UNIVERSITY OF OULU P .O. Box 8000 F I -90014 UNIVERSITY OF OULU FINLAND

A C T A U N I V E R S I T A T I S O U L U E N S I S

University Lecturer Tuomo Glumoff

University Lecturer Santeri Palviainen

Postdoctoral research fellow Sanna Taskila

Professor Olli Vuolteenaho

University Lecturer Veli-Matti Ulvinen

Planning Director Pertti Tikkanen

Professor Jari Juga

University Lecturer Anu Soikkeli

Professor Olli Vuolteenaho

Publications Editor Kirsti Nurkkala

ISBN 978-952-62-1520-4 (Paperback)ISBN 978-952-62-1521-1 (PDF)ISSN 0355-3213 (Print)ISSN 1796-2226 (Online)

U N I V E R S I TAT I S O U L U E N S I SACTAC

TECHNICA

U N I V E R S I TAT I S O U L U E N S I SACTAC

TECHNICA

OULU 2017

C 609

Mirjami Jutila

ADAPTIVE TRAFFIC MANAGEMENT IN HETEROGENEOUS COMMUNICATION NETWORKS

UNIVERSITY OF OULU GRADUATE SCHOOL;UNIVERSITY OF OULU,FACULTY OF INFORMATION TECHNOLOGY AND ELECTRICAL ENGINEERING,DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

C 609

ACTA

Mirjam

i Jutila

C609etukansi.kesken.fm Page 1 Wednesday, February 22, 2017 11:26 AM

A C T A U N I V E R S I T A T I S O U L U E N S I SC Te c h n i c a 6 0 9

MIRJAMI JUTILA

ADAPTIVE TRAFFIC MANAGEMENT IN HETEROGENEOUS COMMUNICATION NETWORKS

Academic dissertation to be presented with the assent ofthe Doctoral Training Committee of Technology andNatural Sciences of the University of Oulu for publicdefence in Kuusamonsali (YB210), Linnanmaa, on 17March 2017, at 12 noon

UNIVERSITY OF OULU, OULU 2017

Copyright © 2017Acta Univ. Oul. C 609, 2017

Supervised byProfessor Tapio SeppänenDocent Tapio Frantti

Reviewed byProfessor Jyri HämäläinenProfessor Jussi Kangasharju

ISBN 978-952-62-1520-4 (Paperback)ISBN 978-952-62-1521-1 (PDF)

ISSN 0355-3213 (Printed)ISSN 1796-2226 (Online)

Cover DesignRaimo Ahonen

JUVENES PRINTTAMPERE 2017

OpponentProfessor Jukka Manner

Jutila, Mirjami, Adaptive traffic management in heterogeneous communicationnetworks University of Oulu Graduate School; University of Oulu, Faculty of Information Technologyand Electrical Engineering, Department of Computer Science and EngineeringActa Univ. Oul. C 609, 2017University of Oulu, P.O. Box 8000, FI-90014 University of Oulu, Finland

Abstract

Communication networks are experiencing a significant growth of data traffic posing newchallenges to the overall systems that should become more reactive and adaptive towardsdynamically changing traffic, connections and network conditions. This thesis examines adaptivetraffic management solutions within heterogeneous communication networks, which can beutilized to improve network performance, provide Quality of Service (QoS) for traffic paths andshare resources in a fair way. The developed adaptive methods include solutions for fuzzy flowscheduling (AWFQ, FWQ) and regressive admission control (REAC) to provide stable networkperformance and efficient resource control. Such techniques for adaptive traffic managementcontinuously balance and control traffic usage and recover from network faults and attacks. Theresults utilize traffic monitoring for estimating the overall network conditions, applying cognitionto learn from previous actions, and adapting to the current traffic conditions for resourceoptimization. The thesis researches how to distribute these computing mechanisms towardsnetwork edges closer to the actual application users for more efficient resource usage, and toprovide better performance for delay-sensitive applications. The methods developed have beenapplied to vehicular communications to assess and improve the messaging between vehicles andvulnerable road users (VRUs). These mechanisms are able to react faster to data traffic changesand guarantee better quality for prioritized traffic and users while at the same time they preservefairness to other flows compared to traditional control and scheduling methods without adaptivecharacteristics. The overall system reacts to changes in the network QoS by determining decision-making procedures on possible flow rejection, marking, or allowed bandwidth weight assignment,thus bringing cognition to the network path.

