developpement d’Applications de Mktrologie pour le

WPAN Bluetooth D, Colloque Francophone sur l’lngknierie des Protocoles CFIP‘2002, Montreal, 27-30 mai, 2002.

13. Val T., (( Contribution a I’ingknierie des systkmes de communication sans fil D, Habilitation a Diriger des Recherches, Universite de Toulouse 11, soutenue le 1 I juillet 2002.

SYSTEM SIMULATIONS OF DS-TRD AND TH-PPM FOR ULTRA-WIDE BAND (UWB) WIRELESS COMMUNICATIONS

SUSAN VASANA, PH.D. Electrical Engineering

University ofNorth Florida 4567 Sf. Johns BluffRoad, South Jacksonville, FL 32224-2645, US

KEVIN PHILLIPS Electrical Engineering

University ofNorth Florida 4567 St. Johns BluffRoad, South Jacksonville, FL 32224-2645, US

Following FCC approval of Ultra-Wide Band (UWB) radio rules for commercial applications in 2002, a UWB radio physical layer standard is currently under development within the IEEE 802 LAN/MAN Standard Committee. Although the UWB transmitter and UWB band are spelled out by the FCC, neither UWB signal forms nor modulation schemes are defined by the regulations [I]. This paper illustrates two types of impulse UWB radio modulation systems and their simulation performance. They are direct sequence UWB (DS-UWB) and time modulated (TM- UWB) or so called time hopping UWB (TH-UWB). Unique ways of coding and positioning the impulses, as in Transmitted-Reference Delay (TRD) modulation and Pulse-Position Modulation (PPM), can result in the simple UWB radio architectures. Systems of TRD for DS-UWB and PPM for TH-UWB are simulated under the same data and under the same AWGN channel conditions. TRD-DS-UWB system presents superior performance over PPM- TH-UWB system, which is shown in the comparison chart of the paper. The simulation was verified with ideal calculated performance for TRD modulation. The UWB applications are also discussed.

1. Introduction

Ultra-wideband (UWB) is a revolutionary wireless technology poised to find use in a broad range of consumer, enterprise, industrial and public safety applications [2]. First developed for military radar, UWB was authorized for commercial use by a ground breaking ruling of the US FCC in 2002. Unlike conventional wireless systems, which use narrowband modulated carrier waves to transmit information, DS-UWB transmits data by pulses generated at very high rates: in excess of 1 billion pulses per second [2].

An UWB signal is defined as a signal that has a bandwidth larger than 5OOMHz. It can coexist with other systems in the same frequency range due to

244

245

its large spreading factor and low power spectral density. UWB technology can be implemented in applications such as short-range high-speed data transmission and also precise location tracking. It can also be used in upcoming applications such as wireless office networks, and applications that require a fast data transfers. The primary commercial application is wireless multimedia personal area networks (PAN). These networks will connect consumer electronics (CE), PCs, and mobile communications devices in the home and office.

One main benefit in using UWB is shown that the frequency range can be shared with other types of systems. Also compared to narrow band systems, UWB uses significantly less RF power. The drawback is that UWB is susceptible to interference from other high power systems that operate over the UWB band that is being used. Hence, choosing a modulation scheme which gives good receiving sensitivity for the UWB signals is important.

UWB radio physical layer standard is currently under development within the IEEE 802 LAN/MAN Standard Committee in Task Group 802.15.3a. Although the UWB transmitter and UWB bandwidth restriction are defined, neither UWB signals nor modulation schemes are defined by the regulations [ 11. Thus two very good solutions have emerged from the standards work (DS-UWB and MB-OFDM) to vie for the standards. DS-UWB achieves its UWB bandwidth from the short impulses that comprise the DS “chips” while MB- OFDM achieves its UWB bandwidth form the aggregation of narrow band carriers. We concern ourselves here with UWB systems that achieve their UWB bandwidth from short impulses. This paper includes the simulation and performance comparison of two types of UWB modulation systems, i:e. Transmitted-Reference Delay (TRD) modulation for a variant of DS-UWB (not the system proposed in 802.15.3a) and Pulse -Position Modulation (PPM) for TH-UWB systems. The simulation is performed in Coware’s Signal Processing Workstation (SPW).

2. TRD Modulation For DS-UWB

Among various proposed UWB systems, direct sequence ultra-wideband (DS-UWB) takes maximum advantage of what UWB has to offer. By using the widest possible bandwidth to produce the shortest possible pulses, DS-UWB supports robust, high rate links in high multipath and offers precise spatial resolution for location detection. By generating continuous smooth white noise at the lowest levels relative to alternative approaches, DS-UWB is the preferred technology in terms of minimizing the potential to cause interference [2].

246

D spacing

An impulse radio (IR) is a system that implements UWB technology. An IR transmits very short pulses with low duty cycles as coded data. The transmitted information can be coded on a train of pulses in a variety of ways. Positions or polarities of the pulses can be coded on different levels [3].

Impulses can be sent with the information encoded differentially [l] and [4]. A method of transmitting and receiving impulses that can easily implement a rake receiver is Transmitted Reference Modulation (TRD). The method employs differentially encoded impulses sent at a precise spacing D (D is the data bit/chip interval). The data value of the pulse is referenced to the polarity of the previously sent pulses. The system is shown in the simplified block diagram of Figure 1. The transmitter sends pulses separated by a delay D that are differentially encoded using pulse polarity. The pulses, including propagation induced multipath replicas, are received and detected using a self-correlator with one input fed directly and another input delayed by D. Long sequences of differentially encoded pulses may be sent in the same manner. The receiver resembles a conventional DPSK receiver [ 11.

encoder generator

Differential Transmitter

"0 1 1 0" "1 1 0 1 1" u/

AID Filter and amplifier integate + 4

Antenna

Mixer Output data: "0 1 1 om\

Differential Reoeiver

Figure 1: A TRD-UWB transmitter and receiver

247

The DS-UWB system with TRD modulation scheme is simulated on the Coware SPW (Signal Processing Workstation). In the simulation system the DS-UWB baseband waveform with bipolar signaling is sampled at 16 samples per symbol. The reference pulse shape defined for the DS-UWB proposal is a root-raised cosine (RRC) pulse with 30% excess bandwidth [ 5 ] . The simulation performance is presented in section IV of this paper.

3. PPM For TH-UWB Or TM-UWB

TRD modulation in section 2 is one way to modulate the IR signal. Pulse position modulation (PPM) is considered as another modulation scheme. To prevent collisions among different users and provide strength against multiple access interference each information symbol is represented by a sequence of pulses. The positions of pulses in a sequence are determined by a random time hopping (TH) sequence specific to each user. This research paper defines an UWB signal and a simple receiver. An introduction of an impulse radio is made to explain a time - hopping pulse position modulation (TH-PPM) scheme. Results are shown gained from simulation software along with findings and conclusion.

In the UWB system, long sequences of pulsed are used with TH - PPM for communication. TH modulation causes a distribution of low power in the frequency domain. This causes the UWB signal to resemble noise in the frequency domain. The advantage causes the UWB to be less detectable, and resistant to jamming. Another advantage allows sharing of the frequency range the UWB signal operates in. The drawback is that the UWB signal is susceptible to interference from other in-band systems [6] and [7].

The transmitter is built to show how data can be transmitted using a PPM scheme. An impulse train is implemented and a delay is set to delay a pulse one half of a cycle or no delay at all. This only shows a transmitter in its simplest form. There could be four different delays in this sequence: one, two, three, and four quarter cycles. AWGN noise is added to resemble a channel for the signal to pass through.

2-ary PPM is simulated by using Signal Processing Workstation (SPW). A clock is added to show how the signal can be coded. When the clock strikes and it does not hear a signal until one cycle later that pulse can be coded as a zero. Otherwise when the clock strikes and the receiver will hear a pulse it can be recovered as a one. Although this is simplified it shows how PPM can be used to code information and transmit.

248

The signal will be sampled with the frame clock where the TH sequence will be implemented to distinguish between different users. When the user is identified the signal will finish the correlation process through integration and a decision that is made after correlation. The demodulated data is then recovered.

This receiver simulated only has one finger of a correlation receiver. More can be added to accumulate more power or as multiple users are added. This receiver accepts the signal after transmitting through an AWGN channel.

The filtered received signal is squared by a square-law device before integrator. The integrated signal clearly shows the PPM scheme that has been coded. Each position a pulse has been shifted the integration changes. The delta detection which is similar to the method illustrated in reference paper [ 5 ] and [6] is used here. The information coded in pulse position is converted to a level coded signal. After this waveform is retrieved it can be compare to a threshold to recover the original data.

4. System Performance Comparison

Two types of impulse UWB radio modulation systems are simulated. Both of them are unique ways of coding and positioning the impulses, results in the simple UWB radio architectures. Transmitted-Reference Delay (TRD) modulation system for DS-UWB and Pulse-Position Modulation (PPM) system for TH-UWB are simulated under the same data and sampling rates (1 6 samples per chip) and under the same AWGN channel conditions. The pulse shape is raised-cosine filtered. Both receivers use correlators implemented as integrator- and-dump. Ideal clock synchronization was assumed in the simulation. Self- correlator was used to demodulate TRD signals. The square-law and delta- detection method, such as in [8] and [9 ] , was used for PPM signal detection.

The simulation results are plotted as bit-error-rate (BER) vs. E D 0 (dB) in Figure 2. The plot shows that TRD-DS-UWB system presents superior performance over PPM- TH-UWB system, about 4 dB gain at 0.1% bit-error- rate (BER). The simple PPM receiver presented in this paper gives better performance when signal-to-noise ratio is low, but it does not perform well at the lower BER.

The ideal TRD performance from calculation in [l] is plotted with simulation performance of TRD in Figure 3. We can see how the simulation performance close to the ideal theoretic performance, where the pulse shape filter may degrade the performance a little. This comparison gives us the confidence in our simulation data.

249

Figure 2: Bit-Error-Rate Performance Comparison

Figure 3: Simulation and Ideal TRD Performance Comparison

5. Conclusion

Since the US Federal Communications Commissions (FCC) approved the use of ultra wide band (UWB) technology in 2002, communication systems that

250

use UWB signal have drawn considerable attention. An UWB signal is defined as a signal that has a bandwidth larger than 500MHz and an EIRP lower than - 41.3 dBm/MHz in the 3.1 to 10.6 GHz UWB band. UWB can coexist with other systems in the same frequency range due to its large spreading factor and low power spectral density. UWB technology can be implemented in applications such as short-range high-speed data transmission and also precise location tracking. It can also be used in upcoming applications such as wireless office networks, and applications that require a fast data rate. Also this type of system will need to operate in an area where there will be little interference from other systems preferable in a small radius where data can be transmitted over short distances in a multiple access environment.

Impulse radio is a form of UWB signaling for low power short-range communication systems. There are considerable findings from research of the DS-UWB and TH-PPM modulation schemes used by impulse UWB radios. A DS-UWB (but not differentially encoded like the system studied here) is the one of the most commonly used signal modes. That 802.15.3a proposal has further chosen to incorporated in the Common Signaling Mode (CSM). CSM, see [lo], is a signaling technique designed to allow different classes of devices (MB- OFDM and DS-UWB) to communicate with each other in order to coordinate their actions and interoperate within the same wireless network. The TH-PPM, also called TM-UWB (Time Modulated UWB), was implemented by Time Domain Corp. in their Radar-vision and P200 series of products.

This paper illustrates two types of impulse UWB radio modulation systems and their simulation performance. Unique ways of coding and positioning the impulses, as in Transmitted-Reference Delay (TRD) modulation and Pulse- Position Modulation (PPM), can result in the simple UWB radio architectures. Systems of TRD for DS-UWB and PPM for TH-UWB are simulated under the same data and sampling rates and under the same AWGN channel conditions. The paper shows that TRD-DS-UWB system presents superior performance over PPM- TH-UWB system, about 4 dE3 gain at 0.1% bit-error-rate (BER). Simulation is verified by comparing TRD simulation results with its theoretic performance.

References

1. Kazimierz Siwiak, Debra McKeown, Ultra- Wideband Radio Technology, Wiley, (2004).

251

2. The UWB Forum, (Online): http://www.uwbforum.org/aboutds/aboutds- uwb . asp

3. http://its.bldrdoc.gov/tpr/2000/its_e/time~hop/time~hop.htl 4. Susan Vasana, A Demodulation Method For Baseband M-Ary FSK Signals

Using Quadrant Crossings And Soft-Decision Detection, Wireless 2004 - The 16& International Conference On Wireless Communications, Calgary, Alberta, Canada, July (2004).

5. Paris Ultra-wide Band Systems, (Online): http://timederivative.con/pubs.html , June (2004).

6. DS-UWB For Wireless Networking, An Improved Direct Sequence Ultra- Wideband (DS-UWB) Proposal For IEEE, March (2004). (Online): http://www.uwbforum.org/standards/specifications.asp

7. Fernando Ramirez-Mireles, (Online): http://click.usc.edu/New~Site/paperslthesisCSI98_Ramirez.pdf

8. Alain SIBILLE, About the role of antennas in UWB impulse radio, United Electronic Information, Athens, Greece, January (2004).

9. Susan Vasana, High Sensitive Detection Method For Manchester Coded Data In Digital Communication Systems, The Florida Research Consortium (FRC) Technology Transfer Conference, St. Petersburg, May (2004).

10. Susan Vasana, Delta Detection Method For Manchester PCM Wireless Data Communication, IEEE EIT Symposium, Indianapolis, June (2003).

11. Y. Bahreini, et al., Requirements for a UWB Common Signaling Protocol: Coexistence of Multiple UWB Physical Layers, IEEE802 document 4 5 - 04-0079-03-003a-requirements-common-signalling-protocol.ppt~ 12 March (2004).

GLOBAL SOLUTION FOR THE SUPPORT OF QoS BY IEEE 802.11 WIRELESS LOCALNETWORKS

ADEL BEDOUI Laboratoire @scorn ENIT Tunis TUNISIE

KAMEL BARKAOUI Laboratoire CEDRIC C N N Paris FRANCE

KARIM DJOUANI Laboratoire LILA Uniwrsite' Paris 12 FRANCE

This paper presents a global solution for QoS improwment m 802.11 Wireless LANs. The proposed approach is based on Medium Access Control (MAC) enhancements, hking mto account informaturn from both the physical and the mtwork layers for packets differenciation scheduling at the MAC level

Keywords- Wireless Local AreaNetworks, accms method, MAC, QoS.

1. Introduction

Wireless LANs are being accepted widely and rapidly due to their simplicity and robustness against f il ure. The standard, adopted by IEEE in 1997 and improved in 1999 to give the version JEEE 802.1 1-1999[1], covers the physical layer specifications and the MAC sublayer responsible for medium control for the wireless local area networks. TheE are two types ofwireless local network architectures [2]: ad hoc architecture and architecture with access point AP (or infrastructure). Actually, 802.11 WLAN can support only best-effort service. Thus, researches continue in order to improve QoS support in 802.11 WLANs. According to 802.1 1 working group, enhancement in MAC protocol may ensure QoS support for the required applications. The principal role of MAC sublayer is to control the channel access in an equitable way. Two access methods are defined: The Distributed Function of Coordination (DCF), which is mandatory, and the Poling Coordination Function (PCF), which is optional. However PCF function, supposed to support real time services, is not implemented in the majority of the commercial 802.11 products. Moreover, the co-operation between PCF and DCF modes leads to performance degradation [3]. Besides, DCF can be used in ad hoc network as in infrastructure network, whereas PCF is used only in infrastructure network. So our choice is done on a distributed control for the channel access ensured by DCF function. The DCF is based on the listen-before-talk scheme according to the Carrier Sense Multiple Access (CSMA). Because collision may occur, than 802.11 define a Collision Avoidance (CA) mechanism to reduce the probability of such collisions.

252

253

RTS/CTS fi-ames are also used to solve the problem of hidden stations. Indeed, they can be detected by the receiving stations and not by the transmitting ones, from where the possibility of collision. To avoid this problem, RTS frame of small size is sent before each data frames transmission. In answer, the receiving station sends a CTS fiame ofsmall size to confirm being ready to receive. Thus, all stations are informed and update their NAV (network allocation vector). If the channel is free during a Distributed Inter Frame Space (DIFS) time, the station send packet and awaits an acknowledgement frame (ACK). If not, the sending is differed according to the backoff procedure as follows: as long as the channel is busy for DIFS time (after a successful reception) or for EIFS time (after a failed reception), Backoff time is decreased. This time is calculated in the following way: BackoflTime = BackoEounter * aSlotTime where BackoEounter is a uniform distribution integer in [0, CW] and CW is the contention window whose minimum and maximum bounds are defined by the standard. The CW value is increased in the case of non availability of the channel according to the following formula:

n t n + l CW(n)=(aCWmin+ 1)*2"- 1

Several approaches were developed in order to improve the MAC level according to the channel conditions indicated by the PHY layer. Barry et al. proposed in [S] a method based on modification of the maximum and minimum bounds of the contention window in order to be able to support two service classes, with high priority and "best effort". The emulation of the MAC sub layer (Virtual MAC) and application of virtual sources (VS) make possible observation and channel state evaluation by following real packets in parallel and in passive way. In [ 6 ] , the approach proposed by Pavon et al. is based on the level of the received signal (RSS) as a decision metric in order to adapt dynamically the transmission rate knowing that the emitted signal level is constant and that relation between RSS and SNR (Signal to Noise Ratio) is linear. In 171, Larnpe, e t al. it is proposed the prediction of Packet Error Rate (PER) as decision criterion for the connection adaptation. This prediction is done thanks to the SNR C/I (Carrier/Interference) ratio, and the temporary channel transfer function. In [8] Qiao et af. use combination of the SNR ratio, the test frames count and average load as metrics for the connection adaptation algorithm based on a preestablished table ofa better transmission rate for future transmission attempts. In [9] , Chevillat et al. took into account the acknowledgment obtained by observation to evaluate the channel quality. Thus, if the number of successive succeeded transmissions exceeds S , then rate increases. Otherwise, if the number of hiled successive transmissions exceeds F, then the rate decreases. In a certain way, the proposed approaches improve the performance of the 802.11 standard but only from a network charge point of view. However, the emergence of new network applications, such as VoIP and video streaming, are characterized by their real time nature which requires their data packets to be delivered within strict time bounds (upper limits on delay and jitter). In addition to such time bounds and regarding multimedia QoS targeb, any proposed approach for QoS improvement must deal with the performance metrics for throughput (physical and network layers) and packets loss.

254

2. Necessity of an integrated approach ofthe channelaccess problem

The majority of current researches concerning IEEE 802.1 1 Wireless Local Area Networks try to improve the performances of the MAC sublayer or the PHY layer without really taking into account their mutual interaction , the interaction of the MAC sublayer with the Network layer and of the PHY layer with environment (channel + wireless stations). Since the MAC sublayer is located between the Physical and Network layers, the improvement of its performances depends on information exchanged with the two indicated layers. In this paper, we propose an integrated approach for MAC QoS improvement based on integrating the information from the PHY layer and information from the Network layer. The following figure represents the interaction between the first 3 layers in order to improve MAC level output and ensure QoS up to PHY layer:

I 0 0 s EDCF Pakme ters 1 /

J 1 1 MW-Netwok

Figure 1. Control of QoS and interaction between layers.

2.1. PHY- MAC m era &on

2.1.1. SW and MVP Parameters

Our goal is to improve the MAC sublayer output by exploiting as well as possible information obtained from the PHY layer. Therefore, it is necessary to distinguish between the measured (or observed) parameters and those with values defined by the standard. The table 1 gives MAC and PHY parameters. The MVP-MAC can be measured according to the values given by the PHY layer. SVP-PHY are fixed by the standard according to the defined physical layer (DSSS, FHSS, IR, OFDM).

255

values defned by IEEE 802.11 standard SVP Parameters with measured vakes MVP

Table 1. MAC and PHY parameters

- DIFS,SIFS, EIFS,PIFS frames - aCWmaxand aCWmin - Maximum packet length - Bit ErmrRateBER - Average waiting - Signal to Nose RatD time to reach the SNR channel

- ACK games - Data frames

- Collision rate - Channeluse rate

1 PHY Level 1 MAC Level Parameters with I - aSlotTime I - RTS and CTS

2.1.2. MIMO fir the W F j Technology

Actually we know that the speeds claimed by WiFi hardware have little relation to actual throughput. Thus, 802.11b never delivers more than half its official 11 Mbitdsec data rate. Elsewhere, and even if 802.11a and 802.11g are little faster, the delivered data rate is far fiom the theoritical one. Hence, WiFi technology needs to improve its performance, both in terms of speed and range, in order to challenge Ethernet. We consider that any improvement of the actual technologies should cater with the physical and MAC layers. Actually, standard 802.1 In is under specification for developing a new physical layer for WiFi aimed at incerasing its maximum TCP/IP throughput to 100 Mbits/sec or more. The MIMO approach (Multiple Input Multiple Output), based on the use of several antennas at the emission end and at the reception end instead ofonly one at each end, appears as a promising solution to improve channel rate and to increase robustness against interference problems. Elsewhere, the MIMO is likely to be 802.1 In's foundation. Because MIMO will improve performance, even when it is used at just one end of a link. An AP that incorporates MIMO can't double the speed of a standard Wi-Fi client, but it can extend the range at which 54Mbitshec is available. The information exchanged between a transmitter and a receiver is classically related to time. In the case of a SISO system (Single Input Single Output), this quantity is calculated between emission and reception signals. For MIMO systems, this information quantity is not only related to the temporal intervals but also to the space volumes. The use ofantenna networks simultaneously in a WLAN emission and reception made it possible, for the first time in the history of radio communications, to exploit space dimension as well as temporal dimension. The construction of the coding and modulation functions generalizes the diagrams of traditional transmission at the space field. The two fundamental mechanisms introduced are the diversity and the multiplexing, respectively associated to the robustness and the output [lo]. The improvement made by MIMO is justified by the reasoning not only on the type of coding/decoding which makes it possible to decrease Packets Error Rate and consequently to increase the rate [4] but also on data packet propagation time given that times occupied by Inter Frame Spaces IFS are very small

256

compared to those occupied by data packets. In fact, the rate D depends on data packet lengths and occupied times:

D = WTt = LJ(TDIFS + TRTS + Tsm + Tcrs + T~ata + TACK) Also, rate can be expressed according to channel bandwidth and SNR:

D = W*log(l + SNR) , with W : channel bandwidth

Our interest relates not only on PHY layerbut also on information it can bring to MAC sublayer since it is the interface with the external environment made up of channel and other stations. Therefore, it can reflect the environment changes and thus allows MAC sublayer to adapt to and offers better performances.

2.1.3. SNRparameter

The SNR parameter is measured on the physical layer (MIMO architecture level). This parameter concerns the usehl rate (2.1.B) computing. In addition, the second measured parameter (table 1) is the Bit Error Rate (BER) which indicates the state of channel and its quality. It is given by the following formula :

BER= Number of erroneous received bits Total number of emitted bits

Although this parameter is measured on the PHY layer level, its value is calculated only at the receiving station. Then, the transmitting station is considered “blind” in meaning of BER and it will be informed of the channel quality only if it will becomes receiving station. Thus, our choice is made on SNR parameter:

SNR = Emittedhe ceived signal power Noise power

Assume that:

(SNR)BE: acceptable limit of SNR for the “Best Effort” traffic; (SNR)@s: acceptable limit of SNR for the “QoS” traffic. The following condition must be observed:

1) If(SNR) < (SNR)BB then the station differs emission (MAC); 2) If (SNR)BE 5 (S/N) < (SNR)Q,s, then the station differs emission i fwe

have QoS traffic, otherwise it emits the Best Effort traffic ; 3) If(SNR) 2 (SNR)qos, then station transmits in both cases ofBest effort

or QoS traffic. This approach clearly shows the narrow dependence between decision to emit or differ the emission on MAC level, SNR ratio on PHY level and traffic classes on Network level.

2.2. Network - MACInteracthn The other part of our integrated approach refers to interaction between MAC level and Network layer. Multimedia applications are supported by IP and high layers. The purpose of Difierv architecture of IF’ layer is to ensure service

257

differentiation in TCP/IP networks. Packets are classified and marked to be treated in a particular way by various network nodes (routers for example). This classification is not known by traditional MAC sublayer since it supposes that traffic flows are not different. To ensure a quality of service at MAC level, we propose to jointly exploit information obtained from PHY and Network layers to deliver packets following their priorities according to desired QoS.

2.2.1. Network Packets - MAC Frames

In order to ensure point-to-point QoS based on Dimerv, frames in WLAN must be differentiated according to priority classes indicated by high layers. When a data frame MSDU (MAC Service Dated Unit) arrives at MAC sublayer, it is encapsulated in a MPDU (MAC Protocol Data Unit) by addition of a "MAC Heading" field and of a "Frame Control Sequence" field [6]. The IEEE 802.1 l e future standard supporting QoS on the MAC level introduces a new function HCF (Hybrid Coordination Function) which defines two medium access mechanisms [ll]: Contention access and Controlled access. The first mechanism refers to EDCF (Enhanced Distributed Coordination Function) which provides distributed and differentiated accesses for wireless channel to users with priority during contention periods. The controlled access is based on election principle with QoS support for free contention periods. It refers to CHCF (Controlled HCF). Our work relates to distributed approaches. Consequently, we are interested in the following to EDCF function. The latter is based on differentiated priorities where the traffic must be delivered according to four Access Categories (ACs) representing virtual DCFs [ 121. Access categories are obtained by differentiation of Arbitrary Inter Frame Spaces (AIFS(i)) and of sizes of minimum (CWmin(i)) and maximum (CWmax(i)) Contention Windows. Thus, for priority category access traffic, access to channel is hster. EDCF function supports eight priorities corresponding to four access categories indicated in the following table:

Tabk 2. Prbrities and access categories correspondewes

2.2.2. RTS/ClSSames

EDCF function ensures QoS requested by a given station. However, the network is composed of a set of stations which can support various types of applications.

258

Bit15 0 1

1 1 1 1

Other stations must be informed of a given station priority flow in order to differ their access requests iftheir flow does not have priority. To solve this problem, we propose that RTS/CTS frames indicate this information. Indeed, being used by MAC level to reserve channel for the required period to data exchange, these control frames arrive at all network stations. They have the following formats [2]:

Octets: 2 2 6 6 4

I FrameControl 1 Duration 1 RA 1 TA [ FCS I Octets: 2 2 6 4

Figure 3 . RTS fiame brmat

B t 1 4 I Bits134 Usage (03.327 67) Duration

0 0 Fixed value within frames

0 1-1 6383 Reserved 1 0 Reserved 1 1-2007 AID in PS-Poll kames 1 2008-16383 Reserved

transmitted during the CFP

[ FrameControl 1 Duration 1 R A 1 FCS I Figure 4. CTS fiame format

The "Duration" field indicates time, in microseconds, needed to transmit the waiting data or management frames, plus a CTS frame, plus an ACK frame, plus three SIFS intervals. This field is of 16 bits (2 octets). Its content (the coding of this field) depending on the type of fiame is given in table 3 [2]. Note that for frames transmitted during contention fiee period (CFP), the field is put at 32768. Each time the content of this field is lower than 32768, duration value is used to update Network Allocation Vector (NAV).

Tabb 3. Contentof 'Duration fiekl"

We propose to use "Duration" field to convey information "Priority" from physical level to MAC level. Our method is based on the idea to subtract or not 1 of the field value to distinguish priority packets fiom non priority ones according to whether the values are even or odd. In another manner, there will be two cases: 1. Priority packets case: 1.a. If the value of "Duration" field is even, "1" is subtracted of this value to obtain an odd number. 1 .b. If the value of "Duration" is odd, this value is kept.

259

2. Non priority packets case: 2.a. If the value of"Duration" field is even, this value is kept. 2.b. If the value of "Duration" is odd, "1" is subtracted of this value to obtain an even number. So, the packet has priority if and only if its "Duration" field value is odd. Thus, "Priority" is memorised (or hidden) in "Duration". In this method, "Duration" field will have the value in p: Tpm = (Tinitial - 1) in 2 cases out of 4.

With T(D, G ) : time in ps of data or management frame. Initial time is always much higher than the sum of CTS, ACK and 3*SIFS times. The subtraction of 1 ps doesn't have notable effect on "Duration" information contained in RTS/CTS frames while it makes this field carrying two information "Duration" and "Priority". Our proposal has double advantage: since RTS/CTS fiames belong to MAC sublayer, it is possible to integrate through them information resulting from Network layer. In the other side, these frames will circulate in wireless network containing information on priority of the data fiame to be sent Thus, stations listening to channel can decide to send or delay their fkames by comparing their priority level to that of the RTS/CTS (which, in fact, indicates the priority of the following data frames). Moreover, the hidden stations, which cannot realize that channel is occupied, will be informed not only that channel is busy but also about priority of packets which will be exchanged.

T. . . - T mlhal- C T S + ~ A C K + 3*TSFS+ Tip,G)'TCI'S+ TACK + 3*TSFS

2.2.3. Synthesis

Thus, our integrated approach is defined by the following steps:

1. 2.

Delivering, by Network layer, packets with priority to MAC sublayer; a. Encapsulation of these packets in fiames according to IEEE 802.11e future Standard (EDCF function); b. Scheduling flow in a differentiated way in accordance with traffic type; c. Measuring metrics on the MAC level; d. Measuring SNR pammeter at physical layer and informing MAC level; Delivering fiames to PHY layer according to SNR required conditions indicated in 2.1.3.; Exchanging RTS/CTS fkames indicating the priority if the station reaches the channel; a. Then sending data frames according to priorities indicated by MAC level; b. Differ sending until reaching the channel if it is busy.

3.

4.

5 .

3. Conclusion A new approach for MAC QoS enhancement for WLANs has been presented. The proposed approach is based on information and metrics delivred by both physical and network layers for frames scheduling at MAC layer.

260

In order to ensure QoS for multimedia applications with real time constraints, dynamic adaptation to the channel characteristics is proposed. Our approach is based on both 802.11 n for the hture physical layer, integrating the MIMO paradigm, and 802.11e for QoS enhancement at MAC level, incorporting services differentiation. Therefore, we propose to encode packets priorities within the RTS/CTS frames. Actually, the proposed‘approach is under implementation on network simulator.

References 1. IEEE Standard for Information technology - Telecommunications and

infbrmation exchange between systems - Local and metropolitan area networks- Specific requirements.

Part 1l:Wreless LAN Medium Access Control (ME) and Physical Layer (PHV Specifications. Adopted by the ISOhEC and redesignated as

P. Miihlethaler, 802.1 1 et les rbeaux sans fil, Eyrolles edition, ISBN 2-21 2-

M. A. Visser, M. E. Zarki, Voice and Data Transmksion over an 802.11 Weless Networks, PIMRC proceeding, Toronto Canada (1995).

M. Barry, A. T. Campell, A. Veres, Distributed Control Algorithms for Service Difirentiation in Wreless Packet Netwrks, IEEE INFOCOM proceeding, Alaska (200 1 ).

6. J. P. Pavon, S . Choi, LinkAdaptation Strategy fir IEEE 802.11 UZAN via Received SigalStrength Measurement, IEEE ICC’03, vol. 2 (2003).

7. M. Lampe, H. Rohling and J. Eichinger, PER-Prediction for Link Adaptation in OFDM Systems, OFDM Workshop, Germany (2002).

8. D. Qiao, S. Choi, K. G. Shin, Goodput Analysis and Link Adaptation for IEEE 802.1 l a Wreless LAN.., IEEE Trans. Mobile Comp., vol. 1 (2002).

9. P. Chevillat, J. Jelitto, A. N. Barreto, H. L. Truong, A dynamic link adaptation algorithm fir IEEE 802.1 l a wireless L N s , IEEE ICC’03, Vol.

10. P. Guguen, G. El Zein, Les techniques multi-antennespour les rkseaux sans Sl, Hermes-Science edition, ISBN 2-74624883-0, Paris (2004).

11. S. Mangold, S. Choi, G. Hierb, 0. Klein, B. Walk, Analysis of IEEE 802.Ile fir QoS support m wireless LANs, IEEE Wireless Communications (2003).

12. Y. Xiao, IEEE 802.1le: QoS Provisioning at the MAC Layer, IEEE Wireless Communications (2004).

ISOhEC 8802-11:1999(E) (1999).

2.

3.

4. wwW.c0msis.fr 5 .

11 154-1 (2002).

2 , pp 1141-1 145 (2003).

Traffic

CROSS-LAYER DESIGN FOR DYNAMIC RESOURCE ALLOCATION IN WIRELESS NETWORKS

JOHN Y. KIM+, ALI SAIDI, AND RANDALL J. LANDRY The MITRE Corporation, 202 Burlington Road

Bedford, MA 01 730-1420

In this paper, a novel analytical cross-layer design framework for dynamic resource management of wireless networks is proposed. First, dynamic bandwidth and time resource allocation policies for a single-user under fading channels that maximize capacity are derived. The analysis is then extended to multi-user environments, where the resource allocation is jointly optimized across both physical and data link layers. A closed-form expression of a QoS measure, mean delay in this case, is derived as a function of layer 2 traffic, multiple access contention from other users, and allocated data-rates at the physical layer. This mean delay expression is then used to efficiently allocate physical layer resources. We also study the effects of various contention mitigation policies on network capacity and average latency under optimum resource allocation strategies.

1. Introduction

Future communication systems will be characterized by high data-rates, diversity in the type of transmitted data, and a rapid evolution towards increasingly network-centric architectures. This places a growing demand on the efficient utilization of system resources including transmitter power, time, and assigned channel bandwidth. Conventional fixed-mode communication systems are typically designed for worst-case channel conditions to combat the effects of deep fading. Adaptive communication systems, on the other hand, provide an important alternative that can achieve considerably higher efficiency compared to non-adaptive systems. The adaptation can be performed by selective variation of the parameters in the physical layer as well as in the higher layers of the protocol stack.

The advantages of adaptive communication systems for operation in widely varying wireless channel conditions are substantial. However, finding an optimum adaptation strategy for a given set of constraints on network resources is not a trivial problem. Maximum efficiency can only be achieved

+J. Kim is now with Nextel Communications, Inc., Reston, VA 20191.

263

264

when protocol layers collaboratively respond to changing network state by dynamically allocating resources subject to the network constraints [ 11. The complexity of this joint optimization problem through cross-layer design (CLD) [2] grows very fast as the number of layers and the parameters in each layer considered for optimization increases. Therefore, it is critically important to consider the feasibility of solving the cross-layer joint optimization problem when selecting variables for adaptation [2-41. Recent work in the area of CLD for wireless networks has focused on rather “ad-hoc” approaches in which state information at one layer is used by higher layer protocols to improve network performance relative to a strictly layered methodology [ 5 ] . The majority of these approaches do not promote a fundamental analytical treatment of the CLD problem.

In this paper we consider a CLD for the allocation of physical layer resources which jointly optimizes across the physical and data-link layers. The proposed analytical approach focuses on the efficient allocation of bandwidth and time, the two primary shared resources in multiple access systems. For Dynamic Bandwidth Allocation (DBA), which is based on Frequency Division Multiple Access (FDMA), multiple users can access the channel simultaneously. The amount of disjoint bandwidth resource being occupied by each user dynamically changes based on its channel condition subject to the total bandwidth constraint as well as each users average bandwidth constraint. In the Time Division Multiple Access (TDMA)-based Dynamic Time Allocation (DTA), on the other rather hand, one user occupies the entire system bandwidth for a dynamically allocated period of time based on channel conditions and subject to each user’s average time constraint. One practical realization of these resource allocation schemes could be based on a multi- carrier system whose bandwidth (carriers) and time resources can easily be managed [ 121.

Our approach to CLD combines information theoretic techniques with queuing theoretic techniques to develop an analytical framework for joint optimization of resource utilization across multiple layers. Physical layer constraints such as average bandwidth and time in our approach are assumed to be variable in deriving the optimum policies at the physical and data link layers separately. The actual values of the average bandwidth and time allocations, as the common variable in both layers, are then computed based on the total resource constraints, the network traffic parameters, and the multiple access contention from other users, to satisfy a soft quality of service (QoS) measure such as average delay.

We begin by deriving the optimal bandwidth and time allocation solutions for the single user case. In [6], the authors compute the optimal power

265

allocation policy under fading channels which maximizes the capacity for a given average power constraint. We employ a similar approach for bandwidth and time allocations, assuming a fixed transmit power level, and obtain the optimal solutions as a function of average bandwidth and time constraints. The analysis is then extended to the multi-user case where contention and contention resolution polices are discussed. A priority queuing system is employed to model the impact of multiple access contention on traffic arrivals for a single user. An expression is derived for mean delay as a function of user traffic, multiple access contention, and average data-rate allocated to the user. Delay constraints are then used to compute the optimal allocation of data-rate and, hence, time or bandwidth resources.

The remainder of the paper is organized as follows. In Section 2, our system model is described and the optimal DBA and DTA solutions are derived. Various contention policies and corresponding capacities are discussed in Section 3. Our CLD approach and delay derivation are presented in Section 4. Finally, the paper is concluded with analysis of the results and some final remarks in Sections 5 and 6, respectively.

2. Optimal Capacity Analysis

In this section, the optimal capacity analysis of DBA and DTA for the single user case is presented. It is assumed that the transmit power level is fixed and the channel encounters flat fading with unit average gain. It is also assumed that the channel information is hown to both transmitter and receiver.

2.1. Dynamic Bandwidth Ajlocation (DBA)

Let B ( y ) be the instantaneous signal bandwidth which is a function of signal- to-noise ratio (SNR) given by y . Since the transmit power level is assumed to be fixed, the SNR distribution is dictated by the channel fading distribution. The optimal solution for B ( y ) in DBA can then be found by maximizing the

average capacity, C,,, , as follows:

subject to the average bandwidth constraint m

266

where P( y ) is the pdf of SNR. This optimization leads to the following Euler- Lagrange equation:

where A is the Lagrange multiplier. Expanding and simplifying the above equation leads to

Letting X = B (y) /g , the above equation simplifies to

1 log, I+- --=A. [ :I x+1 Y

It follows from this expression that since A is a constant, the variable X / y must also be a constant, which suggests that B ( y ) is directly proportional to y . Suppose B( y ) = 67. Then we can solve for E from the constraint equation (2): - . .

B & =-.

[YI

Therefore, the optimal bandwidth allocation that maximizes the average capacity is B y / E [ y ] . Substituting B ( y ) into (1) we get the maximum average capacity

CD,, =Blog,p+E[Y]] 3 (7)

which is the AWGN Shannon capacity for a given average SNR. The optimum DBA policy allocates more bandwidth when the channel fading is less severe and can be viewed as a bandwidth water-filling technique [7] in time.

Two interesting observations are noteworthy. First, the average capacity is independent of channel fading conditions when the optimal BW allocation policy is employed. Second, the spectral efficiency stays constant for a given SNR distribution implying that no adaptive modulation scheme is required for DBA.

The above derivation assumes that the SNR measurements are made with respect to the average bandwidth constraint, B . In order to make relative

267

comparison, the derivation needs to include a reference. Let B,, be the reference bandwidth from which the SNR distribution data is obtained. Then, the instantaneous SNR under instantaneous bandwidth B ( y ) is

y B,,, / B ( y ) . Therefore,

C,,, = Blog, l+E[y]- - [ Bifl

The above equation tells us that as the average bandwidth allocation increases the capacity also increases, while the spectral efficiency diminishes. In practice, the allocated bandwidth will be upper-bounded by the total system bandwidth, B,, . Then, the optimal instantaneous bandwidth allocation

becomes

7 (9) E1/, Y l Y m B(y)={ ‘I’m, Y’Yrnax

where y- = B,, /& , and the average bandwidth constraint is expressed as Y- m

j v P ( y ) d y + J &ym,P(y)dY=B * (10) O Y-

The optimal DBA capacity in the presence of a bounded system bandwidth is then given by

Note that the above bounded result does not deviate much from the << B,, (as it normally would be in the practical unbounded case when

multi-user cases).

2.2. Dynamic Time Allocation (DTA)

In DTA, the bandwidth resource is assumed to be fixed, and the only managedallocated resource is time. Let the time allocation policy, T ( y ) , be an indicator function such that

(12) 1, y E { tx region} 0, otherwise.

Then, the optimal T ( y ) is obtained by solving the following constraint optimization problem:

268

subject to the average time constraint m

p ( Y ) w ) d Y = T 0

Since the instantaneous capacity is an increasing function of y , the average

capacity is maximized if y E { tx region} takes on SNR values that exceed a

threshold, ytx , determined by: m

[P(Y)dY = T YU

The resulting optimal average capacity is simply m

The average time constraint can be directly translated into the average power constraint. Therefore, the optimal DTA policy can be interpreted as a time domain variant of constant power water-tilling proposed in [13]. The optimal DTA performance, unlike DBA, depends on channel fading conditions, and its spectral efficiency changes as a h c t i o n of SNR. This implies that achieving maximum average capacity in the DTA scheme requires adaptive modulation. Again, it is noted that the above analysis is based on the assumption that the S N R is measured with respect to the total system bandwidth and needs to be adjusted when making relative comparisons.

3. Effective Capacity

We have derived the optimal capacity solutions for DBA and DTA given average resource constraints for the single user case. In this section, we study multiple-access systems and the impact of multi-user contention on the performance compared to the optimal single user capacities. Under fixed resource allocations, there is no contention as long as the sum of individual allocations does not exceed the total system resource pool. However, when there are two or more users in the DBA/DTA based systems competing for the same resource could lead to contention. Hence, the actual delivered capacity,

269

or effective capacity, is a h c t i o n of this contention probability and the employed contention resolution policy.

3.1. Contention Resolution for DBA

Let N be the number of users in the system sharing the common resource pool with independent SNRs. In DBA, a contention arises when the sum of individual instantaneous bandwidth allocations exceeds the total system bandwidth. Let PDBA be the contention probability for a DBA based multiuser system. Then,

where yi and 8 denote the instantaneous SNR and average bandwidth constraint for user i , respectively. Given that Bi has the same distribution as yi , the distribution of the sum of Bi over all i can be characterized from the pdf of yi [8][9]. For example, assuming each yi is Rayleigh distributed, the above contention probability can be calculated as follows [8]:

m

PDBA= J P ( S ~ S L 7 .

where P (s) is the pdf I of the sum of the N random variables Bi .The effective capacity for user i, Ci , is a function of contention probability and the contention mitigation policy

C;. =E[c, (yi)lno contention](~-~~,)

+ E[C: ( yi,policy)lcontention]/3,,BA (19)

= E; + e;. Ci ( yi ) is the instantaneous capacity resulting from the optimal DBA policy specified by (6), whereas Ci ( y i , policy) is the instantaneous capacity based on the specific contention resolution policy being employed.

270

The choice of the contention mitigation policy greatly impacts the system performance. Consider the following two different choices of policies and their effects on capacity:

i= l

- B.

C; ( yi , policy 2) = + B,,, log, [I + 7;] c 6 i=l

where ri is the corresponding SNR resulting from altered bandwidth allocation policies. Policy 1 implements a discipline where the allocation during contention is based on optimal instantaneous allocations, whereas Policy 2 divides the bandwidth based on average target bandwidth requirements, analogous to fixed allocation. The effective capacity can be obtained by solving

N

where s ‘ = ~ B j ( y 1 ) , @ , = s ’ E [ y , ] / q and 5, = B m a x E [ y I ] / q . The

contention policy enforcement choice along with assigning resource constraints can be used to satisfy specific QoS needs of individual users in the system. These two additional degrees of freedom in our proposed resource management approach offer a flexible means to control user QoS outcomes under various channel conditions. For example, it is known that the optimal “sum of rate” capacity for FDMA is achieved when each user’s bandwidth allocation is proportional to its instantaneous SNR [ 101. However, the resulting policy does not always adhere to the resource constraints and favors users with “good” channels over others.

1‘1

3.2. Contention Resolution for DTA

In DTA, a contention arises when more than one user is allowed to transmit at a given instance. Unlike DBA, the users can have different contention

271

probabilities depending on assignments. For example, the probability of contention for user i, pi , is

pi = 1 - P [ no contention] N

= 1 -n(l- T ) , j t i

where capacity for user i is defined in (19). For DTA, ci, is simply

is the time constraint for user i. Similar to DBA, the effective

ei! = [’ - pi ] CDTA-optimal ’ (25)

The value of cf again depends on the chosen contention-resolution policy. Assuming a policy based on randomly selecting a user to transmit, the capacity during contention period is given by

N-1 i

* (26) ‘DT A-optimal e,? = xP[contention due to j other users] x j=1 j + l

The policy that maximizes the aggregate capacity for TDMA, however, is to have the user with the highest SNR transmit during contention [ 1 I].

3.3. Residual Resource Utilization

Up until this point, we have ignored the residual bandwidthhime resources during non-contention periods given by:

These residual resources can be utilized to make up for the capacity loss due to contention and/or to service the “best-effort’’ traffic. Therefore, additional residual resource utilization policies can be employed in order to distribute the excess resources among active users. The resulting generalized effective capacity can then be described as - -

(28) ei = E [ Ci (y,,policyR)lno contention]( 1 - p)

+ E[C, (y,,policy~)lcontention]~,

where policyR and policyc are residual resource utilization and contention resolution policies, respectively. Finally, it is aIso possible to employ the same

272

policy for both non-contention and contention periods, in which case the effective capacity simply becomes

E; = E [ C; ( yi, policy)] . (29)

4. Delay Analysis of Cross-Layer Design

In this section, we present an overview of the cross-layer aspects of our work, which demands that we consider the impact of network traffic and the QoS delivered to user i in the presence of multiple access contention. We adopt a two-priority WG/1 queuing model with a Head-of-Line (HoL) service discipline for one of the traffic classes. Let Ai denote the mean arrival rate of the low-priority customers arriving to the queue. These arrivals represent the actual data awaiting transmission from user i . Let Ac denote the arrival rate of contention jobs to the queue. These arrivals are intended to model the impact on local (user i) data from physical layer resources, either bandwidth or time, allocated to other users in the system. We define the service time, S, to be the service time of a single bit and allow it to take a general distribution. The service rate, R, is the inverse of service time and is assumed to remain constant over the service period of a bit. The mean delay, Di , for user i arrivals in the adopted WG/1 priority queue can be found (see [I41 for instance) in terms of the arrival rates and the first two moments of service time, S. By considering the service rate assumptions mentioned above, both the first and second moments for S can then be computed in terms of the average service rate for the queue, denoted b y R . The mean delay expression for user i arrivals then becomes

We can substitute Ci for since it is the average capacity (data rate)

assigned to user i in the absence of contention. Since Ac represents the capacity loss due to contention, it is simply -

A, = c; - c; .

Substituting for Ri and A, in (30) yields the following expression:

273

Our CLD approach is captured in the expression for mean delay given in Equation (32), which incorporates channel statistics, physical resource allocation policies and traffic statistics. Note that the system remains stable in the steady-state, with a bounded mean delay, as long as the average user traffic arrival rate is less than the effective capacity allocated to that user, which is consistent with the queuing theoretic requirement that the ratio of mean arrival rate to mean service rate be less than unity.

Given specific average arrival rates and mean delay requirements, the relationship in Equation (32) can be used to arrive at average data-rate (or capacity) allocations, Ci, to users under specific contention resolution and residual resource allocation policies. These policies, through Equation (28), determine the effective capacity allocated to users. The average capacity allocation per user is then translated into the appropriate physical resource (time or bandwidth) allocation using the analysis in section 2.

5. Results

Figures 1 and 2 illustrate the average capacity performances of the proposed dynamic bandwidth and time allocations compared to static resource allocation schemes. Fixed bandwidth allocation (FBA) assigns static bandwidth while fixed time allocation (FTA) assigns a fixed portion of each frame, given by r , to every user for transmission. Since in both cases resource allocation is independent of channel condition, the average capacities for FBA and FTA are

Both dynamic bandwidth and time allocation schemes clearly provide better average capacity performance compared to their fixed allocation counterparts. However, as it is shown in both figures the capacity gain over the fixed allocation schemes diminishes as the channel variance decreases. As our analysis has shown, the optimal DBA capacity, in contrast to DTA, is independent of the channel fading condition. In order to make a fair comparison between the two schemes, B , B,, and T are chosen such that B / BmX = T . Although this normalization ensures that the average resource (time x bandwidth) allocation is the same for both schemes, DBA still outperforms DTA regardless of the channel condition. This is due to the fact

-

GIVEN BY THE sHANNON'S CAPACITY FOR FADING CHANNELS:

274

E[YI (dB)

Figure 1. Average spectral efficiency comparison between dynamic and fixed bandwidth allocations

Figure 2. Average spectral efficiency comparison between dynamic and fixed time allocations; T=O.l

that both schemes are assumed to use the same constant maximum power when transmitting. The discrepancy in performance stems from employing continuous transmission in DBA vs. on-and-off transmission in DTA. That is, DBA, on average, consumes 1/T times more power than DTA.

The optimal resource allocation schemes for a single user maximize the channel capacity without taking into consideration the other users that are

275

competing for the same resources. Figures 3 and 4 demonstrate the impact of multiple-access contention on average DBA and DTA capacities. It is assumed that all users have the same target resource requirement and SNR distribution and the residual capacities are ignored in both cases. Rayleigh fading with E b ] = 30 dE3 (with respect toB) is used and BIB,,,, is set to be equal to T = 0.05 . The capacity performance in multiple-access systems is a function of contention resolution policies, which is a tradeoff between complexity and efficiency. We consider several contention resolution policies that are summarized in Table 1. Policy A, which prevents all users from transmission when there is a contention, produces the lower bounds on the average performance of our DBMDTA schemes; whereas, the optimal single user capacity curves correspond to the upper bounds.

Table 1 . Contention resolution policies used in Figures 3 and 4

DBA DTA Policy A No transmission No transmission Policy B Allocation based on target requirements Random allocation Policy C Allocation based on both target and SNR Highest SNR transmits

Although multiple-access performance always lies between these two bounds, there exist policies (B and C) that generally outperform the fixed allocation schemes. The figures show that the average capacity performance suffers as the number of users (i.e., contention probability) increases except for fixed allocation schemes where there is no contention. Finally, contention resolution policies are used to mitigate the resource shortage when allocation exceeds availability. However, when allocated resources fall short of the available resources, utilizing the residual resources can further improve capacity performance.

5.1. Cross-Layer Design (CLD)

To demonstrate our CLD approach, presented in Section 4, we use a multiple-access system of ten users. Users 1 through 6 have less stringent delay requirements than users 7 through 10, which would be the case, for instance, if the former were predominantly data users and the latter were voice users. It is assumed that all ten users have the same channel (Rayleigh) and traffic arrival distributions and optimal allocation-based contention resolution policies are employed. Figure 5 illustrates how the proposed cross-layer resource management techniques can be used to satisfy different user QoS requirements by jointly utilizing data and channel statistics. The average

276

0 55

h 05 - r

2 4 6 8 10 12 14 16 18 20 # o f users (N)

Figure 3. Average capacity performance companson among vanous DBA contention policies

- - fixed allocation - DTA Policy A DTA Policy B

-

0.1 I I 2 4 6 8 10 12 14 16 18 20

#of users (N)

Figure 4. Average capacity performance comparison among various DTA contention policies

resource constraint for each user is generated by the network layer, and the physical layer dynamically adjusts the user’s instantaneous resource allocation based on the derived optimal rules. Without cross-layer exchange, if resource allocation is solely based on channel statistics (as in [l l]), all users experience the same average delay in both cases.

277

l 0.055

0.03 * * . .... .*..

0.045

rn

-t- proposed cross-layer RM 0.02

__ .-- .

0.0151 I I 1 2 3 4 5 6 7 8 9 10

user index

Figure 5 . DBA user delay requirement adaptation via cross-layer implementation

6. Conclusions and Future Studies

In this paper, we have proposed a novel CLD approach based on the fundamental relationship between physical layer resource and network performance. In our proposed scheme, the QoS-compliant average resource constraint for each user is determined by the date-link layer taking channel and traffic statistics into account. The physical layer then dynamically adjusts the user's instantaneous resource assignment based on the derived optimal allocation rule. Our CLD-based resource allocation approach along with flexible contention resolution policies allows for a wide range of QoS management plans. Our future studies include extending our approach to cover wider physical channel environments and additional network performance measures. Additionally, we are studying the application of our DBADTA framework to frequency-selective channels. One approach may be to divide the channel into several frequency flat sub-channels and apply DBADTA schemes on an individual sub-channel basis. We are also working to develop analytical models to incorporate additional QoS measures such as outage probability and delay jitter.

278

References

1. I. E. Telatar and R. G. Gallager, “Combing queuing theory with information theory for multi-access,” IEEE JSAC, Vol. 13, No. 6,

2. T. ElBatt and A. Ephremides, “Joint scheduling and power control for wireless ad hoc networks,” Wireless communications, IEEE Transactions on, Vol. 3, No. 1, pp. 74-85, Jan. 2004

3. A. Maharshi, T. Lang and A. Swami, “Cross-layer designs of multichannel reservation MAC under Rayleigh fading,” Signal Processing, IEEE Transactions on, Vol. 51, No. 8, pp. 2054-2067, Aug. 2003

4. S. Toumpis and A. J. Goldsmith, “Performance, optimization, and cross- layer design of media access protocols for wireless ad hoc networks,” ICC ‘03, pp. 2234-2240, May 2003

5 . S. Shakkottai and T. Rappaport, “Cross-Layer Design for Wireless Networks”, IEEE Comm. Mag., pp. 74-80, Oct. 2004

6. A. J. Goldsmith and S.-G. Chua, “Variable-rate variable-power MQAM for fading channels,” IEEE Trans. on Communications, Vol. 45, No. 10,

7. R. G. Gallager, Information theory and reliable communication, Wiley & Sons, New York, NY, 1968

8. Y.-D. Yao and A. Sheikh, “Investigations into cochannel interference in microcellular mobile radio systems,” IEEE Trans. on Vehicular Technology, Vol. 41, No. 2, pp. 114-123, May 1992

9. A. Abu-Dayya and N. C. Beaulieu, “Outage probabilities of cellular mobile radio systems with multiple Nakagami interferers,” IEEE Trans. on Vehicular Technology, Vol. 40, No. 4, pp. 757-768, Nov. 1991

10. W. Yu and J. M. Cioffi, “FDMA capacity of Gaussian multiple-access channel with ISI,” IEEE Trans. on Communications, Vol. 50, No. 1, pp. 102-1 1 1, Jan. 2002

11. R. Knopp and P. A. Humblet, “Information capacity and power control in single-cell multiuser communications,” Proc. IEEE ICC’9.5, Seattle, Wash., 1995

12. D. Kivanc, G. Li and H. Liu, “Computationally efficient bandwidth allocation and power control for OFDMA,” IEEE Trans. on Wireless Communications, Vol. 2, No. 6, pp. 1150-1 158, Nov. 2003

13. W. Yu and J. M. Cioffi, “On constant power water-filling,” Proc. IEEE ICC’2001, June 200 1

14. L. Kleinrock, Queueing Systems: Volume 11: Computer Applications, John Wiley and Sons, 1976

15. A. Safwati, H. Hassanein and H. Mouftah, “Optimal cross-layer designs for energy-efficient wireless ad hoc and sensor networks,” Peflormance, Computing, and Communications Conference, 2003

16. G. Carneiro, J. Ruela and M. Ricardo, “Cross-layer design in 4G wireless terminals,” Wireless Communications, IEEE Transactions on, Vol. 1 1, No. 2, pp. 7-13, Apr. 2004

pp. 963-969, Aug. 1995

pp. 1218-1230, Oct. 1997

COVERAGE AREA ANALYSIS OF SOFT HANDOFF ON

CELLULAR CDMA SYSTEMS

TSANG-LING SHEU Department of Electrical Engineering, National Sun Yat-Sen University,

Kaohsiung, TAIWAN

JAW-HUE1 HOU Department of Electrical Engineering, National Sun Yat-Sen Universily,

Kaohsiung, TAIWAN

In this paper, we present an analytical model to study the impact of changing soft handoff coverage on the system performance for cellular CDMA systems. In the model, we assume a cell is in hexagon shape with two parameters representing the soft handoff region between the inner and the outer cell. By assuming that the new call generation rate per unit area is uniformly distributed over the service area and by adjusting the soft handoff region parameters, we can calculate the blocking and the dropping probabilities of the systems using the Markov model.

1. Introduction

Previous works in soft handoff have been focused on the investigation of handoff delay versus cell coverage. Viterbi et al. [l] demonstrated that soft handoff could result in a larger cell coverage area and a subsequent increase in reverse link capacity. By investigating channel resources of soft handoff in a limited field environment, Cho et al. [2] proposed an analytical model for a DS- CDMA mobile system. Lee and Steel [3] analyzed the capacity loss on the forward link. In [4], Narrainen and Takawira proposed a traffic model for a DS- CDMA cellular network. In their scheme, the teletraffic behavior was investigated by taking into account CDMA soft capacity and soft handoff among cells. Kim and Sung [5 ] derived equations for handoff traffic and introduced a methodology to calculate the capacity increase due to soft handoff. Network capacity versus cell coverage was studied in [6]. Their analysis shows that there exists an explicit relationship between the cell coverage and the number of active users in the cell.