Keywords: adaptive queueing, admission control, edge computing, flow scheduling,fuzzy scheduling, network adaptivity, vehicular communications

Jutila, Mirjami, Adaptiivinen liikenteenhallinta heterogeenisissa tietoliikenne-verkoissaOulun yliopiston tutkijakoulu; Oulun yliopisto, Tieto- ja sähkötekniikan tiedekunta,TietotekniikkaActa Univ. Oul. C 609, 2017Oulun yliopisto, PL 8000, 90014 Oulun yliopisto

Tiivistelmä

Merkittävä liikennemäärien kasvu aiheuttaa tietoverkoille uusia haasteita, minkä vuoksi niidentäytyy tukea reaktiivisuutta ja adaptiivisuutta vastatakseen muuttuviin liikenne- sekä verkko-olo-suhteisiin että yhteyksiin. Väitöskirjassa kehitetään heterogeenisten tietoverkkojen adaptiivisialiikenteenhallintaratkaisuja, joita voidaan hyödyntää verkon suorituskyvyn parantamiseen, tarjo-amaan liikenteen palvelunlaatua (QoS) sekä tasapuolista resurssien jakoa. Kehitetyt adaptiivisetmenetelmät sisältävät ratkaisuja sumeaan logiikkaan perustuvaan skedulointiin sekä regressiivi-seen verkon pääsynhallintaan pohjautuen, jotka takaavat vakaamman verkon suorituskyvyn jaresurssien hallinnan. Nämä menetelmät tasapainottavat ja kontrolloivat liikennettä sekä pyrkivätpalautumaan verkon häiriöistä ja hyökkäyksistä. Tulokset hyödyntävät liikenteen monitorointiaverkon tilan arviointiin, soveltavat kognitiivisuutta oppiakseen aiemmista toiminnoista sekäadaptoituvat nykytilanteeseen resurssien optimoimiseksi. Väitöskirja tutkii, miten kyseisiä las-kentamenetelmiä voidaan hajauttaa verkon reunoille lähemmäksi sovellusten käyttäjiä resurssi-en käytön tehostamiseksi sekä tarjoamaan parempaa suorituskykyä viiveherkille sovelluksille.Kehitettyjä menetelmiä sovelletaan autoverkkoihin autojen sekä suojattomien tienkäyttäjienviestinnän määrittämiseen sekä parantamiseen. Nämä menetelmät reagoivat nopeammin datalii-kenteen muutoksiin, takaavat paremman laadun priorisoidulle liikenteelle sekä samalla tasapuo-lisuutta muulle liikenteelle verrattuna perinteisiin kontrollointi- ja skedulointimenetelmiin. Kehi-tetty järjestelmä reagoi verkon palvelunlaadun muutoksiin määrittelemällä päätöksentekomalle-ja mahdolliseen tietovuon hylkäämiseen, merkitsemiseen tai kaistankäytön painokertoimen mää-rittämiseen, täten luoden kognitiivisuutta verkon reitille.

Asiasanat: adaptiivinen jonotus, autoverkot, laskenta verkon reunalla, sumean logiikanskedulointi, verkkoon pääsy, verkon adaptiivisuus, vuon skedulointi