279

280

In this paper, we build an analytical model to study the impact of changing soft handoff coverage on the new call blocking and the handoff call dropping probabilities for cellular CDMA systems. The assumption of changeable soft handoff area makes our analytical model more realistic and more accurate. The rest of this paper is organized as follow. Section 2 describes the basic traffic model of soft handoff used to build the proposed model for analyzing the soft handoff coverage. In Section 3, a Markov model is built to evaluate the average blocking probability of new calls and the average dropping probability of handoff calls. In Section 4, we present the analytical results and discuss the tradeoffs between enlarging and shrinlung the soft handoff coverage. Finally, Section 5 contains concluding remarks.

2. Soft Handoff Model

The pilot signal strength of base station (BS) that a mobile station (MS) can detect is inversely proportional to the distance between MS and BS. When pilot signal from a new BS is stronger than a threshold value, T-Add, a new link to the new BS is established while the old link to the old BS is still maintained. As shown in Figure 1, a call is said to be in its soft handoff in this situation.

I I Distance t------,

bR,

OReq

Figure 1. SoA handoff model of a CDMA system. An MS in soft handoff can communicate with two BS through two strong pilot signals, respectively. In a normal operation, if a pilot signal fiom the third BS becomes stronger than either of the two existing pilot signals, a handoff may occur and the call will drop the weakest link, so that only two links are available in any given time interval. For simplify, in our model, we do not consider the handoff effect from the third pilot signal. When the pilot signal from either the old BS or the new BS becomes weak and drops below a threshold value, T - Drop, the bad connection is released.

The hexagon cell structure with soft handoff region for a mobile CDMA cellular system is shown in Figure 2. In the figure, R is defined as the distance

28 1

from the hexagon center to any vertex, and Re, is defined as the equivalent radius of the hexagon cell. For convenience, the thresholds of T-Drop and T-Add can be defined as a function of R,. That is, T-Drop = aR, and T-Add =

bReq, where 1 I a I 2 and0 I b I 1. Soft handoff region may vary as the two parameters, a and b, change. Throughout this paper, the circle with radius bR, is referred to as the inner cell. The circle with radius aR,, is referred to as the outer cell, and the hexagon cell with equivalent radius Re, is referred to as the original cell.

Figure 2. The hexagon cell structures and their soft handoff boundaries. The soft handoff rate is the frequency of soft handoff attempts that a call

makes before the call terminates in the cellular system. The soft handoff rate can be determined by a number of factors. For examples, the size and the shape of a cell, the call density, and the moving speed of MS. Under the assumption that traffic is homogeneous and is in statistical equilibrium, the average handoff rate incoming to a cell is equal to the average handoff rate outgoing from the cell. The derivation of the average outgoing handoff rate ( A ) can be referred to Thomas’s formula [2]. Thus, we have

A=---’ pFL (1) 7r

where p , and L, respectively, denote the call density per unit area, the average moving speed of MS, and the perimeter of the cell area.

In the traffic model, we assume new call generations are uniformly distributed over the service area according to Poisson process with a mean rate per unit area ( A.). For simplicity, we make some assumptions for the sake of analytical tractability [2]. That is, the speed and the direction of mobile calls are uniformly distributed in the interval [O,V,,,], and in the interval [0, 27r 1, respectively. In addition, in our model, the speed and the direction of a generated mobile call will remain unchanged during its call interval. The call

282

duration time varies with a random variable, Ted, and it is assumed to be exponentially distributed with a mean 1/ ,u cd. Thus, the probability density function of TCd can be expressed as

In a cellular CDMA system, the call quality is proportional to the signal to interference ratio (SIR). SIR is directly affected by a number of factors, such as the system bandwidth, the code chip rate, the voice activity factor, etc. System capacity of a cellular CDMA system has been discussed in the previous literatures [6], [7]. The cell coverage versus system capacity is illustrated in [6]. The results shown in [6] will be directly adopted to evaluate the blocking and dropping probabilities of our proposed model.


When a call is generated in a certain cell, the cell is referred to as the original cell, and the call is categorized as a new call. When the new call move to the soft handoff region, it may issue a handoff request to a new BS that is selected based on the strongest strength of pilot signals. If the selected new BS can offer a free channel to the mobile call, we say the handoff is successful. Within the soft handoff area, a successful handoff call can transmit data through either the old or the new BS. When this handoff call continues to pass through beyond the outer cell, the services from the old BS will be disconnected. Assuming that the channel distributions and the cell structures are all identical as shown in Figure 2, we can represent the new call generation rate per cell ( R *) as /z ,,= R (I A,,, where A,, is the area of original cell and /z (I is the new call generation rate per unit area.

Handoff attempts may occur in the following two cases. In the first case, we consider that mobile calls are generated only in the soft handoff region, where two neighboring cells have an overlapped area. Since we have assumed that new call generation rate is uniformly distributed over this overlapped area, only half of the new calls generated in one neighboring cell may issue handoff attempts to the other neighboring cell. Let the handoff attempt rate be denoted as R h l , we have

A - - d 0 ( a 2 1 -b’)R,Z,( l -P,)’ (3) h ’ - 2

where Ps is the new call blocking probability. In the second case, we consider that mobile calls are generated within the inner cell. Handoff attempts can occur

283

if the generated mobile calls are passing the inner cell boundary and entering the soft handoff region. This handoff attempt rate is denoted as /z h2. Refer to Thomas’s formula as shown in Eq. (l), we have /z = /z h2, L=2 z bR,, and V is the average speed of mobile calls. Therefore, if the overall handoff attempt rate issued to the neighboring cells is denoted as /z h= /z hlf 1 h2. Since p is defined as the call density per unit area, it can be expressed as

where nc and 2 hc are, respectively, defined as the successful generation rate of new calls and the successful handoff attempt rate of handoff calls. They can be expressed as

P = + ‘ b c > / ( p c h ’ ’ (4)

= an x (1 - P,)

A*, = Ah x (1 - P,) 9

( 5 )

( 6 ) where P, is the failure probability of handoff attempts.

When a mobile call is terminated within a cell or when it passes beyond the boundary of the outer cell, the occupied channel is released. Thus, the channel holding time, Tch, is a random variable and can be expressed as

Tch = minU,,, T , , ) 7

where the two random variables, TCd and To,, are denoted as the call duration time and the call residual time, respectively. By assuming that T c d and To, are exponentially distributed with two mean numbers l/pcd and l/po, , the mean channel holding time (1/ y ch) can be obtained by i /pch = i/(pcd + . Assuming that the call residual time is proportional to the radius of the outer cell, the mean call residual time within the outer cell is given by l / p o , = a/po, . [8]. Sincel/yo, is inversely proportional to the speed of mobile calls, without losing any generality, the mean residual time in the original cell can be

(7)

expressed as l/po, = RwF.

In Figure 3, the Markov model is used to evaluate the system performance. We assume that a cell owns C code channels, and C is a variable that depends on the service area of the cell. The state Ej indicates that there are j code channels in use for a cell. If all the code channels owned by a cell are all occupied, any new call will be blocked and any handoff attempts will fail.

K h 2 . k QCh

Figure 3. State transition diagram of the Markov Model.

284

If we let Pj represent the steady state probability of the state E,, then Pi can be calculated.

If handoff calls do not posses higher priority than new calls, the blocking probability of new calls should be equal to the failure probability of handoff calls. The probability that all the code channels are occupied (Pc) can be derived as

Based on Eq. (1) and Pr, we can derive the average dropping rate of handoff calls ( R hd) as ahd = ( p .v. Lo. p,)/, , where Lo is the perimeter of outer cell. Finally,

the average dropping probability of handoff calls (Phd) is just hd divided by ,? h,

iX., P h d = hd h .

4. Analytical Results and Discussions

According to the relationship between the cell coverage and the number of active users within the cell [6], Figures 4a and 4b show PB and Phd, respectively. Here, we increase the parameter a from 1 .O to 1.6 and observe three different settings of the new call generation rate per unit area (i.e., R (I = 0.05, 0.075, and 0.1). Let b be fixed to 0.8 and assume 7 = 3 0 km/hr. Since increasing the parameter a will enlarge the soft handoff area, R ,, is increased accordingly. As observed from these two figures, a slight increase in R (I can significantly increase PB and Phd.

parameler a

Figure 4a Average blocking probability of new calls when a is varied and b=0.8.

285

parameter a

Figure 4b Average dropping probability of handoff calls when a is varied and b=0.8. Figures 5a and 5b, respectively, show the variations of PB and Phd, when the

parameter b is increased from 0.5 to 1.0. Here, we assume a =1.2 and V = 30 km/hr. Since increasing the parameter b can reduce the soft handoff area, the number of handoff attempts within that area becomes smaller. When R is large (for example, R a = 0. l), increasing the parameter b from 0.95 to 1.0 may adversely increase PM. This is because although increasing the parameter b may reduce the soft handoff region, it also increases the number of new calls per cell. As a result, the number of mobile calls contending for a fixed number of code channels is increased. Therefore, as shown in Figure 5b, the curve of handoff call dropping probability (when R a = 0.1) goes down smoothly first and then goes up, as the parameter b is increased from 0.5 to 1.0.

-

Figure

1041 " " " " ' I 0.5 0.5 0.6 0.a 0.7 0.76 0.0 0.m 0.9 0.35 1

parameter b

5a Average blocking probability of new calls when b is varied and P1.2 .

286

parameter b Figure 5b Average dropping probability of handoff calls when b is varied and ~ 1 . 2 .

5. Conclusions

In this paper, we have presented an analytical model to study the impact of changing soft handoff coverage on the blocking and the dropping probabilities for cellular CDMA systems. In the model, we assume a cell is in hexagon shape with two parameters representing the soft handoff region between the inner and the outer cell. The analytical model was built with Markov chains and Thomas's formula. From the analytical results, we have obtained some significant observations. First, the increase in soft handoff region can cause both the blocking and the dropping probabilities to increase very quickly. On the other hand, however, shrinking the soft handoff area can simply reduce the blocking and the dropping probabilities very slowly and smoothly.

References

1. A. J. Viterbi et al., IEEE J. Select. Areas Commun., vol. 12, 1281 (1994). 2. M. Cho, K. Park, D. Son and K. Cho, IEICE Trans. Commun., vol. E85-B,

3. C. C. Lee and R. Steel, IEEE Trans. Veh. Tech., vol. 47,830 (1998). 4. R. P. Narrainen and F. Takawira, IEEE Trans. Veh. Tech., vol. 50, 1507

5 . D. K. Kim andD. K. Sung, ZEEE Trans. Veh. Tech., vol. 48,1195 (1999). 6. V. V. Veeravalli and A. Sendonaris, IEEE Trans. Veh. Tech., vol. 5, 1443

7. H. Jiang and C. H. Davis, IEEE Trans. Veh. Tech., vol. 52,814 (2003). 8. R. A. Guerin, ZEEE Trans. Veh. Tech., vol. VT-35,89 (1987).

1499 (2002).

(2001).

(1 999).

LMS VS. RLS FOR ADAPTIVE MMSE MULTIUSER DETECTION OVER TIME VARYING COMMUNICATIONS

CHANNELS*

ZARKO B. KRUSEVAC AND PREDRAG B. RAPAJIC National ICT Australia and University of New South Wales

Sydney, NSW 2052 Australia E-mail: zarko.k@student,unsw,edu.au, [email protected]

RODNEY A. KENNEDY National ICT Australia and Australian National University

Canberra ACT 0200, Australia E-mail: rodney. kennedyaanu. edu. au

This paper provides the analysis of the Least Mean Square (LMS) and the Re- cursive Least Square (RLS) adaptive algorithms performance for adaptive CDMA receivers in slowly time varying communication channels in the presence of multipath. We confirm that neither the LMS algorithm nor the RLS algorithm has a complete monopoly over good performance in a non-stationary environment. Rather, one or the other of these two algorithms is preferred, depending on the environmental conditions that are prevalent. We show that the LMS algorithm provides a good tracking if speed of channel change is slow, but LMS stability criteria depend on signal statistics (and so, on channel conditions). The RLS convergence and stability criteria do not depend of signal statistics, but the RLS involves memory and does not explicitly recognize time varying nature of the channel. Presented simulation results confirm the theoretical analysis.

1. Introduction

The adaptive CDMA receivers [l], [2] have provided a contribution removing many obstacles from the practical implementation of the multi-user concept in cellular mobile communications. The common solution uses ad hoc applied Least Mean Square (LMS) or Recursive Least Square (RLS) algorithms and then "tunes" the adaptive algorithm parameters to the best

*This work is supported by the Australian Research Council Discovery project DP0210897

287

288

convergence performance. However, despite a considerable attention in the literature over the last two decades regarding the ability of adaptive algorithms to track time variations in time varying (mobile) communication channels [3], [4], [5], [6] , [7], [8], the theoretical limit of overall performance of this solution remains unclear.

In this paper we provide theoretical and simulation analysis of the LMS and the RLS adaptive algorithm performance for the adaptive CDMA receivers in slowly time varying channels in the presence of multi-path. The performance analysis includes convergence analysis, stability analysis and tracking ability analysis. The main contributions of this paper may be summarized as follows:

- We show that overall performance of the LMS and the U S adaptive algorithms in a non-stationary environment are based on a trade-off between speed of convergence, steady state fluctuation and tracking (excess lag fluctuation). We confirm that neither the LMS algorithm nor the U S algorithm has a complete monopoly over good performance in a non- stationary environment. Rather, one or the other of these two algorithms is preferred, depending on the environmental conditions that are prevalent.

- We show that the LMS algorithm provides a good tracking if speed of channel change is slow, but LMS stability criteria depend on signal statistics (and so, on channel conditions). The U S convergence and stability criteria do not depend of signal statistics, but the RLS involves memory and does not explicitly recognize time varying nature of the channel.

- Our simulation analysis is related to the spreading gain and multi-user interference effects upon performance of the LMS and RLS algorithms for the adaptive CDMA receivers in slowly time varying channels in presence of multi-path. Simulation results confirm the provided theoretical analysis.

The paper is organized as follows. Section 2 is devoted to the adaptive MMSE receiver. Theoretical analysis of the LMS and RLS adaptive algorithms is provided in Section 3 and 4, respectively. Simulation results are presented in Section 5 , with Section 6 concluding the paper.

2. The Adaptive MMSE Receiver

The adaptive MMSE receiver structure for multipath fading channels is presented in [2]. It consists of a bank of adaptive fractionally spaced MMSE FIR filters along with the (ML) detector part of the receiver for data detection, Figure 1. During the receiver training period, a training sequence and an adaptive algorithm [4] are used to obtain the optimal coefficients

289

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

--;KC2 _ _ _ _ _ i .MBI

Figure 1. The Adaptive MMSEML Receiver [2]

Wk = [vc, d:], where vk(n) is the MMSE FIR filter coefficient vector and dk is a vector of tentative decision aided coefficient sequences [2]. Filter coefficients are obtained by minimizing the MSE, E[lek(n)12], where [2]:

Xk(n) and yk(n) are input symbol and discrete time received sample of lcth user at time n, respectively and X D ( ~ ) contains known (training) symbols with unknown symbols (from inter-cell interferers) set to zero' [2]. After the training period, the filter coefficients can be kept fixed during data detection. Alternatively, in the decision directed mode, these coefficients can be updated by tentative decisions.

3. LMS Adaptive Algorithm

The LMS is instantaneous MSE, steepest descent algorithm. The instantaneous MSE methods are model free, computationally simplest but may be slow. The following steps correspond to the LMS algorithm [2]:

vk(n + 1) = vk(n) + ale;(n)Y(n) &(n + 1) = &(n) - aae;(n)x~(n) (2)

where a1 and a2 are the step sizes of the algorithm. Thus, the LMS is computationally very simple and the update needs only latest y(n), wk(n) = [vr(n),d;(n)lT and error ek(n) given by (1).

290

Convergence- To carry out a convergence analysis of the LMS algorithm, time invariant "true" weights W ~ O are assumed [3]. For convergence in mean of the weight error vector w k ( n ) = wk(n) - w k o , i.e. m k ( n ) = E(wk(n) ) -i 0 for n + col it needs [3]:

11 - aA,I < 1, u = 1 , 2 . . .p , i.e. Amin > 0 and dmaz < 2 (3) where A,, u = 1 , 2 . . .p are eigenvalues of R, = E(yk(n)yr(n)). Speed of convergence depends on closeness of 11 - aA,1 to 1 and the time constant for convergence 7, = In,l-aX,, can be recognized.

Stability- The LMS stability analysis deals with the relative steady state excess MSE or misadjustment due to noisy adaptation 131:

-1

where Em = limk,, E(e%), tr(Ry) = trace(R,) is the sum of the eigenvalues of 4 and 0% is the variance of "true noise" ~ ( n ) = yk(n) - x(n)wko. c = O(LYT) means = constant # o as Q -+ 0, T = 1 , 2 ...

M = -tT(R,) +O(Q2) < -

The second order stability requires [3]:

(5) Q 1 2 3

Tracking- The second order LMS tracking analysis, which assumes the time varying "true" weights W O , ~ , relieves [3]:

where QO = cia, 0: is the received symbols variance, 0: is the variance of the weight process, y is true speed of change of time varying weights and p = tr(R,)/a;. Mnoisy adap. = + O ( Q ~ ) is the inflation due to

2 c2a2 noise adaptation and Miag = 2 + f is excess 'lag' terms due to the time varying weights.

Figure 2 shows the total relative steady state excess MSE or misadjustment Mtot for the LMS algorithm. Clearly Mtot has an optimum wrt QO,

namely:

In general, small Mtot needs aO,opt small. Consequently, the true speed of change y must be small (7). Otherwise, if (YO << y (i.e. adaptation is

291

M (misadjustment)

OS,

R

Figure 2.

very slow) then excess lag can be very large. However, small time constant for convergence T~ = & (fast convergence) needs QO = au; large (I1 -aX,I close to 1). Thus, there is a conflict between speed of convergence, steady state fluctuation and excess lag fluctuation. Furthermore, LMS stability criteria depend on the signal statistics CT; (and so, on channel conditions). If the step size QO is set close to the stability limits, then channel time variations may cause the algorithm to blow up.

Relative steady state excess MSE (misadjustment) for the LMS algorithm

4. RLS Adaptive Algorithm

The RLS algorithm is a Newton's Method algorithm. Only one tuning parameter, the forgetting factor Xf, is used, where 0 5 Xf 5 1. Thus, the U S algorithm is almost model free, generally faster than the LMS algorithm, but can be badly biased. Denoting W k = [vc,dTIT, u(n) = [yT(n)x;(n)lT, the steps of the RLS algorithm can be provided as follows:

Xj'Pk(n)u(n + 1)

gk(n + l) = 1 + XfluH(n + l)Pk(n)u(n + 1)

ek(n + 1) = q ( n + 1) - wf(n)u(n + 1)

wk(n + 1) = wk(n) + gk(n + l)e;(n + 1)

Pk(n + 1) = XjlP,k(n) - Xj'gk(n + l)uH(n + 1)Pk(n) (8) By using small 110 approximation, where 110 = 1 - Xf is the forgetting

292

rate, the RLS algorithm can be reduced back to an scaled LMS algorithm [3]. However, the nature of the forgetting rate PO is completely different from the LMS step size ao. The RLS forgetting rate PO involves memory. Furthermore, since the RLS forgetting rate PO does not depend on the signal statistics, the RLS convergence and stability criteria do not depend of the signal statistics either. Thus, local instabilities do not occur.

Convergence- The first order stability condition for the ”small PO” RLS algorithm is given by [3]: 0 < PO < 2 and the time constant for convergence is r = +.

Stability- The misadjustment for the ”small po” RLS algorithm can be expressed [3] :

n l - ~ o l

PO M = + I y ) + O(p!) (9)

Tracking - The relative steady state excess MSE or misadjustment for the RLS algorithm with time-varying weights is [3]:

Furthermore, and Mopt can be found:

for

Figure 3 shows the misadjustment Mtot for two different true speeds of change of weights 72 > y1. Mopt can be small if is small (11). Consequently, true speed of channel change y must be small (1 1). However, a small time constant for convergence T = (fast convergence) needs PO large (close to 1). Thus, an appropriate trade-off between speed of convergence, steady state fluctuation and tracking should be found.

RLS algorithm as a trivial case of adaptive Kalrnan filter- Both, the LMS and the RLS algorithms are based on the received signal observation. On the other hand, the Adaptive Kalman Filter Algorithm is a model based method. In addition to the observation equation, the model based methods use the state equation which describes how parameters, driven by the adaptive algorithm, vary with time. If the trivial, time invariant, state equation for the ”true” weights is used wo,k(n-l) = wg,k(n), then the adaptive Kalman filter is reduced back to the RLS algorithm [3]. Thus, the RLS does not explicitly recognize time varying nature of weights [3] and consequently cannot offer good tracking performance.

293

M (misadjustment) 0.35 I I

Figure 3.

01 ' I , I ! 0,002 0.004 ! 0.m 0.008 0.01 0.012

p 4-1 poopt(Q b p t ( Y 2 ) 0

Relative steady state excess MSE (misadjustment) for the RLS algorithm for two different true speeds of weights time variations, 72 > 71 5 . Simulation results and analysis

Spreading gain G=8 No.of active users=8 Spreading gain G=8

I I I 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000

K (number of iterations) K (number of iterations) a) b)

Figure 4. ference, T/T,,h =

MSE performance of the LMS and the RLS algorithms a) no multiuser inter- b) in the presence of multiuser interference, T/T,,h = lop4

For the purpose of this analysis, we suppose that the channel process is sinusoidal deterministic time varying process. Since both, the LMS and U S algorithms are observation based and they do not recognize the deterministic nature of the channel process, our approach still does enable

294

considerable conceptual insight to be gained. Simulation analysis from Figure 4-a) assumes no multiuser interference

and a slow time varying channel (T/T,,h = where T is the symbol period and Tcoh is the channel coherence time). Fast convergence, usually provided by U S algorithm in time invariant environment, has to be traded- off in order to improve tracking ability. The LMS algorithm exhibits some local instabilities. When the channel is faster (T/T,,h = low4) and in the presence of multiuser interference, Figure 4-b), it is more difficult to find an appropriate compromise between speed of convergence, steady state fluctuation and tracking for adaptive algorithms to provide good performance in a non-stationary environment.

6. Conclusion

The conclusion of the provided theoretical and simulation analysis could be that neither the LMS algorithm nor the RLS algorithm has a complete monopoly over good performance in a non-stationary environment. Rather, one or the other of these two algorithms is preferred, depending on the environmental conditions that are prevalent [4].

References 1. P. B. Rapajic and B. Vucetic, Adaptive Receiver Structures for Asynchronous

CDMA Systems, IEEE J. Select. Areas Commun., vol. 12, no. 4, pp. 685-697, May. 1994.

2. P. B. Rapajic and D. K. Borah, Adaptive MMSE Maximum Likelihood CDMA Multiuser Detection, IEEE Journal on Selected Areas in Communications, Vol. 17, No. 12, December 1999.

3. V. Solo and X. Kong Adaptive Signal Processing Algorithms, Englewood Cliffs, NJ: Prentice-Hall, 1995.

4. S. Haykin, Adaptive Filter Theory, 3rd ed., Englewood Cliffs, NJ: Prentice Hall 1996.

5 . N. R. Yousef and A. H. Sayed Ability of Adaptive Filters to Track Carrier Off- sets and Channel Nonstationarities, IEEE Transaction on Signal Processing, Vol. 50, No. 7, July 2002

6. N. R. Yousef and A. H. Sayed A Unified Approach to the Steady-State and Tracking Analyses of Adaptive Filters, IEEE Transaction on Signal Process- ing, Vol. 49, No. 2, February 2001

7. B. Widrow and S. D. Steams, Adaptive Signal Processing, NJ: Prentice-Hall, 1985.

8. A. Benveniste Design of adaptive algorithms for tracking of time-varying sysytem, International Journal of Adaptive Control and Signal Processing, 1~3-29, 1987.

PERFORMANCE ANALYSIS OF A PREEMPTIVE HANDOFF SCHEME FOR MULTI TRAFFIC WIRELESS MOBILE

NETWORKS*

DIMITRIOS D. VERGADOS University of the Aegean

Department of Information and Communication Systems Engineering Karlovassi, Samos, GR-83200, Greece

ANGELIKI SGORA University of the Aegean

Department of Information and Communication Systems Engineering Karlovassi, Samos, GR-83200, Greece

The next generation wireless and mobile networks need to support a wide range of multimedia applications, such as voice, data, image and video. The crucial issue is to provide QoS guarantees for arriving multiclass traffic. Thus, the need for efficient and bandwidth reservation scheme is compulsory. In this paper, we propose a preemptive priority reservation handoff scheme that categorizes the service calls into four categories according to the QoS guarantees. The system is modelled using a multidimensional Markov chain and a numerical analysis is presented to estimate several performance measurements, in terms of blocking probability of ongoing calls, forced termination probability, and average transmission delay. The simulation confirms the correctness of the analytic model and the good performance of the proposed scheme.

1. Introduction

Multimedia services impose stringent QoS demands on the wireless networks. Therefore, in order to guarantee acceptable QoS in multi traffic wireless networks, effective and efficient resource allocation are extremely important. A great concern must also be given to the call blocking probability of different service types. Also, due to the fact that handoff request calls have a greater importance than the originating calls several prioritization strategies must be considered, to further ensure acceptable QoS. Two mainly strategies have been proposed in the literature: the channel reservation schemes and the queuing

This Research work is funded by the Ministry of Education and Religious Affairs and co-funded by E.U. (75%) and National Resources (25%) under the Grant “Pythagoras - Research Group Support of the University of the Aegean”.

295

296

priority schemes. In the channel reservation strategy a number of channels in each cell are reserved for exclusive use by handoff calls, in predictive or not predictive manner [3], [ 5 ] . Queuing priority handoff schemes follow the principle: when all channels are occupied, either new calls are queued while handoff are queued while handoff calls are blocked, or new calls are blocked while handoff calls are queued, or both calls are queued [l], [4], [6]-[8].

However, in all of the above studies only real time and non real time type of service are considered[5]-[8]. In this paper, we distinguish three types of traffic: real time, semi-real time and non real-time. With this distinction we have a preemptive priority handoff scheme that can achieve guaranteed QoS for a variety of applications.

The paper is organized as follows: Section 2 describes the proposed handoff scheme. In Section 3, we introduce the traffic model. Section 4 presents the mathematical analysis of the proposed handoff scheme. The numerical and simulation results are discussed in Section 5. Finally, Section 6 concludes the paper.

2. The System Model

In this paper, we consider a system with homogeneous cells with a fixed amount of bandwidth capacity of S channels. We focus our attention on a single cell, called the reference cell. In the base station of the reference cell, as figure 1 depicts, there are three queues, Q1, Qz and Q3 with capacities MI, M2 and M3 for typel, type2 and type3 handoff requests, respectively.

sd

297

For the typel users there is an area, called the handoff area, where the mobile user's call in a cell can be handled by either base stations of adjacent cells. On the contrary, for both type2 and type3 mobile users, it is assumed there is a cell boundary between two neighboring cells, which is defined as the locus of points where the average received signal strength of the two neighboring cells is equal. Moreover, in order to give priority to all handoff requests the maximum allowed number of channels for all the types of originating calls in the reference cell is S,.

When a typel handoff request arrives, first checks if there are channels available. If there are no idle channels, the typel handoff request can preempt a type3 or a type2 handoff request. The interrupted handoff request is queued in Q3 or Q2 respectively and is waiting for a channel to be available, based on the FIFO rule. The maximum allowed waiting time for typel handoff requests in Q1 is the dwell time of a mobile user in the handoff area. Otherwise the typel handoff request will be deleted from the queue if the mobile user moves out the handoff area before getting service. A typel handoff request will be blocked by the system, if on arrival at the reference cells finds no available channel and the QI is full. When a type2 handoff request arrives, first checks if the are more than Sd channels available, where Sd is the maximum allowed number of channels for the type2 handoff requests in the reference cell. If there are no channels available, the type2 handoff request can preempt a type3 handoff request. Otherwise, it is put in the Q2. The interrupted handoff request is queued in Q3 and is waiting for a channel to be available, based on the FIFO rule. A type2 handoff request will be blocked if on arrival there are no channels available and 4 2 is full. A type3 handoff request can be served if the are more than Se channels available, where Se denotes the maximum allowed number of channels for the type3 handoff requests in the reference cell. Otherwise, it is put in the Q3. A type3 handoff request will be blocked if on arrival Q3 is full.

3. The traffic model

In order to develop the analytical model we use the Xie and Kuek's fluid flow model [2] as our traffic model. This model assumes a uniform density of users throughout the area under consideration and a user is equally likely to move in any direction with arbitrary distribution of moving speed.

3.1. Cell Dwell Time

Assuming that the dwell time of the mobile users in a cell has an exponential distribution with mean rate pdwell, the average cell dwell time is given by

298

where E[V] is the average of the speed V, L is the length of the perimeter of an arbitrary shape cell and A is the area of the cell.

3.2. Handoff Area Dwell Time

Assuming that the path length and velocity of users are independent, we can get the average dwell time E[Th] is given by

3.3. Channel Holding Time

The call holding time Tci i= 1.. .3 of each type of service calls is assumed to have an exponential distribution with mean l/pci. Therefore, the channel holding time TI (T2 or T3) of a service call is equal to the smaller between Tdwell and Tcl (TC2 or TC3). From the memoryless property of the exponential pdf leads to

3.4. Arrival Process of Service Calls

We assume that the arrivals of originating service calls of each type in a cell are following Poisson process. The arrival rates of originating typel, type2 and type3 service calls are donated by bl, b2 and b3 respectively. For the arrival rate of each type service handoff requests is given by

where E[Ni], is the sum of the type i service calls holding channels. As shown in Figure 1, the total arrival rate h of calls in a cell is

A =Ao, +Ao2 +Ao, +AH1 +AH2 +AH3 ( 5 )

3.5. Performance Analysis

We represent the state of the reference cell by a five dimensional Markov chain (i, j, k, 1, m) where

299

0 i (i E [O,S+M,]) is the sum of the number of channels used by type 1 requests and the number of handoff requests type 1 waiting in Q1.

j (j E [o,sd]) is the number of channels used by type 2 requests k (k i E [0,M2]) is the number of handoff requests type 2 waiting in Q2. 1 (1 E [o,sd]) is the number of channels used by type 3 requests m (m i E [0,M3]) is the number of handoff requests type 3 waiting in Q3. Thus the state space V of the reference cell is given by

0

0

0

V = {(i, j , k , / ,m) I {m = 0,k = 0,O 5 i + j + / < Se}u {k = 0,m >= O,Se 5 i + j + I < Sd , I S S,) u { sd 5 i + j + I I S,{ { I = 0,m t= 0,k t= 0) u { I > 0 , k >= 0,m = M 3 } u { I > 0,m 2= 0,k = O}}u { S < i + j t I <= S + M1,{ { i > S,k t 0,m 2 0 } u { I = 0, k = M2,m >= 0 )

u { I > 0, j = 0, k > 0,m = M 3 } u { ( I > 0, j > 0, k = M2,m = M 3 } } }

(6)

In the state transition diagram there are NT=IIVII. Therefore we can get NT balance equations through the state transition diagram. Since the sum of all state probabilities must be equal to 1, we have

Adding above the normalizing equation and the equation (4) for each type handoff request, we have NT+3 nonlinear independent simultaneous equations. Though NT is usually rather large, all the state probabilities can be obtained by solving NT+3 nonlinear independent simultaneous equations using the SOR iteration method. Based on the above state probabilities we may obtain the following performance measures of the system. The blocking probability of an originating call (type 1,2 or 3) is

The handoff blocking probabilities of each type are respectively given by s, sd- / M 2 M3

BHI =C C P ( S + M l - ( I + j ) , j , M 2 , 1 , M 3 ) + T p ( S + M I , O , k , O , m )

(9) j = l j=1 k=O m=O

/ = I k=O j= l m=O

300

The average length of each queue are respectively given by L,, = c i P ( i - S , j , k , l , m )

( i , j ,k , l ,m)EV

The failure probability of a typel handoff request for single handoff attempt is given by

The handoff probability Ph of a typel call is given by P c-dwell

PI + P c - d w e l l

Therefore, the forced termination probability Pfl of typel calls can be

Ph = (16)

expressed as

3.6. Numerical and Simulation Results

In this section, we present the simulation and the numerical results from the analytic model in order to evaluate the system performance.

For ease of our numerical evaluations, we assume that the shape of the cell is circular with radius r, and the kind of mobile users is pedestrian with following parameters: r=lOO m, E[D]=O. l r m, E[V]=0,5 d s e c , E[TCl]=120 sec, E[Tc2]=60 sec, E[Tc3]=40 sec, S=12, Sd=ll, Se=lO, Sc=X, M1=l, M2=2 M3=2 and &=lo-*.

The ratio of originating typel, type2 and type3 is set to 1 (Aol = hO2 =

ho3).Those numbers are also used for the simulation scenario.

30 1

Figure 2 shows the blocking probabilities BO of originating calls and the forced termination probability Pf of type1 calls versus the original call arrival rate and compares the simulation and the analytical results. From the figure, we can see that the blocking probability BO and the forced termination probability increase as traffic load increases. Moreover, we can also see that the forced termination probability with preemptive priority scheme is lower than that one of the nonpreemptive priority scheme, while the blocking probability of originating calls remains almost unaffected. Therefore, we can conclude that the scheme with preemptive priority and priority reservation of channels is very effective in decreasing the forced termination probability of a high priority handoff call.

1,00E+00 1,OOE-OI 1,00E-02 1,00E-03 1,00E-04 I ,00E-05 1,00E-06 I ,00E-07 1,00E-08

0.01 0.12 0.3 0.6 1.2 1.8 2.4 3 9

h0

8 0 simulation

Pf analytical

o Pf simulation

- 80 - - Pf

nanpreemptive

nanpreemptive

Figure 2. Blocking probability and forced termination probability versus the original call arrival rate

Figure 3. Average queue length versus the original call arrival rate

Figure 3 depicts the average length of each queue versus the original call arrival rate. From the figure we observe that the average queue length of queue Q , is comparatively to the average queue lengths of queue Q2 and queue Q3 too

302

small, due to the fact that type1 handoff requests have the highest priority and consequently more available channels for use.

3.7. Conclusion

The next generation wireless and mobile networks have designed to support multimedia services, with different traffic characteristics and different QoS guarantees. In this paper, we propose a preemptive priority handoff reservation scheme for a three class service wireless network. A new model for the system performance has been proposed. Several measurements for the system performance have been evaluated. These analytical evaluations have been compared to the simulations results. The analytical and the simulation results indicate that the proposed scheme may offer better performance than the conventional handoff schemes in terms of the handoff blocking probability and forced termination.

References

1. S. Tekinay and B. Jabbari, “Handover and Channel Assignment in Mobile Cellular Networks”, IEEE Communication Magazine, vol. 29, no. 11, pp42- 46, November 199 1.

2. H. Xie, and S. Kuek, “Priority Handoff analysis”, Proceedings IEEE Vehicular Technology Conference, pp. 182- 190, Aug. 1993.

3. R. A. Guerin, “Queueing-Blocking System with Two Arrival Streams and Guard Channels”, IEEE Transactions on Communications, vo1.36, no.2, pp. 153-163, Feb. 1988.

4. F.-N. Pavlidou, “Two-Dimensional Traffic Models for Cellular Mobile Systems,” IEEE Transactions on Communications, vol. 42, no. 21314, 1994.

5. D. Hong and S. S. Rapport, “Traffic model and performance analysis for cellular mobile radiotelephone systems with prioritized and nonprioritized handoff procedures,” IEEE Trans. Vehicular Technology, vol VT-35, pp.

6. Q.-A. Zeng and D. P. Agrawal, “Modeling and Efficient Handling of Handoffs in Integrated Wireless Mobile Networks”, IEEE Transactions on Vehicular Technology, vol. 51, no. 6,2002.

7. Li, C.Wu, K. Mukumoto and A. Fukuda, “Analysis and study of a handoff scheme with multiple priority strategies, Science in China(Series E), Vol. 43, No. 4, August 2000.

8. J. Wang, Q.-A. Zeng and D. P. Agrawal, “Performance Analysis of a Preemptive and Priority Reservation Handoff Scheme for Integrated Service-Based Wireless Mobile Networks”, IEEE Transactions on Mobile Computing, vol. 2, no. 1, January 2003.

77-92, Aug. 1986.

PERFORMANCE ANALYSIS of DS-CDMA SYSTEMS in Multiple- CELL with CORRELATED FADING CHANNELS

*JOY TONG-ZONG CHEN, AND NIU CHI-KUANG, and FU-CHAO CHUNG

Department of Communication Engineering, Da Yeh University 112 Shan -Jeau, Rd.Da-Tsuen, Chang-Hua 51505 Taiwan R.O.C.

Tel: +886-4-851-1888 a+:2523 FOX: +886-4-851-1245

*E-mail addr.: jchen(iiimuil.dvu.erlu.tw

.In this paper, the impact of the correlation on the performance of multiple-cell DS- CDMA cellular systems over correlated fading channels is investigated. A new closed- form formula for the joint probability density function (joint pdf) of the diversity combiner with arbitrary correlation coefficients in terms of the generalized Laguerre polynomial and the new expressions of average bit-error rate (BER) for the DS-CDMA system are given in this paper. The results demonstrate that the BER is significantly dependent on the correlation characteristic of diversity branching for multiple-cell environments.

Keywords: multiple-cell, correlated Nakagami-m fading, DS-CDMA, Laguerre polynomial

1. Introduction Whenever one desires to compensate the losses in multipath fading, the diversity-combining technique is one of the effective methods [l]. It is well known that the DS-CDMA (direct-sequence coded-division multiple-access) system with Rake receiver is an effective method for combating multipath fading over a frequency selective fading channels. In earlier studies, the statistics model of a fading channel concentrated on the assumption that the channel diversity branches were statistically independent of each other, whereas, such consideration should base on assuming that the paths are suEciently separated [2]. For a long time, many of the important studies have employed the Rayleigh distribution to characterize the envelope of faded signals [3]. However, Nakagami-m distribution has been thoroughly investigated, since it has been verified as a more versatile model for a variety of fading environments such as urban and suburban radio multipath channels for wireless communication systems [4].

Recently, Aalo [ 5 ] analyzed for both coherent and noncoherent modulation systems with two special correlated Nakagami types and a maximal-ratio combiner (MRC). Lombard0 et al. [2] derived an exact expression for the performance of BPSK and NCFSK with pre-detection MRC in correlated

303

304

Nakagami fading channels. Zhang [6] derived the exact BER expression for BPSK and BFSK systems with MRC over correlated Nakagami channels. Patenaude et al. [7] evaluated the BER of noncoherent post-detection diversity for the dual-branch and three-branch diversity cases in Nakagami-m correlated channels. Recently, A. H. L. Chan et al. [8] investigated the performance of video transmission over frequency-selective, correlated Nakagami fading forward-link channel. The BER performance was derived by Emad K. Al- Hussaini et al. [9], in which the author evaluated the user capacity for DS- CDMA system with SC/MRC Rake diversity over Nakagami fading channels. Yang et al. [lo] analyzed the performance of a DS-CDMA system over correlated Nakagami environments by means of the two special cases illustrated in Aalo's reports [5 ] . A new generic correlated Nakagami fading allowing for arbitrary covariance matrix and distinct real fading parameters was derived by Zhang [ 111. Although that the study of wireless communication systems over Nakagami fading channels has been intensively explored. The analysis of most of the reports focused on single-cell systems with an assumption of statistical independence between branches, and few of them included the DS-CDMA system environments.

In this paper, the impact of correlation on the performance of single-cell DS-CDMA cellular is compared to that of the multiple-cell DS-CDMA cellular systems. This paper is organized as follows. In Section 2, the DS-CDMA system with Rake combining and the statistical model of fading channel are presented first, then, the joint probability density function ('joint pdf) of the correlated fading envelops is derived. In Section 3 the average BER performance of the DS-CDMA system is evaluated, and a special case with dual-branch is given as an example there. Following the analytical results in Section 3, a numerical manifestation and a brief discussion are shown in Section 4. Finally, there is a brief conclusion given in Section 5 .

2. DS-CDMA system models description

2.1. Transmitter and receiver model

The DS-CDMA system is assumed to completely synchronize to the first path of the desired signal. Then the output of the Rake receiver for the reference user can be written as [ 121

where L, denotes the number of taps. The first term of the last equation represents the desired signal, the second term, I$ , is the self-interference (SI) which will be assumed to be negligible in this study by carefully choosing the PN code of the DS-CDMA system, 12 is multiple-access interference (MAI), and the last term AWGN is with zero mean and N,T,/4 variance.

305

2.2. DS-CDMA channel model

The received SNR on the I-th branch of the Rake receiver is expressed as X, = a:Eb INo , where the weight, {a?’}, 1 = 0, 1, ..., LR -1 , is characterized as Nakagami-m distribution [4], ~ ~ = p o q denotes the energy per bit. It can be shown that the pdf of X, is Gamma-distributed with a/ = E , R , / ~ , N , . Furthermore, if we set p, = x,/a, be the normalized SNR then the pdf of p, can be obtained as

Accordingly, the received power at the reference mobile user will be affected by both lognormal shadowing and path loss, which is given as [ 171

where y is the path-loss exponent, d, indicates the distance between the z-th BS and the reference mobile user, and <, denotes a Gaussian random variable with zero mean and standard deviation 0,. The MIP of the multiple-cell environment can be expressed as

(4) 0;’ = a(k’ , e-&’ 10

where a$’ and a:) represent the I-th and the first path signal intensities for the z-th cell channel, respectively, and 6, is the rate of average power attenuation for the z-th cell signal propagation. Assumed that 6, = 6 . The second moment of a single-cell can be expressed as

where 2 ~ ( . , . ; . ; ) stands for the Gaussian hypergeometric function, and Av is the branch power correlation coefficient between x:) and x:) , defined as

. The normalized covariance matrix, C, , has the elements including the set of a,, , i, j = O , I , . . . , L-1 , However, it is more suitable to adopt the envelope correlation coefficient, py , of a,, a, , and which is given as [4]

(7) py = E[a,a , ]=@(m, l ) . 1

306

2.3. Gaussian approximation

Consider now the base station-to-mobile link in a multiple-cell DS-CDMA cellular system, There is MA1 contributed by the additional K, users of each adjacent cell. The total number of cells in the first tier is set as. As a result, the MA1 can be written as

where K, is the number of users, L, represents the number of diversity branches in the multiple-cell DS-CDMA system, and A, denotes the attenuation factor which is induced from the z-th ( z f 0 ) BS, given as

-' (LG) A+) 10 (9)

where cell z denotes any one of the six adjacent cells, as shown in Fig. 1. The SI term is also considered to be negligible, and the AWGN term has its variance as

By using (8) and (lo), the total variance, (of)M, for a multiple-cell system at the output of the Rake receiver can be calculated as

The conditional BER for a multiple-cell system can be easily obtained from (8) by replacing (R), and ( D ; ) ~ with (gsz), and ( u ; ) ~ , respectively, where (s,"), is the conditional mean of the sampled output of the Rake receiver for multiple-cell system, given by

307

2.4. Joint pdf of correlated fading channel

The SNR at the output of the Rake receiver, which is a sum of the squares of the signal strengths a,, is given by

where a, follows the Nakagami distribution. Firstly, one may determine the joint characteristic function of y , and which can be obtained as [ 111

+r(jto,...,jrL-l I U) = E[exp{j(t,a,2 +...+f(L-l,a~L_,,}l =I /de tU-JW" (14)

where T = diug{fo,rl,~..,r(L~l)} is the diagonal matrix. In order to determine the joint pdf of the correlated fading channels, (14) is

expanded and j r , is replaced with s, , I = 0 , 1 , . . . , L - l , then the joint characteristic function is expressed as

+r(so , " 'S (L -~ ) I C J ) = { ( 1 - ~ ~ ) ( 1 - &I)... (1 - E ( ~ - ~ ) ) . G(ao,..., (15)

G ( ' O , " ' , u ( L - l ) ) = l - ( C ~ ~ ~ 'azuk + K, j ,k ' ' , a ~ u ~ +"'+vO.l, .(L-l)'aOul...a(L-l)) (16) '<I KJ<k

where the term making use of the Maclaurin's theorem, (1 5) can be further expanded as

,(L-l) is the determinant of the covariance matrix C, . By

where (m) , = m(m + 1). . . (m + u - 1) = r ( m + u ) / r ( m ) denotes the Pochhammer symbol, I,V is a polynomial in yJ, ..., Vo. I , , (L.l), and 8,. v = o , . . . , ( L - I ) are nonnegative integers of which not more than ( L - 2) are zeros.

Next the inverse Laplace transform was taken, and by means of properties of the generalized Laguerre polynomial, and by using of the definition, a, = EbR,/m,No , and x, = ~ ; E , / N , , the normalized SNR for each diversity branch can be expressed as Po = ma,2/R; ,..., = ma~L_,,/R~L_,, . Hence the generalized joint pdf of the normalized SNR, fa(ma,2/R;,..., ma~L.,,/R~L.,,), can be obtained as

308

where the k -th power of the Laguerre polynomial divided by in can be written as

Finally, by replacing p,(p,) in (18) with (2), the joint pdf of L correlated with Nakagami random variables can be obtained as

where {. + . } k is a symbol of the k-th power of a multinomial.

3. Average BER performance analysis

The evaluation of the average BER in a DS-CDMA system with L-variate correlated Nakagami branches can be accomplished by averaging P,(errorIa,, I = o , I , . . . , L , -1) over such L variates with the joint pdf shown in (8), expressed as

p r n = ' c r . . . c '<(errorla,, / = O , I ; . . , L , - ~ ) (21) x f,(Po,P, >. . .P,- l ) . dP,dPl... dP,,

Hereafter (21) involves L-fold integration and may be computed with the

Consider a dual-branch case as an example. The average BER can be method given by Alouini and Goldsmith [ 141.

calculated as

where the term V,, = det , and p(y,,m,G) can be represented as

where yo = E~R, IN, is the received average SNR per bit per diversity branch. In

309

a similar manner, (23) can be utilized to evaluate the average BER for a multiple-cell DS-CDMA system with the aid of the results shown in section 2.3.

Consider a dual-branch case as an example. The average BER can be calculated as

where the term 5, = det , and v(yo,m,6) can be represented as

where yo = E& 1 N, is the received average SNR per bit per diversity branch. In a similar manner, (21) can be utilized to evaluate the average BER for a multiple-cell DS-CDMA system with the aid of the results shown in section 2.3.

4. Numerical results

The effect of branch correlation on the average BER performance of a cellular DS-CDMA system is examined in this section. The Gaussian correlation model of an equally spaced linear array with an arbitrary correlation coefficient is applied in this paper [2]. It is of interest to note that the correlation matrix followed by the linear array has a Toeplitz form and is revealed as

(26) d av = e x p [ - ~ q ( i - j)’(-)’], i , j=O,... ,L-I a

where q 2 21.4 is a coefficient chosen from setting this correlation model equal to Bessel correlation model [ 131 with -3 dB point, and the parameter d I a is applied to determine the threshold level of correlation. In this numerical analysis, the assigned values of dlaare 0, 0.1, 0.2, andm , in which dla=O and dla --f m represent two extreme conditions, i.e., fully correlated and uncorrelated branches, respectively.

The processing gain is through set to be ~ = 5 1 2 in the numerical evaluation, wherein only the first tier ( z = 6 ) of interfering base stations is taken into consideration for the multiple-cell case. Fig. 2 and 3 illustrate the average BER versus branch SNR with 6=0.3 (the experimental value [2]), m = 3 and L = 4 for single-cell and multiple-cell systems, respectively. It is revealed from both figures that the larger the correlation coefficient, the lower the average BER, and that the performance is degraded when the correlation coefficient approaches the fully ccrrelated condition, d I A= 0.0 . Besides the correlation coefficient, it is obvious that the ave;age power decay rate 6 has great influence on the performance of DS-CDMA systems.

310

The effect of fingers number of the Rake receiver on BER is illustrated in Fig. 4, and 5 , in which the branch number are set with ~ = 2 and ~ = 6 , respectively. Both also indicate that the user number of surrounding cells have a significant impact on the average BER. The reason is that the value of MA1 will be increased in a multiple-cell system. The average BER with m = 4 is better than that of with rn = 2 , due to the fact that the extent of channel fading with rn = 4 is less than that with rn = 2 . It can also be observed from the results in Fig. 4, and 5 , in which the average BER with d / A = 0.1 is better by about 1.28 dB than that of with d l a =0.2 when the BER is set as B E R = I O - ’ . The BER performance of a single-cell CDMA system is always better than that of a multiple-cell system case, as expected. On the other hand, it is of interest to note that with the increase of L, the BER with d l R = 0 . 2 in a multiple-cell system shows a tendency to be better than that with dl a = 0.1 in single-cell system. This phenomenon is due to the fact that the correlation behavior becomes dominant. The average BER versus the number of users K is plotted in Fig. 6. It is obvious that the user capacity is not only deeply dependent on the fading figure, but also affected by the separation in the diversity branch (correlation coefficient).

5. Conclusions

The BER performance of multiple-cell DS-CDMA systems with Rake receiver over a correlated Nakagami fading channel has been analyzed in this paper. There is a new closed-form formula of the joint pdf for correlated branches have been derived in terms of the generalized Laguerre polynomial, and the results can also be applied to generalize Nakagami-m channels.

References

1. R. Price and P. E. Green, Proceeding ofthe IRE, Vol. 46,555, (1958). 2. P. Lombard0 et al., IEEE Trans. on Commun., Vol. 47, No. 1,44, (1 999). 3. B. Natarajan et a]., IEEE Cornmun. Lett., Vol. 4, No. 1,9, (2000). 4. M. Nakagami, Oxford, U.K.: Permagon, 3, (1 960). 5. V. A. Aalo, IEEE Trans. on Commun., Vol. 43,2360, (1995). 6. Q. T. Zhang, IEEE Trans. on khic. Technol., Vol. 48, No. 4, 1142, (1999). 7. E Patenaude, J. H. Lodge, and J. Chouinard, IEEE Trans. on Commun., Vol. 46, No. 8,

8. N. H. L. Chan and P. T. Mathiopoulos, IEEE J. select. Area Commun., Vol. 18, No. 6,

9.Emad K. Al-Hussaini and Iman M. Sayed, Wireless Personal Communications, Vol. 16,

10. Yawp0 Yang, Joy I. Z. Chen, and J. C. Liu, J. Chung Cheng Inst. of Technol., Vol. 30,

11. Q. T. Zhang, IEEE Trans. on Commun., Vol. 51, No. 11, 1745, (2003). 12. N. Eng and L. B. Milstein, IEEE Trans. on Commun., Corn-43, 1134, (1995). 13. W. C. Y. Lee, IEEE Trans. on Commun., Corn-21, 1214, (1973). 14. M. S. Alouini and A. Goldsmith, Proc. IEEE Int. Con$ Commun. ICC’98, Atlanta,

15. A. S. Krishnamoorthy and M. Parthasarathy, Annals Math. Statist., Vol. 22, 549,

985, (1998).

996, (2000).

115, (2001).

No. 1, 143, (2001).

GA, 459, (1 998).

(1951). 16. H. Buchholz, Springer-verlag Berlin Heidelberg N. Y; , Chapter IV, 90, (1969).

311

17. M. H. Fong, V. K. Bhargava, and Q. Wang, IEEE J. Select. Areas Commun., Vol. 14, No. 3, 547, (1996).

: Mobile unit I : Base station

I : Base Station > : Down link signal from BS 0 : Mobile unit

_--------_____

Fig. 1. Configuration of multiple-cell and cell geometry

Fig. 2. BER versus SNR for single-cell system with different values of 6 , d / d , and fixed L = 4 .

312

Fig. 3. BER versus SNR for multiple-cell system with different values of 6 , d l R , and fixed L = 4 .

0 1

h

5 p- 001

n P a L

1E-3 UI c .- a

1 E-4

5 10 15 20 25

7, (dB)

Fig. 4 . BER versus SNR for correlated Nakagam channel comparison between single-cell and multiple-cell with different d / R values. L = 2 .

313

Y, ( d W Fig. 5. BER versus SNR for correlated Nakagami channel comparison between single-cell and

multiple-cell with different d / /2 values. L=6.

, . , . I . , ' , . ,

15 20 25 30 35 40

K (user's number)

Fig. 6. BER versus number of users for correlated Nakagami channel comparison between single- cell and multiple-cell with parameters L = 2, 6 = 0.05 , N = 5 12 , and SNR = 15 dB .

Cellular Networks

MULTIMEDIA TRANSMISSION OVER THIRD GENERATION CELLULAR NETWORKS

ANTONIOS ALEXIOU, CHRISTOS BOURAS AND VAGGELIS IGGLESIS Research Academic Computer Technology Institute and Computer Engineering and

Informatics Department, University of Patras Riga Feraiou 61, Patras, 2622I Greece

The scheme of real time streaming video is one of the newcomers in wireless data communication, raising a number of new requirements in both telecommunication and data communication systems This scheme applies when the user is experiencing real time multimedia content This paper has as main target to study the performance of video transmission over the third generation cellular networks In particular, we examine the performance of UMTS Dedicated Channels (DCHs) for real time MPEG-4 video transmission in Downlink direction Finally, we examine if real time video transmission in conjunction with Internet traffic is applicable in UMTS radio interface

1. Introduction

UMTS constitutes the third generation of cellular wireless networks which aims to provide high-speed data access along with real time voice calls [ 11. Wireless data is one of the major boosters of wireless communications and one of the main motivations of next generation standards [2]. Bandwidth is the most precious and limited resource of UMTS and every wireless network. Video applications produce large amount of data. As a result, video is transmitted in compressed format to reduce the generated data rates. Among the used compression techniques, MPEG-4 is the standard that has recently gained a considerable attention [3] .

This paper has as main target, to examine the performance of UMTS Dedicated Channels (DCHs) for real time MPEG-4 video transmission. The results demonstrate that video quality can be substantially improved be preserving the high priority video data during the transmission. In 131 the authors suggest a dynamic bandwidth allocation scheme for MPEG video sources suitable for wireless networks. The proposed algorithm exploits the structure of the MPEG video stream and allocates bandwidth on a scene basis. This results a high bandwidth gain, which affects the overall network performance.

317

318

The specific contribution of this work is that provides both an overview on MPEG-4 video transmission over third generation cellular networks, and an indication about how friendly is the behavior of video transmission over UMTS towards any other Internet application that coexist in the same channel.

This paper is structured as follows. Section 2 presents an analytical computation of Packet Service Time for MPEG-4 video traffic. Following this, section 3 reviews the main features of the simulation model. Section 4 is dedicated to the experiments results. Finally, some concluding remarks and planned next steps are briefly described.

2. Analytical computation of packet service time for MPEG-4 video traffic

In this section we present an analytical computation of the time required for any packet of a given video sequence to travel from the Radio Network Controller to the mobile user. The video traces we use, are taken from [ 5 ] . The packets of the video sequence do not have a constant size. The size of the MPEG-4 packets which are being transmitted over the UMTS air interface is presented in Figure 1. The video sequence is in QCIF format (176x144 pixels) at the PAL frame rate of 25 frames per second. The average packet size is 758 bytes. The above- mentioned video traffic is going to be also used in the simulation that it is described in the following sections.

~~~~~~~~ w,

Figure 1 MPEG-4 Packet Size Considering an MPEG-4 video transmission from a fixed Internet node to a

mobile user (Downlink) (Figure 4), as the packets leave the RNC and arrive at the Node B they queue up in order to be broken down into smaller size packets. Every PDCP PDU is segmented into multiple RLC PDUs of fixed size. Each of these PDUs fits into a transport block in order to be transmitted over the air. According to 141, the size of the RLC PDUs is 40 bytes.

319

Based on the analysis presented in [4], in Downlink, for any RLC PDU that transmitted over the air the RNC receives a status report of the UE 68ms after its transmission. For the opposite direction, since generally, the uplink TTI is twice as large as in the downlink direction, the UE receives a status report 40ms after its transmission.

In this section we determine the minimum IP packet service time for the MPEG-4 video traffic that we use in the simulation. This time is comparable with the delay in Radio Access Network (Time required for any packet to travel from RNC to UE). This time is calculated as follows. The number of RLC

PDUs an IP packet is segmented into is: NpDU = . The

maximum RLC PDUs that can be transmitted within one TTI is: L,, = 8 . Consequently, the number o TTIs required for the transmission of the packet is:

Packet -size(bytes) 40bytes

N P D U

Lmax N , =-

Since the TTI in downlink is lOms, the minimum time required for the transmission and reception of a whole packet is: T,,, = 40ms + [ N , - 11.10 . For the MPEG-4 video traffic that we use in the simulation, the minimum time required for the transmission and reception of any packet of the video sequence is presented in Figure 2. The y-axis shows the packet service time in msec, while the x-axis presents the packet sequence number.

~~~~~~~~~~~~~~~

-$

w

9% " I_x _ _-_ -1 x1 _" "" - I 4 " I b B ~~~~~ am

Figure 2 Minimum time required for the transmission and reception of any packet of the video sequence

Furthermore, Figure 3 presents the pdf of the minimum time required for the transmission and reception of any packet of the video sequence. The x-axis presents the packet service time, while the y-axis presents the pdf. Since all IP packets do not have a constant size, the minimum time varies and has a maximum at Tmin = 40 msec with value pdf = 0,5087. This means that the

320

50,87% of the packets have a minimum delay in RAN of 40 msec approximately. The average Tmin is 54,39 m sec.

Figure 3. Pdf of the IP packet service time.

3. Simulation model

This Section reviews the main features of the simulation model that has been implemented by using the ns-2 simulator. The performance of DCHs is evaluated for MPEG-4 traffic with different characteristics. Furthermore, in order to exploit the performance of DCHs we consider as background traffic to the system, HTTP and SMTP applications. During the simulations we make the following measurements: a) Delay in RAN (ffom RNC to UE) and b) Throughput in Wireless Link: Bits transferred to UE per unit time in bitdsec.

The addressed scenario comprises a UMTS radio cell covered by a Node-B connected to an RNC. The simulation model consists of a UE connected to DCH as it is shown in Figure 4.

*-

Figure 4. The simulation model In our simulations, we use a 384Kbit-DCH in the downlink and a 128Kbit-

DCH in the uplink direction. The TTIs are lOms and 20ms in the down- and uplink direction, respectively. A real time MPEG-4 video streaming generated in node 2 and through DCH (downlink) heading to UE1 using the UDP as transport protocol (foreground traffic). Also, HTTP and SMTP traffic generated in node 2 heading to UE1, using the TCP as transport protocol (background

321

traffic). The video sequence is following the MPEG4 standard, in QCIF format (176x144 pixels) at the PAL frame rate of 25 frames per second. The video traces we use, are taken from [5 ] . The duration of the video applications is 200 seconds.

4. Experiments

This section is dedicated to describing the results in terms of performance of real time video transmission over UMTS DCHs. Firstly, we present the performance parameters for MPEG-4 video transmission over UMTS DCHs and secondly, we present the results for transmission of MPEG-4 video traffic that coexist in the same DCH with Internet TCP traffic.

4.1. MPEG-4 video transmission without background trafflc

Figure 5, presents the delay in RAN. In other words, the latter represents the packet service time that it has already computed, in an analytical way, in section 2 . The average packet service of the simulation has a value of 61,7 msec while the average packet service time that computed in section 2 is 54,39 msec. As it is shown in Figure 5 around Packet Sequence Number 800 we can see a delay spike. This is due to the fact that these packets have very large size compared to the others and therefore, a great number o TTIs are required for the transmission of the packets.

f

*'.em m<Nn*"w^l -* *% -". * * - "" " ""I x

m ***-*,,' " * u ~ ~ ~ ~ ~ w y l l y I l ~ m I Y ~ y ~ y y y ~ y

- . I % m *a*ee."+-*r*r

u4

Figure 5 RAN delay for the Video Transmission The throughput in wireless link is depicted in Figure 6. The y-axis presents

the throughput in bps while x-axis represents the duration of the simulation. The red line represents the average throughput in the wireless link. The average throughput has a value of 205 kbps while the downlink bit date of the dedicated channel is 384 kbps.

Figure 7 illustrates a comparison of pdfs of the packet service time for both the analytical computation (presented in section 2) and the simulation results. The red line represents the experiment results while the blue one represents the

322

analytical results. As it is depicted in Figure 7, the simulation results are very close to the analytical results. Furthermore, Figure 7 indicates that the majority of the packets achieve a service time lower that 80ms and only a small number of packets seem to achieve a service time higher that 100ms. This signifies that the majority of the packets that reach the mobile user are characterized by low delay and consequently, the quality of the video sequence that this user will see in his terminal is satisfactory in comparison to the sequence originally sent by the transmitter.

x̂II lllllxIl I

Y- rd qe

__I I x "Axxr.-* -XI - - "_L".. l",IAxIIx P *D *D #$ 8s * m raa 1(8 ~ ~ ~ ~ ~ * , ~ ~ ~ ~ ~

Figure 7 Comparison between Analflcal Computation and Simulation Results

4.2. MPEG-4 video transmission with the presence of additional background traffic

Figure 8 presents the throughput in the wireless link during the end-to-end MPEG-4 video transmission. The y-axis shows the throughput (in bps) while the x-axis counts the time. In this figure are shown three elements: a) The Video Streaming throughput (navy blue line), b) The throughput of the SMTP traffic following a Pareto distribution (red line) and c) The throughput of HTTP traffic (green line).

323

The maximum throughput of the SMTP and the HTTP traffic is 50 Kbps. Figure 9 shows the total throughput in the Wireless link, which is the sum of the Video and SMTP traffic during the first time interval (from second 50 to IOO), and the sum of the Video and HTTP traffic during the second time interval (from second 100 to 150). As it is shown in Figure 9, the total throughput in the wireless link is always lower than the downlink bit rate of the DCH which is 384 kbps. Considering the delay in RAN, the presence of additional background TCP traffic produces results similar to the results presented in the previous subsection and in particular in Figure 5.

rma***u****u*

* jl x * I% XB z* w

Figure 9 Total throughput in Wireless Link (wth background traffic)

An obvious observation that comes out from Figure 9 is that in the time intervals from second 15 to 25 and I50 to 180 there is heavy video traffic in the DCH. If we add additional traffic to these intervals, it is possible the network to be congested and this probably could affect the video sequence in the receiver.

Figure 9 indicates that the UDP video traffic shows a friendly behavior towards the TCP traffic and in particular the SMTP and HTTP traffic that coexist in the same dedicated channel. At this point, it has to be mentioned that the above-described scheme is applicable in UMTS as long as the DCH has enough available bandwidth in order to serve the mixed traffic conditions. In situations where the total bit rate of the traffic is higher than the downlink speed of the DCH, the scheme becomes unstable and this causes serious problems mainly in the transmission of the video sequence. Since Internet traffic is mostly

Figure 8. Throught in Wireless Link (With background TCP trafic

324

TCP traffic, if packets are lost the TCP protocol infers that there must be congestion in the network and this leads to the retransmission of the packets until the correct reception in the end user terminal. In addition, concerning the UDP video transmission, if a number of packets of the video sequence are lost then the synchronization between encoder and decoder is broken, and errors propagate through the rendered video for some time.

5. Conclusions and future work

In this paper we present some evaluation results for the performance of UMTS for different traffic types including MPEG-4 video traffic, SMTP and HTTP traffic. In the presented simulations, we use both TCP and UDP as the transport protocols to the system. This paper proves that the scheme of real time video transmission that coexists in the same channel with Internet TCP traffic is applicable in UMTS. However some problems occur when the total bit rate of the traffic is reaching the downlink bit rate of the transport channel. As a consequence, some packets are lost. The seriousness of the situation depends on the number of lost UDP packets and their time variance. To these situations it is necessary for the sender to adapt the transmission rate based on the current network conditions and this will be the step that follows this work.

Furthermore, among the future steps is to evaluate the performance of MPEG-4 video transmissions over High Speed Downlink Packet Access transmissions. HSDPA supports the introduction of high bit rate data services and will increase network capacity, while minimizing operators’ investment.

References

1.

2.

3.

4.

5 .

H. Holma, and A. Toskala, “WCDMA for UMTS: Radio Access for Third Generation Mobile Communications”, John Wiley & Sons, 2004. P. Chaudhury, and W. Mohr, “The 3GPP Proposal for IMT-2000”, TEEE Communications Magazine, December 1999, pp. 72-81. Y. Iraqi, and R. Boutaba, “A Dynamic Bandwidth Allocation Algorithm for MPEG Video Sources in Wireless Network”, DialM99, 1999, pp. 86-92. 3GPP TS 34.108: “3rd Generation Partnership Project; Technical Specification Group RAN; Delay Budget within the Access Stratum (Release 1999)” F. Fitzek, and M. Reisslein, “MPEG-4 and H.263 Video Traces for Network Performance Evaluation”, IEEE Network, Nov.-Dec. 200 1, pp. 40-54.

DISTRIBUTED CONTENT SHARING IN CELLULAR NETWORKS

BALAZS BAKOS, LORANT FARKAS Nokia Research Center, P.O. Box 392, H-1461 Budapest, Hungary

JUKKA K. NURMINEN Nokia Research Center, P.O. Box 407, FIN-00045 NOKIA GROUP, Finland

KALMAN MAROSSY Nokia Technology Platforms, P.O. Box 392, H-1461 Budapest, Hungary

In this paper the Parallel Index Clusters (PIC), a clustered application layer architecture is presented. Based on the performed simulation and analysis it is proposed as an appropriate content sharing solution in decentralized wireless/cellular peer-to-peer environments. In earlier work a basic review of maintenance aspects has been performed [2]. We have also proposed and compared cluster topology alternatives. Here we perform the optimization of these topologies based on the performance metrics: minimal traffic and load balance. The optimal cluster size is determined as a function of metrics in various traffic scenarios.

1. Introduction

Peer-to-peer (P2P) applications, such as KaZaA or Gnutella, are measured to be generating most of the Internet traffic today [7]. In comparison to fixed Internet, cellular networks have less bandwidth and are more sensitive to the amount of transferred data. The mobile devices are frequently switched off or they are outside of network coverage. Therefore enhancements and possibly new protocols are necessary for this type of applications to work in cellular networks. In earlier work [ 11 the performance of a modified Gnutella implementation over wireless topologies has been analyzed.

Gnutella is a brute-force technique with a very simple implementation. Through the introduction of more sophisticated techniques it is possible to reduce the amount of traffic in the network, share the search load evenly between the devices and get reliable results reasonably fast. Gnutella can also be enhanced through clustering. The authors of [ 5 ] propose several clustering

325

326

criteria: interest-based and based on geographical proximity of the nodes. In addition to clustering other approaches in the literature are keeping the topology uncontrolled but exploiting user behavior patterns through query caching, locality exploiting and the like e.g. [6], exploiting more intelligent query strategies instead of flooding [9] and introducing controlled replication in order to increase the efficiency of the search [lo].

Parallel Index Clusters (PIC), another approach based on clustering, has been first proposed in [3] and [4] as a means to reduce overall load in content sharing peer-to-peer networks. In [2] we analyzed the performance of the approach in a cellular network context and proposed candidate topologies. We introduce here more appropriate clustering criteria and perform an optimization of the network traffic both in terms of average load and balance.

Section 2 describes the operation of the PIC network in general terms. Sections 3 and 4 focus on index update, and respectively, query topologies. Section 5 contains analysis and simulation details from which the optimal cluster size can be determined. Section 6 concludes this paper and outlines further possible directions to consider.

2. Parallel Index Clusters (PIC)

Defined in [2], the parallel index clusters architecture is an application layer network in which there are three different kinds of traffic: search, index update and maintenance. Search traffic is used for content retrieval, like in Gnutella: queries are submitted and query hits are sent back to the originator. But instead of Gnutella-type search flood, in PIC the search traffic is routed only on well-defined links and virtually in one step the originator receives all the query hits. In other words, search is controlled in a distributed fashion, on the level of the cluster and the node. To support this, another mechanism is introduced, that of content indexing, which is also controlled in a distributed fashion. There are strict rules to whom a node sends the index update.

/

Cllsbr 1 f Cllsler 2 I I

Figure 1 . Parallel Index Clusters example

327

Figure 1 presents a small PIC network with 2 clusters and 9 nodes. As can be seen, nodes are divided to clusters that are compact regions of the network, being connected through the so-called index links: application layer logical connections between the nodes, which might be permanent or completely random. Within a cluster theoretically the index tables of the nodes (i.e. tables of content information stored in the cluster) is the same. Whenever a node shares or withdraws content, all the co-cluster nodes will be informed about the event through the so called index update messages. The set of index links and nodes are organized in the so-called index update topology, an application layer subnetwork.

Search traffic consists of query and query hit messages. Queries are generated when a user is searching for content. Query hit messages are generated when at least one content is found matching the request. The payload of the query hit messages is a list containing the content names and properties (size, date of creation etc) and the address of the node(s) storing them. A node can have several documents matching the search criterion. It is possible to use wildcards in the query string. In an extreme case the query string could be one wildcard (‘*’), case in which the network would return a complete list of the available content. For small networks this scenario might actually be realistic: a newly joined user could be interested in finding out the kind of information stored on the phones of others.

Index update traffic is generated whenever a node shares new content or withdraws previously shared content. Special cases of index update traffic can occur when nodes join the network: as part of the join procedure an index update is generated containing the joining node’s complete shared content list in some specific cases.

Depending on the index update topology and the delays of the individual links there is an inherent latency in the content information propagation, meaning that there might be a content already available on a node but still non- existent according to those co-cluster nodes to which the update message did not propagate yet.

As a consequence of indexing the search can be achieved in a much easier fashion instead of a “blind search”: a very efficient full search is possible by a node by simply querying one node in each cluster.

The maintenance traffic has been briefly described in [2]. In this contribution its effects will be neglected.

3. Index update topologies

The goal of index update is to propagate the index changes to all nodes within a cluster. This should be done in such a way that a proper balance between several

328

conflicting objectives is reached. The objectives are to minimize the time to propagate the changes, minimize the network load generated by the update traffic, divide the load between the nodes in such a way that on the average all of them are equally loaded and avoid unnecessarily complicated solutions as these are error prone, hard to implement and maintain, and consume phone’s memory and CPU resources.

Star, ring, random N, balancedplanned N alternatives have been considered in [ 2 ] . These are briefly described below.

In a star the node generating an index update sends it to all co-cluster nodes. In a ring the index update message propagates in a ring fashion: every node receiving the message propagates it further. Random N is a topology in which each received index update message is forwarded to randomly chosen N nodes. Finally, balancedplanned N topology works in a way that ensures that the next N neighbors will be “planned” so that there won’t be any duplicated messages. Balanced N adds in addition perfect balance between the nodes in a cluster even for non-uniform index update behavior of the individual nodes.

4. Query topologies

Queries are messages sent to search for content in the network, issued through the query links. The different query links within a PIC network construct a query topology that can have different forms. We have considered star and ring as options.

The query topology has a primary influence on the duration of the search operation. In the case of star query topologies the node generating the query will send the query message to all the clusters except the own one. In case of a large number of clusters this generates high load on the node generating the query. On the other hand, query results are obtained in a fast manner: from each cluster the query results arrive in one hop to the query generator from all recipients.

In the following we will concentrate with our analysis on the star and ring index update topologies, which are the two extreme situations: very fast index updating at all costs (even high traffic allowed) versus very lowhighly balanced traffic at all costs (even slow index updating allowed). In this case the following options are possible, where the notations consist each of an index topology (first term) and a query topology (second term): 0 Ring/Ring: FIL, FSL (2 options) 0 Ring/Star: FIL, NSL

Star/Ring: NIL, FSL ( 2 options) 0 Stadstar: NIL, NSL

329

5. Traffic analysis and simulation: determining optimal cluster sizes

In this section we analyze and simulate a number of scenarios in order to estimate the optimal cluster size from the traffic viewpoint and the scalability of this solution. In this contribution we restrict the focus to the index update and query traffic.

5.1. Index update traffic

We denote in the following with left hand superscript the traffic origin (o=other, wown), with left hand subscript the traffic type (u=index update traffic, s=search traffic), with right hand superscript the cluster index and with right hand subscript the node index.

ti the index update traffic experienced by node i, from which ,“ti is the index update traffic generated by other nodes, ,“ti is the index update traffic generated by itself. Let’s denote with ,“bi the “net traffic” of node i generated by other nodes and ,”bi the “net traffic” of node i generated by itself. We differentiate between traffic and net traffic in the following sense: traffic is the sum of sent and received bytes due to owdother index update in a given time interval, net traffic is the number of bytes of the index update messages that have to be transferred (sentheceivedforwarded) in the network in a given time interval.

In the following we neglect the additional traffic generated by the transport layer in addition to the application layer messages (acknowledge messages, retransmissions etc). The index update traffic experienced by node i member of clusterj in the case of star index update topology is given by:

Let’s denote with

k t i

where cj is the size of clusterj. The first term represents the node generating the index update, the second term signifies index updates generated by other nodes which are only received, not propagated fiuther (NIL-s). Nodes generating large index update traffic have on average larger processed index update traffic than nodes generating lower index update traffic. Own index update traffic in the case of star has (c, -1) times larger influence on the overall index traffic of a node compared to the index update of co-cluster nodes.

Thc indcx update traffic experienced by a node i member of cluster j in the case of ring index update topology is given by:

330

In (2) we assume that the route of the index update message is a closed path, in other words, it propagates back to the originator. The last step can be excluded if the address of the originator is included in the message.

The overall index update traffic of a node in a ring is 100% balanced, not depending on the size of the individual index update traffic. The reason is that the index update messages propagate along the ring, fiom the generator node back to itself, so every node receives it and forwards it (FIL-s). It means that in a ring all nodes in a cluster will cany the same amount of index update, regardless of their own index update traffic.

Assuming that the upstream neighbor of the originator withholds the message, the index update traffic experienced by node i member of clusterj will be the following:

k#i,k#i-1

in which we assumed that node i-I is the downstream neighbor of node i. In this case the index update traffic is not completely balanced: the own traffic and the traffic of the downstream neighbor have smaller influence due to their weigh 1 in (3).

5.2. Query trafflc

Let’s use the same notation for the query topology in which s replaces u, meaning ‘search’ instead of ‘index update’. The search traffic experienced by a node i member of cluster j the case of star query topology is:

where I is the number of clusters in the network and in case of ring query topology is given by:

in which :bk is the own search traffic generated by cluster k. We also assume in (5) that the query messages propagate along a full circle, in other words, they arrive back to the cluster in which they had been generated. The parenthesis in the first term is composed of 2 terms: the outgoing query traffic of node i and the incoming query traffic of node i. The two terms are different since all the outgoing traffic of node i is generated by itself, but the incoming traffic of node i might as well be routed towards any other node member of clusterj with equal

331

probability. The division by cj in the second term of ( 5 ) is performed based on the assumption that the incoming queries arrive equally distributed to node i and other co-cluster nodes and generate on average equal amounts of traffic.

It can be observed here as well that the ring query topology is more balanced than the star topology: own query traffic has (cj + I) times larger weight than other query traffic, whereas in the star case the weight of the own query traffic is (l-f)cj times larger. In other words, even though random ring query topology is more balanced than star query topology, still it is far from being 100% balanced, as in the case of ring index update topology. The other nodes’ search traffic on the other hand is hlly balanced within a cluster, if the queries are uniformly distributed to the nodes of a cluster.

If we assume that the upstream neighbor of the query generator withholds the query message, the search traffic experienced by node i member of cluster j is:

5.3. Traffic objectives and balancing strategies

An application layer architecture of such a highly controlled topology can

Balance the load of the individual nodes as much as possible, without degrading too much the overall traffic, if the two criteria lead to different solutions. Decrease the overall traffic of the network to the smallest possible amount. The second criterion’s interpretation is obvious: we look for the global

optimum of the sum of all kinds of traffic. However, the interpretation of the first criterion can be made in two different ways:

Balance the overall other nodes’ traffic of each node, meaning that there should be on average the same amount of traffic on each node generated by other node traffic; Balance the overall nodes’ traffic on each node, meaning that the sum of incoming and outgoing bytes in a node in a given time period should be on average the same for each individual node. There could be different arguments for both mentioned criteria. But from a

practical viewpoint we have seen above that the only kinds of traffic which are balanced within a cluster are the other nodes’ search traffic and the overall index update traffic, no matter of the used query and index update topology. Therefore in the following the extension of the balance of the other nodes’ query traffic

serve two aims: 0

332

and overall index update traffic over all clusters will be chosen as one of the objectives. Its most general form is the following:

:ti+uti = c, i = 1, n (7)

in which c is a constant. More restrictive than (7), but easier to handle, is the following set of two objectives:

- u t i = u ~ , i =1,n

(9) 0 0 s t i = s C , i =1,n

in which c and ,“ c are constant and

.c+,”c = c (10)

If (8) and (9) are satisfied, (7) holds as well. The second objective, to keep the overall traffic of the network at a

minimum level, can be formulated as follows:

2 ti = 2 ( u ti+,ti = 2 (rt i+it i+:ti+:ti = min i=l i=l i=l

The two formulated objectives constitute an optimization problem in the space of network topologies defined from the following elements:

number of clusters

form as:

distribution of the nodes in the individual clusters. The index update balance condition (8) can be written in a more elaborate

“i . - C ,,t,!=,,c, j = 1,c i=l

where c is the number of clusters, i spans the nodes from clusterj, there are nj nodes in clusterj.

The query balance condition can be understood based on Figure 2. The case of 4 clusters can be viewed: cI...c4. Inside the clusters we have indicated the overall generated own search traffic, ,t , i = 1,...,4. Under the clusters we marked the search traffic experienced by each of the clusters, which is in fact the sum of query traffic of the other clusters. The objective is satisfied, provided:

w i

w 2 w 3 w 4 - w l w 3 w 4 w l w 2 w 4 w 2 w 3 ,t +,t +,t -,t +,t +,t =,t +,t +,t =:t’+,t +,t (13)

333

Reducing one by one the common terms from (13), the only possible solution is the case of equal search traffic generated by each cluster:

that can be translated to the individual nodes as below:

W where I is the number of clusters, i spans the nodes from cluster j, ,C is a constant, representing the own search traffic of a cluster, the sum of queries generated by members of the cluster. In conclusion, a sufficient condition for the first objective is given by (12) and (1 5).

y = : t * t : t 3 t:t4 y = y t : t 3 t 3 4 , O t 3 = y + y + : t 4 ;t4=:tl+:t2+33

Figure 2. Query load of 4 clusters If we reached a solution in which the first objective was satisfied, i.e. the

other nodes’ search traffic and the overall index update traffic is balanced, the first, second and fourth terms of (11) would be fixed. So the only remaining term would be the third term that is the sum of the own search traffic of the nodes, which is an input parameter, depending on the scenario. Results that the global optimum with the constraint of balanced overall index update traffic and balanced other nodes’ search traffic can be obtained through selecting the solution generating the smallest overall index update and other nodes’ search traffic. This is related to the size of the clusters.

5.4. Optimal cluster size and optimal number of clusters

A formal solution for global optimality with the constraint of balanced overall index update traffic and other nodes’ search traffic is provided below. We assume nodes generating equal amount of index update net traffic b and equal amount of search net traffic b .

We derive the optimal cluster size for starlstar index updatelquery topology in the following.

The optimal cluster size for stadstar can be evaluated starting from (2) and (5), the overall traffic being n times the expression, where n is the network size:

334

= 2 ( c - l ) , b + 2 - 1 , b (1-) From (1 6) we can determine the optimal cluster by formally differentiating

as a fimction of c: t t

-=2,b- a i 2n ac c2

More generally:

(18) holds also for the balancedplanned N index update topologies. In other words, we have a general expression in closed form for the optimal cluster size for each topology that depends on the size of the network, the query traffic and the index update traffic.

SWUR - Figure 3. Average load of nodes as a function of the clusters, planned 3

In order to illustrate (18) we have considered a network of 100 nodes generating uniformly update and query load. A total number of 1000 events have been generated, of query and index update type. We used the planned 3 topology, several different cluster sizes and several different queryhpdate loads, as presented in Figure 3.

Figure 3 indicates that for low values of searchhpdate rates the average load measured as the average number of messages on one node of the network with large cluster size is significant and it decreases as updates become more frequent. On the other hand, for small cluster sizes the opposite is true, namely that the average load on the node increases with the increasing searchlupdate

335

rate. However, this increasing behavior is much less significant than the decreasing behavior of the large cluster cases.

Figure 3 reveals another interesting phenomenon as well which we denoted as the ‘square-root law’: the load is almost constant no matter of the searchhpdate ratio if the number of clusters is the square root of the number of nodes. In fact it results directly from (18): introducing in (18) the cluster size equal to the square root of the network size and assuming

11 b+, b = b = const (19)

we obtain it in a straightforward manner:

Traffic indeed does not depend on the SR/UR ratio, provided the sum of the two traffic types is constant. We experimented with a number of networks in all of which the cluster size has been chosen to be the square root of the number of nodes, for planned 3,

Figure 4. Average load of nodes in mesa geshequest as a function of number of clusters, planned 3

Uoptirnai, ring index, star query Eoptimal, ring index, ring query Msqrt (all)

optimal, star index, star query

Figure 5 . Average load, 256 nodes, all topologies

336

The outcome is represented in Figure 4. As we expected, the load per node is uniform when the number of clusters is the square root of the number of nodes for every number of nodes. This is a general rule. The average loadrequest decreases as the cluster size increases, to a certain point.

In order to compare the square root compromise solution with the optimal one, we have conducted a series of simulations with various network sizes and various index update and query topologies (ring, star), shown in Figure 5 . Optimal solutions clearly outperform the square root compromise except when index update and search traffic are equal, case in which the two methods provide the same performance. However, it is to be noted that in order to keep the traffic optimal, maintenance traffic is needed, to increase or decrease the cluster size, move nodes to new clusters, delete or create clusters. We do not deal with these aspects in this contribution.

6. Conclusions and further work

We have analyzed the performance of content sharing PIC networks for uniform traffic and cluster size. We have been able to identify the candidate topologies and analyze their performance. An appropriate cluster size fulfilling the constant traffic has been found.

Deterministic ring is as the most balanced index update topology. On the other hand, balanced N is proposed as an alternative for the case when quicker index updating is a requirement.

Clustering should be done based on traffic minimization. We have identified the square root law as a good compromise when the ratio between the traffic types is not known and the optimal cluster size for the case when we have the precise value of this ratio.

Further work will focus on four different research directions: including the link management traffic in the model, the analysis of non-uniform cluster size and non-uniform node traffic case; definition and analysis of node profiles, and the case of inactive nodes. Later on a comparison with earlier work concerning enhanced brute-force searching techniques [ 11 is also planned.

References

1. G. Csucs, J.K. Nurminen, B. Bakos, L. Farkas, Peer-to-peer Protocol Evaluation in Topologies Resembling Wireless Networks. An Experiment with Gnutella Query Engine, ICON’2003, Sydney, Australia

2. K. Marossy, G. Csucs, B. Bakos, L. Farkas, J.K. Nurminen, Peer-to-peer content sharing in wireless networks, PIMRC’2004, Barcelona, Spain

337

3. B. F. Cooper, H. Garcia-Molina, Modeling and Measuring Scalable Peer-to- peer Search Networks, SIGCOMM’2002

4. B. F. Cooper, H. Garcia-Molina, Ad hoc, Self-supervising Peer-to-peer Search Networks, Technical Report, Stanford University, 2003, www- db . stanford. edd-cooperblpubs

5 . T. Hu, A. Sereviratne, General Clusters in Peer-to-Peer Networks, ICON’2003

6 . K. Sripanidkulchai, B. Maggs, H. Zhang, Efficient Content Location Using Interest-Based Locality in Peer-to-Peer Systems, Infocom’2003

7. Peer-to-peer file sharing - The impact of file sharing on service provider networks, white paper, December 2000, http://www.sandvine.com

8. M. Portmann et al., The Cost of Peer Discovery and Searching in Gnutella Peer-to-Peer File Sharing Protocol, ICON’2001

9. H. Meling, A. Montresor, 0. Babaoglu, Peer-to-peer Document Sharing using the Ant Paradigm, NIK2001, Tromso, Norway

10. E. Cohen, S. Shenker, Replication Strategies in Unstructured Peer-to-peer Networks, SigComm’2002, Pittsburgh, Pennsylvania

ON UMTS HSDPA PERFORMANCE

PAWEL MATUSZ Gdansk University of Technology.

Faculty of Electronics, Telecommunications and Informatics ul. Gabriela Narutowicza 11/12, 80-952 Gdansk, Polamd

JOZEF WOZNIAK Gdansk University of Technology,

Faculty of Electronics, Telecommunications and Informatics ul. Gabriela Narutowicza 11/12, 80-952 Gdansk, Poland

HSDPA (High Speed Downlink Packet Access) is a very important stage in the evolution of UMTS. It allows downlink transmission with peak rates theoretically up to 14.4 Mbps. To enable this, new features and mechanisms were introduced in UMTS physical and MAC layers in release 5 of the standard. The goal of this paper is to analyse some of these features in MAC and suggest solutions to enhance HSDPA performance. We propose a two stage HSDPA scheduler in MAC-hs, consisting of Proportional Fair and dynamically adjusted Weighted Round Robin scheduling algorithms. We also present some configuration guidelines to optimize HSDPA performance and radio resource utilization.

1. Introduction

When the telecommunication market evolves during the next few years, data services are anticipated to have an enormous rate of growth and are likely to become the dominant source of traffic in 3G systems. To satisfy the need for higher bit rates required by these services, new features and mechanisms to increase spectral efficiency have to be added to existing systems in order to increase the capacity of the radio access network (RAN). In release 5 of UMTS (Universal Mobile Telecommunication System) 3GPP introduced such a solution - HSDPA (High Speed Downlink Packet Access) [ l ] - which allows downlink transmission with peak bit rates theoretically up to 14.4 Mbps per cell. The key idea of HSDPA is to increase downlink packet data throughput with methods partially known already from GSMEDGE (Global System for Mobile Communications / Enhanced Data rates for Global Evolution), including fast link adaptation and physical layer (LI) retransmissions.

338

339

A new MAC (Medium Access Control) [2] layer was introduced and placed in Node-B in UTRAN (UMTS Terrestrial RAN) [4], so HSDPA downlink scheduling and retransmission decisions are done as close to the radio interface as possible. On the radio interface (uu), new downlink transport channels are used: HS-DSCH (High Speed Downlink Shared Channel) to carry MAC-hs PDUs (user data) and HS-SCCH (High Speed Shared Control Channel) to carry HSDPA control information. HARQ feedback information and CQI (Channel Quality Indicator) are transmitted by UE (user Equipment) in uplink using HS-DPCCH (High Speed Dedicated Physical Control Channel).

There are many analysis (e.g. in [5], [6], [lo] or [12]) of HSDPA performance concerning the physical level, spectral efficiency etc. but not much work has been done to evaluate the performance of proposed solutions from the protocolar point of view in a real-time system with memory and computational power constraints. The goal of this paper is to provide such analysis of recommended algorithms and propose some enhancements. We propose a 2-stage MAC-hs scheduler, which uses dynamically adjusted Weighted Round Robin (WRR) scheduling to choose data from queues belonging to one UE and slightly modified Proportional Fair (PF) scheduling to decide which UE to schedule in each TTI (Transmission Time Interval). We point out some weaknesses of PF scheduling in case of new HSDPA users entering a cell and in case of not infinitely backlogged MAC-hs queues (which often happens in practice) and describe situations, in which radio resources are inefficiently utilized. A better algorithm is a matter of further study, although we provide some configuration guidelines which should be used to improve PF scheduling efficiency. All proposed solutions were evaluated using a real-time UMTS simulator, capable of interoperating with commercial UMTS test equipment.

2. MAC-hs scheduling algorithm

There has been some investigation concerning an efficient scheduling algorithm for HSDPA (e.g. in [6], [7] and [lo]). The general conclusion is that Proportional Fair (PF) [ l l ] is a good and efficient algorithm, which takes in to account instantaneous radio conditions e.;perienced by each UE and the average service rate already received by them. Using the Exponential Rule (ER) can improve PF fairness in some cases, as described in [6]. We decided not to use ER, because fairness is increased only in some specific conditions and for specific UE movement and fading patterns and is not significant on the average. Of course, it would be good to use the ER since it does provide some improvement, but on the other hand it requires complicated fractional calculations which are usually not easy to perform in a real-time system. The ER does not profoundly change PF, only slightly modifies the UE selection condition so our

340

conclusions can be in general applied also to a scheduler incorporating ER as well. In 171 it has been proven that PF is not stable in some situations - i.e. although queue service rate is equal (or even slightly higher) to the rate of data inserted in the queue, the length of the queue can grow infinitely. In HSDPA, the service rate of each queue changes dynamically on a per-TTI basis, so may the amount of data delivered to queues, and no throughput guarantees are given. Therefore HSDPA performance is marginally (in fact imperceptibly) affected by this problem.

It appears that in a real HSDPA implementation PF scheduling alone is insufficient, because there may be several HSDPA flows with different priorities configured for each UE. Therefore, we propose a 2-stage scheduler; the first stage is a PF scheduler which is used to select the UE that should be served in the current TTI, and the second stage is a WRR (Weighted Round Robin) scheduler used to schedule data from HSDPA flows belonging to one UE (Figure 1).

Stage 1: PF Stage 2: WRR

Flow 1 UE 1

Flown

Flow 1 UE k

Flow m Figure. 1. . 2-stage MAC-hs scheduler

The PF algorithm used in the stage 1 scheduler intends to exploit the time varying radio channel conditions by serving a UE which instantaneous relative channel quality outperforms the one of the remaining UEs in the cell. UE selection is done on a per-TTI basis. When a scheduling decision has to be made at moment t, PF selects UEi which maximizes the following expression:

where &(t) is an expression proportional to instantaneous channel quality experienced by user UEi at moment t and hi(t) is an expression proportional to the average UEi throughput. The classical PF method to average user throughput uses an averaging window equal to the lifetime of the user

341

Ai ( t ) = - t > ti t -ti

where ai(t) is the amount of data successfully received by UEi between times ti and t and ti is the moment when the user started to be served by the PF scheduler (e.g. after entering a cell). In the classical method, the averaging window can be large if the user is active for a long time. The longer the UE is active, the less each new Aai(t), which is calculated every TTI, affects the average. PF can then begin to schedule ineffectively, as it becomes less sensitive to instantaneous changes of user throughput. Therefore, the average should either be calculated from a shorter period of time or an exponentially smoothed average can be used instead of Eq. (2):

where p is the time constant for the exponential filter and ai(n) is the user throughput experienced in TTI n. In order to work efficiently, the process should average out the fast fading variations and therefore should be set to at least 800 to correspond to an averaging window encompassing 1.6s [ 1 I]. Because scheduler time is discrete (decision is made every TTI), n in expression (3) denotes time corresponding to a certain TTI in which the scheduler currently operates. Expression (1) for a scheduling decision made in TTI n can then be written as:

We do not analyse PF scheduler operation in details, because such analyses was already done (the interested reader may find them e.g. in [6], [7], [ 101, [ 1 13). Below, we present our implementation of PF which is efficient in real-time systems, and point out some problems with PF scheduling.

Our implementation of HSDPA stage 1 PF scheduler uses Eq. (4) to decide which UEs should be selected in each TTI. A HS-DSCH channel simultaneously uses up to 15 spreading codes, so the stage 1 scheduler schedules several UEs in a single TTI as long as there are free codes. We define &(n) as the maximum number of bits that can be received by UEi in TTI n, which is proportional to instantaneous radio channel conditions experienced and reported by UEi. hi(0)=l/p for all i and in subsequent TTIs it is updated according to Eq. (3). ai(n) is the number of bits actually sent to UE, in TTI n. In every TTI, only users with data to send (MAC-hs queue not empty) compete to be scheduled, and hi is also updated only for these users to decrease the influence of traffic burstiness on priority computation.

342

Using Eq. (3) to calculate the average may not be a good solution in a real-time system because it operates on Fractional numbers. An averaging process with averaging window limited to p TTIs can be used instead:

j = n Ai(n) = k

k = min(,u,a,)

where q is the number of TTIs for which UEi has been active. The drawback of this method may be that it requires additional memory to hold up to p values of ai for all i. Scheduling results obtained using Eq. (3) and Eq. ( 5 ) are similar.

During simulations we observed two problems with PF scheduling. The first one appears in case a new UE joins a set of already active UEs. The new UE initially has a very low average throughput, so its g(n) may be higher than the one of the remaining UEs. As a result, the new UE may continuously be scheduled until its g(n) falls to a level comparable with g(n) of other UEs. Such a situation is illustrated in Figure 2 for two equal HSDPA category 9 UEs with infinitely backlogged queues (the maximum possible number of bits can be sent to each UE every TTI).

01 , 0 10 20 30 40 50 60 70 80 80

n

Figure. 2. g(n) for two equal users

uE2 becomes active 30 TTIs after UE1 and both experience equal channel conditions so they should be scheduled alternately. Instead, after UE, becomes active it is constantly scheduled for about 30 TTIs (it has a higher g(n) as can be seen in Figure 2). This is somewhat unfair and may lead to a situation, when users which are alternately activated and deactivated constantly “steal” radio resources from users which are active for a longer time. The impact of this problem on PF scheduling can be decreased by setting h(0) of a new user to the average h of all active users.

The second problem is that since PF takes average user throughput into account, users with constantly low throughput are generally privileged (have a higher g(n)). This

343

m- - - P 4w -

203.

a +

may lead to very inefficient usage of radio resources, which can be illustrated in a simple scenario with 16 equal HSDPA category 9 UEs. For simplicity let's assume, that all experience equal radio channel conditions. The queue of UE, is infinitely backlogged, and queues of the remaining U E s are only 1% backlogged.

\ \

\

. , . , , . , . . . , ,

Figure 3 presents the results of PF scheduling in such a scenario. UEI is scheduled approximately once every 95 TTIs and during this single TTI utilizes 100% of radio resources. Each of the remaining UEs is scheduled every other TTI, consuming only 1 spreading code and utilizing 1% of radio resources, so a total or 15% of available resources are used. Therefore, 85% of HSDPA radio resources are not used for 99% of the time, which is highly inefficient. To partially solve this problem, RRC [3] should not allow such a disproportion in user throughput - if traffic for UEi falls below a certain percent of maximum available throughput, the maximum available throughput for this user should be decreased (which affects 6i calculation). RRC may also consider switching the UE from HS-DSCH to DSCH. Another solution would be to take the utilisation of radio resources into account in the scheduler, and lower g(n) of users with low queue occupancies, letting them to have time to queue more data. Detailed analysis of possible improvements to address PF scheduling problems in case of HSDPA is a matter of further study.

The WRR algorithm used in our stage 2 scheduler is a simple approximation of WFQ (Weighted Fair Queuing), which is efficient but hard to implement in real time environments [8]. Using WRR MAC-hs queues belonging to one UE are served in a round-robin manner, but following a defined serving cycle that enables priority differentiation. A serving cycle 1-1-2 for two queues Q1 and 4 2 (with priorities P1 and P2) means that queue Q1 is served two times and then Q2 is served 1 time; thus available bandwidth is shared among queues Q1 and 4 2 with a ratio of 2:1, defining the relationship between priorities P1 and P2. In WRR scheduling, relative queue priorities

Figure 3. g(n) for 16 users: UEI -100% backlogged queue and UE23-16-1% backlogged

344

may be precisely defined by specifying guaranteed shares of bandwidth for each queue [13]. If a queue with a certain priority is empty, its bandwidth is distributed among other queues by the scheduler following the serving cycle. In our implementation, we use WRR with a dynamically adjusted serving cycle [ 141 - each time a new HSDPA flow for a certain UE is added, the serving cycle is adjusted by RRC. The serving cycle may also be adjusted to change relative priorities of queues belonging to existing flows. This enables RRC to effectively and flexibly manage stage 2 scheduler configuration, e.g. to solve temporary congestion problems.

Each time the PF stage 1 scheduler selects a certain UE to serve, the WRR stage 2 scheduler advances to the next index in the serving cycle of this UE, thus selecting a queue to serve. If the queue is empty, the scheduler advances to the next index, until a non-empty queue is found or the index returns to its initial value which means that all queues are empty. In our implementation, a variable containing cumulative occupancy of all UE queues is kept per UE, thus enabling the stage 1 scheduler to quickly filter out UEs for which there is no data to send.

WRR has been analysed and compared with other scheduling algorithms suitable for real-time systems, such as SP (Strict Priority), RR (Round Robin) or DRR (Deficit Round Robin) in some of our previous papers. Details can be found in [ 131, [ 141 and [ 151 so we do not repeat them in this paper. The conclusion is that WRR, especially with a dynamically adjusted serving cycle, provides good priority differentiation. Because in HSDPA only one queue per UE can be served in one TTI, there is no problem with WRR concerning potential unfairness in case of data units of different length in different queues.

3. Conclusions

Performance of HSDPA highly depends on performance of algorithms used in the MAC- hs layer in Node-B. Some algorithms are better suited to work in real-time systems (such as Node B) than other, not only due to their effectiveness but also computational complexity and memory requirements.

We proposed one of the key MAC-hs algorithms - a 2-stage MAC-hs scheduler - and evaluated its efficiency and performance in a real-time system. Proportional Fair (PF) scheduling is used in the stage 1 scheduler and Weighted Round Robin (WRR) in stage 2 scheduler. Our implementation of PF, well suited to work in a real-time system, was presented and analysed. Two discovered weaknesses of PF scheduling in HSDPA were described: first concerning scheduling unfairness in case of new users joining a set of already active users, and second concerning possible inefficient utilisation of radio resources. Some guidelines and solutions of the presented problems were presented, but their detailed analysis is a matter of further study.

345

References

1. 3GPP TS 25.308, HSDPA: Overall Description 2. 3GPP TS 25.321, Medium Access Control ( M C ) protocol specification 3. 3GPP TS 25.33 1, Radio Resource Control (RRC) Protocol Specification 4. 3GPP TS 25.401, UTRAN Overall Description 5. 3GPP TS 25.848, Physical layer aspects of UTRA HSDPA 6 . G. Aniba and S. Aissa, Fast Packet Scheduling Assuring Fairness and Quality of

Service in HSDPA, Proc. of CCECE'04, Canada, May 2004 I . M. Andrews, Instability of the Proportional Fair Scheduling Algorithm for HDR,

IEEE Trans. on Wireless Communications, Vol. 3, No. 5,2004 8. J.C. Bennett, H. Zhang, Why WFQ Is Not Good Enough for Integrated Services

Networks, NOSSDAV, pp. 524-532, April 1996 9. F. Frederiksen, T. Kolding, Performance and modeling of WCDM/HSDPA

transrnission/H-ARQ schemes, Proc. o f VTC 2002-Fall, Vol. 1, pp. 472-476 10. H. Holma, A. Toskala, WCDM for UMTS: Radio Access for Third Generation

Mobile Communications. Second Edition, Wiley & Sons (2002) 11. A. Jalali, R. Padovani, and R. Pankaj, Data throughput of CDM-HDR: A high

efficiency-high data rate personal communication wireless system, Proc. IEEE VTC., Spring, Tokyo, Japan, May 2000.

12. R. Kwan, P. H. J. Chong, E. Poutiainen, and M. Rinne, The effect of code- multiplexing on the high speed downlink packet access (HSDPA) in a WCDM network, Proc. o f IEEE WCNC, New Orleans, USA, March 2003

13. P. Matusz, J. Wozniak, Data Plane Performance Issues Concerning MBMS in UMTS Release 6, Proc. o f European Wireless 2005

14. P. Matusz, J. Wozniak, Performance analysis and optimization of the M C - d scheduler in UMTS, Proc. o f KST 2004, Bydgoszcz, Poland, pp. 3 17-326

15. P. Matusz, J. Wozniak, K. Kaminski, On Performance of MC-c/sh in UMTS, Proc. of IEE 362004, London, UK, pp. 108-1 12

SUPPORTING FLEXIBLE NETWORK OPERATOR POLICIES IN EGPRS TROUGH ADMISSION CONTROL

D. TODINCA, I. SORA University Politehnica Timigoara, Romania

E-mail: { todinca, ioana} @cs.utt.ro

P. PERRY AND J. MURPHY University College Dublin, Ireland

E-mail: perrypOeircorn. net, j . murphy@ucd. ie

Many of the admission control strategies for cellular data networks proposed in the literature allow the network operator to use different policies, depending on the network load, the number of users from each quality of service class, etc. Each policy is applied in a certain region, the regions being separated by thresholds. Those approaches suffer from a lack of flexibility: when the operating conditions change, the values for the thresholds have to be recalculated. Our work supports flexible and adaptable network operator policies, overcoming the drawbacks of the existing algorithms through a fuzzy logic based solution.

1. Introduction

General Packet Radio Service, or GPRS ', is a packetized service implemented over GSM in order to support data transfer applications like e-mail, FTP, WWW or audio and video streaming. EGPRS (Enhanced GPRS 4, is using the EDGE (Enhanced Data Rates for Global Evolution) technology in order to ensure higher data rates. The users of an (E)GPRS network have different quality of service (QoS) requirements, and the network tries to fulfill their demands, using efficient resource allocation strategies. The problem of resource allocation in a GPRS or EGPRS network can be split into two subproblems: admission control (AC), which comprises the techniques used to decide if to admit or not an user (a Mobile Node or MN) into the network, and transmission control (TC), consisting of the algorithms used to share the network resources between the connected MNs. While we have addressed the transmission control problem in our previous works (e.g.lo), in this work we focus on the admission control problem.

346

347

In this work we demonstrate the flexibility and adaptability of the fuzzy logic AC algorithm that we have proposed in * and ’. Dini and Guglilmucci

have developed a fuzzy logic AC algorithm for WCDMA cellular networks, but, as far as we know, we are the firsts to develop a fuzzy logic AC algorithm for GPRS/EGPRS.

The paper is organized as follows. Next section describes our fuzzy logic admission control algorithm for GPRS/EGPRS, section 3 presents the advantages of our AC algorithm in terms of flexibility and adaptability, section 4 contains a set of simulation results, and the paper ends with a section of conclusions.

2. Fuzzy logic for admission control

In ’ we have defined the network load such that it is proportional with the sending delay of uses’ data, given that the scheduling algorithm used for TC is Weighted Round Robin (WRR). Sending delay is the delay encountered by MN’s data across the radio interface (RLC/MAC level) and it has a decisive influence on the total delay of the MN’s data in the EGPRS network. Then, having the desired values for the delays of different QoS class MNs is equivalent with keeping the network load below or close to a target value, called target network load.

We use a Fuzzy Logic Controller (FLC) in order to keep the network load close to the target network load. The inputs of the FLC are the network load and MN’s precedence. The precedence is assigned by the network operator to a MN based on its QoS class and on its mobility characteristics (if it is a handoff MN or not, which means, if the MN comes from another cell or not).

The fuzzy rules are in the form: “IF network load is Low AND MN’s precedence is High THEN admission decision is Strong Admit”. The linguistic variables network load and MN’s precedence have the terms Low (L), Medium (M), and High (H) with linear shapes, while the linguistic variable admission decision has the terms Strong Reject (SR), Weak Reject ( W R ) , Weak Admit ( W A ) and Strong Admit (SA). The fuzzy rules are such that for low network load, all MNs are admitted, while for medium load, the low precedence MNs are marginally rejected, medium precedence MNs are marginally admitted, and high precedence MNs are strongly admitted. If the network load is high, MNs are rejected, except the high precedence MNs, which are marginally admitted. For more details about the linguistic terms and the set of rules, please report to ’.

348

3. The flexibility and adaptability of our fuzzy admission control algorithm

Here we will describe the features of our AC algorithm, that can support flexible and adaptable network operator policies.

No sharp thresholds that separate the AC policies. The admission control algorithms for cellular data networks proposed in 6 ,

allow the network operator to use different policies for different regions. The regions can be based on the network load, like in 6 , ?, or on the number of users from each QoS class, like in 5 . The regions are separated by thresholds, and the main problem of those approaches is to determine or to assign values for the thresholds. Such approaches suffer by the fact that, when the input conditions are changed, the values for the thresholds should be re-calculated or re-assigned.

The fuzzy logic based AC solution that we have developed eliminates this drawback, by replacing the sharp values that separate the regions with fuzzy regions that overlap. In this way, it is not necessary to determine or to assign precise values for the thresholds that separate the regions, giving more flexibility to the AC policies used by the network operators. More- over, the experience and expertise of the network operators can be directly transposed into fuzzy if-then rules, without the necessity to use complex mathematical models (e.g. Markov or queueing models). By its nature, fuzzy logic can tolerate imprecision, and hence it is more robust to changes in the inputs of the problem than other mathematical solutions.

On the fly changing of the target sending delay/network load. We use a function that maps the domain of the linguistic variable network load to the interval [0,63], which is its internal representation (in the FLC). In this way we obtain a very high flexibility and adaptability of our AC algorithm, because, if the network operator decides to change the value of the target network load, he has to change only the mapping function (which is a simple linear function), but the internal representation of the terms of the linguistic variables remains unchanged. This aspect can be crucial if a hardware implementation is used for the FLC, because the changing of the mapping function can be realized without interrupting the normal functioning of the FLC.

Low call dropping probability. The other input linguistic variable used in our AC algorithm, the MN’s precedence, depends on the MN’s QoS class and on the handoff status of the MN. If the MN is a handoff MN, the value of the precedence is higher than for a MN that starts its data transfer

and

349

session in the current cell. The high precedence MNs will be admitted even if the network load is high. Combined with the higher precedence assigned to handoff MNs, this means that the call dropping probability will be very low, or even zero. It is considered that for a user it is less desirable to interrupt a voice call or a data transfer in progress than to block a new call. Because of that, it is very important to obtain low values for the call dropping probability (the probability to drop a call or a data transfer session in progress).

On the fly change of admission strategy. The network operator can chose different AC strategies: for example, it can decide to admit all MNs that come from another cell, situation when the call dropping probability will be zero, or it can decide that, if the network load is high, to admit only the handoff MNs that belong to certain QoS classes, or even to admit the MNs from certain QoS classes no matter if they come from another cell or not. For example, if the QoS classes are based on users’ subscription, the network operator can chose to admit always a call from a premium user. All those different strategies can be obtained without modifying the set of rules or the values of the linguistic variables, only by changing the function that assigns the value of the MN’s precedence.

A general approach to ensure a proper balance between the values of call dropping and call blocking probability (probability to block a new call) is to use guard channels (to reserve resources in a cell for the MNs that come from the neighboring cells) and to combine resource reservation with an attempt to predict the cells that will be reached by the MN during its movement. Our approach could eliminate the necessity to reserve resources for the handoff MNs, treating in a unified way (through the precedence function) the handoff calls and the new calls. Nevertheless, the method that we propose can be used in combination with the guard channels method, maybe by keeping fewer channels in reserve for handoff MNs.

To conclude this section, we recall that the flexibility and adaptability of our fuzzy AC solutions relies on the following aspects:

(1) the intrinsic flexibility and robustness offered by the fuzzy logic, its capability to tolerate changes in the inputs of the problem

(2) the replacement of the sharp values of the thresholds that separate regions with different operator AC policies with overlapping fuzzy sets, eliminating in this way the necessity to determine precise values for the thresholds

350

(3) the use of a mapping function between the values of the network load and the internal representation of the linguistic variable network load, which allows the change “on the fly” of the AC policy without interrupting the functioning of the FLC

(4) the use of the MN’s precedence as the other input variable of the FLC. Being a conventional value, assigned by the network operator based on MN’s QoS class and on its mobility situation, it means that the network operator can easily change its admission control policy, realizing the desired balance between the importance of the QoS class of the MN and its mobility situation.

4. Simulation results

4.1. The simulation model

Our model was presented in and it contains a module for each MN admitted in the cell. There is also a user generator module, which generates MNs at certain time intervals, each MN having a set of attributes, such as the QoS class, traffic characteristics, handoff, etc. After its creation, a MN attempts to enter into the system. The admission controller module decides if the MN is admitted or not, based on its QoS class, its handoff attribute, on the network load, etc. The admission controller models the PDP context activation process in a real GPRS cell. The data transfer is controlled, as in a real GPRS cell, by the.Packet Control Unit (PCU), that runs the scheduling algorithms used for resource allocation at the MAC/RLC level and it has a working period of 20 milliseconds, called block period.

4.2. Simulation conditions

In this work there are three QoS classes of MNs, based on their subscription: 10% of the MNs belong to QoS 3 class (premium), 80% to QoS 2 class (standard) and the remaining 10% belong to class QoS 1 (economy). The weights for MNs in the WRR algorithm used for T C are: WI = 1 for economy class MNs, W2 = 2 for standard MNs and Ws = 4 for premium MNs. Each MN generate 5 files in a session, each file having a length of 20 radio blocks. The file generation mode is interactive: a new file is generated only after the previous one has been sent. The user generation period is an exponentially distributed random variable having a mean value of 180ms. From the MNs generated during the simulation, 25% are handoff MNs, the rest of 75% starts a new data transfer session in the current cell. The

351

simulation stops after the creation of 4000 MNs. All the network resources are allocated for data transfer.

In this work we modify the target value for the average sending delay of premium users from 300 ms to 700 ms and we measure the call blocking probability for all classes of MNs, and the sending delay and the total delay for the QoS 3 class MNs.

MN's precedence is not changed here, being assigned such that all handoff MNs will be accepted (hence, the call dropping probability for all the 3 QoS classes is zero). The precedence value is low to medium for QoS 1 MNs, medium for QoS 2 MNs and medium to high for QoS 3 MNs.

4.3. The results

Figure 1 presents the measured values for the average sending delay of QoS 3 class MNs when the desired sending delay has different values. It can be observed that the values of the sending delays oscillates around the target value.

500

400

300

200

100 0 200000 400000 600000 800000

Simulation time [ms]

Figure 1. 300 ms, 500 ms, and 700 ms.

Measured sending delays for QoS 3 class MNs for target sending'delays of

Table 1 contains the mean values for the average sending delay and for the average total delay of all MNs from the QoS 3 class for the mentioned values of the target delay. It can be noticed that the mean value of the sending delay slightly exceeds the desired value in all situations, but the difference is small. For example, for a target sending delay of 500ms, the measured mean value of 534 ms is only 7% higher than the proposed value.

352

Table 1. Measured mean values for average delays for QoS 3 class users.

Target sending delay Measured sending delay Measured total delay

300 323 364

400 432 479

500 534 586

600 63 1 694

700 735 807

Table 2 shows the call blocking probabilities for all classes of MNs. It is important to note the QoS differentiation: QoS 3 class users have much smaller values for call blocking probability than QoS 2 users, the values for QoS 2 users being also smaller than for QoS 1 users. Only when the proposed network load is very low, compared to the user generation rate, the call blocking probability of QoS 3 class users is unacceptably high: the 0.12 value and even 0.6 value, but when the target delay is reasonable, the blocking probabilities have very good values (3% or less).

Table 2. Blocking probabilities for different target delays.

QoS 1 QoS 2 QoS 3

300 0.61 0.27 0.12

400 0.64 0.28 0.06

500 0.71 0.27 0.03

600 0.73 0.24 0.01

700 0.68 0.23 0.01

In we have compared the performance of the fuzzy AC algorithm (in terms of call blocking probabilities and delays for premium users) with non- fuzzy AC algorithms, based on thresholds, demonstrating the superiority of the fuzzy solution on those aspects too.

5. Conclusions

This paper demonstrates the flexibility and adaptability offered to the network operator by our admission control algorithm, compared to other solutions existing in the literature. The flexibility resides in the fact that the network operator can change very easy the desired values for the sending delay of the users from the most demanding QoS class. Also, the network operator can balance between the importance given to the QoS class versus

353

the mobility status of the user. In this way, it can be obtained the desired trade-off between the call dropping and call blocking probabilities of the AC algorithm. All those changes can be realized without interrupting the functioning of the fuzzy logic controller that implements the admission control algorithm.

Acknowledgments

The support of both the Research Innovation Fund (RIF) and the Advanced Technology Research Programme (ATRP) from the Informatics Research Initiative of Enterprise Ireland is gratefully acknowledged. The first two authors acknowledge also the support of CNCSIS Romania through research grant nr. 663/2005.

References 1. C. Bettstetter, H.-J. Vogel, and J. Ebersacher. GSM phase 2+ general packet

radio service GPRS: Architecture, protocols, and air interface. IEEE Commu- nications Surveys, 2(3), (1999).

2. G. Brasche and B. Walke. Concepts, services, and protocoles of new GSM phase 2+ general packet radio service. IEEE Communications Magazine,

3. P. Dini and S. Guglielmucci. A call admission control strategy based on fuzzy logic for WCDMA systems. In Proc. of IEEE ICC 2004, Paris, France, (2004).

4. T. Halonen, J. Romero, and J. Molero (editors). GSM, GPRS and EDGE Performance. John Wiley & Sons, LTD, (2002).

5. S. Kim, T. Kwon, and Y. Choi. Call admission control for prioritized adaptive multimedia services in wireless/mobile networks. In Proc. of IEEE VTC’OO Spring, Tokyo, Japan, (2000).

6. J. R. Moorman, J. W. Lockwood, and S.-M. Kang. Real-time prioritized call admission control in a base station scheduler. In Proc. of ACM Wowmom, pages 28-37, (2000).

7. P. Stuckmann. Quality of Service management in GPRS-based radio access networks. Telecommunication Systems, 19:3,4:515-546, (2002).

8. D. Todinca, 9. Holban, P. Perry, and J. Murphy. Fuzzy logic based admission control for GPRS/EGPRS networks. In Proc. of CONTI 2004, Timisoara, Romania, May (2004).

9. D. Todinca, H. Graja, P. Perry, and J. Murphy. Novel admission control algorithm for GPRS/EGPRS based on fuzzy logic. In Proc. of IEE sGeOO4, London, UK, October (2004).

10. D. Todinca, P. Perry, and J. Murphy. Algorithms for resource allocation in data transfer over EGPRS networks. In Proc. of ECUMN’O2, pages 246-250, Colmar , France, April (2002).

35(8):94-104, (1997).

PERFORMANCE AND QUALITY OF SERVICE MANAGEMENT IN GPRS NETWORK

DR. OSAMA EL GHANDOUR PROF. SALWA EL-RAMLY MOHAMED FIKRY HELWAN UNIVERSITY A M SHAMS UNIVERSITY HELWAN UNIVERSITY [email protected] [email protected] Mohamed. [email protected]

ABSTRACT Mobile telephony has been for many years the most popular application

supported by mobile systems such as the Global System for Mobile communications (GSM). Recently, the use of mobile data applications has gained popularity. Different protocols and technologies are in place today to provide wireless and mobility communications in various environments. General Packet Radio Service (GPRS) is part of the evolution path towards 3G. So there are many publications to study main parameters that affect the performance of GPRS and to propose techniques that enhance the overall performance. In this paper, the performance and quality of service (QoS) management in the General Packet Radio Service (GPRS) network are examined. To achieve this, simulation results of GPRS performance such as End-to-End delay and throughput are produced under different conditions with a GPRS simulation based on the Optimized Network Engineering Tool (OPNET) simulation package. We examine the effects of the Bit Error Rate (BER) on the signaling plane in GPRS network and how it affects the network performance. We also study the quality of service sensitivity to the change of the packet inter arrival distribution function which is one of the major factors of representing the source applications beside the file size.

KEYWORDS: General Packet Radio Service (GPRS), OPNET, QoS, Delay, Throughput, Performance, GPRS Mobility Management (GMM), Session Management (SM).

I. INTRODUCTION

GPRS has been standardized by the ETSI as part of the GSM phase 2 + development. It represents the first implementation of packet switching within GSM, which is essentially a circuit-switched technology. By adding GPRS functionality to the existing GSM network, operators can give their subscribers resource-efficient wireless access to external Internet protocol-based networks, such as the Internet and corporate intranets. The basic idea of GPRS is to provide a packet-switched bearer service in a GSM network. As impressively demonstrated by the Internet, packet-switched networks make more efficient use of the resources for bursty data applications and provide more flexibility in general [ 1][2]. GPRS is highly suited for most data communications applications due to their bursty nature. It is to offer data rates in the range 14.4 kb/s (single slot) and 115.2 kb/s (multislots) depending on user terminal capabilities and system interference. With the introduction of GPRS into GSM networks, new charging schemes (e.g., based on amount of transported data) for GSM network usage are presented for the data services. The GPRS wireless subsystem consists of mobile stations (MS's) contending for access to a base station (BS) in a radio cell, with traffic generated according to the negotiated

354

355

quality of service (QoS) profiles, defined in terms of precedence, delay, reliability, mean and peak throughputs [3]. The paper is organized as follows. An overview of the features of the GPRS system is presented in Section 11. The signaling plane of GPRS is shown in Section 111. The model is then explained in Section IV with its assumptions. Numerical results are presented in Section V. Finally, Section VI concludes the paper.

11. GPRS SYSTEM OVERVIEW

GPRS or General Packet Radio Service is a hardware expansion of the existing GSM standard providing packet-switched data services to mobile subscribers. It is mainly the internet protocol (IP) which is supported but X.25, which is predominantly used in the European market, is also supported the existing and heavily used circuit- switched services of GSM will not be affected but improved by this upgrade. In order to integrate GPRS into the existing GSM architecture, GPRS network introduces a totally new backbone network based on IP. Two new network elements, namely Serving GPRS Support Node (SGSN) and Gateway GPRS Support Node (GGSN) are added in GSM architecture. SGSN and GGSN are mobile data routers they are interconnected via an IP backbone network. A SGSN is responsible for the delivery of data packets from and to the mobile stations within its service area. Its tasks include packet routing and transfer, mobility management (attacwdetach and location management), logical link management, and authentication and charging functions. A GGSN acts as an interface between the GPRS backbone network and the external packet data networks. It converts the GPRS packets coming from the SGSN into the appropriate packet data protocol (PDP) format (e.g., IP or X.25) and sends them out on the corresponding packet data as in Fig. 1. [4][5].

Fig. 1 also shows the interfaces between the new network nodes and the GSM network as defined by ETSI. The Gb interface connects the BSC with the SGSN. Via the Gn and the Gp interfaces, user data and signaling data are transmitted between the GSNs. The Gn interface will be used if SGSN and GGSN are located in the same PLMN, whereas the Gp interface will be used if they are in different PLMNs [4].

356

HI. Signaling Plane in GPRS Network

The signaling plane in GPRS provides supplementary services [6] and supports transmission plane functions by controlling:

GPRS network access connections Attributes of established network connections Routing paths to support user mobility Assignment of network resources to meet changing user demands

The following paragraphs offer a closer look at the interfaces and protocol layers that comprise the GPRS signaling plane. The GPRS Tunneling Protocol (GTP) on the Gn and Gp interfaces is not only used to transmit user information on the transmission plane between GSN’s but also carries information on the signaling plane, for example the “Create PDP Context Request” of the PDP Context Activation Procedure.

In GSM networks, Mobility Management (MM) and Connection Management (CM) are realized between the VMSCNLR and the mobile station in peer-to-peer communication. The Base Station Subsystem is only used as a relay. The same is true for the GPRS mobility management, which is located above the Logical Link Control layer (LLC) as in Fig.2, and where separate Service Access Points are used to address this management protocol layer:

I Fig.2 Signaling Planes between SGSN and Mobile S ation [6]

GPRS Mobility Managemenu! ent (GMMISM) Mobility Management and PDP on, Modification, and Deactivation Functions are implemented. In tl,, I u r LUIIL~XI Activation process, for example, the mobile station requests a packet data protocol, QoS level. The remaining signaling planes are based on the SS7 protocol stack. SS7 is used for signaling on the Gr, Gc, Gf, and Gd interfaces between SGSN and HLR, GGSN and HLR, SGSN and EIR, and SGSN and SMS-GMSCISMS-IWMSC as shown in Fig.3.

Mobile Application Part (MAP) This protocol layer was already applied in GSM to realize mobile-specific signaling between exchanges and databases (HLR, VLR, etc.). It is enhanced to include GPRS- specific functions.

357

Transaction Capabilities Application Part (TCAP) The TCAP supports MAP and is used as a dialog handler between the databases. (HLR, VLR, etc.)

Signaling Connection Control Part (SCCP) SCCP is an SS7 protocol. It offers connectionless and connection oriented services for switches and databases. It is used to address switches in the worldwide SS7 network (Global Title Translation, GTT)

Message Transfer Protocol (MTP) MTP covers the SS7 signaling on layer 1 to 3. It is responsible for managing signaling links, signaling routes, and signaling traffic. The Gs interface is the only signaling plane interface not based on the MAP. It is used for common GSWGPRS procedures such as location updates. Above the SCCP and MTP there is the Base Station System Application Part + (BSSAP+).

Base Station System Application Part + (BSSAP+) The BSSAP+ is a subset of the A interface’s BSSAP, used for common GSWGPRS procedures including Mobility Management.

I I

Fig.3 GPRS Signaling Planes [6]

IV. SIMULATION MODEL

The simulation model of the GPRS network used in this paper was introduced in [7] that consist of MS, SGSN, GGSN, Internal HLR, and sink as shown in Fig.4.

358

The sink that represents the external data network has two connections each having a different transmission speed, the fast link that supports 28.8 kb/s and the slow link that supports 14.4 kbls. The simulation implements the basic GPRS procedures such as Attach Signaling Procedure, Activation Signaling Procedure, Deactivation Signaling Procedure, Detach Signaling Procedure, and User Data Transmission using the OPNET simulation network version 8.0.C.

To reduce the complexity and the scope of the implementation, the model assumes some assumptions and modifications from the behavior specified in the GPRS standards such as:

- Using a simplified version of the HLR, an Internal HLR, to store the subscriber information of all GPRS users, an internal protocol is used by the SGSN to retrieve the subscriber information from the Internal HLR instead of implementing an HLR running on a Signaling System 7 (SS7) network where the HLR communicates via the Mobile Application Part (MAP) protocol.

- Implementing the highest protocol layer that is necessary to communicate with its neighboring components instead of implementing all the necessary protocol layers. - BSS was not implemented in the GPRS OPNET model because BSS is only a relay agent between the MS and the SGSN. By eliminating the implementation of a BSS, the MS and the SGSN can be modeled as being directly connected.

- Using MS identifier to substitute the IMSI and the TLLI to eliminate the overhead of mapping between the two identifiers.

- Every MS supports only one active data session instead of having multiple active data sessions for transferring user data.

- In the GPRS OPNET model, user data is sent only unidirectionally (from the MS to the external packet data network).

- The GPRS model has only one GGSN where all MS user data will be tunneled through and no selection of the GGSN is necessary.

V. RESULTS

We have created a simulation scenario by adding channel errors with a Bit Error Rate (BER) equal to 0.001 on the GPRS Mobility Management (GMM) link between the MS and SGSN with default error correction technique of OPNET and see its effect on the related procedures such as attach procedure and we have got the below results shown in Fig.5, Fig.6.

359

Fig.6 Num of attach complete received

Flg.5 Num of attach request

From Fig.5 & Fig.6 we can find how the numbers of GMM procedures decreased with (BER = 0.001) on the GMM link between MS and SGSN, this result is due to that there is a percentage of these procedures control signals that fail to reach destination node correctly and need the error correction to handle this failure, so the capacity of system handling for the procedures decreases.

We initiate a new scenario to check the effects of the channel errors on the number of SM procedures by adding (BER=O.OO 1) with default error correction technique of OPNET on the SM link. We got the below results for the number of activate PDP context request receive and activate PDP context accept sent as a two examples of the SM signaling messages shown in Fig.7 8cFig.8.

Fie.8 Num of Activate PDP context Request Received

u - - - - - - - - - - Context Accept Sent

From Fig.7 & Fig.8 we can get the same conclusion and find how the number of Session Management (SM) procedures decreased dramatically with (BER = 0.001) on the SM link between MS and SGSN, which reflects the number of sessions afforded to the users. Now we investigate the effect of channel errors on the user data link in radio interface and how it affects the system throughput and overall delay as in Fig.9 & Fig.10.

360

. . . . Fig. 10 Time averaee of End-to-End under data link with error I

Delay (BEi=0.001)

From Fig.9, we can see how the system throughput decreases as a result of adding channel errors on the user data link with using error correction technique compared to the throughput in the case of error free user data link. For the study of the effect of adding Bit Error Rate (BER) equal to 0.001 in the user data with using error correction technique on the End-to-End delay for the two cases of the slow sink receiver and the fast sink receiver, two results are obtained shown in Fig. 10&11 for the two cases of adding channels errors and the error free link respectively. In both figures End-to-End delay increases with the simulation time for the slow sink receiver, while it nearly remains constant for the fast sink receiver. Comparing Fig.10& 11 we can see how the delay increases by adding the channel errors in the user data with using error correction technique than the case of error free user data link.

Fig. 1 1 End-to-end delay with Fig. 12 End-to-end delay with constant interarrival distribution function and error free link 171.

Pareto distribution function of the interarrival time.

We create another simulation scenario by changing the distribution function of the following parameters: - MS node: Attach Request inter-arrival rate, Detach Request interarrival rate, Activation Request inter-axrival rate, Deactivation Request inter-axrival rate, and User Data interarrival rate in order to check the QoS parameters such as throughput and delay and comparing the results with the constant distribution assumed in the original simulation. We draw the delay curve again but for the Pareto distribution function [8] that

36 1

is suitable for the Internet traffic model that assumes the burstiness of the generated traffic as shown in Fig. 12. In Fig.12 we use shape parameter equal to 1.4 with mean value equal to the mean mentioned in [7] (mean=0.5 for both attach request and user data inter-arrival rate, mean=l for activation request, mean= 1.5 for deactivation request & mean= 2 for detach request). Pareto distribution’s self-similarity property approximates the burstiness of the aggregate traffic of many users, That means that the traffic burstiness only intensifies as the number of sources increases, which was observed to be the true behavior in real world networks. The drawback of a long-tailed distribution function like Pareto is in its demand for long simulation times that would also take into account packet arrivals at the far end of the distribution’s tail. This can be to some extent counteracted by limiting the interarrival time by a maximum allowable value. Such truncation of a Pareto distribution must however be used with caution because the standard deviation may then reduce dramatically [8].

We may note that using Pareto distribution function give much end-to-end delay greater than using constant distribution function for the inter arrival time of the MS node sources which could be accepted in some applications such as E-mail and internet services.

Then we used the slow sink receiver only to get results concerning delay and throughput using different distribution functions like Poisson, exponential, Bernoulli and others as shown in Fig.13 & Fig.14.

Fig. 13Time average of End-to- Fig. 14Time average of Sink traffic End delay with different interarrival distribution

functions

received (packetdsec) with different interarrival distribution

functions.

From Fig. 13 we can notice that Bernoulli distribution gives the lowest values for the End- to-End delay time compared with other functions while Rayleigh distribution gives the highest values. Also from Fig.14, we can notice that all distribution functions give the same curve shape where the constant and Rayleigh distribution functions give the highest values for throughput.

362

VI. CONCLUSIONS

In this paper, we used an OPNET simulation model to create different scenarios which focus on analyzing performance related results captured by the GPRS model that influence QoS in GPRS network, such as delay, number of different procedures in the signaling plane and throughput. We show how the BER even with error correction technique negatively affects the number of signaling procedures that could be handled by network in signaling plane, how the delay increases when we apply the same concept on the user data and how the throughput decreases under same conditions. We then create a new scenario to check how the QoS of the GPRS network affected by changing the interarrival distribution functions and show the sensitivity of the delay and throughput to this parameter.

MI. REFERENCES

[ 11 Christoph Lindemann, Axel Thiimmler, “Evaluating the GPRS Radio Interface for Different Quality of Service Profiles”, University of Dortmund, technical report, Germany, February 200 1. [2] Peter Stuckmann, “Quality of Service Management in GPRS-Based Radio Access Networks”, Telecommunication systems, 19:3, 4, 51 5-546, Kluwer academic publishers, Neitherland, 2002. [3] A. Gyasi-Agyei, S. J. Halme, J. H. Sarker, “GPRS-Features and Packet Random Access Channel Performance Analysis”, IEEE International conference on network, (ICON), Singapore, September

[4] Christian Bettstetter, Hans-J.Rg V.Gel, and J.Rg Ebersp.Cher “GSM Phase 2+ General Packet Radio Service GPRS: Architecture, Protocols, and Air Interface”, E E E Communications Surveys. Third Quarter 1999, vol. 2 no. 3. [5] Karann Chew, Rahim Tafazolli, “Performance Analysis For GPRS with Prioritized and Non-Prioritized Mobility Management Procedures”, University of Surrey, technical report, United Kingdom, 2002.

5-8,2000,

[6] www.tektronix.com/commtest. [7] Ricky Ng, Ljiljana Trajkovic, “Simulation of General Packet Radio Service Network”, School of Engineering Science, British Columbia, technical report, Canada, 2002. [8] Michal Kubik, “Uplink Packet Scheduling in WCDMA Systems”, Chalmers University of Technology (CTH), Master of Science Thesis, December 1999.

Mobile Networks (I)

ENABLING MOBILE IPV6 IN OPERATIONAL ENVIRONMENTS

Xiaoming Fu Hannes Tschofenig, Wenbing Yao Srinath Thiruvengadam

fu @cs. uni-goettingen.de {hannes.tschofenig, wenbingyao @brunel.ac. uk srinath,thiruvengadam}@

@siemens. corn Abstract: Although Mobile Ipv6 allows maintaining transport layer connections alive when an IPv6 node roams to different access networks, certain enabling mechanisms are needed for it to work in large scale network scenarios, including, most notably, issues with Mobile IPv6 bootstrapping and firewall traversal. This paper tries to address these problems by extending the IETF PANA and NSIS protocols to form an extensible framework for wide deployment of a secure, light-weight mobility service in operational IPv6 environments.

University of Gottingen Siemens AG Brunel University

1. Introduction With the recent tremendous penetration of the Internet technology and numerous portable devices, a number of new demands march steadily into view. Among them, two problems needed immediate attention: the insufficient IPv4 addresses and the need for mobility support in IP-based networks. The next generation Internet protocol (IPv6) intends to satisfy these needs. In its mobility support protocol (MIPv6) [ 5 ] , the node’s locator is somewhat separated from the identifier, by introducing a fixed Home Address (HoA) for each mobile node (MN) in addition to its topologically reachable address, the Care of Address (CoA). When the MN is away from its home network, a router on the same link as this address (the Home Agent, HA) redirects any traffic from a corresponding node (CN) to the M”s HoA to its current CoA. Due to its intrinsic mobility support for IPv6, MIPv6 has been regarded as an indispensable component of the next generation Internet infrastructure. However, MIPv6 itself simply describes a signaling protocol which establishes tunnels to change the routing of IP packets and requires pre-configured HoA information, security associations or SAs (e.g., for authentication of the MN and for MIP signaling protection), and assumes there is no firewall or other middleboxes. However, any inability to fulfill these assumptions in a foreseen scenario can cause a number of subsequent problems.

365

366

These problems can be broadly grouped into two categories: a lack of MIPv6 bootstrapping mechanisms [7] or a lack of middlebox traversal mechanisms [6]. In this paper we advocate a new approach to address them: bootstrapping is done by extending the IETF Protocol for carrying Authentication for Network Access (PANA) [3], whereas middlebox states are dynamically configured to allow MIPv6 packets to traverse by extending the IETF NSIS NATFirewall signaling protocol (NATFW NSLP) [ 101. Eventually, control data required for MIPv6 operation can be obtained and their runtime utilization (such as encapsulation of data packets and routing according to binding information, viewed as the data plane function), would be achieved seamlessly, thus forming a universal mobile IPv6 operational framework.

2. Mobile IPv6 Bootstrapping During the MIPv6 initialization phase i.e., when an MN is started in home network or moves to (or is restarted in) a foreign network, some problems may arise, e.g., the node cannot obtain enough information for MIPv6 to work or lacks support for working with the infrastructure (such as AAA). They are identified as bootstrapping functions according to the IETF MIP6 working group [7]. By and large, we can categorize them into two groups: a. Functions related to the acquisition of parameters by the MN for the MN-

HA communication. These include HoA, HA address and the parameters required for IPsec security association (SA) setup between MN and the HA.

b. Functions related to MN’s interaction with the access network, e.g., regarding the MN-AAA policy enforcement point (PEP) or MN-firewall device relationships for run-time control or forwarding. These include e.g., authentication of the MN and key exchange with the firewall device. Basically, 2 approaches for addressing these issues are conceivable: 1) to

address them individually using different, dedicated mechanisms and then combine them in a single system; or 2) to use a more extensible method based on which various bootstrapping issues are addressed, allowing the flexibility to enable certain features. Obviously, the first approach suffers from interworking and extensibility concerns. By contrast, the second one provides more flexibility for (potentially light-weight) bootstrapping mechanisms design.

Unfortunately, most proposals follow the first approach and just limit to certain functions. For example, the IETF defined a dedicated method for MIPv6 authentication and MN-HA key exchange [2], while others suggest (e.g., [4]) to setup MN-HA IKE pre-shared secret based on existing protocols such as PANA or DIAMETER. In [12] we propose to focus on the SA establishment for the recently defined MIP6 authentication [7] based on PANA since PANA already

367

provides mechanisms to bootstrap an IKE pre-shared secret for the establishment of an IPsec SA. This approach, in its first form, also belongs to the first classification as described above. However, due to its additional ability to run PANA in a multi-hop environment, it relaxes PANA assumptions and hence resolves more general bootstrapping issues. Below we shortly review PANA, then present our extension and discuss how it addresses bootstrapping problems.

2.1. Protocol for Carrying Authentication for Network Access PANA is a transport mechanism for the Extensible Authentication Protocol (EAP) [ 11 to enable network access authentication between clients and access networks. PANA carries EAP payload, allowing the use of many authentication methods. By enabling UDP transport of EAP, PANA allows any authentication method to be carried as an EAP method and hence neutral to any link-layer technology. PANA runs between the PANA client (PaC, an end node which uses PANA to authenticate itself to the network) and the PANA authentication agent (PAA, the endpoint of the PANA protocol at the access network) after successful PANA operation. A detailed description of PANA can be found in [3].

2.2. Extensions to PANA for MIPv6 bootstrapping We assume that the h4N acts as a PaC and some agent in the network (most likely the HA) acts as the PAA (which can be co-located with the PEP, or even HA). Our approach requires PANA to traverse multiple PANA hops. After mutual authentication, the PaC SA will be established. Here is a summary how we extend PANA to address various MIPv6 bootstrapping issues: Home Agent Discovery PANA was designed with a focus on network access authentication and the PAA is assumed to be just one IP hop away from the PaC. Thus, in its discovery mechanism, a multicast address is used as the discovery message’s destination [3]. Here, we extend the PANA discovery mechanism to be able to directly address the PAA using the PAA’s unicast address (if it knows this), or (indirectly) a multicast address with the router alert option (RAO) enabled. Such messages are ignored and forwarded further on by routers, until received by a PAA. Once the PAA detects the discovery message with RAO, it responds to the PaC with PANA-start/-answer message including HA’s information. Whenever possible, the MN can also learn about the HA by simpler means, such as manual configuration, DNS or IPv6 anycast mechanism. Obtaining Home Address The payload of a PANA message consists of zero or more Attribute Value Pairs

368

(AVP). We propose a new AVP (HoA AVP) for carrying the MN's HoA; pre- configuring HoA in the MN is not required. Upon receipt of a PANA request in the PAA, the HoA is selected either randomly or based on user authentication, and placed into the HoA AVP which is integrity protected by PANA. Security Association Establishment An SA, so-called MIPv6 SA, is required for subsequent protection of MIPv6 MN-HA signaling messages. It involves at least the following parameters as part of the bootstrapping procedure [7]: Security Parameter Index (SPI), replay protection indicator (e.g., timestamp or sequence number) and cryptographic algorithms. We extend PANA with a new AVP for negotiation of these parameters. Additionally, a session key needs to be derived for the MIPv6 SA. For this purpose we propose to reuse the session key derivation procedure as defined in the PANA SA establishment based on the EAP method [3]. When necessary, use of PANA re-authentication [3] also allows re-keying of SAs (including the PANA SA and the MIPv6 SA).

A further issue here is how the MIPv6 SA state should be maintained, which can impact the scalability and robustness of the bootstrapping mechanism. In principle the state lifetime can be either negotiated (e.g., the PaC proposes a value and the PAA either accepts or modifies it), fully dictated by the MNs home network, or short-lived (which requires periodical refreshes). The short- lived approach additionally deals with failure detection and prevents leaving orphan state at the home agent for a long time. This seems more interesting also because PANA already provides a refresh mechanism.

3. Middlebox Traversal in Mobile IPv6 MIPv6 does not cope with firewalls and other middleboxes. For example, if a CN is located behind a firewall, there needs to be a mechanism to allow MIF'v6 signaling messages (and data traffic) to traverse it successfully. We propose a new solution based on the IETF NSIS protocol for middlebox state maintenance.

3.1. NSIS NAT/Firewall Signaling The IETF NSIS working group is specifying an extensible IP signaling transport protocol GIMPS [9 ] , based on which several signaling application protocols are being developed; NATFW NSLP [lo] is one of them. In addition to NAT bindings, the NATFW NSLP can install, manage and remove firewall states (pinholes) along the flow path, as follows: after a sender learns about the receiving node's IP address and port number (via higher layer e.g., SIP), it constructs a NATFW NSLP CREATE message which includes its flow

369

description and send it towards the data receiver (DR). Upon recept, the firewall authenticates and authorizes the request, and installs the packet filter as specified in the flow identifier. From now on, data traffic will be allowed to traverse it.

3.2. Providing MIPv6 Firewall Traversal by NSIS Signaling As shown in Figure 1, an example scenario is studied in this section where the

r 1 Home Aaent I

i i I I

~ Correspondent node I Mobile node i- I

Figure 1: MIPv6 tirewall traversal (MN as data sender)

CN is behind a stateful packet filter (SPF) firewall and MIPv6 route optimization is used. Other firewall placements under different MIPv6 modes are described in [ 1 11. According to [ 5 ] , when the MN moves out of its home network, it has to perform Return Routability Test (RRT) before sending

a binding update to the CN: it sends a HoTI message through the HA to the CN and awaits a HOT message from the CN; it also sends a CoTI message directly to the CN and awaits a COT message from the CN. The SPF will only allow packets that belong to an existing session and hence both the packets (HoTI, CoTI) will be dropped as these packets are MIPv6-specific packets having a header structure different from normal P v 6 packets. Consequently, RRT fails. It is clear that packet filter rules have to be changed to allow these signaling messages (as well as data packets) to traverse.

The MN initiates the NSIS session by sending a CREATE to the CN. The FW may not need to know the MN, thus may not be able to authenticate the MN. It stores some relevant state regarding this firewall policy installation request and waits for the CNs authorization. Once the CN approves the request, the FW will install the relevant policy requested by the MN. When the MN receives both the messages COT and HOT, it will construct the binding key and perform binding update to the CN. Note the signaling that was aforementioned was only to allow the MIPv6 signaling messages. If the MN wants to continue sending data traffic

370

(MN is the DS), it has to perform another round of signaling (with appropriate flow identifier) to install filter rules for data traffic. Enhancements to NATFW NSLP are necessary for use with MIPv6; more details are discussed in [ 121.

4. An architecture for enabling MIPV6 deployment Although the aforementioned issues have been examined in some recent works, there is no clear architectural consideration for enabling MIPv6 in operational environments. In this section we propose a conceptual architecture that covers (but not limited to) MIPv6 bootstrapping and firewall traversal. It defines the enabling mechanisms through 3 planes of functionalities, namely the management, control and data planes. The management plane involves MIPv6 RRT and bootstrapping mechanisms. The control plane manages control state information for MIPv6 message exchange and data forwarding, including functionalities for MIPv6 registration and middlebox signaling. The data plane uses IP encapsulation and forwarding to handle data traffic.

First, the bootstrapping procedure will be initiated to set up security associations for MIPv6. Once the bootstrapping is completed, MIPv6 signaling can start. Since the MN, the CN and the HA could be behind firewalls or other middleboxes, middlebox need to be configured and maintained to allow MIPv6 messages, where NSIS signaling can be used.

Without architectural impacts, either IPsec or the MIP6 authentication protocol can be used to protect MIPv6 signaling messages. Other components which perform bootstrapping (especially to establish the MN-HA SAs) are based on the PANA extension described in Section 3. Under this architecture, it is easy to develop further necessary mechanisms for enabling successful MIPv6 operation. For example, if certain parameters are required for the control plane, the management plane functionality needs to assist and prepare them in prior to the control plane operation. One realistic example is that in most cases, NSIS signaling for middlebox traversal (a control plane function) requires security association between the MN and a middlebox device, before starting NSIS signaling. Here, PANA-based bootstrapping (a management plane function) can be easily extended for such purposes.

Figure 2 shows such an approach where PANA is used to perform both MIPv6 bootstrapping and NSIS bootstrapping. When the MN is successfully authorized, it receives a response in the PANA messages which makes a token available to the NSIS protocol. When authorization is needed, this token is added to the NSIS signaling authorization procedure.

371

Pi 81

Correspondent node

Figure 2: An example of MIPv6 in operational environments

Several aspects need to be noted with this approach. After finishing the MIPv6 bootstrapping procedure, MIPv6 related information (including keying material) can be sent to the HA (push approach). PANA can be executed either directly with the HA or with another entity, since the PANA framework supports decoupling of the enforcement point (in this case the HA) and the PAA. The latter method can be utilized by a load balancing approach. With NATFW-NSLP signaling a pull approach can work even better since the traversed firewall is unknown to the MN in advance. Note that pushing security information may not be possible in more complex topologies with multiple middleboxes and routing asymmetry, which might result in the signaling exchange towards the HA encounters an unexpected firewall. With a push approach, when the NSIS signaling message arrives at one of the firewalls, it can contact the AAA server for authentication and authorization and fetching the security context.

5. Summary and Future Work In order to successfully enable MIPv6 technology, there are a number of challenges to be addressed. We have chosen two of the most prominent ones: MIPv6 bootstrapping and firewall traversal for investigation and proposed to extend PANA in the management plane and NSIS signaling in the control plane in coordination with the data plane. We believe such a unified, extensible enabling architecture can enhance MIPv6 with envisioned enabling facilities and meet emerging new requirements. Nevertheless this work needs further

,372

investigation, including detailed message flows in different scenarios, tradeoffs between performance and complexity, and support for other features (such as VPNs). We are working on implementations of the NSIS protocol suite (including NATFW-NSLP) and PANA, and plan to integrate them into the proposed architecture and to evaluate the performance and the scalability.

Acknowledgments We would like to thank members of the E T F MIP6, PANA and NSIS working groups for the fruitful discussions. In particular we would like to thank Frank Le, Yoshi Ohba, Alper Yegin, Dan Forsberg, Gerard0 Giaretta, Julien Bournelle, Antonio Gomez-Skarmeta and Riidiger Geib.

References 1. Aboba, B., Blunk, L. and et al., "Extensible Authentication Protocol (EAP)",

RFC 3748, June 2004. 2. Arkko, J., Devarapalli, V. and F. Dupont, "Using IPsec to Protect Mobile

IPv6 Signaling Between Mobile Nodes and Home Agents", RFC 3776, June 2004.

3. Forsberg, D., Ohba, Y., and et al., "Protocol for Carrying Authentication for Network Access (PANA)", Internet draft, work in progress, Oct 2004.

4. Jee, J., Nah, J. and K. Chung, "Diameter Mobile IPv6 Bootstrapping Application using PANA", Internet draft, work in progress, Oct 2004.

5. Johnson, D., Perkins, C. and J. Arkko, "Mobility Support in IPv6", RFC 3775, June 2004.

6. Le,F., Faccia, S . and et al., "Mobile IPv6 and Firewalls Problem Statement", Internet draft, work in progress, Oct 2004.

7. Patel, A., "Problem Statement for bootstrapping Mobile IPv6", Internet draft, work in progress, Oct 2004.

8. Patel, A., Leung, K. and et al., "Authentication Protocol for Mobile IPv6", Internet draft, work in progress, Dec 2004.

9. Schulzrinne, H. and R. Hancock, "GIMPS: General Internet Messaging Protocol for Signaling", Internet draft, work in progress, Oct 2004.

10. Stiemerling, M., Tschofenig, H., and et al., "NATFirewall NSIS Signaling Layer Protocol (NSLP)", Internet draft, work in progress, Oct 2004.

11. Thiruvengadam, S . , Tschofenig, H. and F. Le, "Mobile IPv6 - NSIS Interaction for Firewall Traversal", Internet draft, work in progress, Oct 2004.

12. Tschofenig, H. and S . Thiruvengadam, "Bootstrapping Mobile IPv6 using PANA", Internet draft, work in progress, Oct 2004.

PERFORMANCE EVALUATION OF TUNNEL-BASED FAST HANDOVERS FOR MOBILE IPV6 IN WIRELESS LANS*

HANCHENG LU, JINSHENG LI, PEILIN HONG Department of Electronic Engineering and Information Science, Universiw of Science

and Technology of China, Hefei, 230027, China

The handover delay for the basic Mobile IPv6 (MIPv6) is too long to meet the requirement of real-time services. Therefore, IETF proposed a tunnel-based Fast Handover Protocol for MIPv6 (FMIPv6). We implement the tunnel-based FMIPv6 in Wireless LANs (WLANs) through the combination of using the Link Up trigger (L2-LU) and the Fast Router Advertisement (F-RA). In this paper, we analyze the performance of handovers with respect to the basic MIPv6 and our tunnel-based FMIPv6 implementation in WLANs. In our implementation, we apply a mechanism called Access Router's Information Cache (ARIC) to optimize the handover. Experiment results prove that our implementation can effectively improve the handover performance.

1. Introduction

There have been significant technological advancements in recent years in the areas of laptop, notebook computers and Wireless LANs (WLANs), which have resulted in a movement towards total mobility. Users wanting to access Internet on the move require support for IP mobility. IPv6 [ 13, which is considered to be the next generation Internet Protocol (IP), is becoming more and more popular in commercial and non-commercial services. Mobile IPv6 (MIPv6) [2] provides a mechanism to support mobility in Pv6. As for real-time applications, one of the key issues is providing the Quality of Service (QoS) consistent with the user's expectations. In IPv6-based WLANs, service disruption during handovers of Mobile Nodes (MNs) is one of the principal factors affecting QoS.

Many methods are presented to reduce the handover delay for MIPv6 [3], [4]. Based on the analysis and comparison of these methods, IETF Mobile IP Working Group is developing a Fast Handover Protocol for MIPv6 (FMIPv6) [ 5 ] , which is a tunnel-based fast handover solution. It tunnels packets (destined to the mobile) from the old access router to the new access router. FMIPv6 focuses on the scenario that the MN or the network could know or predict the new access router before the data transport from the old access router to the MN

' Supported by National 863 Project of China, Grant number: 2001AA121041.

373

374

is disrupted. Today, this is possible in IS-95 (CDMA) networks where the soft handover procedure permits the MN to simultaneously receive signals from the old and new base stations. But this capability is neither available nor easily feasible in WLANs. It is not desirable to impose such a capability as a requirement for MIPv6, considering the complexity of predicting the new access router. In order to implement the tunnel-based FMIPv6 in WLANs, we use a method of combination of using the Link Up trigger (L2-LU) and the Fast Router Advertisement (F-RA) [6]. Our implementation is somewhat like the mechanism defined in [7]. It does not need predictive information before handovers; on the contrary it acquires the default router’s information (IP address and MAC address) as soon as it associates with a new wireless access point through the combination of using the L2-LU trigger and the F-RA.

The rest of the paper is organized as follows. In section 2, we analyze the performance of handovers with respect to the MIPv6 protocol and our tunnel- based FMIPv6 implementation in WLANs. Experiment results are discussed in section 3.


2.1. Basic Mobile IPv6 Handovers in WLANs

1: MNdisconnectstoA 2: MNconnectstoAP2 3: MN sends Binding Update to HA 4: HA sends Binding ACKnowledge to MN

Figure 1. Mobile IPv6 handover

MIPv6 provides a mechanism to support mobility in IPv6. The basic operations of MIPv6 handover are showed in Figure 1. The MN hands over from subnet A to subnet B. It contains link layer handover and MIPv6 handover. Link layer handover must precede the MIPv6 handover. With reference to Figure 1, MIPv6 handover takes place after the MN has disconnected to APl and connected to

375

AP2. It includes movement detection and MIPv6 registration. The handover delay (Td) consists of there parts: i) link layer handover delay (Ti), ii) movement detection [2] delay (Tm), and iii) MIPv6 registration delay (7").

Td = Ti + Tm + Tr (1)

Tm + Tris the delay caused by the basic MIPv6 handover. Usually, Ti << Tm + Tr, so we only consider Tm and Tr, ignore Ti.

Tm is decided by the movement detection algorithms. There are two primary movement algorithms defined for MIPv6 [8]: Lazy Cell Switching (LCS) and Eager Cell Switching (ECS). Both are based on the router advertisement. In LCS algorithm, the MN receives router advertisements [9] from other access routers only after the expiration of the current router advertisement lifetime. ECS algorithm tends to work in a way opposite to that of LCS. When the MN receives a different router advertisement, it assumes that MN has entered a new subnet and begins to hand over as soon as possible. The average movement detection delays for LCS and ECS are

TLcs,ave = 0 . 5 ~ + 2x (2)

TECS,ave = 0 . 5 ~ (3)

In Eq. (2) and Eq. (3), the router advertisement interval is assumed to be x, lifetime of the router advertisement is 3x [9]. We can see that the performance of LCS is poorer than that of ECS in WLANs. But both of them are determined by the router advertisement interval x. The only way to reduce movement detection delay is to advertise router advertisements more frequently. But frequent router advertisements may add heavy traffic.

Tr is consists of the new care of address configuration delay (Tdod) [lo] and registration delay ( Treg)

Tr = Tdrrd -t Treg (4)

Tdad is the delay caused by Duplicate Address Detection (DAD) for the new care of address and Treg is the delay caused by MN sending Binding Updates (BUS) to the Home Agent (HA) and Correspond Nodes (CNs) to register it's new care of address.

Reference [ 101 describes the process of DAD. A random delay Trdbetween 0 to Trdmar is applied before sending out a neighbor solicitation [9] for DAD. We denote the number of transmissions of neighbor solicitation and the interval of the transmission of the two consecutive neighbor solicitations as n and Xnt. The average delay caused by DAD is

376

Tdadave = Trd,ave -4 nKnt (5)

In which Trd,ave denotes the average of Trd. In our WLANs testbed, Trd,mox is 1000ms, n is 1, and Tint is 1000ms, which are also default values in [lo]. Therefore,

Tdadave = 1500ms (6)

Eq. (6) shows that the minimum average delay for the basic MIPv6 handover with DAD is more than 1500ms; it’s too long to satisfy real-time services.

2.2. Tunnel-Based Fast Handovers for Mobile IPv6 in WLANs

Because the handover delay for the basic MIPv6 handover cannot meet the requirement of real-time services, IETF Mobile IP Working Group is developing a Fast Handover Protocol for MIPv6 (FMIPv6) [5] to solve the problem. It’s a tunnel-based fast handover solution. In order to implement it in WLANs, we use the L2-LU trigger and the F-RA.

A L2-LU trigger coming from the MN’s link layer contains the newly associated wireless access point’s Media Access Control (MAC) address. Since we cannot get the information (IP address and MAC address) regarding the new access router from the L2-UP trigger, we use the F-RA to acquire this information. Unlike the standard router advertisement defined in [9], which states that a router must delay a response to a router solicitation by a random time between 0 and MAX-RA-DELAY-TIME [9] seconds, F-RA allows access routers to respond a router solicitation sent by the MN immediately.

3FBU I r I I

old Aces Router

5 FBACK

I Mobile Node (MN) I

Figure 2. Tunnel-based fast Mobile IPv6 handover

Figure 2 illustrates our implementation of the tunnel-based FMIPv6 in WLANs. The MN initiates the fast handover process at the time it receives a L2-

377

LU trigger immediately after associated with the new wireless access point. The operations of the handover process are described as follows:

1.

2.

3.

4.

5 .

The MN sends a Router- Solicitation (RS) to all router’s address on the link. The new Access Router (nAR) supporting F-RA responds a Router Advertisement (RA) immediately. Upon receiving the RA, The MN sends a Fast Binding Update (FBU) to the old Access Router (OAR) indicating of the handover. This FBU message associates the MN’s old care of address with the nAR’s IP address. Handover Initiate (HI) and Handover ACKnowledge (HACK) messages are exchanged between the OAR and nAR to set up a bi-directional tunnel. Packets to and from the MN’s old care of address must be forwarded through this tunnel. The OAR sends a Fast Binding ACKnowledge (FBACK) to MN to indicate the completion of fast handover operation.

Once fast handover is completed, the & can perform regular MIPv6 procedures. The handover delay for the tunnel-based FMIPv6 in WLANs is

Td = P-ra f Treg (7)

In Eq. (7), C+o denotes the delay for F-RA acquisition (the MN sends a router solicitation to acquire a router advertisement with F-RA) and Treg is denoted as the FBU registration delay. We propose a mechanism called Access Router’s Information Cache (ARIC) to eliminate Q r a . ARIC maps the AP’s MAC address to the connected AR’s Information (IP address and MAC address). The wireless access point’s MAC address comes from the L2-UP trigger and the access router’s Information can be acquired from router advertisements. The MN caches the relationship as an entry in its ARIC table. Each entry in the ARIC table is set a lifetime and it is updated by receiving a router advertisement. When the MN associates with a new wireless access point, it uses the wireless access point’s MAC address to search in its ARIC table. If a valid entry for the MAC address were found, the MN would use it to acquire the access router’s information immediately without sending a router solicitation. ARIC not only accelerates the handover progress of the tunnel-based FMIPv6, but also reduces the number of router solicitations sent by the MN. It’s very useful when there is a large number of MNs that hand over back and forth among the same wireless access points.

Comparing Eq. (7) with Eq. (l), we can see that the handover delay for the tunnel-based FMIPv6 has nothing to do with the router advertisement interval. Using the old care of address to communicate with CNs after handover also

378

eliminates the DAD delay in the tunnel-based FMIPv6. The FBU is sent to the old access router while the BU is sent to the HA and CNs. Therefore the registration delay for the FBU is much less than that for the BU when the MN is far away from the home network.

3. Experiment Results

3.1. WLANs Testbed Configuration and Components

l a Corespond Node Mobile Node 1-

Figure 3. WLANs Testbed for Mobile IPv6

Figure. 3 shows the configuration of our experimental WLANs testbed. It consists of two access routers (AR1 and AR2), two wireless access points (AP1 and AP2), an MN, HA and CN. The MN equipped with a Lucent “Gold” WaveLAN card to communicate at the speed of 1lMbps to the wireless access point. The wired connections are lOMbps Ethernet links.

The operation system is Linux. The Wireless card driver is modified to support the L2-LU trigger and the F-RA is also implemented in the router advertisement daemons in access routers. We integrate HUT M P L t and our tunnel-based FMIF’v6 code to support mobility and fast handovers.

The CN sends UDP packets to the MN at the interval of 20ms. The payload length of the packet is 32 bytes. The MN hands over from AR1 to AR2. The experiment focuses on the handover delay in MIPv6. Therefore we define the measured delay as the duration between the time when the MN receives a L2- LU trigger and the moment when it’s able to receive UDP packets from the CN. The average delays in experiment results are measured 10 times for each handover case. The movement detection algorithm is ECS.

‘MIPL source code for Mobile IPv6 can be obtained from http:Nwww.mobile-ipv6.0rg

379

3.2. Experiment Results

Experiment results are showed in Figure 4 and Figure 5. In the figures, abscissa is the interval of the router advertisement, which is denoted as (min, max). Min is the minimum value of the interval while max is the maximum. From Figure 4, it can be observed that the performance of the basic MIPv6 handover is significantly poorer than that of the tunnel-based FMIPv6 in WLANs. In the basic MIPv6 handover, the average delay increases from 273 1.9ms to 4047.0ms as the interval of the router advertisement increases. From Figure 5, we can see that it’s mostly caused by the incensement of the movement detection (ECS) delay. The average delay for DAD remains about 1500ms. But in the tunnel- based FMIPv6, the average delay remains about 15ms, which is significantly less than that in the basic MIPv6 handover. And the delay does nothing to do with the interval of the router advertisement. Since we implement F-RA in access routers, the average delay for the router advertisement acquisition is reduced dramatically. Figure 5 shows that it is about 6ms. It’s much less than that of the standard router advertisement implementation, which must apply a random delay between 0 and MAX-RA- DELAY-TIME (default values is 500ms) before it sends a respond to the router solicitation.

0 Basic Mobile IPv6 handover (ECS) + Average delay for movement detection Average delay for DAD

A Average delay for F-RA acquisition Tunnel-based Mobile IPv6 fast handover

500,$Average Delay (ms) 250TAverage Delay (ms)

2046.1

1510.2 1454.2 1491.2

1021.9

6.5 6.5 6.7 6.4 6.5 . , . b - - 0.5s Is 1s IS 1.5s 0 . 5 ~ Is I S IS 1.5s 1.5s 2s 3s 4s 4.5s 1 . 5 ~ 2s 3s 4s 4.5s

Figure 5. Average delay for movement detection, DAD and F-RA acquisition

Router advertisement interval (minpax)

Figure 4. Average delay for handovers

Router advertisement interval (min,max)

With ARIC, the average delay can slightly reduce from 14.8ms to 7.7ms. The reduction is about 7.lms, which is the delay for the F-RA acquisition. Also it reduces the number of router solicitation sent out.

380

4. Conclusion

In this paper, we analyze the performance of handovers with respect to the basic MIF'v6 and our tunnel-based FMIPv6 implementation in WLANs. ARIC is proposed to reduce the handover delay and the number of router solicitation sent out. Experiment results prove that our implementation can effectively improve the performance during MIPv6 handovers.

As a future work, we would like to consider context (Such as QoS parameters, header compression parameters, etc.) transfer between access routers in our tunnel-based FMIPv6 implementation. When the MN hands over to a new subnet, it does not need to associate these parameters with the new access router or HA.

References

1.

2.

3.

S. Deering and R. Hinden, Internet Protocol, Version 6(IPv6) Specification, RFC2460, November 1998. D. B. Johnson and C. Perkins, Mobility support in IPv6, RFC3775, June 2004. G. Krishnamurthi, R. Chalmers and C. Perkins, Buffer Management for Smooth Handovers in Mobile IPv6, draft-krishnamurthi-mobileip-buffer6- Ol.txt, March 2001.

4. K. El Malki and H. Soliman, Simultaneous Bindings for Mobile IPv6 Fast Handoffs, draft-elmalki-mobileip-bicasting-v6-02.tt, June 200 1.

5. Rajeev Koodli (Editor), Fast Handovers for Mobile IPv6, draft-ietf- mipshop-fast-mipv6-03.txt (work in progress), October 2004.

6. J. Kempf, M. M-Khalil and B. Pentland, IPv6 Fast Router Advertisement, draft-mkhalil-ipv6-fastra-01 . t t , May 2002.

7. G. Dommety, Handoff Optimization with no prediction and minimal L2 Trigger information, draft-dommety-mobileip-min-handoffv4and6-0 1 .txt, March 200 1. N. A. Fikouras and C. Gorg. Performance comparison of hinted and advertisement-based movement detection methods for mobile IP hand-offs, Computer Networks, 200 1.

9. T. Narten, E. Nordmark, W. Simpson, Neighbor Discovery for IP Version 6(IPv6), RFC2461, November 1998.

10. S. Thomson and T. Narten, IPv6 stateless address auto configuration, RFC2642, November 1998.

8.

MOBILE IPVG-TYPE ROUTE OPTIMIZATION SCHEME FOR NETWORK MOBILITY (NEMO) SUPPORT

BED P. KAFLE T h e Graduate University f o r Advanced Studies,

Department of Informatics, 2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo 101-8430 Japan,

E-mail: [email protected]

EIJI KAMIOKA AND SHIGEKI YAMADA National Institute of Informatics,

2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo 101-8430 Japan, E-mail: { kamioka,shigeki} Qnii. ac. j p

The network mobility (NEMO) basic support protocol, developed by the IETF, provides uninterrupted Internet connectivity to the communicating nodes of mobile networks. This protocol makes network movements transparent to the mobile nodes. However, delays in data delivery and higher overheads are likely to occur because of suboptimal routing and multiple encapsulation of data packets. To resolve this problem, we propose an extended Mobile IPv6-type route optimization scheme for mobile networks. The proposed scheme is simple and easy to deploy, as it requires only slight modification of the NEMO basic support protocol at local entities such as the mobile router and mobile nodes of the mobile network, leaving entities in the core or in other administrative domains untouched. This scheme enables a correspondent node to forward packets directly to the mobile network without any tunneling, thus improving the packet delivery efficiency and reducing end-to-end packet delay.

1. Introduction

In recent years, a number of wireless computing devices may move together, such as on trains and aircrafts. In this type of environment, managing the mobility of individual devices via the devices themselves would be inefficient (and sometimes unfeasible) [l]. Addressing this issue, the IETF (Internet Engineering Task Force) Network Mobility (NEMO) working group [2] has developed a conceptual architecture for mobile network [3] and NEMO basic support protocol [4].

A mobile network is composed of one or more mobile routers (MRs)

381

382

and mobile network nodes (MNNs) that move together as a single unit in the Internet topology, as shown in Fig. 1. The MR performs two types of functions: (a) acts as a gateway to provide Internet connectivity to the MNNs, and (b) manages the mobility of the entire network by using the NEMO basic support protocol. The NEMO basic protocol makes network mobility transparent to MNNs by enhancing the operation of Mobile IPv6 (MIPv6) [5] at MRs and their home agents (HAS). The operation of the protocol is illustrated in Fig. 1. When an MR moves away from its home link, it configures a care-of adddress (CoA) in the visited link and performs a mobile network prefix (MNP) scope binding update (BU) with the HA. Following the BU, the HA intercepts data packets addressed to any MNNs that have obtained an address from the MNP, and tunnels the packets to the MR’s CoA.

MR Mobile Router CN Correspondent Node

4 - + MR regisbakw *- . -h VlDiting ncde reg t--, Vlsitlng nade data +--* LDcal nade date

LOCa node

Figure 1. Network mobility (NEMO) architecture.

The MNNs are categorized into three groups: local fixed nodes (LFNs), local mobile nodes (LMNs), and visiting mobile nodes (VMNs). The LFNs and LMNs have their home addresses (HoAs) associated with the MNP whereas VMNs have CoAs associated with the MNP. Mobile networks can occur in nested form: a mobile network, such as a PAN (personal area network) accessing another mobile network, such as a vehicular network. The MRs of a nested mobile network form a hierarchy with parent-MRs providing connectivity to sub-MRs.

The data packets from a correspondent node (CN) addressed to a VMN first arrive at the VMN’s HA, which tunnels packets to the VMN’s CoA. Since the VMN’s CoA belongs to the MNP, these packets go to the MR’s home agent, which again tunnels them to the MR’s CoA. The MR decapsulates packets and forwards them on to the interface to which the VMN is

383

attached. Similarly, outbound packets originated by the VMNs follow the same route in the reverse direction. Clearly, the process of data delivery in the NEMO basic support protocol is full of overheads because of multi- angular sub-optimal routing and multi-layer encapsulation. This problem is worse if the mobile network is a multi-level nested mobile network where packets need to be tunneled through every MR’s HA. We know, from experience with MIPv6, that route optimization (RO) can resolve the problems of sub-optimal routing. However, the use of MIPv6-type RO without any enhancement is inefficient because of different nature of the MNN’s CoA [l]; in MIPv6, a mobile node’s CoA gives the actual location of the mobile node, whereas in NEMO, the MNN’s CoA gives the location of the home link of the mobile network.

Various approaches to RO in NEMO have been proposed in the literature [7, 6, 8, 91. However, most of them require additional network entities or functionalities to be implemented outside the mobile network domain, which may hinder their deployment in the existing Internet. Moreover, these approaches provide only partial solutions by optimizing only some sectors of routes between CNs and MNNs. To overcome these problems, we propose an extended MIPv6-type RO scheme for NEMO support network. Our scheme is simple to implement, as it requires only slight modification of the NEMO basic support protocol at the MRs and MNNs, leaving the network entities in core network or in other administrative domains untouched. This scheme uses the well-established functionality of MIPv6 to enable a CN to route data packets directly to a mobile network without any encapsulation. The rest of this paper is organized as follows. We describe our proposed scheme in Section 2. We present an evaluation of the scheme’s performance in Section 3. Finally, we conclude the paper in Section 4.

2. Proposed Route Optimization Scheme

We describe the proposed scheme by considering a nested mobile network with multiple levels of MRs as shown in Fig. 2. In our scheme, all MRs keep a binding cache, which we call the MR-binding cache (MRBC) for all the sub-MRs behind them. Additionally, the root-MR (i.e. the top level MR) keeps another binding, which we call the RO-binding cache (ROBC) for all active MNNs that have ongoing communication sessions with CNs.

The MRBC is used to store bindings between CoAs of all sub-MRs and their MNPs. As illustrated in Fig. 2, when the sub-mobile router MR1 is attached to a parent mobile router MR2, the former registers its CoA and

384

Figure 2. optimization in NEMO support network.

MR-Binding Cache (MRBC) and RO-Binding Cache (ROBC) used for route

mobile network prefix (MNP1) with the latter. MR2 uses this information to forward a packet addressed to the MNN’s CoA, which is associated with the MNP1. Similarly, the MR2 registers its CoA and all mobile network prefixes (MNP1 and MNP2) that are accessible through the MR2 with the MR3 (the root-MR) so that the MR3 can forward data packets to the MNNs that have CoAs associated with MNPl and MNP2. Every entry in the MRBC has a lifetime, which must be renewed by re-registering if the sub-MR is going to stay in the parent-MR longer than the time allocated in the previous registration. Moreover, the root-MR keeps updating the sub-MRs and MNNs of the root-MR’s CoA and HoA.

The ROBC is maintained by the root-MR for two purposes: (1) to obtain the CoAs of active MNNs to enable data packets to be tunneled to them, and (2) to perform an RO with the CNs on behalf of active MNNs when the network moves. The ROBC includes bindings between the MNN’s HoA and other information such as the CoA, CN, authorization data, and so on. The information required for creating the binding is supplied by the MNNs in RO-inform message (explained below) when they first perform RO with the CNs. To keep the ROBC up-to-date, every entry in the cache has a lifetime, which is renewed by packets traveling between the MNN and CN.

2.1. Route Optimization Operation

The RO process is triggered when an encapsulated data packet destined for an MNN arrives at the MR1 via the MRl’s HA. The MR1 decapsulates the packet and forwards it to the MIPv6-enabled MNN.” Then, as in MIPv6 RO, the MNN performs a Return Routability address test to ex-

=In the case of MIPv6 not supporting local nodes, the MR1 performs RO on behalf of the nodes.

385

change the authorization data with the CN. After the test, the MNN forms a new control message, RO-inform, containing the MNN’s HoA, CoA, CN, authorization data, etc., and sends it to the root-MR. The root-MR uses RO-inform to create an entry for the MNN in the ROBC and then replies with an RO-accept to the MNN. The MNN then formulates a correspondent binding update (CBU) message using the root-MRs CoA in the alternate care-of address option, and sends the CBU to the CN. On arrival of the CBU, the CN creates a binding between the MNN’s HoA and the CoA mentioned in the alternate care-of address option field. The CN then sends packets directly to the root-MR’s CoA using Routing Header (RH) type 2, an extension header defined in MIPv6 [5 ] . The root-MR checks the home address option field of the RH type 2 header to get the HoA of MNN that the packet is addressed to. The MNN’s HoA is used to search for the corresponding CoA in the ROBC, which in turn is used to tunnel the packet to the ‘MNN. Similarly, MNN-originated outbound packets have the root- MR’s CoA and address of CN in the source and destination address fields, respectively. These packets are tunneled to the root-MR using the root- MR’s HoA as the destination and MNN’s CoA as the source address in the outer IP header. The root-MR decapsulates and forwards packets normally to the CN.

When the root-MR performs the handover to a new link and configures a new CoA, it sends a CBU to all CNs that have active sessions with the MNNs. As the root-MR has already stored the addresses of the CNs and binding authorization data for all active MNNs in the ROBC, it uses this information to send a CBU on behalf of the MNNs. The root-MR also notifies all sub-MRs and active MNNs of the CoA change so that they can use the new CoA as the source address in the outbound data packets.

2.2. Advantages and Disadvantage

This scheme has following advantages over other schemes. (i) Data packets are forwarded through the shortest path between CN and MNN. (ii) Little modification of NEMO basic support protocol at the MR and MNN is required. (iii) No additional entities in the core network or in other domains are required. (iv) End-to-.end semantics of the IP packet transportation is preserved. That is, packets are not modified on the way. (v) Network movement transparency to MNN is preserved. (vi) Faster reaction to handovers is achieved because the root-MR can inform CNs of CoA changes faster than any other network entities can.

386

A possible disadvantage of the proposed scheme is the addition of the CBU function to a root-MR which could result in the root-MR becoming a bottleneck when serving a large number of MNNs. Our evaluation, however, indicates that the improvement in system performance due to lessening packet delivery overheads exceeded the burden imposed by CBU signaling at the root-MR. In addition, the performance of a root-MR can be improved by using multihoming, which enables the root-MR to access the Internet through different interfaces [3].


We developed analytical models to evaluate the performance of the p r e posed scheme in terms of packet delivery efficiency and delay. While performing the evaluation, we compared our scheme with the path control header (PCH) [6] and reverse routing header (RRH) [7] schemes.

Packet delivery efficiency measures how effectively the network resources are utilized for delivering the IP payload. The efficiency is computed by dividing the average traffic load when an IP packet is routed normally (without using mobility support) from a source to a destination by the tr&c load when the same packet is routed with mobility support from the same source to the same destination. As explained in the previous section, the mobility management functions introduce additional overheads in the form of signaling and extra headers such as encapsulation headers and extension headers that need to be transported over sub-optimal paths. We define the traffic load as the cumulative product of the size of packetb and the distance (in terms of IP hops) it travels while moving from a source to a destination.c The traffic load is measured in units of ‘octet x hop’ [lo].

Figure 3 compares the packet delivery efficiency of different schemes against packet size for nested mobility level of two (i.e. an MR serving VMNs). In terms of efficiency, our scheme performs best. For smaller packets such as TCP acknowledgments, the packet delivery is less efficient because mobility-specific headers consume a comparatively larger share of network resources. The comparison shows that the efficiency of the basic support protocol is very low (about 40%) for any packet size. On the other hand, the proposed scheme has efficiency of more than 80% when the packet size is around (or larger than) 200 bytes. This indicates that our scheme

bPacket size may change on the way due to mobility management operation. =Due to space limitation we are unable to present detail analysis, we have instead chosen to present the numerical results only.

387

+ Proposed

-m- Basic

A n am A A A I d , ,A ,A Packet size (Bytes)

Figure 3. Packet delivery efficiency.

increases the packet delivery efficiency of the NEMO basic support protocol by more than 100%. Furthermore, it is obvious that increasing the nested mobility level further reduces the efficiency of the basic support scheme, whereas our scheme is the least affected.

PPN Utililzation

Figure 4. One-way packet delivery delay from CN to MNN

Figure 4 compares the results of the analysis of packet delivery delay versus utilization (or load) on the mobility agents such as HAS and MRs, for nested mobility level of two.d As expected, the maximum delay is occurred in the basic support protocol and the minimum in our scheme. The PCH has slightly less delay than the RRH because the PCH allows packets to be routed from a correspondent router to the root-MR without going through an HA.

dFor delay estimation, we referenced the network parameter values from a paper [lo].

388

We also evaluated the CBU signaling overheads at the root-MR when it performed RO on behalf of active MNNs during handovers. We found that the volume of the root-MR-originated CBU packets was less than 2% of the total data packets flowing out through the root-MR. When compared to the gain in packet delivery efficiency (more than loo%), the increment in overheads by the CBU signaling was insignificant.

4. Conclusion

We presented an extended Mobile IPv6-type route optimization scheme for NEMO support networks. Our scheme enhanced the mobile network operation by significantly improving the packet delivery efficiency and reducing the end-to-end packet delay, without imposing any modification of protocol implementation in the correspondent node, home agent or other Internet entities. In future work, we will further investigate computational overheads in mobile routers, and hence the scalability of the scheme.

References 1. E. Perera, V. Sivaraman, A. Seneviratne, “Survey of network mobility s u p

port,” ACM Mobile Computing and Commun. Review, vo1.8, no.2, pp.7-19, April 2004.

2. IETF Network Mobility (NEMO) Charter. http://www.ietf.org/html.charters /nemo-charter . html.

3. T. Ernst and H-Y. Lach, “Network mobility support terminology,” IETF In- ternet Draft, draft-ietf-nemo-terminology-02.txt (work in progress), Oct. 2004.

4. V. Devarapalli, R. Wakikawa, A. Pertruscu, P. Thubert, “Network mobility (NEMO) basic support protocol,” IETF RFC 3963, Jan. 2004.

5. D. Johnson, C. Perkins, J. Arkko, “Mobility support in IPv6,” IETF RFC 3775, June 2004.

6. J. Na et al., “Route optimization scheme based on path control header,’’ IETF Internet Draft, draft-na-nemo-path-control-header-0O.txt (work in progress), April 2004.

7. P. Thubert et al., “IPv6 reverse routing header and its application to mobile networks,” IETF Internet Draft, draft-thubert-nemo-reverse-routing-header- O5.txt (work in progress), June 2004.

8. E. Perera et al., “Extended network mobility support,” IETF Internet Draft, draft-perera-nemc-extended-00.txt (work in progress), July 2003.

9. T. Ernst et al., “Extending mobile-IPv6 with multicast to support mobile networks in IPv6,” ECUMN’OO, Colmar, France, Oct. 2000.

10. S.-C. Lo, G. Lee, W.-T. Chen and J.-C. Liu, “Architecture for mobility and QoS support in all-IP wireless networks,” IEEE J. Sel. Areas Commun., v01.22, no.4, pp.691-705, May 2004.

COMPARATIVE ANALYSIS OF HANDOFF DELAY OF MIFA AND MIP

ALI DIAB, ANDREAS MITSCHELE-THIEL, RENE BOERINGER Chair for Integrated HW/SW-Systems, Ilmenau University of Technology, Gustav-

Kirchhoff-Str. 1 , Ilmenau, 98693, Germany

Latency during handoffs affects the service quality of real-time applications. Mobile IP (MIP) presents the standard mobility management protocol. However, Mobile IP is not adequate for delay sensitive applications. Mobile IP Fast Authentication protocol (MIFA) is proposed to avoid the problems of Mobile IP and to match the requirements of real- time applications. In this paper we present an analytical model to evaluate the performance of MIFA. Our performance study shows that the handoff latency of MIFA is independent of the distance between the current FA and the HA. Thus, MIFA highly reduces the handoff latency for most cases. MIFA clearly outperforms Mobile IP (MIP) with respect to the handoff latency and the number of packets dropped due to handoffs.

1. Introduction

Latency during handoffs affects the service quality of real-time applications. Therefore, the development of fast mobility management solutions is a big challenge in future IP-based mobile networks.When the Mobile Node (MN) notices that the current Access Point (AP) is no longer reachable, it starts to scan the medium for other available APs. After that the MN authenticates and re-associates itself with the newly discovered AP. These procedures are called layer2 handoff. No additional procedures are required if the new AP belongs to the same subnet as the old one. However, the MN must discover the new Foreign Agent (FA) serving this subnet, register and authenticate itself with the Home Agent (HA) or another special agent through this FA, if the new AP belongs to another subnet. These additional procedures are called layer3 handoff.

In order to implement Layer3-Handoffs, several protocols have been proposed. With Mobile IP Version 4 (MIPv4) [I], the MN must be registered and authenticated by the HA every time it moves from one subnet to another. This introduces extra latency to the communication, especially when the HA is far away from the FA.

In order to avoid these sources of extra latency, an approach to use an Anchor FA (AFA) has been proposed [2]. If the MN is away from the home network, it will be initially registered by the HA. During this registration a shared secret between MN and FA (KMN-FA) is established. The FA then acts

389

390

as AFA. Thus, in subsequent registrations, the MN is registered at this AFA instead of the HA as long as it remains in the same domain, which the AFA belongs to. In this approach there is no need to establish a tunnel between HA and FA. Instead, an additional tunnel from the AFA to the current FA is established. However, the forwarding delay on the downlink as well as the uplink, i.e. the path from HA via AFA and current FA to MN and vice versa, increases compared to MIP. An additional reverse tunnel is needed from the current FA to the AFA.

In order to reduce the problem of temporarily discontinued communication between HA and MN during handoff, regional registration [3] has been proposed. With this approach the HA is not aware of every change in the point of attachment. This is due to the fact that the MN is registered and authenticated by the GFA or the Regional Foreign Agent (RFA) instead of the HA. Thus, the MN communicates with the HA only if it changes the GFA. As a result, the handover latency known from MIP is incurred in rare cases only.

Proposals for low latency handoffs use a trigger originating from layer2 (L2-trigger) to anticipate handoffs prior to a break of rado link. In [4] methods for pre-registration, post-registration and a combined method have been proposed. Thus, a layer3 handoff is triggered by a LZtrigger. With the pre and post-registration method, the MN scans the medium for other APs if the strength of the signal received from the current AP deteriorates or if the error rate increases. If another AP is available and this AP belongs to another subnet, a L2-trigger is fired. This trigger contains the IP address of the new FA or another address from which the IP address can be derived, e.g. the MAC address. This prompts the MN, when employing pre-registration, to register with the new FA through the old one. Thus, the layer3 handoff is performed while the MN performs layer2 handoff. However, with post registration the MN performs only a layer2 handoff when the L2-trigger is fired. If the link between the current FA and the MN breaks down (receiving Layer2 Link Down trigger (L2-LD) trigger), a bidirectional tunnel is established between the old FA and the new one. As a result the packets destined to the MN will be forwarded to the nFA through the old one. Thus, the MN receives the packets before the registration. With the combined method, the MN first tries to use the pre-registration method when a L2-trigger is received. If this fails, the MN employs the post-registration method.

Performance studies and an implementation of the pre-registration and post- registration method are presented in [5 ] and [6] respectively. [7] present a comparison between the two methods. The simulation results indicate that the timing of the trigger has a major influence on the latency of the handoff methods as well as the packet lose rate. If the L2-trigger for Pre-registration is delayed, increased latency results. In case the Registration Request (Reg-Rqst) is dropped, the method resorts to the standard layer 3 method, e.g. MIP. In addition, the causes for latency of MIP still remain which is due to the

391

forwarding delay between the FA and the HA. Even though Post-registration is faster than Pre-registration, the impact of delayed LZtriggers with Post- registration is the same as with Pre-registration. Due to the missing MIP registration with the Post-registration approach, the packet delay is larger (uplink and downlink). The combined method inherits the problems of the both approaches.

2. Mobile IP Fast Authentication Protocol (MIFA)

In order to avoid the problems of MIP without needing to insert intermediate nodes between the FA and the HA, Mobile IP Fast Authentication protocol [S] has been proposed. The basic idea of MIFA is that the HA delegates the authentication to the FA. As a result the MN authenticates itself with the FA and with the HA. However this happens in the FA. Thus the MN sends RegRqst to the FA, which in turn directly replies by sending a Registration Reply message (RegRply) to the MN. After receiving the RegRply, the MN can resume the transmission on uplink. In downlink a tunnel is established to forward the packets, arriving at the previous FA, to the new FA until the HA is informed about the movement and a tunnel from the HA to the current FA is established to forward the packets directly to the new FA. Thus the delay experienced from the communication between the new FA and the HA is eliminated, similar to the micro mobility protocols, see [S].

The local authentication by FAs relies on groups of neighboring FAs. Each FA defines a set of neighboring FAs called a Layer3 Frequent Handoff Region (L3-FHR) [9]. These L3-FHRs can be built statically by means of standard algorithms (e.g. neighbor graph [lo] or others [9]), or dynamically by the network itself, by observing the movements of MNs. Every FA defines its own L3-FHR. The L3-FHR doesn’t necessarily comprise all of the adjacent FAs, e.g. in the case of physical obstacles between the areas that prevent a move between the adjacent FA areas.

There is a security association between the FAs in each L3-FHR. This security association can be established statically, e.g. by the network administrator, or dynamically, e.g. by the network itself as described in [I 13,

Figure 1 depicts the basic operation of MIFA. While the MN communicates with the current FA, this FA sends notifications to all of the FAs in the L3-FHR the current FA belongs to. These notifications contain the security associations between the MN and the FAs in this L3-FHR on one side and between the FAs and the HA on the other side. These security associations are recorded in soft state and will be used by one of the FAs in the future and deleted from the others. Additionally these notifications contain the attributes of the HA and the authentication values (between the MN and the HA) the MN has to generate in the next registration with the next FA. These notifications are authenticated by

1121.

392

means of the security associations established between the FAs as described in detail in [ 131.

(Ei)HAh-pemaa (a) Pmlouv FA ;\ck

Figure 1 : Basic operation of MIFA

When the MN moves to one of the FAs in the L3-FHR, to which the previous FA belongs to, it sends a RegRqst message to this FA. This new FA first checks the authentication between itself and the MN. The authentication is checked employing the security association sent from the previous FA with the notification. Subsequently, the new FA checks the MIFA information, which presents the authentication information between the MN and the HA. The new FA then checks if the requirements requested from the HA can be satisfied. This can be done employing the attributes of the HA received with the notification. If the authentication succeeds, the FA builds a previous FA Notification message to inform the previous FA that it has to forward the packets, sent to the MN, to the new FA. After that the new FA sends Registration Reply to the MN. At this time the h4N can resume transmission in uplink and receiving in downlink. Additionally, the new FA sends a HA Notification message to inform the HA about the new binding. In turn the HA establishes a new tunnel to the new FA. After that, the HA tunnels the packets to the new FA. Due to the notification of the previous FA, the time to inform the HA about the new binding and to establish a new tunnel is hidden from the application.


3.1. Basic Assumptions

We have designed a simple analytical model to compare the proposed protocol with Mobile IP. The network topology used for our analysis is shown in Figure 2.

393

Figure 2: The network topology

Consider that the MN moves from the subnet served by FA1 to the subnet served by FA2. In order to model the handoff procedure we define the following terms: to: the time at which the handoff begins. &: the time required for a message to pass through the link from node x to y. R,: the response time a packet incurs at router x. PM: the processing time required by the MN during the registration. TM.": the time required by the MN to resume transmission in uplink. Tw.D: the time required by the MN to resume receiving in downlink.

All delays the packets encounter at the different network elements as well as at the links are assumed to be deterministic. This assumption is motivated by simulations of IP networks with mixed traffic (short voice packets in addition to large data packets) with ns2 [14]. Our studies for a managed IP network show that the delay jitter is minimal compared to the overall delay. Especially, the delay is far from being exponentially distributed as assumed by other studies on mobility management protocols [5] [6] [7]. Thus, a deterministic model represents a much better approximation of the problem. Throughout the paper we assume a constant bit rate UDP stream of packets originating from the Corresponding Node (CN) and destined to the MN. We assume that one packet of the UDP stream arrives every T = 10 ms at the HA.

3.2. MIP Model

If The handoff delay in the case of MIP is given by Eq. (1):

where 4*tm,FAr accounts for the agent discovery. The time at which the HA is informed about the new binding is given by Eq. (2):

Tm.u= T ~ . D = 2*Pm -k 4*tM,FAr ~ * R F A ~ + 2*tFAr,HA RHA (1)

(2) tl = PMN + 3*tw,~,42 + R F A ~ + tFA2,HA + R H A Assuming the constant UDP stream described above, each packet arriving

at the HA belongs to one of the following classes: Class 0: Packets arriving at the FA1 before to; these packets are forwarded

394

directly to the MN. Class 1: Packets arriving at the FA1 after to; these packets are lost. Class 2: Packets arriving at the HA after t l ; these packets are forwarded to the

FA2, which forwards them directly to the MN.

3.3. MIFA Model

In the case of MIFA, the time required to resume transmission in uplink by the MN is given by Eq. (3) :

The time required to resume receiving in downlink is given by Eq. (4):

The time at which the FA1 is informed about the new binding is given by Eq.

Tm-u= 2*Pm + 4 * t W , F A Z + RFAZ ( 3 )

(4 )

(5)

(6)

TMN-D= PMN + 4 * t W , F A 2 + 2*RFAZ + 2*tFAZ,FAI + RFAl

(5): t 2 = PMN + 3 * t W , F A 2 + RFAZ tFA2,FAI + RFAI

t 3 = PMN + 3*tMN,FA2 + RFAZ + fFA2,HA + RHA The HA is informed at the time t 3 , which is given by the Eq. (6):

Each packet arriving at the HA belongs to one of the following classes: Class 0: Packets arriving at FA1 before to; these packets are forwarded directly to

Class I : Packets arriving at FA1 after to and before t 2 ; these packets are lost. Class 2: Packets amving at FA1 after t 2 ; these packets are forwarded to the FA2,

Class 3: Packets arriving at the HA after t 3 ; these packets are forwarded to the

the MN.

which forwards them directly to the MN.

FA2, which forwards them directly to the MN.

3.4. Performance Comparison of MIFA with MIP

It is also While the packets are forwarded to the MN, each packet follows a specific path of routers according to the class it belongs to. In our mathematical model, this path is the sum of a set of constants. The constants represent the times required for a message to pass through the links between the nodes participating in the handoff procedure and the delay incurred by the MN as well as the routers.

From the Eq. (1) we notice, that TMN-U and Tm-D for MIP depend on tFAZ,HA,

while Eq. ( 3 ) and Eq. (4) indicate that TMN.U and T M N - ~ for MIFA are independent of tFA2,HA. Thus, the increase of the distance between the HA and the current FA produces more undesirable latency when using MIP. In contrary, this increase has no influence on the handoff latency when using MTFA. In order to evaluate the performance of MIFA compared to MIP, we use the following values for the parameters used in this model: tMN,FAZ = 2 ms, tFA2,FAI = 1 ms, tFA2,HA = 0.. .lo0 ms. The estimated queuing delay incurred at each router participating in this handoff procedure is 0.2 ms.

395

Figure 3 depicts the time after which the MN resumes receiving and sendmg data for both cases, MIP and MIFA. From this figure we can notice, that the two alternatives are comparable when the HA is close to the current FA. However, the latency experienced in the case of MIP will increase when the delay tcuventFA,HA increases, while the latency remains unchanged in the case of MIFA. Thus, the main advantages of MIFA are the independence of tcurentFA.HA

and the approximately constant handoff delay without restriction on the location and the topology of the network.

Dl&ncs bsawasn tU and FA7 MStMfe be(WIM HI and FA2

Figure 3: Latency experienced in downlink Figure 4: Expected number of dropped packets

Considering that the first packet of the constant UDP stream originating from the CN arrives at the HA at tbegin = -80 ms and the MN loses the communication with the FA1 at to= 0. Thus, the time at which the first dropped packet arrives to the HA is given by Eq. (7):

where link-latency is the delay on the link to the FA1 and given by the Eq. (8):

From Eq. (7) and Eq. (8), the number of dropped packets in FA1 can be derived for varying tFA,HA (0 . . . 100 ms). In figure 4 we see that the number of dropped packets in the case of MIP increases when the delay between the HA and the current FA increases, while in the case of MIFA this number varies within a small range. The fluctuation in the number of dropped packets for MIFA results from the dependence of the link-latency on tFA,HA. As a result, the link-latency and t ~ ~ ~ ~ d r

pa& does change when tFA,HA changes.

GrSt droppedpocket = (to - link-latency) - ((to - link-latency) mod T)

link-latency = Rm + tHA.FAI + RFAI

(7)

(8)

4. Conclusion

In the paper, we have analysed the performance of MIFA compared to MIP. Our performance study , focusing on deterministic UDP traffic, shows that the handoff latency of MIFA is independent of the distance between the current FA and the HA. Thus, MIFA highly reduces the handoff latency for most cases. MIFA clearly outperforms Mobile IP (MIP) with respect to the handoff latency and the number of packets dropped due to handoffs.

396

Currently, we implement the MIFA protocol with the Network Simulator (NS2) in order to do a more detailed performance study of the protocol, especially to study the behaviour of MIFA for TCP traffic.

References

1. 2.

3.

4.

5.

6.

7.

8.

9.

Charles E. Perkins. MOBILE IP - Design Principles and Practices. (1998). Gopal Dommety and Tao Ye. Local and Indirect Registration for Anchoring Handoffs. < draft-dommety-mobileip-anchor-handoff-0 1 .txt >. July (2000). Eva Gustafsson, Annika Jonsson and Charles E. Perkins. Mobile IPv4 Regional Registration. < draft-ietf-mobileip-reg-tunnel-08.txt >, November (2003). K. El Malki et al. Low Latency Handoff. in Mobile IPv4. draft-ietf- rnobileip-lowlatency-handoffs-v4-04.txt7 June (2002). C. Blondia et al. Low Latency Handoff Mechanisms and Their Implementation in an IEEE 802.11 Network. Proceedings of ITC18, Berlin, Germany, (2003). 0. Casals et al. Performance Evaluation of the Post-Registration Method, a Low Latency Handoff in MPV4. Proceedings at ICC 2003, Anchorage, May (2003). C. Blondia et al. Performance Comparison of Low Latency Mobile IP Schemes. Proceedings at WiOpt’03, INRIA, Sophia Antipolis, pp. 115-124, March (2003). A. Diab, A. Mitschele-Thiel and R. Boringer. Extension of Mobile IP for Fast Authentication. 49. Internationales Wissenschaftliches Kolloquium, Ilmenau, Germany, September (2004). S. Pack and Y. Choi. Fast Inter-AP Handoff using Predictive- Authentication Scheme in a Public Wireless LAN. Networks 2002, Aug (2002).

10. S.K. Sen et al. A Selective Location Update Strategy for PCS Users. ACM/Baltzer J. Wireless Networks, September (1999).

11. Charles E. Perkins and Pat R. Calhoun. Generalized Key Distribution Extensions for Mobile IP. draft-ietf-mobileip-gen-key-0O.txt. July (200 1).

12. David B. Johnson and N. Asokan. Registration Keys for Route Optimization. < draft-ietf-mobileip-regkey-03.txt >. 14 July (2000).

13. A. Diab and A. Mitschele-Thiel. Minimizing Mobile IP Handoff Latency. Proceedings at HET-NETs’04, Ilkley, West Yorkshire, U.K., July (2004).

14. R. Boringer et al. MPLS/RSVP-TE-based Future UMTS Radio Access Network. Proc. 3rd Polish-German Teletraffic Symposium (MMB & PGTS 04), Dresden, Germany, Sept. 2004.

INTEGRATION THE PROTOCOLS HMIPV6 AND DIFFSERV

NETWORK MOBILITY OVER M-MPLS IN ORDER TO PROVIDE QOS IN IP

JESUS HAMILTON ORTIZ University Polytechnic of Madrid-Spain

Absfracr- this paper presents the integration of protocol hierarchical mobile IPv6 with multiprotocol label switching (M-MPLS) and the protocol of quality of service Diffserv in mobility scenarios. We use for it simulation TCP traffic. The goal and integrated framework is to facilitate efficient and reliable network operation and to provide optimal Quality of Services (QoS).

1. Introduction

One of the main differences between fixed and wireless networks is: in wireless networks it’s necessary to consider that terminal mobility produces an impact on QoS. We mean to reduce this impact by integrating mechanisms to supply QoS like RSVP and MPLS, to wireless networks. We have taken into account the limits related with topology and mobility for our simulation [2] [7]. Problems related with topology: We will talk about the different roaming and how they affect the supply of QoS. The different types of roaming create different loads of signalisation in the access network. Roaming implies the variation of the route followed by the data in order to arrive at the mobile terminal. Any level of QoS and any level of reserve will be interrupted in this situation. The object will be to guarantee a minimal interruption of the traffic during roaming. While roaming, RSVP for example, has problems guaranteeing reserves because the reserve, routing and transmission of data are independent phases of the process. When a change in the route is produced, the packages receive only a Best Effort service until state of reserve introduces itself in the nodes of the new route. The longer it takes to detect the change in the route, the longer the degraded QoS will remain.

397

398

Problems related with mobility: The use of micro- portability protocols helps to diminish the service interruption time that roaming produces due to the necessity of introducing rerouting information and QoS in the new route. The outlines of micro-mobility used can use different mechanisms based in tunnels, multicast, or adaptable routing algorithms. The movement of terminals inside the access network affects the architecture in a different way and also the existing QoS mechanism. In this manner the integrated service mechanism stores states in each router for the flow of data because one route change makes it necessary to realize said states in differentiated there are not states to realize in the routes but level of service can change with the movement. When you try to combine portability and QoS in a network there is a series of aspects of protocols of micro-mobility that need to be taken into account: The use of tunnels that hide information of the original packet, the change of external addresses during lifetime of the connection, the use of multicast to various access routers consumes lot resources, to have one fixed exit point to external networks (one single gateway to the exterior) is less scaleable, the adaptability of the protocol and the techniques used for the actualisation of routes, to have one optimal route between the gateway and the access router and the support of rerouting of the QoS[2]. In our simulation, we have selected an access networks architecture that is characterized by completely based IP; it is adapted perfectly to mobile environment. This architecture has been used to analyze the effectiveness in QoS and mobility in both quality and quantity. The parameter as metrics that we used in our simulation to provide QoS are: packet delay, loss rate, delay jitter, and throughput. This paper proposes the analysis of the parameters to provide QoS with and without integration the protocols in the similar form with and without reserve of resources. Section 2 presents the scenarios of use. Section 3 selected traffic models .Section 4 description of simulation. Section 5 results of the simulation and finally section 6 provides conclusions of all idea presented.

399

2. Scenarios of use

Figl. - Topology wireless LAN.

3. Traffic models selected

In order to simulate the TCP traffic we will use a traffic source of file transfer protocol (FTP). The FTP arrivals can be molded as a distribution of Poisson with a fixed rate per hour. An FTP session is divided in two parts: Connections and spaces. The connections represent commands or data transfer. The spaces represent the time between the end of a connection and the beginning of the next. For the simulation a package size has been chosen of 510 bytes, and a maximum window size of 20 packages. From the different types of FTP transmitters implemented in NS-2, the Sack 1 is a transmitter with selective retransmission has been chosen, which. Among the possible receivers one has been chosen with a combination of selective ACK and ACK delay (Sackl/Delay). In the TCP traffic losses are not produced; the TCP packages are confirmed by the receiver which sends an acknowledgment message called ACK. If no ACK is received corresponding to a package the transmission is sent again from that package on (or only that package is resent depending on the implementation). In TCP a slippery window is defined that consist in the maximum numbers of packages that can be sent without waiting for their corresponding ACK. The measurements done for this

400

type of traffic are: Throughput, Delay endpoint to endpoint, Jitter, Slippery window, and Sequence number [7].

4. Description of trials done

For each one of the three types of traffic, four different trials have been done. The object of said trials is to demonstrate the improvements in the behaviour of the network when resource reserves and intergrading of protocols are applied. These trials are: No interfering trafic: With interfering trafic with interfering trafic and reserve of resources (without integration), with interfering trafic, reserve of resources and integration. The reserve of traffic is done at the beginning of the simulation. The object of the trial is not to study the different characteristics of the QoS protocol, but the interaction of QoS with the micro-mobility protocol: for which there is done only one reserve during the entire simulation, knowing that said reserve can be done now that sufficient resources exist throughout the entire network [7].

5. Description of the simulation

The start of each simulation, the 13 mobile nodes register in the network, 1 in each FA except MH (0) that registers along with MH1 in FA (0) and that later on moves passing from FA to FA. In t=3s, the remote node (host (0) or host (3), depending on if the traffic is TCP, trial or UDP) begins sending the traffic flows to the mobile nodes. In our case UDP traffic. The traffic that is sent is only traffic of interfering interest, it is to say, that which corresponds to the nodes that are sharing bandwidth with MH (0). For the simulation in which there is no interfering traffic, in t=3s only traffic corresponding to MH (0) is sent. In t=9s the MH (0) begins its movement towards the extreme right of the network, which begins to register in the first FA. Throughout the simulation, the MH (0) will go and register in each one of the access routers of the network. Each 40 seconds the MH (0) registers in a different FA.

401

6. Results of simulation

To improve the quality of service offered the solution applied has been the weak integration of protocols. This solution produces an improvement in the results. The losses are shown as well as the throughput for the received by MH (0). The improvement obtained is evident: in the throughput there appears practically no change in access point and in regards to the losses, the number of packages dropped is reduced and the period of time produced in integrations is very small in comparison without integration.

When integration is produced between HMIPv6 and Diffserv over M-MPLS, it is ob served that the bandwidth does not diminish in the roams and it appears much like what we would obtain in the ideal case, around 120 Kbps. The bandwidth received by the MH (0) node is compared in the 4 trials done. It is observed that the corresponding traces to the ideal case and to the case in which integration are applied are practically the same. The only noticeable difference is the absence of the trace corresponding to the case with integration of the peak of the throughput which in the ideal case appears around t=214 seconds. This is due to the fact that a reserve has been made of a 12OKbps.

The packages that in the previous graphs reflect how losses are really packages lost andlor disorganized. If we consider only the packages that are really lost, we can compare the result with the case in which we apply integrated protocols with ideal case (no interfering traffic).We can see, in the ideal case, without interfering traffic, in case of roaming very few packages are lost (2.4 and 3). In the case where integration is used, we obtain that only a few more packages are lost (3 ,9 and 10).

In regards to the packages lost/disorganized we see that the interval of time in which they are produced decreases.

7. Conclusions

1) The integration of the protocols HMIPv6, M-MPLS and Diffserv is a solution to the degradation in the quality of services caused by the movement of the terminals in mobile IP networks.

402

2) The results obtained from the integration of these protocols allow the improvement of the quality of service provided to the network. In Appling the integration between Diffserv and the quality of service protocol is observed that the bandwidth that the mobile makes use of is required, that the losses of packages diminish and that they are found in the instant of the roam that the delay that the packages suffer diminishes, as well the variation in the delay of packages will diminish.

References

[ I ] Vasos Vassiliou, Henry L. Owen, David Barlow, Joachin Sokol, Hans-Peter Huth, Jochen Grimminger, “ M-MPLS: Micro- mobility-enabled multiprotocol label switching”, /CC 2003- IEEE International conference on Communications, vo1.26, no. 1, May 2003 pp 250-255.

[2] Jesus Hamilton Ortiz, Hyldee Ibarra, Sergio Miranda” Limitations the protocols and architectures for to providing quality of service in the big scale”, CIIIEE 2004, November Aguas calientes, Mexico.

[3] Jesus Hamilton Ortiz, Luis Javier Garcia Villalba. “Integration the protocols for to provide QoS in IP network mobility. Springer LNCS August 2004, Aizu Japan.

[4] Ante-project thesis: Integration the protocol HMIPv6 and Diffserv over M- MPLS for to provide quality of service in IP network mobility with bottleneck scenarios. University Polytechnic of Madrid Author: Jesus Hamilton Ortiz. October 28-2004. Without to publish.

[5] H. Soliman, et al., “Hierarchical MIPv6 Mobility Management. “IETF Internet Draft, Nov. 200 1 . [6] J. K. Choi, M.H. Kim, and T.W. Um, “Mobile IPv6support in MPLS,” Internet Draft, aug.2001 [7] A. Bueno. “Evaluacion de 10s mecanismos de calidad de servicios en redes inalambricas de tercera generacion. Proyecto de fin de camera. Marzo del 2003. Universidad Politecnica de Madrid. ETSI de telecomunicacion.

AN AGENT-BASED FRAMEWORK FOR MOBILE MULTIMEDIA SERVICE QUALITY MONITORING AND DIAGNOSIS

MAN LI Nokia Research Center, 5 Wayside Road

Burlington, M 01 803, USA

In recent years, we have witnessed the proliferation of commercial multimedia services being offered over wireless mobile networks. The limited bandwidth and higher data loss and error ratio of mobile network air interface make the monitoring of end-to-end service quality experienced by mobile users a critical element in service management. Equally critical is the ability to quickly diagnose the root causes when service performance degrades. Few studies have been reported on monitoring service and traffic performance of wireless mobile networks, partly due to the difficulty in accessing or installing third party monitoring applications on mobile devices. Though many challenges remain, the increasing openness of mobile device platforms has opened the door for service quality monitoring and diagnosis. This paper proposes an agent-based framework for mobile multimedia service quality monitoring and diagnosis. Monitoring and diagnostic agents can be dynamically dispatched, installed, and activated on end user mobile devices. The proposed framework effectively turns thousands of mobile devices into service performance monitoring stations. It also provides an accurate, efficient way for end-to-end mobile service quality monitoring and diagnosis.

1. Introduction

Advanced wireless technologies such as wide band CDMA enable mobile operators to provide a wide range of new multimedia services to end-users. These services when offered commercially must meet certain performance standards in order to attract and to sustain subscribers. On the other hand, the air interface in wireless mobile networks has relatively low bandwidth and high bit error ratio. As a result, application traffic is subject to higher packet loss and error ratio. These challenges make end-to-end service quality monitoring and diagnosis vital for managing multimedia services over mobile networks.

While there have been many research on traffic monitoring of the fixed Internet, few studies have been reported on monitoring service and traffic performance in wireless mobile networks, partly due to the difficulty in accessing applications or installing third party monitoring applications on mobile devices. The situation has improved significantly as mobile device platforms become increasingly open. It is now possible to

405

406

develop and deploy third party applications on many mobile devices. Still many challenges remain. There is a need for a unified framework to monitor and to diagnose existing and future services since it is not practical to have specific mechanism for each individual service. Any measurement mechanism must take into account the fact that mobile devices have limited memory, processing and battery power, and that mobile wireless networks have limited radio bandwidth.

Existing monitoring solutions for wire line networks include the IETF SNMP [1] and IPFIX [2] infrastructure where network elements such as routers or gateways collect traffic performance data and make them available to central management servers. Unfortunately, network elements usually lack the knowledge of application information. The ongoing work at IETF IPPM [3] group is also more suited for assessing network layer traffic performance.

Previous studies have produced a good understanding of service performance targets for different services, e.g., web browsing [4][5]. Impacts of packet loss, delay, and jitter on service performance have also been studied extensively [6][7][8][9][10]. The importance of monitoring individual service at user terminals is pointed out in [ 111 . However, most of these studies focus on a particular service. There will be numerous services provided over a wireless mobile network but little has been studied about a framework for monitoring and diagnosing many different types of multimedia services offered over the same wireless mobile network.

There has been research on the application of intelligent agents in network managements. A good survey of the topic is provided in [12]. An efficient and lightweight mobile agent platform is proposed in [16] that can play a useful role in network management toolbox. The lightweight approach can be adopted for the implementation of agents described in this paper. Intelligent agents are employed at network elements in [ 131 to process information collected in the S N M P framework and to proactively detect network anomaly. Mobile agents are used in [14] for monitoring large-scale networks. A mobile-agent based monitoring infrastructure for enhanced IP network level services is described in [15]. Once again, these studies focus on network level traffic monitoring by installing agents on or close to network elements. End user- perceived service quality needs to be measured close to end users in order to obtain accurate estimations.

In this paper, we propose an agent-based framework for monitoring and diagnosing mobile service quality. Monitoring agents and diagnostic agents are dynamically dispatched, installed, activated on end user mobile devices for service quality monitoring and diagnosis. Being close to end users, these agents can accurately monitor user perceived service performance, detect performance problems, and help diagnose root causes. The proposed framework effectively turns thousands of mobile devices into

407

service performance monitoring stations. It also provides an accurate, efficient way for end-to-end mobile service quality monitoring and diagnosis.

The organization of the paper is the following. Section 2 discusses the proposed agent-based framework for mobile service quality monitoring and diagnosing. Section 3 discusses future work and Section 4 concludes the paper.

2. An Agent-based Framework for mobile service quality monitoring & Diagnosing

Although research on intelligent agent has been ongoing for decades, there has not been a formal agreement on the definition of an intelligent agent. In the context of this paper, we adapt the definition from [ 171 and define an intelligent agent as a software code that is situated in a computer device and that is capable of flexible autonomous action inside the device in order to meet its design objectives.

Our proposed mobile service quality monitoring and diagnosing framework consists of intelligent agents on mobile phones and central management servers in network operation centers. The framework is shown in Figure 1 . The key elements within the framework are discussed in details in the following subsections.

Central server Mobile network

Figure 1 . Agent-based monitoring & diagnosing framework

2.1. Monitoring Agent

The responsibilities of a monitoring agent are to measure mobile multimedia service quality, to produce and report performance statistics to central management servers, and to activate a diagnostic agent on board the same device when performance degradation is detected. A monitoring agent can conduct both active and passive measurements. For active measurements, the agent actively initiates services and records the performance of each service instance. For passive measurements, the agent observes the services initiated by an end user and records the service performance.

To simplify service management, we propose a common set of performance parameters to be measured by a monitoring agent for all services. Since all services must be evaluated by success rate, response time and the quality of involved media, we define the following parameters for service quality assessment:

408

Service success ratio: The percentage of service instances that are complete successfully Service response time: Measures the speed of response to an end user’s request for service. Applicable to successfully complete services only. Bearer failure ratio: The percentage of bearers (e.g., GPRS) that either cannot be established or are established but prematurely disconnected before the end of a service. Media quality is measured for successfully completed services only. There is no

single metric that can completely characterize media quality. Instead, we recommend the use of multiple parameters to evaluate different aspects of media quality and they are discussed below.

As packets travel through a mobile network, they experience different end-to-end delays. For real time applications, a display buffer is implemented at the receiving end to smooth out the delay jitters so that the media can be displayed continuously. It may happen, for example due to low throughput of the network, that the buffer becomes empty temporally. This causes a play-out break - a period during which the rendering of streaming data stops. A very short break may not be visible to an end user. However, when a break lasts longer than a threshold, it becomes noticeable performance degradation. Some device problems may also cause play-out breaks. We propose the following two media parameters for measuring real time services: 0 Media play-out break ratio: The ratio of the sum of media play-out break durations

that are longer than a pre-defined threshold over the total service time. Number of play-out breaks: The number of media play-out breaks that last longer than a pre-defined threshold While play-out breaks directly measure user perceived play-out quality, it may also

help to record the number of times and length the display buffer content becomes empty, if the buffer content is observable. This will help to identify if play-out breaks are caused by network problems (resulting in empty display buffer) or by device problems.

In addition, we also propose the following three parameters for media quality assessments: 0 Packet loss & error ratio: The percentage of packets that are either lost or detected

as erroneous. Throughput: Throughput of the media involved in the service. Round-trip delay: round-trip delay from a mobile device to a server or to another device. It should be noted that real time applications that use RTPRTCP protocols may

already conduct performance measurements as required by RTCP. Monitoring agents should make use of those measurements whenever possible.

0

0

0

0

409

A monitoring agent periodically submits reports to a central server. To reduce the amount of traffic introduced into the network, an agent shall report only service quality statistics. Detail measurements for each service instance are sent only when necessary, for example, when requested by the server. Since mobile devices move around networks, when recording measurements, an agent also attaches time and location information with each monitored service instance.

A monitoring agent is also configured with a set of thresholds for each service. When the statistics for a service exceed the thresholds, a monitoring agent will activate a diagnostic agent on board to detect the causes of the problem and to fix the problem if possible. Installing monitoring agents on end user mobile devices effectively turns thousands of mobile devices into service performance monitoring stations and provides an accurate way of monitoring end user service performance.

2.2. Central Management Server

The responsibilities of a central management server include deriving key performance indicators out of the reports from monitoring agents, and managing monitoring and diagnostic agents - dynamically dispatching, activating, and deactivating them.

Key performance indicators (WI) are performance statistics that are crucial for evaluating the quality of services provided over a network. The statistics computed by each monitoring agent is not considered as KPIs because they reflect the service experience of a single user only. However, a central management server can derive KPIs from reports made by multiple agents.

These KPIs show the level of the service quality within a given time and space. When KPIs for different areas are computed, a “service weather map” can be effectively constructed with different colors indicating the performance in each area. Such a map would be a very convenient tool for monitoring and displaying overall service quality.

Another responsibility of a central server is to manage the agents - dynamically dispatch, install, activate or deactivate the agents on the phones. For example, when some of the KPI statistics are below thresholds, the server may dispatch diagnostic agents, if it has not done so before, to some of the mobile devices whose monitoring agents have reported low performance.

2.3. Diagnostic agents

The responsibilities of a diagnostic agent are to analyze information collected by a monitoring agent and to detect the root causes of service quality degradations. A diagnostic agent is dispatched by the central management server and is activated either by the server or by a monitoring agent when it senses service quality problems. A diagnostic agent can be considered as a tiny expert equipped with rules or logics to find

410

out the causes of problems. If the source of a problem is local to a mobile device, a diagnostic agent will try to fix it if possible. If the diagnostic results indicate that the degradation is caused by network or server problems, the agent will report the diagnosis to the central server.

Once being dispatched and installed at a mobile device, a diagnostic agent registers with the monitoring agent in the same device. It is activated when the monitoring agent detects performance degradations. The first step in diagnosis is to request the statistics produced by the monitoring agent. An analysis of the statistics will prompt the agent to fire one of the many diagnostic rules. For example, if the service success ratio is below a predefined threshold, the agent may first check if the configurations on the phone for this particular service are correct. It can reset the configurations if that is the problem. Otherwise, the agent will ask the monitoring agent to provide the “status” recorded for the services to find out the reasons for the high rate of service failure. The status typically states a reason, for example, no response from server, unable to establish connection, gateway time out, etc. The agent may need to drill down if necessary. For example, no response from the server may be a result of many faults: the server is down, not able to set up a PDP context, or IP layer connectivity is down. The agent can look at the records at different protocol layers to identify the problem. If necessary, the agent can use Ping or Trace Route utility.

Having an agent conducting diagnosis on a mobile device is likely to be more efficient than having a central management server remotely retrieving information and finding the root causes. Remote diagnosis by a central server depends on the availability of a network connection and also introduces additional traffic into the network. In comparison, a diagnostic agent can be activated by a monitoring agent and can conduct diagnosis regardless if a network connection is available. There is no additional network traffic introduced. Further more, when many devices request diagnoses at the same time, distributed agents can work more efficiently than a centralized server.

Overtime, a diagnostic agent can learn new rules to be more efficient and effective. A concern is the complexity and footprint of such agent because the diagnostic rules could become complex over time. We plan to develop a prototype to evaluate this in detail.

2.4. Agent-sewer Communications

Instead of defining new communication protocols, we propose that the communications between the mobile agents and the central management servers are through the OMA SyncML for device management (DM) protocols [ 18][ 191. These are XML-based client server protocols that allow client and server to exchange additions, deletes, updates and other status information. The information, e.g., WIs , to be exchanged between client and

411

server is organized with a tree structure where all nodes can be uniquely addressed with a URI. The central server can also dispatch, install, activatelde-activate agents through the SyncML DM protocol.

3. Future work

We are currently in the process of prototyping the above proposed framework, including developing agents on mobile phones. The plan is to conduct measurements and analyses of multimedia service quality with the developed agents. In addition, we also plan to test and analyze the impacts of the proposed framework on wireless mobile network capacity and mobile devices.

4. Conclusions

We have proposed an agent-based framework for mobile multimedia service quality monitoring and diagnosing framework. Intelligent monitoring agents and diagnostic agents can be dynamically dispatched, installed, and activated on end user mobile devices for mobile multimedia service quality monitoring and diagnosis. The proposed framework effectively turns thousands of mobile devices into service performance monitoring stations. It also provides an accurate, efficient way for end-to-end mobile service quality monitoring and diagnosis.

References

1.

2.

3.

4.

5 .

6 .

7.

8.

9.

J. Case et al., “ A simple network management protocol (SNMP),” May 1990, RFC 1157. G. Sadasivan et al., “Architecture for IP Flow Information Export,” IETF draft-ietf- ipfix-architecture-03, June 2004. V. Paxson et al., “Framework for IP Performance Metrics,” IETF RFC 2330, May 1998. R. Chakravorty, I. Pratt, “WWW performance over GPRS,” 41h international workshop on Mobile and Wireless Communications Network, September 2002, pp.

J. Schlaerth, B. Hoe, “A technique for monitoring quality of service in tactical data networks,” IEEE MILCOM ‘96, Oct. 1996, pp. 576 - 580. 0. Verscheure et al, “User-Oriented QoS in Packet Video Delivery,” IEEE Network Magazine, NovemberIDecember 1998. J. Jassen et al, “Assessing Voice Quality in Packet Based Telephony,” IEEE Internet Computing, MaylJune 2002. T. Kostas et al, “Real-Time Voice Over Packet-Switched Networks,” IEEE Network Magazine, Januarymebruary 1998. L. Sun et al, “Impact of Packet Loss Location on Perceived Speech Quality”, Internet Telephony Workshop, April 2001.

527 - 531.

412

10. ITU-T Recommendation P.862, “Perceptual evaluation of speech quality (PESQ), an objective method for end-to-end speech quality assessment of narrowband telephone networks and speech codecs,” February 200 1.

11. A. Kajackas et al., “Individual QoS rating for voice services in cellular networks,” IEEE Communications Magazine, June 2004.

12. A. Bieszczad et al., “Mobile agents for network management,” IEEE Communications surveys, Fourth Quarter 1998.

13. M. Thottan et al., “Proactive anomaly detection using distributed intelligent agents,” IEEE Network Magazine, September/October 1998.

14. A. Liotta et al., “Exploiting agent mobility for large-scale network monitoring,” IEEE Network Magazine, May/June 2002.

15. M. Gunter et al., “Internet service monitoring with mobile agents,” IEEE Network Magazine, MayIJune 2002.

16. W. de Bruijn etal. “ Splash: S N M P plus a lightweight API for SNAP handling,” IEEE Network Operations and Management Symposium (NOMS), April 2004.

17. G. Weiss, “Multiagent systems: a modem approach to distributed artificial intelligence,” The MIT Press, 1999.

18. “SyncML Representation Protocol, Device Management Usage, Version 1.1.2”. Open Mobile AllianceTM.

19. “SyncML Device Management Protocol, Version 1.1.2”. Open Mobile AllianceTM.

NEURAL NETWORK AND SELF-LEARNING BASED AUTONOMIC RADIO RESOURCE MANAGEMENT IN

HYBRID WIRELESS NETWORKS

CHONG SHEN, DIRK PESCH AND JAMES IRVINE Centre for Adaptive Wireless Systems

Cork Institute of Technology Cork, Republic of Ireland

E-mail: cshenOcit.ie

The dramatic increase in the number of mobile subscribers has put a significant strain on resource and service provisioning current cellular networks in particular in terms of multimedia and high-data rate service provision. Hybrid wireless networks, which is a novel scalable and adaptive wireless network architecture utilizing a mixture of cellular and ad hoc multi-hop routing, facilitates cellular network design with small cell systems without having to wire a large number of base stations into a core network. In this paper, a self-management architecture for base stations in hybrid networks based on autonomic computing is presented. Neural Network training and Self-learning within the IBM autonomic element are introduced in order to enable self-configuration and self-optimisation at the radio resource and routing layer. Simulation results indicate the advantages of our proposed approach in utilizing wireless radio resource, optimizing system performance and reducing the administration complexity.

1. Introduction

With the increase in demand for wireless communications, traditional cellular network architectures encounter difficulties in supporting both high data rate and ubiquitous coverage. An architectural approach to remedy this problem is called Hybrid Wireless Networks (HWN), in particular HWN with dedicated relay. A number of problems arise from this new wireless architecture, in particular the large number of access points or base stations (BSs) in the system cannot be efficiently managed and configured through a centralised interface anymore as is the case in current cellular networks. In order to remedy these difficulties, a more flexible, self-management approach for radio resource management, multi-hop radio relaying, and routing needs to be developed.

Based on earlier research on radio resource management (RRM) algo-

413

414

rithms we propose a possible solution to self-nianagement of RRM in such HWN utilising Autonomic Computing (AC) principles. Self-management of wireless BSs in HWN is critical to its adoption as a next generation wireless network architecture. Broadband wireless services will not be available in a widespread fashion without such HWN and provides the rationale behind this study.

Previously, Artificial Intelligence (AI) methods have been applied extensively to Engineering problems. Neural Networks (NN) are especially useful for function mapping and module prediction, hence we study individual RFtM algorithms in a trained feedforward NN to develop a better high level policy output. The IBM Autonomic Element (AE) concept is used on BS. Self-Learning is a suggested method for AEs to do explore, learn, and explore. Therefore, self-learning for managing total system gain after NN translation on system behavior implication are deployed by each BS to achieve self-configuration and self-optimization principles.

2. Hybrid Wireless Networks with Dedicated Relay

HWN l , which is a novel scalable wireless network architecture utilizing a mixture of cellular and ad hoc multi-hop routing, facilitates cellular network design with small cell systems without having to wire a large number of base stations into a core network. Although advances in communication technology make mobile terminal based ad hoc relays feasible, However, two reasons make mobile based ad-hoc relays unacceptable. First of all, recal- culating and reconfiguring on ad hoc topology introduces volatility into the communication system and makes quality of service provision a challenge that has yet to be overcome. The second is security. In HWN without dedicated relays, mobiles, wireless enabled laptops, PDAs and other portable wireless devices are involved in the routing process. This begs the question: “DO you want your data routed via somebody’s mobile or laptop?”. There- fore, adaptive and scalable wireless network architectures with dedicated relay stations are more suitable. However, this new network architecture also has drawbacks similar to pure cellular networks in terms of routing complexity, radio resource heterogeneity and network infrastructure design growth. Centralised system administration methods become too inflexible and result in an unmanageable system. A possible solution is to give total freedom to individual BSs, which means let BSs make autonomic decisions without central planning.

415

3. Using Autonomic Computing

A recently introduced concept - AC2, based on stimulation from biological systems, may provide a remedy to the network management complexity. AC interrelates capabilities of self-optimizing, self-configuration, self-protecting and self-healing within the self-management scope. Self-configuration involves automatic incorporation of new components and autonomic component adjustment to new conditions 4. Self-optimization involves autonomic components coordination to achieve optimum behavior based on a set of goals. Due to the system complexity, configuration and optimisation are difficult to achieve in one step, hence self-learning becomes a possible approach to achieving self-optimisation. Here we investigate the use of NNs to implement self-learning behavior. Using supervised training we believe that self-optimisation in radio resource allocation can be achieved.

The autonomic system architecture we consider here is composed of an interactive collection of AE, Figure 1. AEs manage both their relationships with other AEs and internal behavior according to established policies. An AE typically consists of one or more managed elements coupled with a single autonomic manager that controls and represents them. The Autonomic manager is responsible for monitoring and managing the knowledge-based processes such as monitoring, analyzing, planning, and executing in managed elements.

Figure 1 Figure 2

We investigate how neural network based self-learning approach embedded into AEs can achieve optimum radio resource in hybrid wireless

416

networks using a flexible combined TD/CDMA, system. In prior work, we have investigated radio resource management algorithms that focus on distributed decision making in BSs ’. As part of this work we proposed a framework for an AC based infrastructure to facilitate HWN. Figure 2 shows a section of a typical protocol stack in the framework. Adaptive application services in upper layers provide for autonomic service creation and management. We assume that data is sent using a reliable network layer, we also assume routing is also based on autonomic behavior which can achieve self-optimization and self-configuration. Figure 3 illustrates network control algorithms mapped onto autonomic features within the overall framework for autonomic hybrid wireless network operation.

Figure 3 Figure 4

A Feedforward Trained “-is introduced as a means of giving the AEs in each BSs (whether core wired or relay) more autonomic decision making ability. Each feedforward NN contains three layers with a total of 49 neu- rons. Learning in NN is achieved using the back propagation algorithm. Sufficient historical data is provided as the network evolves its operation. TDD-CDMA is selected as a radio access technology in our study because routing algorithms and RRM are all important aspects in the operation and lend would need to be made adaptive to achieve autonomic behavior. A simple model of the TDD-CDMA system is implemented in our simulation environment, which allows us to experiment with HWN and also to generate sufficient data for NN training. We translate accumulated data about the state and operation of the radio access system, the ‘‘past experience”, from simulations results into three hierarchical policy classes, Goal Polices, Action Polices and Utilities Polices based on behavior impact implication (here, behavior impact implication means the impact on system performance parameters such as throughput, Quality of Service etc). Finally,

417

several policy rules are trained by NNs Figure 4 to predict the optimum resource allocation pattern.

Here we select as an example a distributed DCA algorithm where autonomic channel allocation decisions are made by AE based BSs. The autonomic decisions are conducted in the following steps (refer to Figure 1 ): 1. Monitor: Input for training neural networks is obtained from monitoring the state and operation of the system in terms intra-cell and inter-cell interference level, available channel count for the system awareness, and current throughput; 2. Analyze: Based on the current training status neural networks analyze as to whether the channel allocation status and QoS meet system requirements (for example, system requirements can be divided into three categories which are Golden, Silver and Normal) or not. If the desired behavior is not achieved, changes in the behavior must be made; 3. Deci- sion: Based on whether requirements are met or not, a decision as to which channel allocation may suit the requirements better are made and a new learning cycle is initiated; 4. Execute: The channel allocation decisions are executed by each BSs individually.

4. Self-Learning in Autonomic Elements

The range of resource management methods have different impact on the system as shown in Figure 5. Individual resource management algorithms not only make positive impact to whole system and AE performance, but can also have a negative impact. For example, load balance directly affect system Cost, Throughput, and etc. and indirectly influence the decision making of distributed DCA and adaptive routing algorithms. Therefore, self-learning is conducted to achieve self-configuration and self- optimization. It helps an AE to refine the behavior implication vector (QoS, Throughput ...) in the form of:

j= l ,n

Where S[G] stands for total system gain after self-learning and the composite function (9) is learnt by neural network, stands for current behavior value, y is workload characteristics when the state change was invoked, v is the current value, and n is the number of individual algorithms we deploy in each AE. For example, if load balance is deployed, P can be presented as LB = (0.4)Throughtput + (-0.3)Cost + O.1QoS + (O)Reliability+ (0.2)Latency where y is calculated through P’s effect on n - 1 algorithms and 7 is a new value after P, and y change. Here, behaviour implication

418

factors are defined ranging from 0 to 1. For example, the Max-Min gain of Throughput is between 0 and 1. It is easy to see that Max gain of p is 5.

The behavior implication vectors we analyze here can be defined as: Throughput: The amount of communication sessions accommodated in a BS simultaneously; Cost: The total spend for managing the system; QoS: The quality of service during a communication session; Reliability: A vector which depends on the call dropping rate; Latency: We take the latency of call admission and system response time into account in this vector.

Figure 5 Figure 6

5. Simulations

In order to evaluate our ideas we carried out a range of computer simulations using Omnet++. Distributed BS channel allocation decision making and centralized channel allocation (FCA) are compared in Figure 6. The difference among three Distributed Dynamic Channel Allocation (DDCA) algorithms is their different levels of signal quality requirement. In other words, increase by degrees is in the sequence of Violate DDCA < Aggres- sive DDCA < Timid DDCA. The performances of all three distributed algorithms are superior than the FCA on the call blocking probability. It also indicates that distributed algorithms are better in system resource utilization than centralized ones. If we set 2 % of call blocking probability as an acceptable level, we can assume the average system Throughput in this HWN is llerlang/cell. From Figure 5, we know Load balance approach allows the system to exploit the availability of the network resources to increase Throughput and reduce end-to-end Latency. It provides additional bandwidths through dedicated relays to alleviate the degrading signal quality, so QoS is slightly improved. However, the increased Cost in installing

419

new dedicated relays should be taken into account. Therefore, different levels of Load balance gain p by Throughput, Latency, QoS and Cost are evaluated. System Reliability is not considered Figure 7. One and a half times the traffic load, which equals 16.5erlanglcel1, is assumed in a hot spot case. As can be clearly seen, that when 39 % Loading balance is used, the system achieves maximum gain.

Another important feature of the TDD-CDMA system is Soft Han- dover. The QoS is improved and Reliability is slighted improved due to the diversity provided by the extra channel path at the cell edge. How- ever, more operational Cost is required to support this procedure and system Latency is potentially increased due to the potential "Ping Pong" effect. Furthermore, guard channels are required in individual BSs for both inter-cell and intra-cell handover. This reduces Throughput. We investigated different percent of guard channels for handover over 128 available system channel. Figure 8 shows the levels of Soft Handover gain p by Threghput , Latency, QoS, Reliability and Cost. The maximum gain occurs when 11% of all channels are reserved as guard channels.

Figure 7 Figure 8

Since distributed DCA, adaptive routing and adaptive power management can improve almost all of the behavior implication factors, it makes sense to monitor the effect of those resource allocation methods with respect to traffic load, signal quality, etc. After behavior implication data accumulation, the neural network is trained to allow for more robust resource prediction and optimization. Last Figure shows the system gain after NN training of two algorithms. 41% of Loading balance and 10% of channels reserved as guard channels are recommended as optimum by the NN. This demonstrates both the optimization ability and accurate prediction degree achieved by our proposed NN and self-learning based radio

420

resource self-management.

. . , . .-. . : . .. . ...

CIIM.!.

6. Conclusion

Implementing A1 based AC on HWN has been proposed. A certain level of self-organization and self-optimization in distributed BS has been achieved based on the NN training and self-learning approach. Simulation results indicate that self-learning and NN training show both the optimization ability and accurate prediction ability.

There is a final aspect that we must consider with regard to real-world use of our approach. Parameters such traffic load, signal quality and etc are constantly changing in a real network. These facts can change the weights of the behavior implications vectors. Hence, we must incorporate recurrent re- training of the NN. For future work we consider defining high level policies for training of individual NN or fuzzy NN based RRM algorithms.

References 1. J. Hubaux D. Nagel and B. Kieburtz, “Provision of Communication Services

over Hybrid Networks,” IEEE Comms magazine, Vol. 37, pp. 36-37, July 1999. 2. J. 0. Kephart and D. M. Chess, “The vision of autonomic computing,” IEEE

Computer, vol. 36, No.1, pp.4150, 2003. 3. C. Shen, D. Pesch and J. Irvine, “Distributed Dynamic Channel Allocation

with Fuzzy Model Selection, ” ITT Conference, Limerick, Ireland, Oct. 2004. 4. X. P. Jing, “A Policy-Based Self-evolving Protocol for Wireless Heterogeneous

Networks,” Parallel and Distributed Computing, Spring, 2004. 5. A. Lozano and D. C. Cox, “Distributed Dynamic Channel Assignment in

T D M A Mobile Communication Systems, ” IEEE Trans, On Vehicular Tech., Vo1.51, pp.1397-1406, Nov. 2002.

WEBBEE: AN ARCHITECTURE FOR WEB ACCESSIBILITY FOR MOBILE DEVICES

KRIAN UPATKOON, WENJIE WANG AND SUGIH JAMIN Department of Electrical Engineering and Computer Science, The University of Michigan, Ann Arbor, MI 48109-2122, USA

E-mail: { kupapatt, wenjiew, jamin} Oeecs. umich. edu

In this paper, we introduce WebBee, a client-proxy architecture that combines a web scraping proxy with a Java client to make a platform-independent gateway between small mobile devices, such as mobile phones, and the vast information available on the World Wide Web. The key technology behind WebBee is a web scraping engine that executes “scraping scripts,” a domain-specific scripting language we designed for the purpose of fetching web pages and selectively extracting information from them. By transmitting to and displaying only this extracted information on the client device, WebBee gives mobile devices with small displays and low network bandwidth clean and quick access to web content customarily designed for desktop browsers.

1. Introduction

Building bridges between the world of small mobile devices and the Internet remains an open area of research. While support for some applications such as e-mail have matured, the ability to serve world-wide-web content on mobile devices have been limited by screen size and network bandwidth.

Current mobiie web browsers (e.9. 0pera)l render each web page to fit the limited display size of the target device, even if the web pages were not designed with small displays in mind. These browsers force the device to download each web page’s HTML file in its entirety, burdening the limited wireless bandwidth of the device.

As an alternative to the generic web browser approach, the mobile- specific solution requires providers to create content specifically made for mobile devices. Examples are web sites written in mobile markup languages such as WML2 or NTT Docorno’s cHTML for i - m ~ d e . ~ However, such a solution relies completely on the content provider to create and maintain these sites.

421

422

We introduce WebBeela a user-driven approach to mobile web interaction. WebBee is a client-proxy architecture that combines a web scraping proxy with a Java client to make a platform-independent gateway between small mobile devices and the vast information available on the World Wide Web. The key technology behind WebBee is a web scraping engine that executes “scraping scripts,” a simple scripting language we designed for the purpose of fetching web pages and selectively extracting information from them. Because scripts can be customized for each web site, the client will display only information of interest to the end user. In this way, WebBee gives mobile devices with small displays and low network bandwidth clean access to web content originally designed for desktop browsers. F’urthermore, the use of a Java client ensures that WebBee is a portable solution among modern mobile devices.

Our prototype implementation of WebBee includes commonly-used services such as directory services and weather forecasts. Trial users have found web access to these services on their mobile phones to take far less time using WebBee than it would using a generic browser like Opera. In addition to the convenience, network traffic is significantly reduced, resulting in cost savings for users with a pay-per-kilobyte Internet access plan. For example, a generic web browser needs to download a total of about 419KB for the full amazon.com web page and to retrieve the price of a book searched by its ISBN. With our WebBee prototype, the mobile client only needs to retrieve about 150 bytes of data. The WebBee mobile client program is compact and computationally inexpensive, expanding Internet accessibility to a large range of mobile devices.

1.1. Overview

At the core of WebBee is a web scraping engine that executes mini-scripts, referred to as “scraping scripts.” Web scraping is the process of extracting information from a web page; the scraping scripting language we designed allows for pattern matching to locate the desired information on e x h web page. Scraping is typically useful for extracting information from sites that have changing content, such as weather reports and news. Combined with user input, scraping can also be used for querying applications such as online dictionary or online auctions.

aA full technical report and sample WebBee applications are available at: http://webbee.eecs.urnich.edu/

423

The WebBee server runs on a machine with a high-speed Internet connection. This server contains a web server as well as the scraping engine itself. The web server allows the WebBee server to act as a proxy between the mobile device and the web sites the user wants to view. The scraping engine is a web server module that is executed to process incoming requests from clients.

The target client device can be any small mobile device with some form of Internet access, such as a mobile phone with an Internet plan. Another common client may be a PDA with wireless access like GPRS or WiFi (802.11 b, etc . ) .

When a client on the mobile device communicates with the WebBee server via an HTTP POST request, the scraping engine module is invoked to service the request. The client indicates which script the user wants to run (or supplies its own script, if necessary), and the scraping engine begins to execute the script. The script instructs the WebBee server to download and scrape the appropriate web page(s). The scraped information is then returned to the mobile device via an HTTP reply.

The advantage of the proxy approach is that all the unnecessary information is eliminated before any data is transferred to the mobile device, reducing transfer time and network traffic costs. Also, because the work of scraping is done off the device, the client program can be kept small and simple. This saves memory and processing power on the device, thus improving interaction response times for the user, as well as extending the device’s battery life.

We illustrate how our system works with a simple example. When a mobile phone user wants to look up the price of a book on amazon.com, she starts a Java client on her phone and enters the ISBN of the book (Fig. 1). The client contacts a WebBee server, which fetches the appropriate script from its cache and begins to execute it. According to the script, the WebBee server then posts a search request to amazon.com for that particular book by ISBN, and downloads the page with detailed information regarding the book. The title and price are scraped from the page, formatted, and sent back to the phone to display. On the surface, the user sees only the end result: a screen showing the price of the book.

Using HTTP as the method of client-server communication brings several advantages. First, in the case of mobile phones, some service providers block access to all remote ports except port 80 (HTTP). Additionally, HTTP connection classes are provided with Java MIDP,4 which handles most communication error handling, simplifying the Java client and keep-

424

Figure 1. Searching for a book’s price from amazon.com.

ing its footprint small. Another important consequence of using HTTP is that the WebBee server appears as if it were just another web server to the clients; the fact that external web sites are being downloaded for scraping is hidden from the perspective of the client.

1.2. Scraping Scripts

In our prototype implementation, WebBee scraping scripts are manually written for each web site to scrape. The scripting language is simple enough that a user with some HTML experience can pick it up in an hour. Creating a basic script to scrape a page typically takes around five to fifteen minutes. The scripting language includes functionality such as fetching a web page, forwarding form data, and pattern searching for locating information.

Similar to sites with mobile-friendly pages, there is a need to customize scripts for each site. However, we argue that the use of scraping scripts provides big advantages over these specialized mobile content pages, and while requiring less work to create.

The first is portability: any mobile device capable of running Java MIDP applications will have access to all the content scraped by any script, regardless of whether the device’s built-in browser understands HTML, WML or cHTML. The content author need only produce a normal HTML site and a scraping script for it, rather than creating multiple versions of each page. If the content author includes “scrape-friendly” tags (described later in this paper) in their pages, no changes to the scraping script will be necessary when the author updates the site’s layout; thus, maintenance of scripts will

425

be kept to a minimum. The second advantage is ease of implementation. In the traditional

solution of creating separate mobile-friendly versions of a site, it is the content provider’s responsibility to author the mobile-friendly pages, and to tie in their functionality to their server back-end. Using the scraping approach, any user (not just the content provider) can create a mobile- friendly interface to an existing web site without requiring any additional access to the server back-end functionality. By opening the development of such scripts to all users, we are confident that WebBee will speed the expansion of mobile-accessible content.

2. Implementation Details

2.1. The WebBee Server

We have implemented a working prototype of our architecture. The scraping engine is written in P y t h ~ n , ~ and runs on any machine that has the Apache web server6 with the mod-python module. Mod-python is an Apache module that embeds the Python interpreter within the server. We chose mod-python because it is proven capable of efficiently handling large volumes of concurrent request^.^

The Apache web server listens for incoming client requests on the standard HTTP port 80. When a connection from a client is established, Apache forwards the request to mod-python for handling. Values of the form fields submitted by the client are extracted and sent as parameters to an instance of the WebBee scraping engine. The most significant parameter specifies which scraping script to run. As shown in Fig. 1, the WebBee engine loads the appropriate script from disk and begins to execute it. The script will in- struct the WebBee engine to download one or more web pages into memory. The downloaded web page is then scraped according to the instructions in the script, and the formatted output is sent through Apache to the waiting client. Once the client has received all the output data, the connection is closed.

Whenever a new user request is received while all existing WebBee engine instances are busy serving other requests, mod-python creates a new subinterpreter to service that request. Subinterpreters ensure that different instances of the WebBee engine do not interfere with each other. Once created, a subinterpreter will be reused for subsequent requests. Thus, the multiple WebBee engine instances remain persistent in memory, keeping overhead minimal for future requests.

426

2.2. Web Scraping Scripting Language

We have designed a preliminary version of our scraping script language. The script is interpreted line-by-line, and each line contains a command to be executed. When the last command has run to completion, the script terminates and the resulting output is sent back to the client. Some important script commands are listed in Fig. 2.

Description 1 (Script commands)

1. fetch {getlpost} <url>

2. seek {startlcurrentlend} <offset>

3. search {fwdlback} {startlend} <exp>

Fetches a web site into an internal buffer.

Moves the cursor.

Searches for the first instance of a string (exp) within the file.

Sets the start/end selection markers to the current cursor position.

Prints the current selection to the output stream

Prints a string to the output stream.

4. select {startlend}

5. select print

6. print <string>

7. if {passlfail} else endif

These three commands are used to test if a command (e.g., search) succeeded, and branch accordingly.

Figure 2. Basic WebBee script commands.

The way information is extracted is based on manipulating a cursor through the HTML file (or files) downloaded from the web site to be scraped. The movement of the cursor is influenced by search commands that attempt to locate the specified patterns in the HTML file that encase the information to be extracted. As the cursor is moved to its desired positions, the script allows for the setting of starting and ending markers at the current cursor position. These markers serve to indicate blocks of text to be extracted. An additional command copies the indicated block of text to an output buffer, removing HTML tags if desired. The contents of the output buffer is returned to the client at the end of the script. A simplified example of how to extract the title from an HTML document is shown in Fig. 3.

In Fig. 3, line 1 specifies the web page that the script should fetch. The second line searches for the first instance of “<title>”. If there is a hit, as indicated in line 3, the script continues to scrape the page. Otherwise, the script stops and prints an error message (lines 9-11). Line 4 marks the

427

start of selection to current cursor position, which is right after the first instance of “<title>”; line 5 search for a following instance of “</title>”; line 6 marks the end of selection to the new cursor position, which is before “</title>”. Lines 7 and 8 then print out the result.

I Description 2 (Script Example)

1. fetch get http://www.some-url.info 2. search fwd end <title> 3. if pass 4. select start 5 . search fwd start </title> 6 . select end 7. print The title is: 8. select print 9. else 10. print No title found!

Figure 3. An example scraping script that extracts the title of a page.

Because all the scraping is done on the proxy, the client program can be kept small and simple.

2.3. Making Pages “Scrape-Friendly ”

An important issue with the use of scraping scripts is that once the layout of a site changes, the patterns that surround each piece of information of interest are likely to change too. This will likely cause scripts to fail.

A content author who creates a scraping script for his own site (or wants to aid script writers) can include “scrape-friendly” tags in their site content. Such tags preserve the validity of scraping scripts throughout any site revisions, and require almost no effort to add and maintain.

The syntax for scrape-friendly tags is simple:

<scrape id=” dat a-element-name ‘I >

</scrape> interesting information goes here

Desktop browsers that do not recognize the “scrape” tags will simply ignore them, rendering the pages without any visible difference. To the WebBee script author however, pattern matching has been made trivial, and they can also be assured that their scripts continue to work correctly after any page layout changes. Content authors who are interested in allowing mobile-friendly access to their content can include these tags at almost no additional effort. In contrast, the current practice of making explicit

428

mobile-accessible content would require the content author to create multiple versions of each page in each mobile markup language (WML, etc.) .

Scrape-friendly tags are not required for WebBee to function normally; they are just an optional means of ensuring scripts remain robust in the face of site layout changes.

3. Conclusion

In this paper, we have presented WebBee, a client-proxy architecture that describes a platform-independent gateway between small mobile devices and the World Wide Web. The design of WebBee encourages each application to specialize in a single task, allowing the user interfaces to be made as simple as possible; simple UIs are more suitable for the small displays of mobile devices. Each WebBee application minimizes the use of the network, requesting as much information from users as possible before connecting to the network to resolve their queries. Web scraping scripts, which are an important part of WebBee, tell the server how to extract information of interest from each web site. Since only extracted information is returned to each mobile client, less data is sent over the mobile network, and the information can be fitted to the small displays of the mobile device. Fi- nally, instead of requiring users to go through a network portal, WebBee applications reside on the users’ mobile devices, further reducing network usage. The WebBee application repository is open to public contribution, requiring no support from each web site’s administrators in order to grow.

References 1. “Opera Web Browser,” http://www.opera.com/. 2. “Wmlscript tutorial from wireless developer network,”

htt p://www.wirelessdevnet .com/channels/wap/training/wmlscript . html. 3. “NTT Docomo I-Mode,”

http://www.nttdocomo.com/corebiz/imode/index. html. 4. “Java Mobile Information Device Profile (MIDP),”

http://java.sun.com/products/midp/. 5. “Python Programming Language,” http://www.python.org/. 6. “Apache Web Server,” http://www.apache.org. 7. “MoinMoin Performance Proposals,”

http://moinmoin.wikiwikiweb.de/PerformanceProposals/.

EMPOWERING WIRELESS UPNP DEVICES WITH WEBPROFILES

JUAN IGNACIO VAZQUEZ AND DIEGO LOPEZ DE IPII?A Deusto University,

Avda. Universidades, 24, 48007 Bilbao, Spain

E-mail: { ivatquez,dipina} Oeside.dewto. es

Our surrounding environment is changing day after day. Almost in an unperceivable way, even though steadily, more and more little computing and communicating devices are populating our homes, workplaces, clothes, streets or cars. All these devices need a common architecture to communicate, self-organise and cooperate, being one of such architectures Universal Plug and Play (UPnP). Wireless UPnP is the appropriate technology for mobile devices that roam around, creating and partaking in ad-hoc networks emerging everywhere. But UPnP still uses an interaction model with the environment not suitable for present users’ needs that require more intelligence around them, as stated in the concept of Ambient Intel- ligence (AmI). Issuing commands continually such as open the door, turn on the light, or play the movie using some kind of universal controller such as a PDA or mobile phone, only relieves the user from having physical contact with the device, but not from completely free him from directing and coordinating the action. In this paper we propose the use of the Webprofiles model to extend UPnP capabilities enabling wireless UPnP devices to act in response to user’s preferences, adapting the environment without being explicitly commanded, and so, getting closer to the new, more subtle interaction model with the activated world.

1. Wireless UPnP

Universal Plug and Play [l] is a standard that describes an architecture for connecting and communicating devices, most of them wireless-enabled. It is strongly based on TCP/IP and Web technologies, mainly HTTP, XML and derived protocols such as SSDP (Simple Service Discovery Protocol), GENA (General Event Notification Protocol) and SOAP (Simple Object Access Protocol). HTTP over UDP, both in unicast (HTTPU) and multicast (HTTPMU) flavours, is also used as a substrate for SSDP communication.

UPnP relies on a zero-configuration auto-descriptive model where no drivers are needed to interact with the devices, but discovering and stan-

429

430

dard interaction mechanisms are applied to achieve a universal ”invisible” networking system. Wireless technologies are at the core of UPnP since they are the selected communication alternative for user-agents (PDA, mobile phones) and for many other UPnP powered devices.

1.1. Adapting the UPnP environment

In a wireless UPnP scenario, a user can make use of a control point powered wireless PDA to discover surrounding devices and interact with them switching the TV channels, checking the heating, validating the identity at the door and so on. The control point acts as a user-agent or proxy that represents the actual user when interacting with an UPnP powered environment. The main UPnP mechanism perceived by the user when performing these tasks is control: a very active control where the user has to command the actions via the PDA interface, graphical or voice-sensitive. This kind of interaction only relieves the user from physically performing the task over the involved devices, but all the previous phases of thinking what to do, which devices are involved, selecting them and invoking the actions must be performed both mentally and physically over the PDA interface.

The outcome is that the whole process of adapting the environment for user’s preferences slows down making highly undesirable to use UPnP wireless technology when he enters home and want to have the present devices (heating, TV channel, lights) configured for his profile. Of course, for concrete actions the user must interact and invoke concrete operations over devices explicitly, but it would be desirable to find a way for automatically configure the environment, and thus, achieving a true Ambient Intelligence ( AmI) scenario: interactions become invisible and unperceivable for the user, but they exist and tasks are performed silently. In an Am1 scenario, the user enters a room and is identified, heating is automatically configured for his preferences and, if present, the TV switches to his preferred show at this time. No action has been explicitly commanded but adaptation has been performed.

2. Passive Influence and Context-Aware Scenarios

The previous examples illustrate how a concrete agent can influence the environment, and thus, its constituent agents’ state (devices), via active or passive methods. Active methods are those in which the agent explicitly commands other agents to change their state or perform an action. Exam- ple: as a user enters the building, a sensor identifies him and commands

43 1

the elevator to come and get him in. When the user stops at the room door his mobile phone commands the electric lock to open.

Active methods can be implemented using any of the well-known distributed computing technologies such as CORBA [2], SOAP (Simple Object Access Protocol) [3], OBEX, etc. In UPnP, strongly based on XML technologies, SOAP over HTTP is used for representing invocations back and forth between control points and devices.

Passive methods to influence the environment are those in which an agent disseminates certain information, expecting that other agents change their state or perform an action at their discretion to create a more adapted environment. Using passive methods an agent does not command the target agents to do anything concrete, it simply publishes/broadcasts information preferences expecting the others react changing their state in a positive way. We can state that passive mechanisms are not intrusive, but they are less predictable. The particular set of information to disseminate by the agent is dependant of the configuration of the environment in which is going to be published. Example: a user behavioral profile can be formed by thousands of different parameters, but only a subset of those are required to adapt an hotel room (with TV set, telephone, temperature and lights) to his preferences.

Anyway, an agent must be aware of the surrounding environment to identify and disseminate the proper information that can influence the neighbor devices in the desired way. Active and passive methods are com- plementary. Active methods perform in a master-slave way, where advanced smart features in agents are not required except for authorization processes. Usually, smart environments are based only on ”command and control” mechanisms that centralize intelligence in only one or few agents that control a greater number of ”dummy” entities.

2.1. UPnP Passive Interaction

UPnP covers quite well the active methods functionality using SOAP over HTTP to implement active control for devices and adaptation. No passive alternatives are provided, which in most of cases would simplify user’s behavior, without worrying about how to interact. Passive methods can be also coordinated with active ones to provide additional information for the device when performing a task. That additional information creates some kind of background for performing the desired process, not forcing but suggesting.

432

For example, an active invocation such as ”switch on the TV” can be complemented by passively disseminated information representing ”these are my favorite shows”. The active command is the former and the passive suggestion for a better adaptation of the task is the latter. Passive interaction mechanisms allow devices to know user’s profile when carrying out an action, probably performing it in a more adapted way. In our research we have found that UPnP can be extended by passive mechanisms enriching its features, without interfering with the existing behavior and creating and interaction model fully compatible with traditional UPnP devices.

3. Wireless UPnP Extended With Webprofiles

3.1. Introduction to WebProjiles

In order to add passive influence capabilities to the HTTP protocol we have developed the Webprofiles interaction model. It is a non-intrusive mechanism that enriches HTTP with passive interaction capabilities if supported by the communicating entities.

The goal of the WebProfiles model is to provide an HTTP-based mechanism to negotiate and exchange contextual information that can be used for the client to obtain more adapted web results. The client is the unique entity that manages the contextual information repository, providing the authorized services with the appropriate subset to generate adaptation. The client repository stores user-related profiles on different knowledge domains, being several profiles allowed and applicable to different scenarios. The point with the Webprofiles model is that the context information is not statically structured and composed, but it is dynamically generated depending on the situation by selecting and grouping the convenient profiles and forwarding them to the service provider. The elements that define the situation and, thus, influence the selection of profiles are the profiles themselves, the service provider data, and the user’s established permissions about profile information access.

All these entities’ data serve as criteria to negotiate and exchange the context information with the service provider, and so, set up the environment for further services execution. The main involved structure is the Webprofile: an XML document representing user preferences on some domain(s) under certain conditions from the same or other domain(s), via a language called WPML (Webprofiles Markup Language).

It is out of the scope of this paper to detail thoroughly the conformation of Webprofiles and how they are generated or changed either by the client

433

(user) or the server (service). It is quite evident that they constitute the context information about user preferences available to authorized services for adaptation.

For example, in a simple Ambient Intelligence scenario] the client disseminates appropriate Webprofiles to present servers, which use them to adapt different services (light and temperature control, TV programs presentation, presence information, among others) in order to provide a more suitable user experience. The servers probably follow a passive influence model as detailed in [4]. The information contained in the Webprofiles is very dependent on the services involved, since they represent the context information understandable by those services, but it is expected to be standardized via XML Schemas or Semantic Web technologies.

3.2. WebProfiles Negotiation

The Webprofiles model defines an HTTP-based negotiation mechanism that allows both client and service providers to set up the context in which further interactions can be performed. The most remarkable phases within this negotiation process involve notification of negotiation capabilities] knowledge domains, Webprofiles selection and delivery] and service adaptation. The figure 1 illustrates the negotiation process at a higher level, stressing the sequence of tasks each party must accomplish.

The detailed description of each step is: (1) The client issues a normal request to get some resource from the service provider. (2) The service provider processes the request and sends back the resource along with information about the types of adaptation available for this and future requests, indicating the supported domains structures about which preferences can be processed to generate a more adapted response. Service Credentials are sent, so the client can verify whether the service provider is authorized to receive the Webprofiles information in order to perform adaptation. If the client does not support Webprofiles, or it does not validate credentials or it does not require adaptation for this service, the negotiation process ends at this point as if it was a normal finalization without Webprofiles. (3) If the client demands service adaptation, it checks the presence of suitable Webprofiles with preferences about the declared domains, in order to create a candidate list of Webprofiles for the service. (4) The client filters the list of candidate Webprofiles against the Service Credentials supplied by the service provider] and thus obtaining the final list of validated WebPro- files suitable for that concrete service adaptation. (5) The client issues the

434

(client) Prouder

2. Response (generic)- [+Domains]

[+ Service Credentials ]

3. WebProfiles Selection

4. Authorization

5. Request +Web Profiles _2_2_$

7. Response (adapted) I- ----- - - - -User-Context establis hed - - - - Figure 1. The Webprofiles negotiation process.

original request adding the validated Webprofiles. (6) The service provider uses the information conveyed in the received Webprofiles to better know the client and adapt the responses. (7) The service provider generates the corresponding response to the request, conveniently adapted by means of the Webprofiles. Now, the contextual information between the client and the service provider is established for further interactions, allowing even dynamic modification by sending Webprofiles updates.

This interaction model illustrates the process of contextualization via Webprofiles. In the case context information is not needed or Webprofiles are not supported either by the client or the service provider, the interaction finishes at step 2 and the overload is minimal in relation to the normal process.

Only if Webprofiles are applicable and agreed by both parties, a further interaction is required where Webprofiles are exchanged in an overall process that resembles HTTP Basic Authentication [5 ] , in the sense that the client is the responsible for resending the original request extended with

435

additional information to obtain a preferred response (client-driven negotiation). In fact, this resemblance is not casual. The Webprofiles model has been designed in such a way that shares many similarities with existing HTTP mechanisms in order to be easily integrated within the hypertext protocol. Nevertheless, the Webprofiles negotiation model does not follow an strict client-driven or server-driven negotiation model as specified in [6 ] , but it shares hybrid characteristics with both of them.

3.3. UPnP Messages with WebProfiles

Webprofiles can be applied mainly during two different processes of the UPnP interaction: description and control. During description, the user agent acting as a Webprofiles client can suggest adaptation from the device when obtaining its description information. The interaction follows the general Webprofiles negotiation process described previously and the interaction is illustrated by the following (not complete) messages:

UPnP User Agent (Control Point) Request -+

UPnP Device +- Response

GET description-uri HTTP/1.0 WP-Version: 1.0

HTTP/1.0 200 OK WP-Version: 1.0

WP-Accept: text/vnd.webprofiles.wpml+xml; cnf-l="http://www.webprofiles.org/dataschemas/ambient"



POST description-uri HTTP/1.0 WP-Version: 1.0 WP-Activate: urn:uuid:f8ld4fae-7dec-lldO-a765-OOaOc9le6bf6

--multipartseparator Content-Type: text/vnd.webprofiles.wpml+xml WP-Content-URI: urn:uuid:f8ld4fae-7dec-lld0-a765-00aOc9le6bf6

 --multipart -separator--

436

HTTP/I.O 200 OK UP-Version: 1 .O

UP-Collection: urn:uuid:f8ld4fae-7dec-lldO-a765-OOaOc91e6bf6; max-age=300



The user agent asks for the device description fulfilling the UPnP description process. The device generates the HTTP response normally, but it adds a WP-Accept header indicating the list of WebProfile configuration domains from which Webprofiles are accepted (in this example, it accepts adaptation of ambient conditions: temperature, music, lights, . . . ). If the user agent keeps Webprofiles with information about those domains in the repository and user’s permissions allow delivering them to the involved device, a second interaction is carried out, where the user agent re-sends the original request along with the appropriate Webprofiles to perform adaptation. The unique identifiers of those Webprofiles are listed in the WP- Activate header and the Webprofiles contents are embedded in the body of the HTTP POST request (in this case it contains preferences about temperature). Finally, the device sends back the response again confirming that adaptation has been performed using the Webprofiles whose unique identifiers are listed in the WP-Collection header, and it will last for 300 seconds (max-age parameter), allowing renewal. Example: when the user enters a room, his PDA (user agent) acting as an UPnP control point discovers a surrounding UPnP heating service using the UPnP defined mechanism and obtaining the description for it. During this process, the PDA negotiates adaptation with the heating service re-sending the description request with the appropriate Webprofiles containing user preferences about temperature. The perceived result is that the heating service meets user preferences with neither explicit human intervention nor action invocation.

During control, the user-agent can invoke an operation on the device supplying Webprofiles in a second message if supported by the device. Again the mechanism is similar to that on the description phase, but now the adaptation scope (the request URJ) is only the invoked service and not the overall device state as with description. Example: the user interacts with his PDA in order to turn on the TV invoking the appropriate action. The PDA (user agent) acting as an UPnP control point negotiates with the UPnP device (TV set) the Webprofiles that can contextualize the action, maybe sending Webprofiles with user’s TV preferences information. The

437

perceived result is that the TV turns on at the appropriate channel to meet user preferences, without human explicit intervention.

4. Conclusions and Future Work

Webprofiles are a suitable mechanism to extend the wireless UPnP architecture in order to create passive influence scenarios, fully compatible with traditional UPnP mechanisms. The user interacts with the environment in a completely free manner, while his user agent adapts the surrounding devices in an unperceivable way, preparing them for further active invocations.

The Webprofiles model is an extension to HTTP with minimal interference with traditional HTTP parties, since the whole negotiation is carried out by the means of added HTTP headers that can be silently ignored by no-supporting entities. In this way Webprofiles enabled control points can interact with traditional UPnP devices as well as traditional control points can communicate with Webprofiles enabled devices.

Wireless UPnP architecture can incorporate Webprofiles as a passive influence mechanism, creating smart and adaptable environments that take advantage of the flexibility provided by wireless communications extending user influence around him in an invisible way.

Acknowledgements

This work has been partially supported by the Cathedra of Telefonica Moviles at Deusto University, Bilbao, Spain.

References 1. UPnP Forum. UPnP Device Architecturel.0. UPnP Forum (2003). 2. Object Management Group. Common Object Request Broker Architecture

(CORBA/IIOP). Version 3.0.2. Object Management Group (2002). 3. World Wide Web Consortium. Simple Object Access Protocol (SOAP) 1.1.

World Wide Web Consortium (2000). 4. J. I. Vazquez and D. Lopez de Ipiiia. A n Interaction Model for Passively In-

fluencing the Environment. Adjunct Proceedings of the 2nd European Sym- posium on Ambient Intelligence, Eindhoven, The Netherlands (2004).

5. J. Franks et al. RFC 2617: H T T P Authentication: Basic and Digest Access Authentication. IETF RFC (1999).

6. R. Fielding et al. RFC 2616: Hypertext Transfer Protocol - HTTP/1.1. IETF RFC (1999).

UICC COMMUNICATION IN MOBILE DEVICES USING INTERNET PROTOCOLS

BINH HOA NGUYEN HONGQIAN KAREN LU

Axalto. Smart Cards Research 831 I North FA4 620 Road, Austin, Texas, 78726, USA

tel: +I 512 331 3350 fax: + I 512 331 3059 email: [email protected], kareiilu(rila.~alto.com

This paper presents new methods for Subscriber Identification Module (SIM) card communication using Internet protocols. These methods enable the handset user to access the SIM card as a network node, for example by using a Web browser on the phone. More importantly, the methods enable the SIM to establish reliable and secure end-to-end connection with any remote Internet server and to secure Internet online transactions for mobile commerce.

1. Introduction

A subscriber identification module (SIM) is a small microprocessor card inserted into a mobile device, such as a cell phone or PDA phone. The SIM authenticates the user to the cellular network; associates the user with an account with the telecom operator; and authorizes the user to use the network’s voice, data, and other services. SIMs for 3G networks are called USIMs (Universal SIM). A SIM or a USIM card is also called a Universal Integrated Circuit Card (UICC). In this paper, we use the terms (U)SIM, SIM, smart card, and UICC interchangeably.

Network Local PC

1. Network SIM access from the mobile 2. Network SIM access from the local PC

3. Remote server access from Network SIM

Figure 1. Network SIM communications

438

439

Until recently, there was no direct, reliable, and secure communication channel available between the SIM and a remote server in the outside networks. In addition, different mobile devices may have different ways for a user to invoke the same or similar SIM application. To improve the user experience, telecom operators have requested Web servers on the SIM cards and the ability to access the SIM via a Web browser in the handset. In the future, a SIM card will be able to store services that belong to different service providers. Currently, there is no direct, independent, and secure channel for a service provider to update or customize a service in the SIM card without passing by the mobile operator.

The methods presented in this paper for SIM communication using Internet protocols address these issues. The methods enable the SIM to establish a reliable and secure end-to-end connection with a remote Internet server for secure online transactions. They also enable the user to access his SIM using a Web browser on his mobile device or on his computer. This not only enhances the human interface to the SIM, but also allows the use of standard applications, such as a Web browser, to communicate with the SIM. Figure 1 illustrates secure communication between the network SIM and other entities.

The paper is organized as follows: Section 2 introduces some related works. In Section 3, we describe a method to enable the SIM card as a network node with TCP/IP stack and communication between the SIM and the phone using Internet protocols. Section 4 presents the method to access the SIM in the phone using a Web browser on the desktop. Section 5 describes how the SIM can establish a direct, independent, and secure connection to any Web server on the Internet. Some application scenarios are outlined in Section 6, and Section 7 concludes the paper.

2. Prior Arts

2.1. Network Smart Card

A network smart card is a smart card that is an Internet node [ 5 ] . Internet protocols (TCP/IP) and security protocols (SSL/TLS) are built into the network smart card. The network smart card can establish and maintain secure Internet connections with other Internet nodes. The card does not depend on a proxy on the host to enable Internet communications. Moreover, it does not require local or remote Internet clients or servers to be modified in order to communicate with the card. As with other smart cards, user information is stored in the network smart card. The smart card only gives out information to the trusted client or server at the user’s authorization. The network smart card can be used

440

to secure Internet online transactions and to provide other Internet applications, such as a Web server.

2.2. WebSIM

WebSIM [l] is a solution to integrate a Web server on the SIM card that permits any network node in the Internet to access this Web server, based on the phone number of the mobile device that holds the SIM. The approach consists of a Web proxy that intercepts HTTP requests from a Web client and encapsulates them into SMS messages to send to the mobile phone over mobile networks. There is no TCP protocol stack in WebSIM; therefore, there is no end-to-end secure communication between a WebSIM and a Web server on the Internet based on standard security protocols such as SSL or TLS. In addition, the maximum length of SMS messages is 160 bytes, so performance is not acceptable.

2.3. CAT-TP

Until recently, there has been no direct and reliable communication between a (U)SIM and a remote entity, such as an authentication server on the Internet. The handset that hosts the (U)SIM relays the communication between the (U)SIM and the remote entity. This not only posts communication reliability issues, but also raises security concerns because there is no end-to-end security. To solve this problem, 3GPP has recently standardized a new protocol called CATTP, which provides reliable data transmission between the (U)SIM and the remote entity. However, there are drawbacks to this approach. First, it increases the communication overhead, because it is another transport layer added on top of the UDP or TCP transport layer. Second, all Internet remote entities have to add this transport layer, which is not an Internet standard, and have to treat wireless clients differently from other Internet clients.

3. Network Connection between Network SIM and Mobile Device

We have implemented SIM functionalities in a network smart card. We call such a smart card with both TCP/IP networking and SIM functionalities a network SIM.

3.1. Method Outline

We use PPP [4] as the data link layer protocol for the mobile device and network SIM communications. The network SIM behaves as a PPP server for the mobile device to initiate the communication. The PPP protocol is a full- duplex protocol, however the SIM and the mobile device communicate with

44 1

each other according to the smart card standard IS0 7816 communication protocol that specifies a half-duplex serial interface. In addition to that, Internet protocols are peer-to-peer, meaning a node can talk at will, while IS0 7816 specifies commandresponse operation where the smart card only responds to a command issued by the terminal. To solve these two problems, we use the Peer UO protocol [S] to enable the half-duplex commandresponse IS0 78 16 protocol to support full-duplex, peer-to-peer Internet protocols. The network SIM card implements the Peer VO client. We have developed a serial driver for the mobile device. This serial driver implements the Peer UO server and acts as a virtual serial port in the mobile device. One can dial through this virtual serial port using PPP to establish a network connection with the network SIM.

After the initial PPP handshaking finishes, the mobile device and network SIM start PPP negotiations. When the PPP connection is established, the mobile device updates its IP routing table to add an entry for the network SIM. The mobile device and network SIM can now communicate using Internet protocols, such as TCP/IP. The network SIM hosts a Web server on the card. The user can use the Web browser on the mobile device to access the network SIM without requiring changes to the Web browser. The network SIM is a true network node supporting Internet protocols.

3.2. Communication Protocol Stacks

I UICC I I TCP I

I I I Peer I/O I I IS07816 1 APDU

Mobile device

I TCP I

physical link network

Figure 2. Protocol stacks on UICC and the mobile device

Figure 2 illustrates the protocol stacks in the network SIM and the mobile device. They have similar protocol stacks in order to communicate with each other. At the lowest level, the UICC and the mobile device communicate according to the smart card standard IS0 7816. To enable the half-duplex commandresponse protocol IS0 78 16 to handle full-duplex, peer-to-peer Internet protocols, Peer I/O protocol is layered on top of IS0 7816. The mobile device uses PPP, which carries Internet protocols such as TCP/IP to dial into the

442

network SIM. The mobile device may have more than one physical link layer and, hence, more than one data link layer, to connect to the outside network, such as GPRS, Bluetooth, Wifi, and USB. Such outside network links make it possible for the network SIM to connect to another Internet node, such as a remote server, over the Internet.

3.3. Peer I/O

The Peer I10 protocol enables the half-duplex commandhesponse protocol IS0 7816 to handle full-duplex, peer-to-peer Internet protocols. The Peer UO essentially consists of two finite state machines: the Peer I/O server running inside the mobile device, and the Peer I10 client running inside the UICC. The Peer VO server forwards the data between the half-duplex commandhesponse communication module (IS0 78 16) and, the full-duplex, peer-to-peer communication module (PPP). It buffers and controls the data flow, but does not examine or change the data passing through it.

3.4. Data Link Layer Connection

To bridge between PPP and the IS0 7816 modules, on the mobile device, we implemented a virtual serial port driver that includes the Peer UO server and the serial port interface that is defined by the operating system of the mobile device. The PPP connection is established through this virtual serial port.

PPP frames

I SIM module I I Peer I/O ciient 1 I I APDl

IS0 7816 I

Mobile device

I RAS I ....................

-b IS0 7816

Figure 3 . Low level communication modules in UICC and the mobile device

Figure 3 illustrates the data link layer connection between a network SIM and a mobile device in order to support network connection between the two. The mobile device is normally an end point of a network. It typically does not have PPP server support, which means it can only dial out to a network and cannot be dialed into. Therefore, to initialize the PPP connection, the mobile device dials into the network SIM by initiating the PPP handshake with the network SIM. The PPP negotiation starts after the PPP handshake finishes. The

443

PPP connection enables the mobile device and the network SIM to communicate using Internet protocols, such as TCPAP.

3.5. IP Addresses and Routing

The PPP negotiation includes IP address assignment. In our case, the mobile device requests an IP address from the network SIM during PPP negotiation. The network SIM also needs an IP address. One way to assign IP addresses is to use private IP addresses, such as 192.168.0.1.

The networking module on the mobile device updates the IP routing table after the mobile device obtains an IP address via the PPP negotiation. This is an existing fknctionality of the mobile device. One way for the user to access his network SIM through the Web browser on the mobile device is to use the IP address of the network SIM, for example, httvs://l92.168.0. I/.

4. Access Network SIM from a Computer

This section describes the method that enables the mobile device user to use the Web browser on hisher desktop computer to access the network SIM in hisher mobile device: the proxy method. In this way, the user has a good human interface and keyboard to manage the network SIM without passing through the mobile network. In addition, the network SIM can secure communication between the computer and other networks (e.g., Wi-Fi networks). This method uses two proxies: a Web proxy server running on the computer and a TCP proxy server running on the mobile device. With mobile devices (PDA or mobile phones) running the Microsoft Smartphone [2] operating system, we use two tools provided by Microsoft: Activesync, [3] which enables the connection and data synchronization between the computer and the mobile device using a USB cable or Infrared port, and Remote API (RAPI), an application programming interface and library that enable an application running on the computer to communicate with other applications on the mobile device through Activesync.

The proxy server on the computer uses the RAPI to invoke the TCP proxy server on the mobile device. The proxy on the mobile device opens a TCP socket connection to connect to the Web server on the network SIM. Thus, the HTTP requests and responses will be transmitted between the Web server on the network SIM and the Web browser on the computer with the help of the two proxies. This proxy method, while enabling network connection between the computer and the network SIM in the mobile device, does not prevent the user from utilizing features of Activesync, such as data synchronization and Internet

444

access between the mobile device and the computer. It does not require manual reconfiguration of the routing tables.

5. Internet Access from the Network SIM

A. Luotonen has described in [6] how to use the CONNECT method of HTTP 1.1 standard to tunnel TCP-based protocols through a Web proxy service. The TCP-based protocol can be HTTP, or HTTPS with SSL and TLS.

We have implemented such a Web proxy tunneling service on the handset. The Web proxy waits on port 8080 for the Web client on the network SIM. When the Web client in the network SIM wants to establish an SSL connection with a remote server over the Internet, for example www.axalto.com Web server, it opens a TCP connection to the handset on port 8080, where the Web proxy listens. The Web client then issues the following request:

CONNECT www.axalto.com:443 HTTP/1 .O User-agent: Mozilld4.0

Upon receiving this request, the Web proxy knows that the Web client in the SIM wants to connect to the www.axalto.com server on port 443. The proxy opens a new TCP connection to the Axalto server on port 443, which is the port for SSL connections. If the TCP connection with the Axalto server is established successfully, the Web client can begin the SSL session with www.axa1io.com. The Web proxy on the handset moves data back and forth between the Web client on the SIM and the remote Axalto.com server. This tunneling proxy service moves encrypted data and does not affect the secure transaction.

With the HTTP proxy tunneling service described above, the Web client in the network SIM can establish a direct secure SSL or TLS session or a normal HTTP session to any Web server in the Internet. In the case of an SSL/TLS connection, security is the same as it would be without the tunneling service.

6. Applications

Mobile payments play an important role in e-commerce because of the popularity of mobile devices and the integration of mobile networks into the Internet. In addition, the mobile device is a personal device that users carry with them all the time. In these mobile devices, the SIM card is the most secure place to store important personal information. With a network SIM card on the mobile device, the user can use the Web browser on the device itself or on the desktop to buy goods or services on the Internet, but the payment can always be done directly between the network SIM and the merchant server. This provides a secure mobile payment method.

445

Users can use the Web browser on the mobile device or on the computer to access the network SIM to customize personal information, such as phone book, SMS messages, or passwords of other services. Mobile network operators can use the HTTPS protocol to update or modify services and information in the SIM of the mobile device. Deploying applications on the network SIM card will be the same as deploying normal Web server applications.

The network SIM card can carry a personal identity server for the mobile user and this identity server acts as an identity provider participating in a network identity federation such as Liberty Alliance Federation [7]. This will allow mobile users to protect the privacy and security of their network identity information and enable access to their services over the Internet by authenticating with the network SIM card only one time (single sign-on standard).

7. Conclusion

This paper has presented a technology that turns the SIM card in the mobile device into a network node. The network SIM card, with its own IP address and an IP/TCP/HTTP protocol stack, is a true network node in the Internet. It enables the use of network security protocols for secure communication between the SIM and other servers on the Internet. This opens a new paradigm for applications in SIM cards and mobile devices.

References

1. Guthery, S., et al., “HOW to Turn a GSM SIM into a Web Server”,

2. Windows Mobile-based Smartphone,

3. Activesync,

4. Carlson, J., PPP Design, Implementation, and Debugging, second edition, Addison-Wesley, 2000.

5 . Montgomery, M., et al. “Secure network card - Implementation of a Standard Network Stack in a Smart Card”, CardisO4, Toulouse, France,

6. Luotonen, A., “Tunneling TCP based protocols through Web proxy

7. Liberty Alliance Project, http://www.proiectliberty.ord. 8. Lu, H.K, “New Advances in Smart Card Communications,” International

Conference on Computing, Communications and Control Technologies, Austin, TX, August, 2004.

CARDIS 2000, September 20-22,2000, Bristol, UK.

http://www .microsoft.com/windows~nobile/ur~ducts/s~~ia~hon~/~e~auIt.ms~x.

http ~/www.m1crosoR.co1~~v1ndowsmobilc/rcsourccs/downloads/pockctnc/act1vcs~c37.1ns~x.

August 23-26,2004.

servers”, Internet Draft, Work in Progress.

MODULAR PROXIES FOR SERVICE ADAPTATION AND SESSION CONTINUATION OVER HETEROGENEOUS

NETWORKS

TIM SEIPOLD AND THITINAN TANTIDHAM Informatik 4 (Communiact ion Systems) RWTH Aachen University of Technology

{ seipold, thit inan} @informatik. rwth-aachen. de

Current middleware systems provide convenient access mechanisms to services for wire bound stationary terminals, but not for all special needs of mobile terminals. Dynamic adaptation of services to varying network quality and support for het- erogenous terminals are not integrated, but must be configured manually. Service persistence and session continuation are great challenges in the presence of mobility across network borders and service adaptation. In this paper, we present an architecture that inserts a node between mobile clients and wired service providers. This node bridges between the service providers and the clients. Thus, it acts as a gateway between the different access networks and we call it Middlegate (MG).

1. Introduction

The ascent of wireless data communication technologies brings a whole lot of new mobile business applications into life. One of these applications are telematics systems. These systems combine IT and mobile communication for the delivery of information towards vehicles and mobile users, improving road efficiency, safety, comfort, asset management and vehicle utilization. This business case of the FAST Integration project is used to present a comprehensive set of problems that are found in mobile enterprise applications together with the proposed solutions. Refer to Fig. l(a) for a simplified overview of a telematics system’s scenario. The user terminals used to access the systems range from desktop workstations to embedded systems.

The terminal nodes’ heterogeneity in operating systems, interaction capabilities and available resources has to be overcome. All those systems employ several communication technologies that differ in terms of bandwidth, QoS, coverage and cost of data transmission: LAN, WLAN, UMTS,

449

450

(a) Classic (b) with Middlegate

Figure 1. Telematics systems business case; entities and connections

TETRA etc. This set of access networks - including non-IP networks - needs to be unified. Several of the access networks may be concurrently available; a smart selection which network to use for what kind of application is mandatory. All highly mobile terminals suffer from frequent interruptions in communications due to roaming etc. Important data like the network identifications changes during this process. These interrupt the communication on the application level between the terminal and its service providers. Further, changes of the access network usually inflict changes of the access quality and the access costs, between which the optimization must be performed. The integration of legacy applications is a major issue in telematics systems. Therefore a system may require the combined use of CORBA, web services, etc. to provide services using an existing infrastructure. Since a telematics system are customized to the needs of every customer, a solution providing excellent composability is required.

None of the existing middleware systems fulfills all the requirements mentioned above. The integration of non-IP-networks, concurrent access networks, the usage over very low bandwidth networks and the suitability for resource-limited embedded systems are especially cumbersome. More- over, the underlying transport paradigms fail to provide transparent support for dynamic network addresses - and dynamic networks. Our solution is to deploy an intermediate hub between the terminals and the service providers, splitting the communication into two individually optimizable halves, see Fig. l(b). The client half depends on the access network and terminal type, the server side half depends on the used middleware. En- abling access to diverse middleware technologies on one side and being a gateway to different networks, this intermediate node was named Middle- gate (MG). Several management functions have to be provided by the MG: Tracking available networks and network addresses of a node, selecting the

451

best suited access network for a service and supporting dynamic network changes while maintaining continuous service provision.

2. Related Work

Middleware was designed as a transparency solution which shields application developers from the heterogeneity of distributed systems. Conventional middleware such as CORBA, DCOM, and RMI does not cover the adaptation to different wireless devices, various communication protocols and diverse application services. Since bandwidth is limited in wireless communications mobile CORBA allows CORBA messages to be transferred over WAP and supports disconnect operations. As this is not sufficient to solve the problems of constrained resources and capabilities of mobile terminals, dynamic content transcoding becomes a more substantial function for middleware [3]. This concept is called reflective middleware or adaptive middleware [9]. To achieve this transcoding, a repository of user, terminal, network and service profiles is required. The profiles are usually represented as an XML metadata describing the characteristics of resources and content formats, e.g. CC/PP [8].

The concept of users being able to continue services while moving across network boundaries or changing terminals is called Virtual Home Environ- ment (WE) [l]. The VHE defines the basic requirements for the MG: Personalization of service environment, dynamic service adaptation and service mobility in order to enable the user to access his home services from a foreign network.

In wireless and mobile communications an application may disconnect before its session completes. This is similar to a case where a user suspends a job at one terminal and resumes it a t another one - so called session mobility which the main issue concerns with suspend/resume operations [4]. The investigation of the mechanisms of session state tracking and logging for services recovery after a disconnection or after environment changes is ongoing research. The conventional techniques such as checkpoint [2] and migration mechanisms [5] will be used for our session management and service persistence [6].

Several academic and commercial projects have addressed middleware as a solution in mobile computing environment and for service adaptation, some even for telematics applications [7]. However, none of these approaches fulfills all requirements of our business case which covers IP and non-IP services’ runtime adaptation and session mobility management.

452

- . _____.

Terminal (Client) 1 Application I

r--

Middlegate Server

Control channel- I

Figure 2. Middlegate, data adapted by adaptation modules for each link

Architecture overview; link between front- and back-end intercepted by the

3. Middlegate

The MG is embedded in a complete framework for telematics systems, where it fulfills a central role by providing various management services and especially by granting control over the communication processes.

In order to improve the connectivity between mobile terminals and middleware systems, it is essential to have access to the underlying transport network to accommodate the data to the current resources. Following the layered approach to networking, conventional middleware inhibits this approach. As the complexity and variety of middleware systems disallows modifications of them, inserting the MG as an intermediate node between the front- and the back-end becomes the solution. Refer to Fig. 2 for our design that splits the communication link into two individually optimizable halves. This change in end-to-end semantics should be done transparent to the application. Therefore it is presented a logical direct link between front- and back-end that is divided into two network links. The MG’s role is to couple the two network links. Within the MG, every network/application combination is associated with an optimized adaptation module (AM). They encode/decode the application data according to the currently used network and translate it to an XML-based meta language, that is used for data transfer between the two AMs that form a direct logical link. On deployment, a MG instance is configured to support a number of applications and network types known beforehand. It is supplied with an adapter repository containing the individual AMs. For a given application, the MG dynamically joins the network dependent adapters to support any combination of front- and back-end networks. At the end nodes, the data processed by the MG’s AMs is handed to the application after recoding it to

453

a common format. To integrate legacy services into the framework, adaptation and network layer on the back-end link may be replaced completely by a custom design. Skipping the extended framework on the server, the MG will immediately interacts with the service itself.

Two types of communication are provided. An end-to-end link from the front- to the back-end is used on the application layer, whereas the network layer (NL) establishs point-to-point links between client (or server) and the MG. The framework shields the transport layer (TL) links used by the application from the underlying network. The NL insulates the framework itself from the operating system and network dependent differences in the available transport mechanisms. Further, it adds the functions that are not natively provided. To minimize the effort required to use the framework, the access paradigms of both layers are designed very similar to the socket model known from IP-programming. Sockets are used as handles for the application to access different network connections, and ports are used for multiplexing multiple connections between a pair of endpoints. This way, the developer is presented familiar mechanisms that are easily mapped to most existing operating systems’ APIs. Transmitted data undergoes several steps of transcoding during the transport. Every step shall not change the semantics of the data, but lossy transcoding is admissive and delib- erate. The application must be known to distinguish between essential data that must not be changed, and other data, that may be transmitted using a simplified representation or skipped completely. Using semantic knowledge about the data transferred, lossy compression may be applied to yield maximum efficiency. Knowledge about the destination terminal is also important for the encoding process. Fitting the data to the terminal’s capabilities minimizes the data transmitted over the wireless interface, saving time and cost while preserving service quality. Profound knowledge about the underlying network is essential for an optimized transmission performance. Changing network parameters have different implications in different networks and require different action to be taken.

Adaptation management works on two different levels. The adaptation manager selects then the proper AM for the current application/network combination and thus optimizes transmission on a coarse level. The AMs themselves then further optimize data transmission while the connection stays within the same network.

Session management is a fundamental feature for maintaining services’ continuation from request to graceful termination. It is a challenge to keep session persistence while users roam across network domains causing service

454

Service Provider A Service Provider B

1 1 - Resume A

GSMIUMTSMILAN

Limited Bandwldlh, Device A Capabilities

GSMIUMTSMILAN

Figure 3. Session Management Framework

disruption. In our framework the MG will establish a session whenever a user logs in. Each session is identified based on the user account and specifies the services which are currently used. Hence every user has to be registered at a home MG. His profile contains the user’s permissions for accessing different services at different service providers, both the home MG and the MGs visited, the data for the VHE etc.

Figure 3 depicts the use of a user’s home MG. Before the user requests any service, he logs into his home MG. It checks his user profile and forwards a list of suspended services to the user’s terminal. This allows users to resume services from any terminal. When users cannot conclude the services usage on one terminal, suspend and resume operations will be supported. A manual solution provides an API that is explicitly called on the user’s wishes to suspend services. With the manual solution consistency can be ensure for secured applications. It is equivalent to the case of the user manually logging out, causing the system to automatically issue suspend messages to the MG for all suspensible applications. An automatic solution is required since a connection may be terminated before the application gracefully terminates. To solve this, timeout mechanisms as well as service and session policies to maintain the sessions are required.

To maintain sessions across different terminals we do not allow the user to access the same service from more than one terminal at the same time, preventing service inconsistency. Therefore a terminal ID (TID) is used for specifying the terminal, not a network address. As the user logs in,

455

the TID is transmitted to the MG, which checks whether the user requests the same services from another terminal and keeps those in a suspended state. These services may explicitly be transferred, or any other services started. The MG must be able to support the disconnect operations; once it receives a service request, a service session is created as a record and identified by a session ID. From this time the operations at the back-end can be kept running while the user is logged out. Once a response message becomes available, the MG checks whether the user is logged in and whether a communication channel is available before it transmits the message.

In order to support service persistence the MG must provide user profiles. Service policies as e.g. Authentication Authorization and Accounting information used to control user access must be contained herein. Service and session maintenance policies are required since the conditions and criteria for handling the sessions from establishment until termination need to be defined. This policy should cover all service states during the service’s lifetime to optimize resource allocation. A session cannot be kept forever since the resources are limited and service results may become obsolete. Stateless services such as an information retrieval can ignore the feature of suspend/resume operations. In our MG environment, session management module places on top of the TL, see fig. 2. However it must ensure that no data is lost; the state information includes the current applications’ states and the TLs’ buffers containing unacknowledged data. The applications have to allow this information to be used as parameters for service recovery on resumption. In order to support session persistence and service availability in the presence of errors, most events causing service interruption need to be handled and hidden from applications. Performance parameters are used to detect upcoming problems and for the checkpoint creation. Con- cerning service performance, the monitoring has to cover the whole session for resource binding such as directory services, fail-over and load balancing services, etc. On service resumption, the MG checks whether the suspen- sion period is being valid for service consistency. After adaptation and resource setup, the checkpoint information is used for service restoration.


We have introduced an architecture that bridges betveen conventional middleware systems and the proprietary applications by inserting an intermediate node, the Middlegate. Located between the wired/wireless mobile terminals and the service providers, it splits the application’s communica-

456

tion link transparently into two individual optimizable halves. A specific transport layer decouples the applications from the trouble of mobility in the presence of changing networks. An asynchronous method for remote procedure calls is offered, improving error resilience on changing access networks. Additional functions to support personal mobility and session continuity are provided as well.

The design of the system’s components is finished and its implementation is ongoing work. Once the core facilities are operational, synthetic benchmarks will be performed in our laboratory setup to compare the MG approach with conventional middleware. The measurement will be focused on the throughput, processing time, transfer speed, service availability to the application before and after user moves in the case of changing terminals and across different network domains.

Acknowledgements

The research results presented in this paper were developed within the FAST Integration project, which is funded by the German Federal Ministry of Education and Research within the ” Softwareoffensive 2006” program.

References 1. 3GPP. TS22.121: Service aspects: The virtual home environment, 2002. 2. K. M. Chandy and L. Lamport. Distributed snapshots: determining global

states of distributed systems. ACM Trans. Comput. Syst., 3(1):63-75, 1985. 3. B. Knutsson, H . Lu, J. Mogul, and B. Hopkins. Architecture and performance

of server-directed transcoding. A CM B-ans. on Internet Technology (TOIT) ,

4. M. Kozuch, M. Satyanarayanan, T. Bressoud, C. Helfrich, and S. Sinnamc- hideen. Seamless mobile computing on fixed infrastructure. IEEE Computer,

5. D. S . Milojicic, F. Douglis, Y . Paindaveine, R. Wheeler, and S. Zhou. Process migration. ACM Comput. Sum., 32(3):241-299, 2000.

6. C. U. D. of Computer Science. Zap: Migration of legacy and network applications home page. http://www.ncl.cs.columbia.edu/research/migrate/, 2004.

7. C.-F. Soerensen, M. Wu, T. Sivaharan, G. S. Blair, P. Okanda, A. Friday, and H. Duran-Limon. A context-aware middleware for applications in mobile ad hoc environments. In Proc. of the 2nd workshop on Middleware for pervasive and ad-hoc computing, pages 107-110. ACM Press, 2004.

8. W3C. Composite capabilities/preferences profile public home page. http://www.w3.org/mobile/ccpp/, 2004.

9. S. S. Yau and F. Karim. An adaptive middleware for context-sensitive communications for real-time applications in ubiquitous computing environments. Real-Tame Systems, 26(1):29-61, 2004.

3(4):392-424, NOV. 2003.

37(7):65-72, 2004.

A CHANNEL PREEMPTION MODEL FOR MULTIMEDIA TRAFFIC IN MOBILE WIRELESS NETWORKS

TSANG-LING SHEU AND YANG-JrNG WU Department of Electrical Engineering, National Sun Yat-Sen University

Kaohsiung, TAIWAN

This paper presents a channel preemption model for multimedia traffic in mobile wireless networks. In the proposed model, a mobile call is distinguishable with four parameters, call type, traffic class, channel requirement, and preemption ratio. Different classes of multimedia traffic possess different priority levels. To effectively reduce dropping probability, high-priority handoff calls are allowed to fully or partially preempt low- priority ongoing calls when the mobile network becomes congested. Performance measures, including the dropping probability of handoff calls, the call-interruption probability of fully preempted calls, the bandwidth reduction ratio of partially preempted calls, and the preemption cost paid for supporting high-priority handoff calls are investigated through a multi-dimensional Markov model.

1. Introduction

Preemption policies, purposely designed to give more radio access opportunities to higher-priority traffic, have been introduced to increase the utilization of network resources. Scenarios for preempting a portion of network resources employed by an ongoing traffic can be divided into two categories: full preemption (FP) is defined when the preempted call is completely disconnected; and partial preemption (PP) is defined when the preempted call continues to operate with degraded quality of service (QoS).

Research on channel preemption schemes for multiple traffic classes in mobile wireless networks has attracted more and more attention in recent years [ 1-31. Considering K (K >2) traffic classes but only one call type, Shi et al. [ 13 derived K-dimensional steady-state balance equations for the scenario that a class-i call may preempt one of the ongoing calls belonging to classj (j > i). Instead of assuming that only one channel is required for each traffic class, Yao e t al. [2] used a parameter, rk, to denote the channel requirement for every call of class k. However, they have oversimplified the full preemption model by assuming simultaneous call disconnection due to preemption is not allowed in any two different traffic classes of either group. Considering the effect of PP, Das et al. [3] introduced an unprecedented framework of bandwidth degradation.

457

458

In their scheme, a call employing i channels (classified as a class-i call) may be degraded by decreasing the number of its channel requirement by one. The degraded call consequently converts itself to a class-(i-1) call after PP.

This paper presents a generalized channel allocation scheme by simultaneously investigating the effect of full and partial preemption. With adequate identification of preemption parameters and little modification, the proposed model can be easily adapted to analyze any previous works with FP and/or PP. The generalized model consists of different traffic classes individually corresponding to different priority levels. The number of channels initially requested by a call is randomly generated, while the number of channels that can be preempted from a call is obtained by multiplying the call’s channels with a factor, referred to as preemption ratio. The preemption ratio is a real number in [0, 11. In other words, an ongoing call could be fully preempted (FP) if the associated preemption ratio is one and partially preempted (PP) otherwise. To effectively reduce the overall dropping probability, a handoff call with class i is allowed to perform full or partial preemption on any ongoing calls with class j given thatj is larger than i.

The remainder of this paper is organized as follows. Section 2 elaborates our generalized channel preemption model for multimedia traffic in mobile wireless networks. In Section 3, we build and solve the proposed model through multi-dimensional Markov chains. The analytical results and discussions are presented in Section 4. Finally, concluding remarks are given in Section 5 .

2. The Generalized Channel Preemption Model (GCPM)

Each cell has a physical capacity of C channels. Multimedia traffic are classified into A4 classes according to their priority levels in a decreasing order, i.e., the priority of class ml is higher than that of class m2 if ml m2, We assume that the initial bandwidth requested by a call, denoted as r, can be expressed as an integer number of channels and mapped into [ 1, R ] , where R is the maximum number of channels initially requested by a call. A parameter referred to as the preemption ratio, denoted as q, is introduced to specify the percentage of bandwidth to be released when preemption is invoked. Specifically, performing preemption on an ongoing call will release 1.41 channel(s) to the requester. The preemption ratio can be any values in a discrete set {q,, q2,. . ., qs}.

A mobile call is identified with its call type (new or handoff call), traffic class (m), channel requirement (Y), and preemption ratio (4). A new call of (m, r, q) is admitted to the system if there are r free channels; otherwise it is rejected. On the other hand, even if the number of free channels is less than r, a handoff

459

call of (m, r, q) may still get the service by fully or partially preempting any ongoing calls with larger m. The preemption follows the discipline: an ongoing call of class m may be selected for another round of channel preemption, if a higher-priority handoff call (ie., with class index smaller than m) is still hungry for gathering its required channels after a series of channel preemption on the other calls whose associated class indices are greater than m. Since more than one ongoing call may exist in a lower-priority traffic class m and they could be different in r or q, our model assumes that from the same class, the ongoing call with the largest r and the largest q is always preempted first.

3. Analytical Model of the GCPM

The GCPM assumes a traffic model similar to the previous works [4-61:

1. The new and handoff calls of (m, r, q ) generated in each cell follow independent Poisson processes with mean arrival rates am,r,q and Pm,r,q, respectively. The call duration time Dm,r,q and cell residence time S,,,,,, are exponentially distributed with means l /~ , , , ,~ ,~ and l/v,,,,, respectively. Besides, the call duration time is independent of the cell residence time. The overall call arrival rate of (m, r, q) is (am,r,g+/l,,,,r,q). The channel holding time Tm,r,q, defined as the minimum of Dm,r,q and Sm,r,q, is also exponentially distributed with the mean l / ~ , , , ~ , ~ = l / ( ~ ~ , ~ , ~ + v , , , , ~ , , ) .

The proposed GCPM can now be built with multi-dimensional Markov

2.

3.

chains in which each state is denoted as

'l,l,I ~ d l , l , Z ~ ~ ~ ~ ~ ' l , l , S ~ d l , Z , l ~ ~ ~ ~ ~ d l , ~ , ~ ~ ' l , 3 , l ~ ~ ~ ~ ~ ~ d M , R , ~ ]

where nij,k (dij,k) represents the number of non-preempted (preempted) calls. Let p(n) be the steady-state probability of n in the Markov chains. We have

811 c n p ( n ) = l (1)

From the rate-equality principle [7], we can obtain the general balance equation:

460

where X<,,ij,k> (x<dij,k>) has the same size as n and contains all zeros except an element being one corresponding to the position of n;j,k (d;j,k) in n (d ) . I(t1) is an indicator function with its value being one if tl is true, and zero otherwise. n' is defined as a preemptable state in the Markov chains. Let C, (C',) be the total number of channels employed by Mtraffic classes at state n (n'). For 1 I I I 6

1. t , o C , I C - j

2. M

i r = i + l t2 a ( c, I c - j ) OR ( C n ir ( r , - r 2 ) 2 j - C + c, )

3. t3 nj , j ,k > 4. t4 o C , IC- j

5 . t5 e ( k e { l , S } ) A N D (c, ~ ~ - r j ( l - g k ) l ) ~ N D ( ~ i , j , k < m a x I d i , j , k } ) M

i t = i + l 6. t , a ( C : > C - j ) A N D ( C n i f ( r 1 - r 2 ) 2 j - C + C : ) A N D @ ( i o , j o , k o )

where Q ( i 0 , '0, ko) is true if the following two conditions hold. 0, f o r i , < i t I M , l I j t < R , l < k t < S 0 , for it = i, , j , < j t I R, 1 I kt I S 0, fo r i t= i , , j t= j , , k , < k t I S n 'il,il,kl - 7, for it = i,, j t = j , , kt = k,

n'il,Jl,kl+ 1 , for it = i , j t = j , k t = k

n',l , i l ,kl , otherwise

d ' i l , i l ,k l+ n ' i l , J l , k l , for i, < it < M , 1 < j t I R, 1 < kt < S

d ' i l , i l , k l + n ' i l , i l , k l , for it = i,, j , < j t I R, 1 < kt < S

d 'il,il,kl + n 'il,i,,kl, for it = i,, j t = j , , k , < kt < S

d ' , l , , l ,kt+ 7, for it = i,, j t = j , , kt = k, ( k , e (1, S } )

d ', l , i l ,k l , otherwise

Note that io > i, 1 I j , I R , 1 I k, 2 S, and 0 I q I n ',o,,oco .

Thus, the new-call blocking probability Bij,k and the handoff-call dropping probability Hij,k can be expressed as, respectively,

Bi,j,k = C p ( n E S B i , j , k ) and H;,j,k = C P ( n E S H i , , , k ) , (3)

where M

ir=i+l S B ~ , ~ , ~ = ( n l c 0 > ~ - j ) , SHi,j,k = C, >c- j , c nj f ( r , - r 2 ) < ~ - c + c ,

Under the equilibrium condition [4], the mean arrival rate of handoff calls can be expressed as

461

where h i j , k denotes the handoff-attempt probability that an ongoing call of (i, j , k) will need further handoffs. If h i j , k = Prob[Dij,k>&j,k] [ 6 ] , we have h i j , k =

The call-interruption probability P ; j , k that an ongoing call of ( i j ,k ) is forced to termination due to fill preemption and the bandwidth partial-preemption ratio Dij,k used to measure the average percentage of bandwidth reduction on the partially preempted ongoing calls of (i,j,k), are derived as follows.

v i j , d ( r i j , k + v i j , k ) .

R 1 i - l R S

' i , j , k ( i + l , k = S ) = - c c c P i p , j p , k p C P ( n " i , j . k ) j r = l r ( j r ) ip=l j p = j r k p = l

L j ~ q k ] ~ i , j , k Di , j ,k ( i t l ,k+l ,S) = 1 . P ( n sDi,j ,k)

J ( " i , j , j + d i , j , k )

M M

if=ip+l i f=i+l SR . = n C, = C - j r + 1, C ni,(rl - r2) 2 j p - C + C,, C nil (rl - r2) + r,J,k { 1

The expected preemption cost (EPC) in terms of the number of ongoing calls being suffered from full or partial preemption is defined as

Epci, j , k = 1 N , p ( n sE;, j , k ) , (7)

The expected channel wastage (ECW) due to excessive preemption is defined as

ECwi, j , k = c Cep(n sEi, j , k ) , (8)

4. Analytical Results

This section presents a numerical example showing how to evaluate the performance of the proposed GCPM based on the equations derived in Section 3. Setting the parameters as shown in Table 1, the analytical results of performance

462

metrics are presented from Figures 1 to 5 for the six traffic types of (m, Y , 4): (1, 1,0), (1,2,0), ( 2 , 2 , 0 . 5 > , ( 2 , 2 , I), ( 3 , 2 , 0.9, and ( 3 , 2 , 1).

Figure 1 shows the hand-call dropping probabilities of the six traffic types. It is demonstrated that the handoff-call dropping probabilities of class 1, Hl , l , l and HI,^,^, are both smaller than those of classes 2 and 3 . This is because the highest-priority class-1 calls could perform partial or full preemption on the other two classes. Similarly, the handoff-call dropping probabilities of class 2 , H2.2.2 and H2,2,3, are smaller than those of class 3. Besides, HI,^,^ and H1,2,1 exhibit significant differences even when the two types of traffic are in the same class.

Figures 2 and 3 illustrate either of the bandwidth partial-preemption ratio and call-intemption probability increases as the new-call arrival rate increases. The ongoing calls of class 3 experience more bandwidth degradation or have higher interruption probability than those of class 2 . On the other hand, an interesting phenomenon appears in Figure 3 : the call-interruption probability increases rapidly to a peak first and then decreases slowly as a increases. This is because the number of ongoing calls that can be fully preempted becomes smaller for larger a than that for smaller a.

Figures 4 and 5 show the EPC and ECW paid for supporting higher-priority handoff calls, respectively. The EPC for the traffic type of (1, 1, 1) is lower than that of (1, 2, 1) since the average number of ongoing calls required to be preempted simply depends on the initial channel requirement of the requesters. It is also observed the EPC for the traffic types of ( 2 , 2 , 2 ) and ( 2 , 2 , 3 ) is the lowest. This is because the number of ongoing calls on which the lower-priority traffic can perform preemption is always smaller than the higher-priority traffic. The ECW for the traffic type of (1, 1, 1) is higher than that of (1,2, 1). Actually, it is true only for the example, which has too small R (the maximum channel requirement of a call). Similar to Figure 3 , some curves of the EPC and ECW do not monotonically increase with respect to the increase of a. One logical explanation for this phenomenon is made; when a is small, the number of low- priority calls that can be preempted is still large, but when a is increased to a certain point, the low-priority calls admitted to the system becomes small.

5. Conclusions

We have presented a channel preemption model for multimedia traffic in mobile wireless networks. In addition to the new-call blocking and handoff-call dropping probabilities, the call-interruption probability and the bandwidth partial-preemption ratio, are derived. To quantitatively measure the side effect of full and partial preemption, we also defined the expected preemption cost and

463

Parameters Total number of channels (C) Total number of traffic classes (M) The maximum channel requirement of a call (R) The set of preemption ratios ({qr, 4 2 , q 3 ) ) New-call arrival rate of each traffic type (a) Mean channel holding time of each traffic type (Up) Call completion-to-mobility ratio of each traffic type (dv)

the expected channel wastage. From the analytical results, we have demonstrated the proposed model not only effectively differentiates the handoff- call dropping probabilities for different traffic classes, but also provides some statistical averages for preemption costs based on a certain preemption rules.

Values 8 3 2

{O, 0.5,Il 0.5 -5 calls/minute

1 minute 1

References

1.

2.

3.

4.

5 .

6.

7.

Victor T.-S. Shi, Wang Chu, and William Perrizo, IEEE Trans. Communications, vol. 46, no. 6, pp. 743-746, June 1998. Jianxin Yao et al., IEEE Trans. Vehicular Technology, vol. 53, no. 3, pp. 847-864, May 2004. Sajal K. Das et al., IEEE J. Selected Areas in Communications, vol. 21, no. 10, pp. 1790-1802, Dec. 2003. Romano Fantacci, IEEE Trans. Vehicular Technology, vol. 49, no. 2, pp. 485-493, Mar. 2000. Y.-R. Haung, Y.-B. Lin, and J.-M. Ho, IEEE Trans. Vehicular Technology, vol. 49, no. 2, pp. 367-378, Mar. 2000. D. Hong and S . Rappaport, IEEE Trans. Vehicular Technology, vol. VT-35,

Sheldon M. Ross, Introduction to Probability Models, 7th Edition, San Diego: Academic Press, 2000.

pp. 77-92, Aug. 1986.

1

0.0

0 8

0 7

0 8

0 5

0 4

03

02

Figure 1. Handoff-call dropping probabilities of the six traffic types

Table 1. Parameters Used in the Performance Analysis

464

O ? l r . . . . , , , , , ,

Figure 2. Bandwidth partial-preemption ratios for partially

0 5 7

preempted calls.

Figure 3. Call-interruption probabilities for fully preempted calls.

Figure 4. Expected preemption costs paid for supporting higher-priority handoff calls.

Figure 5 . Expected channel wastages paid for supporting higher-priority handoff calls.

NONUNIFORM-DETECTION-BASED FAST MOBILE IP HANDOFF FOR WIRELESS LANS

BO SHEN', HONGKE ZHANG, YUN LIU, YINGSI ZHAO School of Electronics and Information Engineering, Beijing Jiaotong Universiw

Beijing 100044, P. R. China

The handoff makes the primary contribution towards the latency of mobile IP. Lots of fast handoff issues are already being touched in IETF Mobile IP Working Group. Unfortunately, there is no solution that can be used in the current wireless LANs directly because the link-layer handoff triggers are needed by all of those fast handoff solutions. However, those triggers are not supported by the current wireless LANs standards. In this paper, we propose a nonuniform-detection-based handoff scheme that can provide fast handoff ability to the mobile nodes of wireless LANs with no changes to the hardware and standards. In our scheme, MAC addresses of access points are compared at nonuniform intervals. That enables the mobile node know whether the handoff occurs rapidly. By comparing with the existent solutions, we show that the scheme can reduce the latency of mobile IP handoff and the cost of handoff detection for wireless LANs.

1. Introduction

Along with the increasing demand for mobile communication, the service that has simply offered voice and message transmission services is required to provide multimedia services, and wireless Internet access is becoming more and more significant. Comparing with other manners such as GPRS and 3G technology, IEEE 802.11 [ 11 wireless LAN can provide higher channel bandwidth and flexible network architecture, so it is playing important roles in wireless Internet access. Recently, IETF distributed the standard of mobility support in IPv6 (RFC3775) [2] which would promote the popularization of wireless Internet access technologies. However, the actual mobile IPv6 protocol (MIPv6) only solves the problem of permanent connectivity to the network while mobile nodes (MN) are moving around. Because the latency brought out by link-layer (L2) handoff and network-layer (L3) handoff is biggish, MIPv6 can not provide satisfying performances to the time-sensitive applications yet. Especially in the infrastructure mode

' This work is supported by the High Technology Research and Development Programme of China , and by the Beijing Education Committee Fund for Emphasis Laboratory (No. sYs100040408~.

465

466

of IEEE 802.1 1, L3 can not know the occurrence of L2 handoff in time, because L2 does not offer enough information to L3. This instance induces more latency to the whole handoff process.

In this paper, we proposed a novel mobile IPv6 handoff scheme for IEEE 802.11 wireless LANs. The scheme uses nonuniform detection model to reduce the latency and the cost of mobility detection. Differ from more actual methods, our scheme can provide fast handoff performance to MNs in IEEE 802.11 LANs and lower the burden of mobility detection. Meanwhile, it does not need to modify the L2 hardware and standards of wireless LANs, and it does not need L2 triggers support also.

2. Related Work

At present, several proposals for mobile IP handoff are in existence. Literature [3] presented three methods to achieve low-latency handoff. Those methods assume that a

MN can keep the connectivity to the old access point (oAP) when it is attaching to the new access point (nAP). The research in [4] indicates that those methods can improve the performance of mobile IP handoff, but one wireless interface of a MN can not establish two L2 links to the APs in IEEE 802.1 1 environment. So the above methods can not be used in the wireless LANs.

The fast handoff scheme for MIPv6 (FMIPv6) [ 5 ] gets the message that handoff is on hand through the beacon of L2, and then sends the fast binding messages to the old access router (OAR) or new access router (nAR). The study indicates that FMIPv6 can decrease the latency of binding update effectively. But this scheme does not impact the handoff latency in a significant way since the main component of handoff latency comes from discovery of nAR [6].

Literature [6] addresses another method that optimizes the handoff in L3. It periodically detects the change of MAC address of the AP that a MN is attaching to. The change of MAC address indicates that the handoff occurs. The method reduces the latency of handoff detection and router discovery process. It does not depend on the L2 trigger, and the modification to L2 protocols is not needed. The limitation of this method is that periodical MAC detection may increase the burden of a MN, and the discovery of the nAR needs L2’s support yet.

3. Proposed Handoff Scheme

3.1. The Basic Scheme

As an analytical result of the MIPv6 handoff process, the handoff detecting action of a MN is a passive behavior when the MN can not know whether the L2 handoff occurs. And that it occupies more wireless resources because the AR sends multicast RA

467

messages on wireless interface. Because the hardware of IEEE 802.1 1 wireless LANs can offer the MAC address of the AP which the MN attaches to currently, our scheme detects the handoff by probing the variety of MAC address of the AP during the movement of the MN. When a MN is in an ESS, it gets the MAC address of each AP in the ESS and catches those addresses. During the movement, the MN marks the MAC address of the AP which the MN currently attaches to, and then detects it every other interval time. If the MAC address is different from the marked MAC address in the catch, it shows that the MN has attached to a new AP. By comparing with the addresses in the catch, the MN can know if it has been in a new ESS.

3.2. Nonuniform Handoff Detection Model

The interval of detection is the key factor, which impacts the performance of mobility detection in the above detection process. Larger interval will lead to more latency, whereas smaller interval will bring much burden to MN. In order to solve those problems, we propose a nonuniform model.

It is assumed that the max overcast radius of the AP is d, as show in figure 1. Let I ( x ) be the handoff intensity when the distance between MN and AP is x. The nonuniform detection interval can be presented by equations (1) and (2).

A1 = P A (2)

--______----- Fig. 1. The mobility detection of a MN handofling between ESSs

In the above expressions, (1) is called compression characteristic of detection interval, where /I is the proportion of interval, and p is the compression parameter. Equation (2) calculates the handoff detecting interval, where ~t is the interval, and A is the max detecting interval. The handoff intensity I ( x ) is the token of handoff probability when MN is located in the place x. It indicates that larger the handoff intensity is, larger the probability of handoff is.

468

Figure 2 shows compression characteristic curve of detection interval defined in equation (1). Obviously, the interval proportion becomes larger when the handoff intensity of MN is smaller. So selecting the interval according to equation (2), a MN can slow the detecting frequency to lighten the burden on the location where the handoff probability is small, and quicken the detecting frequency to reduce the handoff latency on the location where the handoff probability is large. Further, the changing trend of interval is diverse when p takes different values. When p is enlarged, the compressing effect is obviously enhanced.

0.0 0.2 0.4 0.6 0.8 1.0 HandoffIntensity

Fig. 2. Compression characteristic curve of detection interval

3.3. Address Mapping and Care-of Address Configuration

A MN needs to obtain the IF’ address of nAR after detecting handoff. In our scheme, a map mechanism from IP address to MAC address is used for finding the IP address of nAR quickly. The basic idea is that the AP of a ESS gets the IP address of the AR; the AR gets the MAC addresses of the APs in its neighbor ESSs by exchanging advertisement messages with the ARs in its neighbor ESSs; and then the AR builds an one-many mapping relationships from IP address of the AR to MAC addresses of APs which belong to the same ESS.

In the handoff process described in section 3.2, let y be the threshold of pre- registering care-of address. When p s y , a MN sends pre-registering request to the OAR. The OAR uses the method of Stateless Address Autoconfiguration to construct the care-of address basing on the MAC address of the MN and the subnet prefix of its neighboring ESS, and then processes duplication detection of care-of address and pre-registion to the ARs in neighboring ESSs. If successful, OAR return pre-register acknowledge message to the MN. The ARs of neighbor ESSs mark the care-of address and sets a pre-register timer. When the MN enters a new ESS, the MN uses the MAC address of itself and the subnet prefix of the new ESS to construct the care-of address, and then connects to the

469

AR of the new ESS using this address. The AR changes the mark of pre-register address of MN to engrossed status, and informs the old AR of canceling the pre-register of the MN. If the pre-register timer is not timeout at this time, the old AR informs its neighbor ESSs of canceling the pre-register of the MN.

Pre-register mechanism can reduce the latency of handoff process because it finishes the care-of address constructing and handoff detecting synchronously.

4. Analysis and Simulation

4.1. Performance Analysis of Handoff Detection

In the above scheme, handoff intensity function I ( X ) describes the probability change of a MN handoff. Because the handoff intensity of a MN is only used for calculating the handoff detecting interval, and it does not effect the operation of L2 handoff, the change trend of handoff probability can be approximately described by selecting appropriate handoff intensity function. Let I be the calculated value by equation (3) when MN handoffs, and assume that 1 obeys negative exponent distribution showed as equation (4).

% 1 =--,0 < x < d d - x (3)

(4) -u

f , ( l ) = Ae , I > 0

Where d is the max radius covered by the AP, and x is given by equation ( 5 ) .

( 5 ) RTT

2 x = v-

Where v is the radiating speed of wireless signal, and RTT is the Round-Trip time of signal which is measured by detecting packets.

Let the distribution function of I be the handoff intensity, then the handoff detecting interval of a MN can be gotten by equations (1)(2)(3), where the MN is in the location x .

To compare with uniform handoff detection scheme, the average value of ~t is calculated as following. According to equation (6), we can get the probability distribution function of interval proportion p :

470

So its mathematical expectation ~ [ p ] is:

If A = OWI IS , according to equation (2) and (8), when p = 100, ~ [ p ] = 0.7926, the average value of handoff detecting interval is = 15.852ms ; when p = 30, ~ [ p ] = 0.7421,

the average value of handoff detecting interval is ;iT = 14.842~1~. Obviously, the average value of handoff detecting interval of nonuniform detection is larger than that of uniform detection, so the average detecting times of nonuniform detection is fewer than that of uniform detection in a certain time.

As shown as in figure 1, let A = OWI IS , ,U = 30, 2 = 4 , d = 20m . Assume that the MN starts off from AP1, and moves to AP2 along with beeline at a speed of 2 d s . When the distance from the h4N to AP1 is 18.5m, handoff occurs. According to equation (6), the detecting interval is Af = 7.64ms at this time, and the average detecting latency is 3.82ms. The MN has aggregately performed about 560 times detecting until the handoff is over. In the same condition, the detecting interval of the method in literature [6] is lOms, so the average detecting latency is 5ms. Before handoff, the MN needs to perform 926 times handoff detection. If it is not considered that the cost of other calculation, the efficiency of our scheme is enhanced 40%, and the average handoff detecting latency declines 24%. If we assign the threshold of detection beginning to 0.5, the detection in our scheme is only performed about 96 times on the above condition. The comparing results are shown in figure 3.

20 18

h 16 14

c -

Ncnunfam Detection - -

- U n f m Detectlon -

- Handoff

1000 c I

s *0° 1600 c

.i 400 H 200

Unifam Detectm

Handoff 0

0 2000 4000 6000 8000 0 2000 4000 6000 8000

Fig. 3. Detection interval and times of uniform and nonuniform detection Time (m) Time (m)

4.2. Simulation and Results

The proposed scheme has been simulated on the extended ns-2 presented by [7], meanwhile, FMIPv6 [5] and uniform handoff detection presented by [6] have been simulated on the same condition also.

471

In the simulation, the MN starts off from the OAR, and moves to the nAR at speed of 3 d s . The method of calculating handoff latency is: from the time of the MN receiving the last packet from the OAR to the time of the MN receiving the first packet from the nAR. The pre-register threshold V/ is assigned to 0.75.

The figure 4(a) shows the compare of three handoff detection methods. The value of handoff latency in the graph is the average of 50 results. The graph indicates that the handoff latency of the proposed scheme is small than that of uniform detection obviously, but is larger than that of FMIPv6. The results also illuminates that using L2 trigger information is propitious to reduce the latency of handoff. But FMIPv6 needs the modification of E E E 802.11 and 802.3 specifications due to the absence of trigger mechanism in those specifications.

2o

r"i

Fig. 4. (a) Average handoff latency, @) Protocol load

In figure 4@), the top three curves show the amount of information packets, which are sent and received during the process from begin of the simulation to end of handoff. For the proposed scheme and uniform detection, the amount also includes the times of detecting. The nether curves in figure 4(b) show the amount of protocol message packets that is sent on the wireless interfaces by the proposed scheme and uniform detection method. It indicates that the times of detecting action needed by the proposed scheme is far less than that needed by FMIPv6, so the proposed scheme can save wireless link resources effectively. Moreover, the times of detecting in the proposed scheme are less than that of uniform detection. That is to say, the proposed scheme has great detecting efficiency and small detecting load.


In this paper, a new fast handoff scheme is presented for IEEE 802.11 wireless LAN environments. The scheme aims at improving the handoff performance of MN in wireless LAN and solving the problem that most fast MIPv6 handoff methods can not be used to

472

reduce the latency of handoff in wireless LAN environments due to its dependence on the trigger of L2.

Comparing with other methods, our scheme has the following characteristics: First, it does not need the support of L2 beacon or trigger, so it can fleetly be applied in the wireless LAN environments which is very popular now; second, it introduces the nonuniform detection method to L2 handoff detection for the first time. The method reduces the latency of mobility detection efficiently, and decreases the M"s burden; third, it makes use of the cache in which the map from IP addresses to MAC addresses is stored to implement the pre-register of MN care-of address. This manner reduces the latency induced by the process of care-of address construction. The simulation results indicate that our scheme is very effective for reducing the handoff latency of MN, and that it saves wireless resources used for location detection of MN.

Because the scale of network and the distance between MN and HA effect the latency which is brought by the process of binding update, the method for reducing this part latency will be studied in the future.

References

1. IEEE standard for Wireless LAN Medium Access Control (MAC) and Physical

2. D. Johnson, C. Perkins and J. Arkko, IETF standard, RFC 3775 (2004). 3. K. Malki, Internet DraJ, draft-ietf-mobileip-lowlatency-handoffs-v4-09.txt

4. C. Blodia, 0. Casals, et al. Proceedings ifITCI8,5,971 (2003). 5 . R. Koodli, et al. IETF standard, draft-ietf-mobileip-fast-mipv6-0S.txt (2003). 6 . S. Sharma, N. Zhu, T.Chiueh, IEEE Journal on Selected Areas in Comm. 22-4, 643

(2004). 7. The CMU Monarch Project, Computer Science Department, Carnegie Mellon

University. http://www.monarch.cs.cmu.edu/.

Layer (PHY) specifications, ISO/ZEC standard. 8802 ( 1 999).

(2004).

ANALYSIS OF ACD: &JTONOMOUS COLLABORATIVE - DISCOVERY OF USER AND NETWORK IFORMATION

TAO ZHANG, SUNIL MADHANI One Telcordia Drive, Telcordia Technologies,

Piscataway, NJ 08854, USA

SHANTIDEV MOHANTY Georgia Institute of Technology,

Atlanta, GA 30332, USA

Absrrad - Mobile users and devices need to discover and share a growing range of information. Users may want to discover, for example, the existence and locations of local networks, network services and applications, and information contents. Emerging mobile devices may want to know, for example, the existence of neighboring networks and information needed to access these networks. This could reduce handoff delay and the energy expended on attempting to connect to the wrong networks. Existing service discovery frameworks cannot easily support the discovery of dynamic information (e.g., traffic conditions), information regarding neighboring networks, or other location-based information. This paper uses a new approach for real time collection, discovery, and sharing of user and network information, which we refer to as Autonomous Collaborative Discovery (ACD). It analyzes the information reporting and retrieval costs of the proposed methods.

I. INTRODUCTION Mobile users and their user devices (e.g., mobile phones, Personal Digital

Assistants, notebook computers, communication devices in automobiles) need to discover a growing range of information [ 1][2][3][5][6][7]. Mobile devices may want to discover in real time network information they can use to improve mobility support and mobile applications. Several fiameworks exist for discovering network services and devices and they are often referred to as service discovery approaches [ 1][2][3]. However, they cannot easily support the discovery of dynamic information, such as trafic conditions, and accidents. They allow a device to discover only other devices, services, applications that also implement the same service discovery platform. Furthermore, they do not support discovery of neighboring networks, such as the existence of neighboring networks, the addresses of key network elements or user content in neighboring networks.

We advocate that future approaches for information discovery will allow mobile users and devices to discover information about neighboring networks in a timely manner; allow them to discover information they want, not what network providers think they might want; allow them to discover dynamically changing information and location-based information; have zero or little dependency on local networks, except for transporting user packets and support heterogeneous radio systems, which may belong to competing network providers who may not be motivated, or may not have established the necessary agreements, to share network information with each other.

473

474

A new approach for real-time collection, discovery, and sharing of user and network information, referred to as Autonomous Collaborative Discovery (ACD), has been presented in [6] [7] to meet the above requirements. ACD is built upon the principle that all users and devices provide information, which is significantly different from most other service discovery methods where few information providers provide information for most users and devices to consume. In particular, with ACD, mobile users and their devices act as scouts to collect, in an autonomous fashion, the user and network information they want during their routine use of networks. Afterwards they make the information available to other users and devices by reporting the information to a functional entity called Knowledge Server, which all users and devices may query.

11. PERFORMANCE ANALYSIS This section presents analytical models and results on the performance of the

proposed ACD methods focusing on 1) information creation cost and 2) information query cost.

Analytical models and results on the time it takes for scouts to discover all networks of interest and how this discovery time is impacted by the number and mobility patterns of the scouts can be found in [6] [7] .

A. Information Creation Cost

second. C, can be calculated using We denote by C,, the signaling cost of information reporting by the scouts per

where E and o are the unit transmission costs in the wired and the wireless portions of the network path between a reporting scout and the Knowledge Server, N, is the number of times information reporting occurs per second, and dBs-, is the average distance between the wireless base stations (BS) the scout is communicating with and the Knowledge Server. PCs and PC, are the processing costs for information processing at the scout and at the Knowledge Server, respectively.

For analysis purpose, we assume that the scouts are uniformly distributed in the geographical region of interest and that the Knowledge Server is located at the center of the region. Let R be the radius of the geographical region and S be the number of scouts. Then, the average distance of a BS from the Knowledge Server (dBs.m) is 2S*R/3.

To determine the number N, of information reporting per second, we consider three types of networks in the region of interest: wireless LANs (WLANs), wide- area radio networks such as 3G cellular systems, and global-area radio networks such as a satellite networks. Each WLAN may contain multiple radio access points. We assume that a satellite network has global coverage, each 3G network has a coverage area of A,, and each WLAN has a coverage area of A,,,.

475

To calculate N,, we assume that each scout reports the information about each new network it visits every time it moves from one network to another that we refer as inter-system handoff. Therefore, the number of information reporting per second is equal to the number of inter-system handoffs (ISHO) performed by the scouts. This represents the worst case in reporting traffic volume. The reporting traffic volume could be reduced in many cases using approaches such as the one described in [6] .

We assume that the scouts prefer WLAN to 3G, and prefer 3G to satellite networks. Hence, the possible types of ISHO are between satellite networks and 3G, and between 3G and WLAN. The number of 3G-to-satellite and 3G-to- WLAN ISHOs are given, respectively, by [9]

where Lg and L, are the perimeters of a 3G and WLAN network. Ng and N, are the number of 3G and WLAN networks. Furthermore, pw andp, are the density of the mobile users in a WLAN and 3G network. Parameters v and v1 are the average user speed in 3G and WLAN, respectively. We assume that 3G networks have coverage over the same region and scouts are uniformly distributed among these 3G networks. Therefore, the user density of each 3G network is pBN,. We assume that the number of satellite-to3G and 3G-to-WLAN ISHOs are respectively, equal to the number of 3G-to-satellite and WLAN-to-3G ISHOs. Now using (2), the total number of inter-system handoffs per second, hence the number of information reporting per second, is given by

B. Information Query Cost Mobile devices can query the Knowledge Server for information about their

neighborhood when they anticipate the need for an inter-system handoff. We follow an analysis similar to Section 1I.A to derive the cost, C,, associated with the information queries as follows,

where Nq is the number of queries by the mobile users per second, and PCUQ and PCmQ are the processing costs for the query message at the mobile terminal and the Knowledge Server, respectively. Other parameters have similar meaning as defined in Section 1I.A. The expression for N, is given by,

476

where pw and pg are the density of the mobile users in a WLAN and 3G network. We assume that the density of the users is identical in different types of wide area networks such as 3G and satellite networks, whereas the user density in a local area networks such as WLAN is higher than that of wide area networks, i.e., pw > Pg-

111. PERFORMANCE EVALUATION This section presents numerical results of ACD’s performance analysis based

on Section 11.

A. Information Creation Cost

0.12

0.1

0.08

5 0.06 0.04

0.02

0

-5=20 +5=40

10 20 30 40 50 v (kmlhr)

Figure 1 : Effect of v on Nr.

12

10

8

5 6 4

2

0

-5=40 -%- 5=60

10 20 30 40 50 v (kmlhr)

Figure 2: Effect of v on Cu.

Figure 1 shows the value of N, versus the average velocity (v) for different number of scouts. We assume that the average velocity of scouts inside a WLAN (v,) remains constant and equal to 5 kmhour. Hence, we only vary their average velocity in 3G and satellite networks (v). We assume that a WLAN network has a coverage radius of 500 meters and 3G has coverage radius of 50 km. We assume the values of 2, 1, 25, and 15 for co, E, PCs, and PCKS, respectively [9]. We

477

assume Ng = 10 and N, = 100. We assume that the population density of the region of interest is 10,000 per square kilometers, which is a population density close to that of New York City. We then consider the number of mobile users as a percentage of the total population. For our simulation, we have varied this percentage from ZO% to 80%. This results in different value of the mobile user density in a typical wide area network, i.e., pg. Moreover, we assume that

Figure 1 further shows that for a futed number of scouts, the value of Nr increases with v. This is because faster moving scouts move across more networks. Figure 1 also shows that Nr increases with the number of scouts. From Equation (l), it is clear that the information processing cost Cu is proportional to Nr. This is because all the terms in Equation (1) except Nr are independent of the number of scouts and v. Therefore, Cu increases when Nr increases as shown in Figure 2.

The above analysis shows that there is a tradeoff between how fast we want to discover all the networks of interest and the associated reporting cost. When most subscriber mobiles are interested in highly dynamic changes in the networks or highly dynamic application layer contents, more frequent information reporting by the scouts will be needed, leading to higher C,,. On the other hand, when most mobiles are interested in static or infrequently changing information such as the existence and locations of networks and addresses of key network elements (e.g., IF' address servers, authentication servers, and content servers), the rate of information reporting can be low, resulting in low C,. With the reporting rate reduction method in [6], the rate of reporting will decrease naturally, as the Knowledge Server accumulates more information about the networks of interest.

Pw. =lop,.

B. Information Query Cost

-L-ho-g = 0.0033 - h p g = 0.005

Figure 3: Effect of v on Nq.

To determine the information query cost C,, we consider identical values for N, N@ L , Lg, p,, pg, and VI as in our analysis for N, and C,, in Section 1II.A We also assume that PCUQ = PCs and P C ~ Q = PCm.

Figure 3 shows the value of Nq versus the average velocity v of the mobile users for different value of pg. The results show that for a particular value of pg, Nq increases as the average users' velocities increase. This is because when users move faster they come across more number of networks. Figure 3 further shows that Nq increases with p. This is because when more mobile users are present, the

478

number of queries will increase. Therefore, Cq increases when Nq increases as shown in Figure 4.

2000000 1800000 1600000 I400000 1200000

800000 600000 400000 200000

0

0” 1000000

10 20 30 40 50

v (kmlhr)

-rho-g = 0.0017 +rho-g = 0.0033

-x- rho-g = 0.005 rho-g = 0.0066

Figure 4: Effect of v on Cq.

IV. CONCLUSIONS This paper uses a new framework for mobile users and their devices to

collect, discover, and share information they want. It provides analytical models and results on its information reporting and retrieval costs.

V. REFERENCES Sun Microsystems Inc., http://www.sun.com/jini/ Microsoft Corporation, “Universal Plug and Play Device Architecture”, http://www.upnp.orgl Salutation Consortium, “Salutation Architecture Specification”, June 1, 1999, http://www.salutation.orgl Marco Liebsch and Ajoy Singh (editors), “Candidate Access Router Discovery”, Internet Draft draft-ietf-seamoby-card-protocol-06.txt, Internet Engineering Task Force, June 2004. Tao Zhang, Eric van den Berg, Sunil Madhani, “Peer-to-Peer Network and User Information Discovery and Sharing for Mobile Users and Devices”, the 61st IEEE Vehicular Technology Conference (VTC), 2005. Tao Zhang, Sunil Madhani, Eric van den Berg, Yoshihiro Ohba, “AC-CDS: Autonomous Collaborative ColJection, Discove , and Sharing of User and Network Information”, the 3‘ International Znference on Information Technology: Research and Education (ITRE 2005). Tao Zhang, Sunil Madhani, Ashutosh Dutta, Eric van den Ber , Yoshihiro Ohba, Kenichi Tauiuchi, Shantidev Mohan9, ‘LAutonomous 8ollaborative Discovery of Network Information”, the 3 International Conference on Information Technology: Research and Education (ITRE 2005). R. Durrett, “Probability: Theory and Examples”, Third Edition, Duxbury Advanced Series, 2004. S. Pack and Y. Choi, “A stud on performance of hierarchical Mobile IPv6- based cellular network,” I E d E Trans. Commu., Vol. E87-B, No. 3, March 2004.

[1][2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

EASYMN: AN EFFECTIVE IP MOBILITY SOLUTION FOR HIGH-MOBILITY NETWORK*

LI WANG, MING-WE1 XU, KE XU Department of Computer Science and Technoloa, Tsinghua University

Beijing, 100084, P.R. China

Cellular SCTP (cSCTP) and Mobile SCTP (mSCTP) are two extended SCTP protocols which all support handover management. The existing cSCTP/mSCTP-based IP mobility solutions all must use third party methods to realize location management. In the high- mobility networks applied these schemes, some problems will emerge, e.g. (1) MN's load is relatively heavy. (2) More network resources are used. (3) May result in handover blind zone and handover delay. In this paper, we propose a novel cSCTP-based scheme- EasyMN, it is especially fit for high-mobility network. EasyMN has some notable advantages, e.g. (1) It reduces the load of MN and abrogates location registration mechanism. (2) It avoids handover blind zone and handover delay. (3) It needs not to use third party location management methods. (4) Its location management load is in proportion to the busyness degree of calls, and is irrelative to the number of handover.

1. Introduction

SCTP inherently provides the multihoming feature, this feature makes it possible to support IP mobility [l]. Cellular SCTP (cSCTP) and Mobile SCTP (mSCTP) are two similar extended SCTP protocols with the dynamic address reconfiguration extensions [2,3,4,5]. Both cSCTP and mSCTP support handover management, but cSCTP provides better handover performance than mSCTP [3]. The existing cSCTP/mSCTP-based IP mobility schemes all must use third party methods to realize location management.

The high-mobility network is the network in which most MNs migrate frequently. In the high-mobility networks applied these schemes, some problems will emerge, e.g. (1) M"s load is relatively heavy. (2) Relatively more network resources are used. (3) May result in handover blind zone and handover delay.

In this paper, we propose a novel cSCTP-based solution-EasyMN. It is especially fit for high-mobility network, and has some notable characteristics: (1) It reduces the load of MN and abrogates location registration mechanism. (2) It

* This work was supported by the National Natural Science Foundation of China (NO 6037301 0, NO 60473082) and National Key Fundamental Research Plan (973) of China (NO 2003CB3 1480 I).

479

480

avoids handover blind zone and handover delay. (3) Its location management load is irrelative to the number of handovers.

2. Related Work

The existing cSCTP/mSCTP-based solutions are “mSCTP + Mobile IP”, “mSCTP + SIP”, “mSCTP + DDNS” and “cSCTP + SIP” [2,3,4]. They all depend on third party methods to realize location management. There are some problems exist in the high-mobility networks applied these schemes.

More network resources are used [6,7,8,9]. The load of MN is heavy. In high-mobility network, most MNs must send the location update messages frequently [6,7,8,9]. Aiming to reduce the number of signaling messages, and also to reduce the delay, some variants of Mobile IP are proposed [ 10,111. But, MN’s load is still unreduced. May result in handover blind zone. Location registration mechanism exists delay. To CN and MN, the new IP address of MN is not available for data transmission until the registration mechanism is finished [6,7,8,9]. Take MIPv6 for example, there will be some time period, when both MN is not reachable from its previous IP address and the binding update procedure is not completed. We say this phenomenon as handover blind zone. May result in handover delay. For MN is migrating frequently, the delays during the location registration procedures will be high.

Proposed Solution-EasyMN

In the solution EasyMN, MN does not need location registration mechanism, so the load of MN is reduced. and MN is in a much more easy state.

BS[O] BS-Server

Figure 1. The EasyMN illustration. Figure 2. The SCTP-AIRA mechanism

48 1

SCTP-DEST UDP Header

I SCTP-SOURCE I

Figure 3. The format of BS-SIM Figure 4. SCTP-INIT-PAYLOAD

EasyMN is illustrated as Figure 1. Our assumptions are as follows:

The wireless IP network consists of N IP subnets, each subnet has one base station (BS). These BSs are numbered BS[O] to BSW-11. MN has a unique identifier MN-Name. CN’s IP address is Address-CN. A BS is also the DHCP server of its subnet. Every BS maintains a User-Table which stores IP address and identifier of every MN currently locates in its subnet. When MN enters or leaves one BS’s subnet, this BS’s User-Table will add or delete corresponding records. Every BS is also an IP node, and it has an IP address. Physically adjacent wireless IP subnets overlap each other.

3.1. Some Concepts

Host-BS. We regard the BS in which MN currently locates as the Host-BS.

M-SIMT (Multicast-like Session Initialization Message Transfer). EasyMN appoints a special BS as BS-Server, its IP address is well-known. All other BSs are regarded as a logical multicast group. BS-Server creates and sends BS-SIMs (BS session initialization messages) to BS multicast group in multicast-like mode. Note: The multicast-like mode here is actually the multiple unicasts. BS-Server has a sending timer (the suggested timer period is 200ms).

BS-SIM (BS Session Initialization Message). It is shown as figure 3. Its UDP source and destination port numbers are all UDP-PORT-NUM1 (the suggested value is 8888). The payloads encapsulated within the UDP datagram are 1-N (the default value is 15) SCTP-INIT-PAYLOADS (SCTP association initialization request payloads). The value of N must ensure that BS-SIM’s IP packet length is not greater than the PMTU. Then in default cases (PMTU =

l500), EasyMN can transfer 75 (5*15) calls to MNs in one second.

SCTP-AIRA (SCTP Association Initialization Request Agent). Because cSCTP supports seamless handover, and it enables CN to keep track of MN’s location changes [3], so the essential issue is how to get the current location

482

information of MN when CN wants to establish a session with MN. In this paper, we propose the SCTP-AIRA mechanism, it is shown as figure 2.

As shown in figure 2, only the first SCTP association establishment message (SCTP-INIT) is processed and transmitted by the SCTP-AIRA mechanism, other three messages are exchanged directly between CN and MN.

3.2. Operation Principle of CN

1.

2.

CN's cSCTP component creates an SCTP association initialization packet, it consists of one SCTP common header and one SCTP-INIT chunk [l]. CN's UDP-Agent component constructs a SCTP-INIT-PAYLOAD, this payload is shown as figure 4, its some fields are defined as follows: - SCTP-DEST: (16 bytes). It contains the identifier of MN. - SCTP-SOURCE: (1 6 bytes). It contains the IPvWIPv6 address of CN. - SCTP Header: (12 bytes). T h s field is the SCTP common header. - SCTP-INIT CHUNK This field is the SCTP INIT chunk. CN's UDP-Agent component encapsulates the payload within a UDP datagram, and sent it to BS-Server. The UDP source and destination port numbers are all UDP-PORT-NUM2 (the suggested value is 7777). CN starts its retransmission timer (the suggested timer period is 500ms).

3.

4.

3.3. Operation Principle of BS'Server

1. Upon receiving a UDP datagram with the source and destination port numbers are all UDP-PORT-NUM2, BS-Server extracts the identifier of MN from the SCTP-INIT-PAYLOAD encapsulated in the UDP datagram. If BS-Server is the Host-BS of MN, then it does the succeeding work according to the step 3 of subheading 3.4. Otherwise, it places the SCTP-INITPAYLOAD into the buffer. When sending timer expires every time, BS-Server checks its buffer to see whether new calls have arrived. If no new calls, then BS-Server stops sending BS-SIMs, otherwise, it takes out N (default value is 15) SCTP-INITPAYLOADS from queue. BS-Server encapsulates these payloads within a UDP datagram. The UDP source and destination port numbers are all UDP-PORT-NUM1. Then BS-Server constructs a BS-SIM IP packet and sends this packet to each BS of BS multicast group through unicast mode.

2.

3.

4.

3.4. Operation Principle of BS Multicast Group

1. Upon receiving an IPRJDP packet with the UDP source and destination port numbers are all UDP-PORT-NUM1. BS processes each SCTP-INIT-PAYLOAD encapsulated within the UDP datagram.

483

BS extracts the identifier of MN from the payload. If BS is not the HostBS of MN, then BS does the work according to the step 4. BS's SCTP-Agent component constructs a new IP/SCTP packet according to MN's IP address Address-MN and the SCTP-INIT-PAYLOAD. Then BS sends thls packet with the source IP address Address-CN to MN. Then BS extracts the next SCTP-INIT-PAYLOAD from the received IP/UDP packet, and processes it according to step 2. When all SCTP-INITPAYLOADS are processed, then BS discards the packet.

3.5. Operation Principle of MN

1. MN will receive one or more BS-SIMs during BS-Server's one timer period. We suppose that the time difference of these BS-SIMs' arriving times is smaller than DALTA-TIME (the suggested value is 1Oms). Upon receiving the first BS-SIM, then MN waits DALTA-TIME. This period later, MN does the following steps. If only one BS-SIM is received, then MN does the succeeding work according to the SCTP association initialization mechanism. If multiple BS-SIMs initialized by the same CN are received, then MN processes these BS-SIMs according to one of three address selection policies below. - Policy A: MN selects the IP address assigned by the Host-BS according to the movement direction of MN itself. - Policy B: MN selects the IP address assigned by the HostBS according to the comparison of the multiple HostBSs' different wireless signal strength detected by MN's hardware. - Policy C: MN selects the IP address assigned by the Host-BS from which the first BS-SIM came. Then MN uses this selected IP address as the primary address to establish the SCTP association directly with CN.

2.

3.

4. Analysis and Comparison

Because EasyMN and the existing solutions almost have no difference in functions besides the location management, so we need only analyze their differences in location management loads. For simplification, we regard the number of the location management messages as the estimation metrics. We consider that network consists of four parts: MNs, BSs, CNs and the other parts. The location management loads of these four parts are FM, Fc, FB and Fo, then the total location management load FL can be expressed as below:

FL = FM + Fc + FB + FO (1)

In this paper, we evaluate the EasyMN and the solution "mSCTP + Mobile IPv6". We consider that the network contains M M N s , C CNs and N BSs.

2 .

3 .

4 .

484

4.1. Analysis and Comparison of the Location Management Loads

We suppose that network has occurred following events during a time period T.

0

All MNs have experienced K handovers. All CNs have initialized H calls (session requests) to MNs. BS-Server has started Q multicast-like BS-SIMs sending operations.

4.1.1. FM - the location management load of MNs

0

0

"mSCTP + MIPv6": All MNs have sent K binding update messages, and have received K binding ACK messages. EasyMN: MNs need not to do special processes during handovers.

FM(mSCTP+MIPv6) = 2 *K FM(EasyMN) = 0

4.1.2. FB - the location management load of BSs

"mSCTP + MIPv6": BSs need not to provide special message processes. EasyMN: BS-Server has received H session requests, and it has started Q M-SIMT operations. BS-Server must send out N-l BS-SIMs to the BS multicast group during every multicast-like sending operation period.

FB(mSCTP+MIPv6) = 0 FB(EasyMN) = H + (N-l)*Q + (N-l)*Q = H + 2*(N-l)*Q

4.1.3. FC - the location management load of CNs

"mSCTP + MIPv6": CNs need not to provide special message processes. EasyMN: CNs need not to provide special message processes.

Fc(mSCTP+MIPv6) = 0 Fc(EasyMN) = 0

4.1.4. FO - the location management load of other parts

"mSCTP + MIPv6": HAS have received K binding update messages, and have sent K binding ACK messages. HAS also have received H call requests sent by CNs, and have forwarded these messages to MNs.

0 EasyMN: The call request messages sent to the BS-Server need transmission support of the network. The BS-SIMs sent by the BS-Server also need transmission support of the network.

F0(mSCTP+MIfi6) = (2*K) + (2*H) -+ (2*K) + (2*H) = 4*K + 4*H Fo(EasyMN) = H + (N-1) *Q

485

4.1.5. The location management loads of two solutions

According to the formula (l), then we get formulas as follows:

FL (mSCTP + MIPv6) = 6*K i- 4*H FL (EasyMN) = 2*H + 3*(N-l)*Q

(2) (3)

Because the number Q only relates to the number H of the calls, and the maximum value of Q is 5 every second in default cases, then Q can be expressed as Q = f(H). Then the formula (3) can be expressed as below:

FL (EasyMN) = 2 *H + 3 *(N-1) *f(H) (4)

Through comparison of the formula (2) and formula (4), we get the conclusion:

0 The EasyMN's location management load has not any relations with handovers' number K , it only relates to calls' number H and BSs' number N (BSs' number N can be regarded as a constant). The solution "mSCTP + MIPv6"'s location management load is related to handovers' number K and calls' number H.

0

We consider that the EasyMN has some advantages over the solution "mSCTP + MIPv6" in the location management load. Our opinions are as follow:

MN usually powered by battery and has relatively low computing capacity, the length of MN's work duration is influenced by the degree of MN's load. Because the busyness degree of communication is related closely to the calls' number, and it has not any relations with the handovers' number. So EasyMN's load is in proportion to the busyness degree of communication.

4.2. Features of the Solution EasyMN

0

MN's load is reduced. MN's battery cost is also reduced. Avoiding handover blind zone and handover delay. Firstly, cSCTP provides seamless handover management. Secondly, EasyMN abrogates location registration mechanism. So EasyMN reduces the negative influence of data delay and data losing during handover to a great extent. Its location management load is in proportion to the busyness degree of calls, and is irrelative to the number of handovers. The success rate of session establishment is improved. MN can select the most proper IP address from all its current addresses according to association address selection policies, and uses this selected IP address for association establishment. In other solutions, MN must use the destination IP address contained in the received SCTP-INIT message IP packet for association establishment even though it may be going to be unreachable.

0

486

0 It is fit for high-mobility networks, and it is also fit for mobile node-to- mobile node communication. Firstly, because EasyMN abrogates location registration mechanism. Secondly, because CN needs not to know MN's current location information when CN wants to establish a session with MN.


We propose a cSCTP-based solution-EasyMN for the hgh-mobility networks. Because it abrogates the location registration mechanism, so MN's load is reduced, and it also avoids handover blind zone and handover delay. In addition, EasyMN's location management load is independent of handovers' number.

EasyMN is not fit for the large-scale networks. It will be interesting to introduce the hierarchical methods or the regional methods into the EasyMN. In the future, we will separate all BSs into several BS regions, and use M-SIMT mechanism in the special BS region according to MN's movement trend.

References

1 .

2.

3.

4 .

5 .

6. 7 .

8 .

9 .

R.Stewart, Q.Xie, K.Moreault, Stream Control Transmission Protocol, RFC2960,2000. Seok Joo Koh, Mee Jeong Lee, Mary Li Ma, etc, Mobile SCTP for Transport Layer Mobility, draft-sjkoh-scp-mobility-04. txt, June 2004. lllcnur Aydin, Woojin Seok, etc, Cellular SCTP: A Transport-Layer Approach to Internet Mobility, ICCCN2003, (2003), pp. 285-290. R.Stewart, M.Ramalho, Q.Xie, etc, Stream Control Transmission Protocol(SCTP) Dynamic Address Reconfiguration, draft-ietf-tsvwg-addip- sctp-09, June 2004. Seok J. Koh, Mee Jeong Lee, Maximilian Riegel, etc, mSCTP with Mobile

IP for Transport Layer Mobility, draft-sjkoh-mobile-sctp-mobileip-O3.txt, June 2004. C.Perluns, IP Mobility Support for IPv4, RFC 3344,2002. D.Johnson, C.Perkins, and J. Arkko, Mobility Support in IPv6, RFC3775, 2004. M.Handley,H. Schulzrinne, E. Schooler, and J. Rosenberg, SIP: Session initiation protocol, RFC2543, 1999. P.Vixie, S.Thomson, Dynamic Updates in the Domain Name System (DNS UPDATE), RFC2136, 1997.

10. R.Koodli, Fast Handovers for Mobile IPv6, draft-ietfmobileip-$ast-mipv6-

1 1. H.Soliman, C.Castelluccia, etc, Hierarchical Mobile IPv6 mobility O8.txt7 October 2003.

management (HMIPv~), draft-ietf-mobileip-hmipv6-08. txt, June 2003.

Signalization

PROPOSAL OF PAPR REDUCTION METHOD FOR OFDM SIGNAL BY USING DUMMY SUB-CARRIERS

PISIT BOONSRIMUANG, KAZUO MORI AND HIDE0 KOBAYASHI Department of Electrical and Electronic Engineering

Faculty of Engineering, Mie University, Japan 1515 Kamihama-cho, Tsu-shi, Mie, Japan, 514-8507

PONGSAKORN BOONSRIMUANG AND TAWIL PAUNGMA Faculty of Engineering and Research Center for Communication and Information

Technology, King Mongkut's Institute of Technology Ladkrabang, Thailand. 3 M2 Chanlongkrung Rd., Lamplatue, Ladkrabang, Bangkok, Thailand, 10520

One of the disadvantages of using OFDM is the larger peak to averaged power ratio (PAPR) of its time domain signal. The larger PAPR signal would course the fatal degradation of bit error rate performance (BER) due to the inter-modulation noise occurring in the non-linear channel. In this paper, we propose the PAPR reduction method for the OFDM signal by using the dummy sub-carriers of which phases are optimized on the basis of the time-frequency domain swapping algorithm and flipping technique. The proposed method can achieve the better PAPR performance and BER performance in the non-linear channel with less complexity of processing at the transmission side. This paper presents various computer simulation results to verify the effectiveness of proposed method as comparing with the conventional methods.

1. Introduction

The OFDM technique has been received a lot of attentions especially in the field of wireless communications because of its efficient usage of frequency bandwidth and robustness to the multi-path fading. One of the limitations of using OFDM technique is the larger peak to averaged power ratio (PAPR) of its time domain signal [l]. The larger PAPR signal would cause the severe degradation of bit error rate (BER) performance due to the inter-modulation noise occurring in the non-linear amplifier.

Recently, various kinds of PAPR reduction methods were proposed such as the selected mapping method (SLM) [2], partial transmit sequence method (PTS) [2] and dummy sequence insertion method (DSI) [3]. The SLM and PTS methods control the phase of data sub-carrier and the DSI method controls the

' This work is supported by the Hitachi Scholarship Foundation (HSF).

489

490

phase of dummy sub-carriers at the transmission side. First two methods are required to inform the phase information controlled for the data sub-carriers to the receiver as the side information (SI). The DSI method is required no side information. However, the conventional DSI method proposed in [3] employs the flipping algorithm where the phases of dummy sub-carriers are optimized by using the certain number of discrete predetermined phase values.

In this paper, we propose a novel phase optimization method for the DSI method, which can achieve the better PAPR performance with almost the same complexity as that for the conventional DSI method. In the following of this paper, Section 2 presents the system model. Section 3 presents the proposed phase optimization method for the dummy sub-carriers. Section 4 presents the various computer simulation results, and we draw some conclusions in Section 5 .

2. System Model

Figure 1 shows the block diagram of OFDM system to be used in the following evaluation. In the figure, the modulated signal in the frequency domain is converted to the time domain signal by IFFT. The time domain signal is given by the following equation.

1 N - l .2*

xk = - C X , e N JT n=O

I -

where N is the number of IFFT points and X, is the modulated data at k-th subcarrier. The time domain OFDM signal is input to the non-linear amplifier after adding the guard interval (GI). The output of non-linear amplifier can be expressed by the following equation.

where, Yk is the input signal of non-linear amplifier and F[ ] represent the AM/AM conversion characteristics of non-linear amplifier. The non-linear amplifier assumed in this paper is the Solid State Power Amplifier (SSPA) of which input and output relationship is modeled by the following equation.

P [I + ( p / A ) 2 r ] ' / 2 r F [ P I = (3 )

where, p is the amplitude of input signal, A is the saturated output level, and r is the parameter to decide the non-linear level. The phase conversion of the non- linear amplifier is assumed to be linear in the following evaluation. The

491

operation point of non-linear amplifier is defined by the Input Back-Off -(IBO), which is given by the following equation.

where Pi, is the average power of input signal to the non-linear amplifier and Po is the input saturation power. The PAPR is defined by the following equation.

PAPR = lolog( k)

‘k k IFFT Input

MOD

AWGN -@ Output

DEMO

Figure 1. Block diagram of OFDM system.

3. Proposal of Phase Optimization Method

This section proposes the phase optimization method for the dummy sub- carriers so as to reduce the PAPR performance. Fig.2 shows the structure of proposed OFDM symbol represented in the frequency domain. The OFDM symbol consists of M data sub-carriers and L dummy sub-carriers that are placed at the both ends of data sub-carriers as shown in Fig. 2. The transmission efficiency of proposed DSI method becomes M/(M+L). In the proposed method, the phase values of dummy sub-carriers are optimized by using the time- frequency domain swapping algorithm and flipping technique, which can achieve the better PAPR performance and the less required number of iterations.

Figure 3 shows the block diagram of proposed phase pptimization method. In the figure, N sub-carriers consisting of L dummy sub-carriers and M data sub-carriers in the frequency domain are converted to the time domain signal by IFFT. Then, the “Phase Optimization” module processes the time domain signal so as to reduce the PAPR performance symbol by symbol. The “Phase optimization” module employs the time-frequency domains swapping algorithm [4]-[5] to optimize the phases of dummy sub-carriers. Since this algorithm is

492

usually required a large number of iterations to achieve the optimum results, “Phase optimization” module also employs the flipping technique [ 5 ] to reduce the number of iterations.

Dummy Sub-carriers DATA Sub-camers Dummy

Sub-carriers L I 1 L I A

u2 M -I- L/2

N (=M+L) Subcarriers

Figure 2 Structure of proposed OFDM symbol in the frequency domain.

rn

OFDM Signal

Figure 3 Block diagram of phase optimization method.

In the proposed method, the target PAF’R and the maximum number of iterations are firstly set. The time domain signal consisting of N sub-carriers, which corresponds to the signal before optimization, is given by Eq.(l). In Eq.(l), the initial dummy sub-carriers are given by the following equation.

n=O;..,(L/2-1) n = (A4 + L / 2), * * I, (N - 1) n

x = eien

where, Xn has the constant amplitude and 6, is given by the random phase. The basic principle of this algorithm is to find higher peak level in the time

domain signal of Eq.(l) than the reference level of S and calculate the error signal defined by the following equation as the i-th iteration.

The reference parameter S is decided on the basis of the average power of input signal. The time domain error signal given by Eq. (9) is converted to the frequency domain signal by FFT, which is given by the following equation.

493

By using Eq.(lO), the phase values are calculated only for the dummy sub- carriers. The obtained phase value is subtracted from the original frequency domain signal given by Eq.(8). The phase optimization is processed only for the dummy sub-carriers and the data sub-carriers are kept the same as the original signal. The frequency domain signal to be used at the next iteration (i+l) is given by the following equation.

In the time-frequency swapping algorithm, Eqs.(8) to (1 1) are repeated up to reach the optimum results. In the proposed method, the flipping technique is also employed to reduce the iteration numbers as similar to the conventional DSI method. In the flipping technique, the PAPR is calculated by changing the phase of each dummy sub-carrier. If the new PAPR is lower than the previous result, the new phase will retain as part of the final phase sequence. Otherwise, the phase reverts to its previous value. These processing will perform for all dummy sub-carriers. The time-frequency domain swapping algorithm and flipping technique will repeat up to reaching either of the predetermined target PAPR or the maximum number of iterations.

4. Performance Evaluations

This section presents various computer simulation results to verify the effectiveness of proposed method. The simulation parameters to be used in the following evaluations are shown in Table 1.

Table 1. Simulation parameters.

Modulation 16QAM or 640AM Demodulation

Allocated bandwidth Number of FFT points Number of sub-carriers

Symbol duration Guard interval

Number of dummy sub-carriers Non-linear amplifier

Non-linear parameter of SSPA

Coherent 5MHz

256 64

12.811s 1 .28~s 2-16

SSPA I-2

494

Figure 4 shows the PAPR performance for the conventional DSI and proposed DSI methods when changing the number of dummy sub-carriers. The PAPR performance for the conventional OFDM is also shown in the figure as the purpose of comparison. In the figure, the PAPR performance is evaluated by using the Cumulative Distribution Function (CDF). From the figure, it can be observed that the proposed DSI method shows the better PAPR performance as increasing the number of dummy sub-carriers, while the transmission efficiency decreases. It can be also observed from the figure that the proposed DSI method can achieve the better PAF'R performance than that for the conventional DSI method and the conventional OFDM.

PAPR(dB)

Figure 4 PAPR performance of proposed DSI method.

Figure 5 shows the PAPR performance both for the conventional and proposed DSI methods when changing the number of iteration. From the figure, it can be seen that the proposed DSI method with the larger dummy sub-carriers requires slightly larger number of iterations for the convergence of PAPR performance than that for the conventional DSI method. It can be also observed that the proposed DSI method can achieve the better PAPR performance than the conventional DSI method by the acceptable small number of iterations with less than 20 iterations even for the larger number of dummy sub-carriers.

Figure 6 shows the BER performances for the proposed and conventional DSI methods in the non-linear channel when the modulation methods are 16QAM and 64QAM. In the simulation, the number of total sub-carriers is 64 including 48 data and 16 dummy sub-carriers. As for the modulation method of 16QAM, the proposed DSI method can achieve much better BER performance than those for the conventional OFDM and DSI methods when IBO is -2dB. The proposed DSI method shows almost the same BER performance as that of Ideal performance. As for the modulation method of 64QAM, the proposed DSI

495

method when IBO is -6dB can achieve the very close to the Ideal performance, while the conventional OFDM and DSI methods show the worse BER performance even at the higher CNR.

8.5 ----. h p r c d DSI 1 + Dummy = 2 subcsrriers --~!n??tiF! DSL 4 Dummy= 4 subcsrriers

+ Dummj. = 8 subcsrriers -S Dummy- 16 subcarriers

0 5 10 15 20 25 30 35 40 Number ofIteration

Figure 5 PAPR performances versus number of iterations.

5. Conclusions

In this paper, we proposed the PAPR reduction method for the OFDM signal by using the dummy sub-carriers. The feature of proposed method is to employ the time-frequency domain swapping algorithm and flipping technique for improving the PAPR performance with less complexity at the transmission side. From the various computer simulation results, we confirmed that the proposed DSI method could achieve the better PAPR performance and better BER performance in the non-linear channel than that for the conventional DSI method.

496

(b) 64QAM Figure 6 BER performance of proposed DSI method in non-linear channel.

Acknowledgments

The authors would like thank to the Hitachi Scholarship Foundation(HSF) has supported this research.

References

who

1.

2.

3.

4.

5.

D Dardari, V. Tralli and A Vaccari, “A Theoretical Characterization of Nonlinear Distortion Effects in OFDM Systems,” IEEE Trans. on Corn . , Vol. 48, no. 10, pp.1775-1764, Oct 2000. S. H. Muller and J. B. Huber, “OFDM with reduce peak-to-average power ratio by optimum combination of partial transmit sequences,” Electron. Lett., vol. 33, no. 5, pp. 368-369, Feb. 1997. Heung-Gyoon Ryu, Jae-Eun Lee and Jin-Soo Park, ‘‘Dummy Sequence Insertion (DSI) for PAPR Reduction in the OFDM Communication System,” IEEE Transactions on Consumer Electronics, Vol. 50, No. 1, Feb 2004. M. Friese, “Multitone Signals with Low Crest Factor”, IEEE Trans.

E. V. Der Oudera, Schoukens and J.Renneboog., “Peak Factor Minimization Using a Time-Frequency Domain Swapping Algorithm,” IEEE Trans Instrum. Meas. vo1.37, pp145-147, Mar 1988.

Comm~n., V O ~ . 45, pp. 1338-1344, Oct 1977.

ADAPTIVE SCHEDULING FOR HETEROGENEOUS TRAFFIC FLOWS

IN CELLULAR WIRELESS OFDM-FDMA SYSTEMS*

S. VALENTIN’, J. GROSS’, H. KARL’, AND A. WOLISZ2 University of Paderborn, Warburger StmJ3e 100,

33098 Paderborn, Germany E-mail: { stefan,valentin 1 holger. karl} Oupb. de

T U Berlin, Einsteinufer 25,

10587 Berlin, Germany, E-mail: {gross I wolist)@tkn.tu-berlin.de

In this paper we study the performance gain achieved by introducing dynamic subcarrier scheduling schemes to OFDM systems in the presence of heterogeneous traffic streams. Dynamic subcarrier scheduling copes with channel state variation by dynamically assigning different subcarriers to terminals, e.g. in the downlink of a cellular system. In addition to the channel state variations, also the arrival of data is highly variable for most applications. This calls for schemes that exploit both sources of variation. In the homogeneous traffic scenario, dynamic subcarrier scheduling in combination with semantic queueing mechanisms can significantly increase the capacity of an OFDM cell. In a heterogeneous traffic scenario, however, these mechanisms do not simply carry over. Therefore, in this work we study first the performance increase achieved for heterogeneous flows by purely switching from static to dynamic subcarrier scheduling and then introduce a semantic traffic management scheme that improves the achieved performance further by sacrificing the transmission of semantically unimportant video packets for important video packets and for packets related to HTTP.

*The main part of this work has been done while the authors were with the T U Berlin. It has been partially supported by the German research funding agency ‘Deutsche Forschungsgemeinschaft (DFG)’ under the program ‘Adaptability in Heterogeneous Communication Networks with Wireless Access (AKOM)’.

497

498

Traffic bit-rate Upper layers

_ _ _ - _ _ _ _ _ - - _ - ...-_______._

I I Cross-layer optimization scheme 1 fink layer

t Physical layer Channelstates h+L-

Figure 1. Sources of Variability and the related layers

1. Introduction

Delivering high quality multimedia streams via a wireless link is still a challenging task for mobile communication systems. In addition to high transmission speed, quality aspects, such as bounded delays, are required for the transmission. In wireless systems multi-path propagation and movement leads to Inter-Symbol Interference (ISI) and fading. IS1 can be minimized by modern modulation schemes, e.g. Orthogonal Frequency Division Multiplexing (OFDM) , where the spectrum is separated into narrow bands, the so-called subcarriers. Fading, however, leads to drastic changes in signal attenuation and Symbol Error Probability (SEP) and cannot be neutral- ized on the physical layer (PHY). Thus, the available channel resource is a random variable with rather large dynamics. In addition, the resource (bit rate) demanded by the packet streams, which have to be sent by the transmission system, varies. This variation is common in packet-switched networks and depends on traffic characteristics which are originally defined at the sender’s Application layer (APP).

In multi-terminal environments these random processes-channel and traffic variation-add new challenges for the management of the channel resources. With static scheduling both fluctuations lead automatically to over- or under-provisioning of the channel resources per packet stream. A solution is the dynamic distribution of resources according to the wireless channel and traffic stream parameters. This adaption can be done in the APP, e.g. by the video codec, as well as in lower layers. Performing the adaptive scheduling on the Logical Link Control layer (LLC) enables fast access to the information provided by the lower and higher layers, e.g. PHY and APP. As illustrated in Figure 1 , this so-called cross-layer optimization approach adapts the Medium Access Control (MAC) to the two sources of variability, traffic bit rate and channel state. In this paper we discuss the cross-layer optimization scheme separately for each source of variability. The first discussed component is called dynamic subcarrier scheduling. It

499

follows the channel-state dependent scheduling approach [l] and distributes the available capacity (e.g. OFDM subcarrier) to the terminals according to the measured channel state. The second component, called trafic management [2], controls the bit rate of the source according to application layer parameters. We discuss both optimization schemes in greater detail in Section 3.

Performance studies were done in previous work for both schemes considering only one type of application layer traffic. While dynamic subcarrier scheduling was typically studied for bulk or HTTP traffic [3], for traffic management streaming media is relevant [4]. In these homogeneous traffic scenarios each type of APP traffic is considered separately, which clearly does not reflect the nature of the Internet. Here, typically a combination of HTTP, streaming, and bulk related traffic flows is present [5]. Due to the different demands of the flow types on the transmission system (e.g. Quality of Service (QoS) guarantees for streaming, fault-tolerant and fast transmission for HTTP) especially the combination of HTTP and video streaming is interesting. For cross-layer optimization approaches in wireless systems this combination has not been investigated so far.

For this so-called heterogeneous traffic scenario we first study how the combination of the different traffic characteristics affects the performance of the cross-layer optimization scheme. Second, we propose a new traffic management scheme, called Video Queue Management (VQM) , which executes a filter that removes video data from the access point queue. While the filter considers semantic parameters in order to minimize the quality decrease due to packet removal, parameters from the link and the physical layer are evaluated for load estimation.

The model of the considered dynamic OFDM-FDMA system is discussed in the next section of this paper. Section 3 introduces the basic approach and algorithms of VQM. Here the two major parts-load estimation and video filtering-are discussed in greater detail. In Section 4, the results of the performance study for dynamic subcarrier scheduling and the proposed traffic management scheme are presented. In the last section, we conclude the paper.

2. System model

We consider a single wireless cell of radius R in which a certain number of J terminals as well as one access point are located. The terminals are connected via a radio link to the access point, which controls all wireless

500

transmission. Since the optimization of the uplink is not the topic of this paper only the downlink transmission is investigated. Data is transmitted to the terminals using the available bandwidth B around the center frequency fc. Typical effects which occur during the terrestrial transmission of radio signals are considered in the system model. Specifically, path loss and, due to multi-path propagation, frequency-selective fading are modeled. The terminals are independently moving within the cell. Due to this movement, also time-selective fading occurs. The OFDM transmission scheme separates B into S subcarriers. Each symbol is transmitted during the symbol time T, using constant transmission power Pt, per subcarrier. An adaptive modulation scheme selects one modulation type out of five (BPSK, QPSK, 16-QAM, 64-QAM, and 256-QAM) in order to stay below the SEP bound P,.

Subsets of subcarriers are distributed to the terminals using a static or dynamic Frequency Division Multiple Access (FDMA) scheme. The dynamic subcarrier scheduling scheme, applied in case of dynamic FDMA, is discussed more detailed in Section 3. The subcarrier assignments are valid for one MAC cycle of length T f . The terminals in the cell are receiving data flows from external sources that are connected to the access point via a cable link. The flows consist of packets and are queued at the access point until they are forwarded to the terminals via the wireless link. At the access point transmission functions and queueing are handled for each terminal separately. A fraction of the flows is related to Web pages. All other packets are related to MPEG-4 coded, Variable Bit Rate (VBR) video streams. The Web pages are transmitted using TCP/IP, for the video streams UDP/IP is employed. For the video streams a traffic management scheme is used in the link layer at the access point. We will give an overview of this scheme in the next section.

3. Optimization Approach

3 .I. Dynamic subcarrier scheduling

Since channel-state dependent scheduling was proposed by Bhagwat et al. [ 11 several scheduling algorithms are under discussion. Performance studies [6] have shown that the proportionally fair and the exponential rule are good candidates for the support of real-time and non-real-time applications with high data rates. However, all these algorithms consider an optimal solution for the subcarrier assignment problem, an optimization problem typically considering the parameters link state and queueing time. In scenarios with

501

many terminals the calculation time for the optimal solution will exceed the assignment cycle (i.e. MAC-frame), which is typically short in high data rate scenarios.

The dynamic subcarrier scheduling scheme employed in this paper, called advanced Dynamic Algorithm (aDA), belongs to the class of proportionally fair scheduling algorithms. It provides a slightly suboptimal but fast solution for the assignment problem [7] depending on the most recent subcarrier states. With the state of each subcarrier, which depends on the chosen modulation type, aDA selects subcarriers with the highest possible states. This selection of subcarriers according to channel states and the further adaption of the number of subcarriers to the queue length leads to a continuous variation of the downlink capacity per terminal.

3.2. h f i c management for video streams

For the transmission of videos in real-time application scenarios, such as video conferences or the broadcasting of live events, small, predictable transmission delays are required. This is even the case if the state of the wireless channel is temporarily bad or the sender rate is temporarily high. For the considered system and the chosen optimization approach this means that the queueing delay at the access point should be minimized for each packet. It is beneficial if the scheduler is aware of application layer information regarding the packets on the link layer. With this knowledge those video-related packets whose transmission time has exceeded an estimated deadline D can be removed from the queue.

The temporal compression of modern video codecs leads to three different types of video frames. While the so called I-frame contains a whole (spatial compressed) video picture, P- and B-frames contain only temporal differences to the preceding frames (P-frame) or to the preceding and following frames (B-frame). Since I-frames have the highest entropy and B-frames contain the least amount of information, the deadline of the video packets can be weighted according to the semantic relevance of the frame type. One example is the so-called semantic-aware traffic management scheme [8]. Here I-frames are only removed from the queue after the transmission delay for the related packets has exceeded the full deadline D. B-frames are removed first, followed by P-frames.

In this paper we propose an extension to the semantic-aware scheme that is called Video Queue Management (VQM). With VQM an additional removal cycle, called packet filter, is invoked if the traffic load exceeds the

502

Downlink capacity

I I

I

I Packet filter measurement

t Dynamic subcarrier scheduling

Queue

Figure 2. General architecture of the considered dynamic MAC scheme

threshold Q after the semantic-aware scheduling is performed. The traffic load is estimated once per MAC cycle n by the ratio R(n) = &, where F stands for the downlink capacity, which varies with dynamic subcarrier scheduling, and s for the sum of bits already removed by the semantic-aware traffic management scheme. Since the traffic management scheme has to be performed before dynamic subcarrier scheduling, the value F ( n - 1) of the preceding downlink phase is used. The idea to use R(n) as a representation of the system load is, that s increases with rising system load due to more missed deadlines or that the capacity F decreases. The packet filter removes whole packets from the access point queue and considers semantic features in order to minimize the additional distortion of the video. Only B-frame related packets are removed during the iteration of the packet filter over the whole queue. The filter is executed until the sum of the bits removed reaches a certain quality threshold p.

The resulting combined cross-layer optimization scheme is illustrated in Figure 2. All components denoted by the blocks rely on cross-layer information, which is extracted by the functions marked by the ellipses. The provision of the extracted data to the optimization schemes is illustrated by the dashed arrows. Dynamic subcarrier scheduling considers channel states for its scheduling decision. VQM relies on the scheduling decision and on semantic features extracted from the video stream.

4. Performance study

4.1. Metrics and methodology

Two types of traffic are considered to be transmitted via the wireless link: Web pages and MPEG-4 coded video streams. On the user level both traffic types are differently affected by LLC capacity variation. Thus, we use two metrics, one for each traffic type. For the transmission of Web pages

503

the latency per page is measured on the application layer of each receiver. The quality of the transmitted video is measured using the Distortion In Interval (DIV) metric [8]. This metric provides an approximation of the subjective Mean Opinion Score (MOS) grade by using the Peak Signal-to- Noise Ratio (PSNR) value, which can be calculated for each video frame. PSNR and MOS are widely used for the evaluation of picture quality. The video quality is rated to be unacceptable if more than 20% of the decoded video pictures in a 20 s interval have a lower MOS grade than the original pictures [8]. Note that this threshold is somewhat subjective, it could also be put at 18% or 22%, but definitely not at values like 40% or 60%.

Both metrics are investigated in the performance study for four different setups. Using the notation subcarrier scheduling/trafic management scheme for each setup, the first setup is called Static/Static. Here no optimization is done at all. Each terminal receives a set of subcarriers at the beginning of the simulation and FIFO is used for the packet queueing. The dynamic assignment of the subcarrier sets is considered in the Dynamic/ Static setup. Then we combine dynamic subcarrier scheduling with traffic management. The performance of the semantic-aware approach in combination with the dynamic subcarrier scheduling is studied in the Dynamic/ Semantic case. Finally, we investigate dynamic subcarrier scheduling together with the proposed VQM scheme in the Dynamic/VQM setup.

For all scenarios and metrics the average of all values measured during the simulated time is shown versus the number of video-receiving terminals. Due to their small size the 0.95-confidence intervals are not shown.

4.2. Simulation parameterization

For the simulation we chose the following settings. The total bandwidth of 16.25 MHz in the 5.2 GHz band is separated into 48 subcarriers, which is equivalent to 802.11a. In order to assign at least one subcarrier per terminal J 5 48 terminals are moving with a maximum speed of 1 m/s in a cell of radius R = 100 m. Path loss, shadowing and fading are modeled according to the channel model proposed in [9]. The delay spread is set according to the ETSI C large open space model [lo] to 0.15 ps. Data is transmitted using a constant power of -7 dBm. The SEP threshold for the dynamic modulation scheme is For TCP the NewReno flavor and the common maximum segment size of 1460 Bytes is chosen. Since two types of application layer traffic are considered in this paper, Jh out of J terminals are receiving Web pages while video streams are simultaneously

504

, . . , . . . . . , . . . . . ........................... ._....__ . . . , . .

4 8 12 16 u) 24 28 32 36 Number of video receiving terminals (I,)

Figure 3. Mean DIV of the video streams for four optimization scheme setups

transmitted to J, = J - Jh terminals. In the performance study we set Jh = 12 and J, is varied. The HTTP traffic is modeled according to the corporate environment in [ll]. For the mean inter-session time we chose 3 s to assure an adequate utilization of the system. The MPEG-4 coded video streams have mean bit rate of rv = 951 KBps, using the common Group of Pictures (GOP) IBBPBBPBBPBB, which provides good quality on hand- held terminals. In the Dynamic setups during each MAC cycle of 2 ms dynamic subcarrier scheduling and traffic management are performed. The traffic management schemes semantic-aware scheduling and VQM use the (heuristic) deadline D = 100 ms. B-frames are removed at 0.50 , P-frames at 0.750 and I-frames at D. In addition to these weights VQM makes use of two additional thresholds. The threshold for the load-rate a was chosen to be 5. For the video quality we chose the threshold p = 12.

4.3. Simulation results

In this section we present the results of the performance study of the cross- layer optimization schemes versus increasing numbers of video stream receiving terminals in the cell (J,). With the metrics shown for the two traffic types better results are achieved for smaller values. Because the overall capacity of the wireless channel is limited, both metrics degrade with increasing J , due to the contention increase of the traffic flows.

As shown in Figure 3 and 4 the dynamic subcarrier scheduling clearly outperforms the static case for both types of traffic. For an upper DIV bound of 20% gains of up to 62% can be achieved with the dynamic scheduling compared to the static scheme. This means that up to 17 video streams

505

, . . ...... < ........... i ......................

. . . . . . . , . . . ..__.___.__I .....___..__.___.... <..

2 ' I 4 8 12 16 XI 24 28 32 36

Number of video receiving terminals (J,)

Figure 4. Mean Web page transmission time for four optimization scheme setups

can be sent with dynamic subcarrier scheduling while no stream of this quality can be sent without any optimization scheme. Furthermore, the average transmission time of the simultaneous transmitted Web pages is decreased by up to 47%.

The results for the traffic management scheme show small increases of the video quality near the upper DIV bound at J,, = 18. Compared to the Dynamic/Static setup the semantic-aware scheduling increases the performance gain by only 4% for video and by 4% for Web page transmission. With the Dynamic/VQM setup a gain of 12% for video quality and of 6% for Web pages can be reached. Furthermore, semantic weighting and the additional filter of VQM do not significantly decrease the video quality.

5. Conclusions

In this paper we have investigated the combination of the two cross-layer optimization approaches channel-state dependent scheduling and traffic management for a wireless OFDM-FDMA system in heterogeneous traffic scenarios. We have shown that channel-state dependent scheduling increases the performance even with heterogeneous traffic in the downlink. Here gains of up to 62% were found for the transmission of MPEG-4 coded video streams and up to 47% for the Web page transmission time, respectively. This is comparable to the performance increase found for streaming video in [8].

Furthermore, we have found that traffic management schemes do not lead to a major performance increase in heterogeneous traffic scenarios. Here only a part of the data is related to video streams, whose redun-

506

dancy can be exploited by the semantic weighting schemes. Thus, the achieved gains are quite low if compared to the improvements of semantic- aware scheduling in homogeneous scenarios [8]. Semantic filtering such as VQM, however, further increases the performance at feasible complexity (one queue iteration). This performance increase rises with the traffic load that can be filtered. This makes VQM applicable in cases of congestion. Here the semantic traffic management provides a “buffer” until the traffic load is lowered otherwise, e.g. by admission control schemes, dropping some flows, or invoking a hand-off of some terminals.

References 1. P. Bhagwat, P. Bhattacharya, A. Krishna, and S. Tripathi, “Enhancing

throughput over wireless LANs using channel state dependent packet scheduling,” in IEEE INFOCOM, 1994.

2. M. Smirnov, E. Biersack, C. Blondia, 0. Bonaventure et al., Quality offuture Internet services, Springer-Verlag, 2003.

3. T. LeNgoc, N. Damji, and Y. Xu, “Dynamic resource allocation for multimedia services over OFDM downlink in cellular systems,” in Proc. of IEEE Vehicular Technology Conference Spring, 2004.

4. M. Hemy, P. Steenkiste, and T. Gross, “Evaluation of adaptive filtering of MPEG system streams in IP networks,” in Proc. of IEEE International Conference on Multimedia & Expo, 2000.

5. C. Fkaleigh, S. Moon, B. Lyles, C. Cotton et al., “Packet-level traffic measurements from the sprint IP backbone,” IEEE Network, vol. 17, 2003.

6. S. Shakkottai and A. Stolyar, “Scheduling algorithms for a mixture of real- time and non-real-time data in HDR,” in Proc. of 17th International Tele- tm f i c Congress, 2001.

7. J. Gross, H. Karl, F. Fitzek, and A. Wolisz, “Comparison of heuristic and optimal subcarrier assignment algorithms,” in Proc. of International Confer- ence on Wireless Networks, 2003.

8. J. Gross, J. Klaue, H. Karl, and A. Wolisz, “Cross-layer optimization of OFDM transmission systems for MPEG-4 video streaming,” Computer Com- munications, vol. 27, 2004.

9. A. Aguiar and J. Gross, “Wireless channel models,” Technical Report TKN- 03-007, Telecommunication Networks Group, April 2003.

10. J. Medbo and P. Schramm, “Channel models for HIPERLAN/2,” document no. 3ERI085B, ETSI/BRAN, 1998.

11. A. Reyes-Lecuona, E. GonzBlez-Parada, E. Casilari, J. C. Casasola et al., “A pageoriented WWW traffic model for wireless system simulations,” in Proc. of 16th International Teletmfic Congress, 1999.

THE POWER SPECTRAL DENSITY OF THE H-TERNARY LINE CODE: A SIMULATION MODEL AND COMPARISON

ABDULLATIF GLASS Department of Planning and Training, Technical Studies Institute

P 0 Box 26833, Abu-Dhabi, UAE Tel: +9712 5045785, Fax: +971 2 5854282, e-mail: aplassCa),ieee.org

NIDHAL ABDULAZIZ College of Information Technology

University of Wollongong in Dubai, P 0 Box 20183, Dubai, UAE Tel: +971 4 3900604, Fax: +971 4 3664585, e-mail: nidhuluh~iila:izCu~~~~~~~~~

EESA BASTAKI Department of Education, Training, Research, and Development

Dubai Silicon Oasis, P 0 Box 491, Dubai, UAE Tel:+971 4 2027742, Fax: +971 4 29941 16, e-mail: eesu(cidso.ue

A verification for the accurate calculation of the power specfral density of the H-Ternary line code is carried ouf in this paper. The power spectral density is evaluated through the use of a Matlab simulation model. The model considers the generation of a P-N sequence and encodes it to the H- ternary code. The FFT techniques are used to determine the power spectral density of the line code. The results obtainedfi-om the simulation model reveal a total agreement with that of the theoretical model. Both results show a favourable spectrum of the proposed line code compared with other line codes.

1. Introduction

Coding is the technique used in many applications of telecommunications and digital signal processing. Coding generally is divided into three main technologies, which are source coding, channel coding, and, finally, line coding. The first technique is used to remove redundancy from a signal source. In the second technique, redundant bits are added for the purpose of information reliability when transmitted or stored via a noisy channeymedium. The third, however, is used to adapt the signal waveform to the transmission or storage medium.

Line coding is the process of adapting a digital signal to a transmissiodstorage medium for the purpose of efficient spectrum utilisation and reliable recovery. Line codes have to fulfil some of the many desirable requirements. The requirements include a small amount of low frequency components with no dc content. This desirable feature reduces the bandwidth

507

508

needed for the transmission or the storage of the encoded signal and hence more messagesldata can be accommodated by either multiplexing several signals or by increasing the data rate for the same bandwidth occupancy. While direct coupling is not desirable in most data transmission circuits as well as for storage, line codes are thus required to contain no dc component. Line codes should also contain sufficient transition in signal levels to ensure a continuous clock signal recovery and thus provide a means for synchronism. There are many other desirable features that are also recommended to exist in line codes such as immunity against noise, its ability to detect and correct errors, its power efficiency, its transparency and the complexity of the encoding and decoding procedures and hence circuit complexity and cost. A short summary of a survey of the different line codes are given in references [ 1-41 for different applications in telecommunication networks such as LANs, MANS, WANs, DSL and wireless networks.

Line codes are available in several forms and their classification methods depend on the number of signal levels at the encoder output. Binary line code is the most common where the encoder output is a two-level signal. This is the most simple for encoding and decoding processes. The decoder has to decide between two received signal levels and hence make its decision for the binary 1 s and 0s. This is, however, the least efficient from a capacity point of view and bandwidth occupancy. Examples of binary codes are RZ, NRZ, Manchester, differential Manchester, Miller (delay), ... etc. The second type is the ternary line code, where the output of the encoder takes three signal levels. Ternary line codes are generally more efficient compared to binary line codes from a bandwidth occupancy point of view. The decoder circuit however has more complexity compared to binary codes. Polar RZ, AMI, HDBz, xByT and MLT3 are examples of ternary codes that are used in ISDN, FDDI and other systems. The third is the quaternary line code where the signal levels at encoder output take four levels.

Ternary and quaternary line codes are considered as different classes of the general multi-level line encoding scheme. These are the most efficient codes with regards to bandwidth occupancy but also the most difficult to encode and decode. These are currently the least in usage where they find applications in systems such as ISDN and DSL, however, the fkture prospectus predicts that they will find wide applications.

In this paper, a simulation model for the H-ternary line is given. The next section provides a description of the procedure by which a sequence of binary data is encoded to a H-Ternary line code and then decoded back to its original sequence. A review of the mathematical model of the power spectral density (PSD) of the new line code together with its counterparts is given in section three. Section four is devoted to a description of the simulation model and how it has been achieved. Section five gives a discussion of the results and a

509

comparison between mathematical and simulation models. In the final section the conclusions are given.

2. Encoding and Decoding Principles

The H-ternary line code operates on a hybrid principle that combines the binary NRZ-L, the ternary dicode and the polar RZ codes and thus it is called hybrid- ternary. The H-ternary code has three output levels for signal representation; these are positive (+), zero (0), and negative (-). The following subsections give a description of the procedures for the encoding and decoding principles.

2. I . Encoder Operation The states shown in Table 1 depict the encoding procedure. The H-ternary code has three output levels for signal representation; these are positive (+), zero (0), and negative (-). These three levels are represented by three states. The state of the line code could be in any of these three states. A transition takes place to the next state as a result of a binary input 1 or 0 and the encoder output present state. The encoding procedure is as follows: 1. The encoder produces + level when the input is a binary 1 and whether the

encoder output present state is at 0 or - level. 2. The encoder produces - level when the input is a binary 0 and whether the

encoder output present state is at 0 or + level. 3. The encoder produces 0 level when the input is binary 1 and the encoder

present state is + level or when the input is binary 0 and the encoder present state is - level. Initially, the encoder output present state is assumed at 0 level when the first binary bit arrives at the encoder input.

4.

Table 1. Encoding principles and output states

The procedure gives the reader sufficient information about the operation of this new line code scheme. Further details and comparison together with design and modelling of the encoder can be sought from references [4,5,7,8]. The variation of this new line code is that it violates the encoding rule of NRZ-L

510

and dicode when a sequence of 1s or 0s arrives. In the latter case, it operates on the same encoding rule of polar RZ but with full period pulse occupancy.

2. 2. Decoder Operation Table 2 shows the input states of the H-ternary decoder and its decoding procedure for an output binary. It is a reverse process of the encoding operation given in the previous subsection. The decoder has only two output states (binary) whereas the input is three states (ternary). The decoding procedure is as

1. The decoder produces an output binary 1 when the input ternary is at + level and whether the decoder output present state is a binary 1 or 0.

2. The decoder also produces an output binary 1 when the input ternary is at 0 level and the decoder output present state is at a binary 1.

3. Similarly, the decoder produces an output binary 0 when the input ternary is at - level and whether the decoder output present state is a binary 0 or 1.

4. Finally, the decoder produces an output binary 0 when the input ternary is at 0 level and the decoder output present state is a binary 0.

It is clear that the decoding process at the receiver is quite similar to that of the NRZ-L code when the + and - levels are received. The difference arises when level 0 is received. In which case, the decision is made depending on the decoder output present state.

Table 2. Decoding principles and output states

fOllOWS [6-81.

Input Ternary output Binary Present State Next State

I I

3. Review of Mathematical Analysis Models The power spectral density (PSD) of a line code can be evaluated using either deterministic or stochastic analysis techniques. Since in our case the input data sequence is random, the second approach is therefore adopted.

The general expression of the PSD of a digital signal is given by [2, 8-

113.

511

where So is the Fourier transform of the line code pulse shape s(t) of amplitude A and duration Tb, and Rfi) is its autocorrelation function. It is evident that the above equation shows that the spectrum of the digital signal depends on two things: the pulse shape used and the statistical properties of the encoded signal. Equation (1) can also be rewritten in a simpler series form as follows.

The pulse shape of the H-ternary line code is a pulse o f f unit amplitude with a duration Tb. The Fourier transform of the H-ternary pulse is given by [8,111.

The above sinc function has a spectrum that extends to infinity for both the positive and negative frequencies.

The statistical properties of the data are referenced to the autocorrelation function of the line code that is given by

i =1

where A,,, and Am+k are the signal levels that correspond to the mth and m+kth symbol positions that represent the H-ternary line code respectively, and Pi is the probability of having the ith Am and Am+k product.

To calculate the autocorrelation hnction of the line code for a different combination of symbols, we first calculate the R(0). This means the autocorrelation function of the line code pulse/symbol with itself. From [7], the probability of occurrence of each symbol of the H-ternary line code is equal. This means the probability of the three transmitted code levels are equal, i.e. P+=Po=P-=%. Substituting these values together with their respected unit amplitude symbols into Eq. (4) for N signal symbols and averaging the same signal symbols results in,

R(0)=+[+(+1)2 ++(-1)2++(0)2]=5.

512

The calculation of Rfi) for all other values of k excluding k=O can be determined using a tabulation method [8,9]. The values of R(I), R(2), R(3), ... Rfi) can be found using the probabilities of all possible states. The probabilities of each case also depend on the number of H-ternary symbols that are considered. For example, the probabilities of each symbol are PI=& Pr=%, Pj=1/16, ... and so on. The autocorrelation function for each case, using Eq. (4), is thus, R(I)=-4 R(2)=& R(3)=-% and so on for other values of Rk. The overall autocorrelation function for all values of k excluding k=O is thus given as follows [8].

Substituting Eqs. ( 5 ) and (6) together with (3) into (2) gives the PSD of the H-ternary line code that is [8],

The PSD of the H-ternary line code is a re-shaped form of the Fourier transform of a rectangular pulse having L4 amplitude and duration of Tb. The re-shaping function is the autoconelation function of the H-ternary line code.

The normalised PSD results of the above derived formula for the H- ternary code versus the normalised frequency is shown in figure 1. Figure 1 clearly shows that the bulk of the high-weight frequency components of this line code are centred at about half of the normalised frequency (signalling rate). The spectrum also shows the line code has no dc component.

To compare the H-ternary line code spectrum with other line codes spectra under consideration, the derived spectra formulae are given [2,8-111. 1. Polar Non-Return-to-Zero (NU-L).

2. P c ( f ) = ATb sinc2(fTb) (8)

P, (f ) = A ?b sinc2(fTb )sin2(n-jTb) Bipolar NRZ or Alternative Mark Inversion (AMI).

(9) 3. Manchester.

P C ( f ) = A V b sinc2VT, /2)sin2(n-jTb /2) (10) Figure 1 also shows the PSD comparison between the different line codes.

The figure clearly shows that the PSD of the H-ternary line code overperforms NRZ-L and lies between AM1 and Manchester line codes spectra.

513

“I> 1 .z

Normalized Frequency

Figure 1 . Power spectra of different line codes.

4. Simulation Model

A simulation model has been implemented to evaluate the power spectral density of the H-ternary line code using Matlab. The model contains routines that represent a random binary P-N sequence generator, an encoder and power spectral density analyser. The P-N sequence generator generates random binary data in the form of a matrix. The binary data is then encoded into H-ternary code having three levels. The encoding routine mimics the procedure given in section two. Finally, the FFT algorithm is used to determine the power spectral density of the H-ternary line code. The magnitude of the complex spectrum is then normalised to its highest- amplitude frequency component. The frequency ordinate has also been normalised to the basic rate of the original binary sequence frequency.

Figure 2 shows the results of the normalised power spectral density values obtained from the simulation model for a different normalised frequency of up to a value of 2. The results displayed are only for a single seed of the P-N sequence. Figure 3 shows the same results for another seed. From figures 2 and 3, it is clear that different seeds give different weights of the power spectral density components for the different normalised frequencies. To minimise the difference, the simulation model has been run for different seeds (200 different seeds) and the results are then averaged over the number of runs. Figure 4 shows the average value of the normalised power spectral density of the H- ternary line code for different P-N sequence seeds.

514

1

0.9

0.8

f 1::

I ;;;

0.5

0.4

1 0.3

8 0.2

0.1

0 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Nonnaiiaed Frequency

Figure 2. Normalised power spectral density of the simulation results for a P-N seed 12345.

1

0.9

0.8

p 0.4

9 0.3

8 0.2

0.1

0 0 0.2 0.4 0.6 0.6 1 1.2 1.4 1.6 1.6 2

Nonnalised Frequency

Figure 3. Normalised power spectral density of the simulation results for a P-N seed 54321

Figure 4. Normalised avearged power spectral density of the simulation results for different seeds

515

5.

The mathematical analysis has been accurately modelled using the statistical properties of the H-ternary line code under consideration. The results shown previously in figures 1 shows a power spectral density spectrum that has many desirable features. These include low frequency components that peak at around 50 percent of the signalling rate of the original sequence with no dc component. It also shows that the concentration of the line code bandwidth spans less then 40 percent of the original signalling frequency with insignificant frequency components at around 1.5 times the original signalling rate. Practically, if the last part is removed or suppressed by the transmitter filters or the channel characteristics, this would only lead to rounding the comers of the transmitted three-level pulses with almost no effect on the detection process at the receiving side.

Comparing the simulation results obtained from the simulation model with that of the mathematical model [8] can be seen in figure 5 . In this figure the analytical results are compared with the simulation results for only a single seed. The results show an insignificant deviation from each other in both models. To further eliminate this minor deviation between the simulation and analytical results, the simulation model power spectral density needs to be run many times for different values of seeds. The simulation results from these many runs are then averaged by the number of runs and normalised as in figure 5. Figure 6 shows a comparison where both results are very closely fit with each other except for very low frequency components.

The H-ternary line code considered here is a modification of other predecessor codes that are used for base-band and pass-band data transmission. The new code exploits the merits of these codes and eliminates their deficiencies. The reshaping process of the H-ternary pulses show preferred spectra with no dc component that enables better use of the allocated spectrum. It provides a signalling rate of around 50 percent of the original data bit rate, provides better timing information for encoder-decoder synchronisation and many other desirable features.

The H-ternary line code has a noise performance, which is superior to its predecessor line codes at low-level signal-to-noise ratio [7]. This superiority enables the code to operate at lower power and perform better than the other line codes.

The code has also got the property of single-error detection [5-81. A property which is very much desirable in line codes. This came into effect due to the fact that encoding has a correlative relation between the adjacent line code symbols.

The new line code hardware is relatively more complex in implementation of the encoder and decoder. This complexity however meets a relative simplicity

Comparison of Results and Discussion

516

in the clock recovery circuit at the receiving end. This is due to the fact that for every transmitted H-ternary symbol there is a change in signal level.

t- meoreticat


Figure 5. Results comparison of the theoretical and the simulation models (single seed 12345)

-----__ -____, -0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 l.E 2


Figure 6. Results comparison of the theoretical and the simulation models (200 seeds)

6. Conclusions A new H-ternary line code which belongs to the multi-level (ternary) family has been proposed. The potential advantage of this code is that it takes the merits of other line codes while subtracting their undesirable features. The line code power spectral density is modelled using mathematical and simulation techniques. The results obtained from both models show a clear PSD superiority

517

compared to other similar codes. The simulation results show a relative change in the normalised power density spectra for the different P-N sequence seeds. Replication runs of the simulation program for different seeds and then averaging these different values reduce the difference significantly and hence approximates the simulation values with that of the theory very closely. The simulation model results however still drift slightly away from that of the mathematical model at the very low frequency of the spectrum. Further work is under progress to generate long P-N sequences to overcome this problem.

References

1. W. Cattermole, International J. of Electronics. 55-1, 3-33 (1983). 2. F. Xiong, Digital modulation techniques, Artech House, 17-83 (2000). 3. H. Dutton and P. Lenhard, High-speed networking technology: an

introductory survey, Prentice Hall, 2/1-5 1 (1 995). 4. A. Glass and E. Bastaki, ICCCP'OI, Muscat, Oman, 107-1 10 (2001).

5. A. Glass, B. Ali, and E. Bastaki, ICII2001, Beijing, China, 503-507 (2001). 6. A. Glass, B. Ah, and N. Abdulaziz, DSPCS2002, Sydney, Australia, 149-

153 (2002). 7. A. Glass, E. Bastaki, and N. Abdulaziz, MESM2002, Sharjah, UAE, 146-

150 (2002). 8. A. Glass, N. Abdulaziz, and E. Bastaki, 2"d IEEE-GCC, Manama, Bahrain,

9. P. Lathi, Modern digital and analog communication systems (3rd Ed), Oxford University Press, 294-353 (1988).

10. L. W. Couch, Digital and analog communication systems (5Ih Ed), Prentice Hall, 127-225 (1997).

1 1. S. Haykin, Digital communications, John Wiley, 234-272 (1988).

112-1 16 (2004).

AN OPTIMIZED CPFSK-RECEIVER BASED ON PATTERN CLASSIFICATION IN COMBINATION WITH THE VITERBI

ALGORITHM

DIETER BRUCKMANN

Faculty of Electrical, Information and Media Engineering, University of Wuppertal, Wuppertal, 0-421 19, Germany

A number of modern wireless systems use modulation schemes like CPFSK, GMSK and DPSK, where the information to be transmitted is exclusively coded in the phase change of the RF-signal. Thus the detection of the received data can be performed by using only the zero crossings of the hard-limited IF-signal. With respect to implementation costs and power consumption this is a very attractive receiver concept. A superior performance compared to conventional zero crossing detectors can be obtained by applying methods of pattern recognition for the signal detection. A straight forward implementation of such a detector however is considerably more complex with respect to the requirements for the digital signal processing. In this contribution it will be shown that the complexity can be drastically reduced by combining the pattern classification scheme with the viterbi algorithm. No performance losses have to be taken into account when using this more efficient implementation.

1. Introduction

Continuous Phase Frequency Shift Keying is quite often used for modulation in modem short-range wireless systems, e.g. systems according to the Bluetooth-, HomeRF- or DECT-standard [l , 21. The reason for this is its inherent noise immunity and the possibility to use high efficiency non-linear power amplifiers on the transmitter side. Whereas classical receivers for CPFSK-modulated signals are based on an FM-demodulator with a limiter discriminator [3], due to the advancement of digital signal processing, architectures using a low intermediate frequency (IF) became more and more popular in the last years [3- 61. In the respective implementations the analogue IF-signal must be digitized by additional A/D-converters. FM demodulation and signal detection can be performed then by methods of digital signal processing [4]. The performance improvements thus obtained, must be paid for by higher implementation costs and higher power consumption for the AD-converters, however. Implementation costs and power consumption can be drastically reduced by replacing the Am-converters by simple limiters.

518

519

The concept for symbol or sequence detection of CPFSK-modulated signals using a limiter is shown in fig. 1. The received analogue IF-signal with modulation index q and symbol sequence {dk} is given by:

where fIF stands for the intermediate frequency, T,, is the symbol duration, g(t) is the pulse shape of the gaussian filter, and cpo stands for the initial phase. This signal is transformed into a square wave signal by the zero crossing detector. Even though the information about the amplitude A(t) of the received signal gets lost by this operation, the transmitted bits can be still recovered, since the respective information is exclusively contained in the phase of the transmit signal. Such limiter receivers have the inherent advantage of being insensitive with respect to dynamic range problems.

Analoa CPFSK- I I Zero crossing

Figure 1 . Data-symbol detection in the receiver using the zero crossings of the hard-limited IF- signal.

The time intervals Di between successive zero crossings can now be used for symbol detection. This estimation problem can be also considered as a problem for pattern classification [7] by interpreting the measured zero crossings as a pattern, which has to be classified according to the transmitted, hypothetical symbol sequence. A defined number of values Di are used for detection of the respective bit sequence (dk}. The values Di can be measured by a counter clocked with frequency fo. The frequency fo of the counter must be much higher than the symbol period, since fo determines the time resolution and thus the quantization noise in the values Di.

Under ideal noise-free conditions it is possible to detect the transmitted symbol sequence using only one value Di per bit interval. Assuming binary symbols this can be achieved by performing a comparison with a reference

520

value during each symbol or bit period. If Di is larger than the reference value, the lower instantaneous frequency was used for transmission in this bit interval, if it is smaller the higher frequency was used. For CPFSK-modulated signals these frequencies correspond to bits 1 and 0, respectively. Thus the implementation of the detector is very simple, but it is also very sensitive with respect to signal distortions and noise.

The performance can be considerably improved by using several successive zero crossings for symbol detection, as has been shown in [7]. When using CPFSK-modulated signals, Inter-Symbol Interference (ISI) is generated due to the pulse shaping filter. Thus each bit affects a time interval of several bit periods. For improved performance the whole zero crossing sequence, which is effected by a certain bit, should be used for detection.

The pattern classification detector has the advantage, that nonidealities and mismatches of the receiver can be also taken into account when determining the training patterns and thus can be compensated. This also holds for nonidealities of the transmitter and for distortions on the transmission link, as long as they can be assumed as not time-varying.

2. Receiver configuration with zero crossing detection

On the transmit path the CPFSK-modulated signal is corrupted by co-channel- and adjacent channel interferers and other noise sources. Figure 2 shows the block-diagram of the receiver based on zero crossing detection. The RF-front end, not shown in fig. 2, performs down-conversion of the received RF-signal to a low intermediate frequency. For Bluetooth typical values for the low IF are 1 or 2 MHz [4, 51. In the next step the complex IF-signal is band limited by the channel selection filter. Then the limiters, operating with an over-sampling ratio of OSR with respect to the symbol rate, generate binary signals from the filtered I- and Q-components. In order to limit the quantization noise of the values Di, the sampling rate of the limiters must be considerably higher than the intermediate frequency. A reasonable value for the over-sampling ratio is 104.

1 DSP functions

Figure 2. Receiver for CPFSK-modulated signals with hard-limiting of the IF-signal and a DSP- module for symbol detection.

521

This was confirmed by investigation based on simulation models. The values Di can be determined by a counter, which is reset every time a zero crossing occurs. Before reset the counter value must be stored in a shift register.

Since the values Di still contain the information about the phase @(t) of the received signal, in principle a perfect phase reconstruction based on the zero crossings is possible. Such algorithms for phase reconstruction are described in [8, 111, it turned out however, that quite complex digital signal processing may be required.

Therefore we will consider a simpler method for data recovery in the following, which is based on pattern classification of the zero crossing sequences. Such a concept was already described in [7]. It had been shown, that for a Bluetooth system with an intermediate frequency of 1 MHz about 16 zero crossings are effected by each bit. These 16 zero crossings are effected, however, not only by the bit to be detected but also by two preceding and two successive bits. Thus 32 different zero-crossing patterns exist and must be considered for the detection of each bit. These training patterns represent the reference patterns of the zero crossings for each hypothetical 5-bit pattern and are stored in a memory. The detection is performed by a pattern classifier [7, 81, which determines the minimum distance to the respective training patterns yi.

The smallest distance dst between the received N-dimensional input vector D (corresponding to the values Di) and the stored pattern yn is obtained by computing the metrics over all yn and taking the smallest value according to following relationship:

Investigations based on simulation models showed that the performance degradation is negligible when using the city block distance, which is obtained for g=1, and when using only four out of the 16 values Di for the metrics computation. Thus 4*3*32=384 additionslsubtractions and 3 1 comparisons have to be carried out during each bit period. The computational load can be however considerably further reduced by combining the pattern classifier with the viterbi algorithm. This is described in the following chapter.

3. Pattern Classification Combined with the Viterbi Algorithm

The output data corresponding to the time differences between zero crossings suffer from Inter-Symbol Interference (ISI), as already described. Using an

522

Figure 3. State transition diagram for the Viterbi algorithm.

efficient search algorithm the optimal path through the trellis diagram can be determined. We have used the well-known viterbi algorithm for our application [9, 101. The Gaussian filter for Bluetooth has a length of about 3 bit periods, thus a viterbi algorithm with 2 state variables should be sufficient. The respective 4 states and the possible state transitions are shown by the diagram in figure 3.

Every bit period the viterbi algorithm must calculate 8 branch metrics. These branch metrics BMi are obtained e.g. from the Euclidian distance between the received N-dimensional input vector D (corresponding to the values Di) and the eight possible zero crossing patterns yn corresponding to the 3 consecutive bits {dk-2, dk-,, dk}. The values BMi are obtained according to following relationship, where g equals 2 for the Euclidian norm:

Adding the values BMi to the 4 accumulated branch metrics ABMi, 8 new accumulated branch metric values are generated. As shown in figure 3, two paths enter each state. According to the viterbi algorithm only the path with the smallest accumulated metric has to be saved, the other one can be discarded. Thus for each state an accumulated branch metric value is stored. In the next step the state with the smallest accumulated metric is determined and a decision about bit dk-2 is made.

Even though each 3 bit sequence is represented at least by 12 zero transitions, it turned out that it is sufficient to store only 3 values Di of each

523

sequence D. Thus only 3*2*8=48 additions/subtractions have to be performed during each bit period for the computation of the values BMi. Furthermore 8 additions and 7 comparisons have to be carried out for the computation of the accumulated branch metrics ABMi and for the determination of the minimum accumulated metric. This adds up to a total of 56 additions/subtractions and 7 comparisons. Thus the computational load for the detector can be reduced by more than 80% by combining the pattern classifier with the viterbi algorithm.

4. Performance results

Simulations confirmed the improvements which can be obtained by the proposed estimation and detection method. In order to analyze the performance of the receiver a simulation model for a Bluetooth system with gross bit rate 1 Mbit/s and a nominal modulation index of 71=0.3 15 has been used.

Fig. 4 shows the results for an AWGN-channel. The performance of the combined classification and viterbi detector seems to be slightly better than that of the classification detector considering a sequence of 16 zero crossings. For a bit error rate of for both a signal-to-noise ratio of about 12,5 dB is obtained. The performance has been also compared with that of a conventional system. For this a simulation model was generated with a receiver using a limiter

.. + .

102

5rn

. \.; ................ 4 ................ ; ....... & ..... i ............... .......... ..................... ........................ ..... c ............... ............................. .I ............................. .b ............... I -:--------:. ---- -c .---- .-.--L -----.. i --_. ....... .. ................................. -. ........... ........... ............... d ............ * ................ , ................ L..

............... d ............. * ................ , ................ L ............ .......................... ,...*................I................ c ............

............... i ............. 7 ................ i ................ i .............. ............... ............. .I ................................. + ........... . : ............... ; ........... t . 4 ................ i ................ & ............. . . c :

1 ............................. .- ............... -:--------:. ---- -c .---- .-.--L -----.. i ---. ....... .. ................................. -. ........... ........... ............... d ............ * ................................. L.. ............... i ............. 7 ................ i ................ i .............. ............... ............. .I ................................. + ........... ............... * ............ .+ ................! ................ L ............ .............................. *... .............................. ............ 1 ............... i ........... t . 4 ................ i ................ ....... \...I ; I..-; ............... i ............. hj ................. i ................ + ............

+-.....d ............-. I ..........-.. :: ::::::::::::::::F::::::::::::::: ...I.........# c .......- ................................ v ................................................ ...............I... ............. +... .... ......... $ ................ ............... ................ I ................ 6 ..... ................ : .................

i

12 14 16 18 20 I lo5

10

Figure 4. Signal to noise ratio versus bit error rate for the limiter receiver, pattern classification combined with viterbi detection (curve I), pattern classification over 16 zero crossings (curve 2), conventional limiter discriminator detector (curve 3).

524

I I I I I I 10 11 12 13 14 15 16

SNR [dB]

Figure 5. Signal to noise ratio versus bit error rate for the novel limiter receiver, for nominal modulation index and lower and upper bound of modulation index, q=0.35 (curve I), q=0.315 (curve 2), q=0.28 (curve 3).

discriminator and an FM-AM-converter for demodulation. The simulation results confirmed the superior performance of the classification detectors, which perform better by more than 5db with respect to the receiver sensitivity. Figure 5 shows the bit error rate versus SNR for different modulation indices. The reference patterns had been determined with the same modulation index as the transmitted data. The good performance of the novel detector based on pattern classification and viterbi detection is thus confirmed. If always the nominal modulation index is used for the reference patterns, the performance is only slightly degraded compared to curves 1 and 3.

5. Summary

In this contribution a novel method for data detection in phase-modulated signals was presented. Phase-modulation schemes are used in several wireless communication systems such as Bluetooth, DECT and GSM. The proposed concept is based on the idea, to detect the modulated data directly from the time intervals between the zero crossings of the limiter output signal. A very good performance with respect to distortions is obtained by using methods of pattern classification for symbol detection. It has been shown that the computational load can be considerably reduced by combining the pattern classifier with the

525

viterbi algorithm. No performance losses must be traded for the reduced implementation costs.

The performance of the proposed method has been verified by using simulation models. It was confirmed that the concept has a superior performance compared to conventional limiter-discriminator demodulation and detection schemes and that the sensitivity of the respective receiver can be made better than those of classical analogue demodulators. For the digitization only comparators and no multi-bit AD-converters are needed, thereby making the implementation simple and cost effective.

The method can be completely implemented in the DSP-part of the receiver. The concept furthermore enables the complete integration of RF and base-band functionalities into a single integrated circuit in a pure CMOS technology.

References

1.

2. 3.

4.

5.

6.

7.

8.

9.

ETSI, European Telecommunication Standard, ETS 300 175 Digital Enhanced Cordless Telecommunications (DECT). Bluetooth specifications, Bluetooth SIC at httu://www.bluetooth.com. H. V. Thomas, J. Fenk, S . Beyer, A One-Chip 2 GHz Single Superhet Receiver for 2Mbh FSK Radio Communication, Proceedings IEEE lnternational Solid-state Circuits Conference, 42, (1994). Ch. Diirdoth et. al., A low-IF RX Two-Point XA-Modualtion TX CMOS Single-Chip Bluetooth Solution., IEEE Trans. on MTT, MTT49, no. 9, 153 1 (200 1). H. Darabi et. al., A 2.4 GHz CMOS Transceiver for Bluetooth, IEEE JSCC, JSSC36, no. 12, 2016 (2001). M. J. Crols and M. Steyaert, CMOS WIRELESS TRANSCEIVER DESIGN, Boston, MA, Kluwer, (1997). D. Briickmann, A. Neubauer, Signal detection based on pattern classification for use in wireless CPFSK receivers, Proc. IEEE Int. Conf. Circ. and Systems, Vancouver Canada, (2004). D. Briickmann, M. Hammes, and A. Neubauer, A CPFSKPSK-Phase Reconstruction Receiver for Enhanced Data Rate Bluetooth Systems, Proceedings IEEE Int. Conf. On Electronics, Circ. and Systems, Tel-Aviv, Israel, (2004). M. Ba R. Schiphorst, F.W. Hoeksema, and C.H. Slump, A (simplified) a posteriori probability (MAP) receiver, SPAWC 2003, IEEE Workshop on advances in wireless communications, Rome, Italy, (2003).

10. J. G . Proakis, Digital Communication, Mc Graw Hill (1995). 1 1. A. Neubauer, Irregulare Abtastung - Signaltheorie und Signalverarbeitung,

Berlin, Springer, (2003).

personal wireless communications: pwc 05 - proceedings of the 10th ifip conference

Documents