Preface

The research work reported in this thesis was conducted at the VTT Technical ResearchCentre of Finland, from 2008-2016. I joined VTT in 2004 to do my Master’s thesis, andsince then I have been involved in a large variety of different national and internationalresearch projects receiving funding from Tekes - the Finnish Funding Agency forTechnology and Innovation, the EU and VTT. The research work was started inEureka/ITEA Easy Wireless Project in 2008 that developed solutions for servicecontinuity and tools for enhancing QoS. The work was continued in the Tekes fundedICT SHOK Future Internet project with development of Adaptive Weighted FairQueueing (AWFQ) and Fuzzy Weighted Queueing (FWQ) methods for link sharing andpacket scheduling. The Regressive Admission Control (REAC) method was introducedin the Celtic Easy Wireless 2 project, and the work was continued in Eureka/ITEA2DICOMA project. In the latest work, the network monitoring and adaptive solutionshave been applied to Internet of Things (IoT) in the Digile IoT project and to IntelligentTransportation Systems in the Celtic CoMoSeF and EU FP7 VRUITS projects. Mythesis has also been supported by scholarships granted by the foundations of UllaTuominen, Riitta ja Jorma Takanen, Tauno Tönning and Emil Aaltonen.

First, I am deeply thankful to my thesis supervisor Dr. Tapio Frantti who has givenme patient guidance through all these years. Without your support and knowledge Iwould not have been able to accomplish this thesis. I also thank my other supervisorProf. Tapio Seppänen, and Prof. Jyri Hämäläinen and Prof. Jussi Kangasharju for thepre-examination and valuable comments to improve the quality of the thesis.

I want to thank my colleagues at VTT and elsewhere, who have been workingwith me in these projects. I give special thanks to Mr. Jyrki Huusko for guiding me towork as a project manager in the DICOMA project. I am grateful for the support thatI have received from the co-authors of my papers Dr. Jarmo Prokkola, Dr. DespinaTriantafyllidou, Dr. Johan Scholliers, Mr. Mikko Valta and Mrs. Kaisa Kujanpää. I amalso thankful to Dr. Jukka Mäkelä for encouraging me to start the writing process of mythesis. I want to give special thanks to my room-mates Mr. Markus Luoto and Mrs.Kaisa Kujanpää as well as to Dr. Esa Piri, Dr. Martín Varela and Mr. Olli Mämmelä forthe useful discussions, technical assistance and for keeping up the working spirit.

7

I like to thank my parents, Raija and Jorma, for their endless encouragement andsupport in my life. I am also grateful to my brother Sakari, other close relatives andfriends for being there for me. Finally, the three supermen at home, my dearest Matti,and sons Akseli and Severi, what would I do without your funny jokes and the cheerfulmoments that provide a nice counterpart to the researcher’s life.

Oulu, January 12th, 2017 Mirjami Jutila

8

Abbreviations

3GPP 3rd Generation Partnership Project

5G Fifth Generation Wireless Systems

AC Admission Control

ACR Absolute Category Rating

AF Assured Forwarding

AWFQ Adaptive Weighted Fair Queueing

BB Bandwidth Broker

BE Best-Effort

BLE Bluetooth Low Energy

CAM Cooperative Awareness Message

CBQ Class-Based Queueing

CBR Constant Bit Rate

CSMA/CA Carrier Sense Multiple Access with Collision Avoidance

C-ITS Cooperative Intelligent Transportation Systems

CoA Center of Area

DENM Distributed Environmental Notification Message

DHCP Dynamic Host Configuration Protocol

DiffServ Differentiated Services

DCF Distributed Coordination Function

DPI Deep Packet Inspection

DSCP DiffServ Code Point

DSSS Direct Sequence Spread Spectrum

DWFQ Dynamic Weighted Fair Queueing

EDCA Enhanced Distributed Channel Access

EDGE Enhanced Data rates for GSM Evolution

EF Expedited Forwarding

ETSI European Telecommunications Standards Institute

FAN Flow-Aware Networking

FCFS First Come First Served

FHSS Frequency Hopping Spread Spectrum

FSA Flow-State-Aware

9

FIFO First-In First-Out

FSM Finite State Machine

FWQ Fuzzy Weighted Queueing

GPRS General Packet Radio Service

HSPA High-Speed Packet Access

HCCA HCF Controlled Channel Access

HCF Hybrid Coordination Function

IANA Internet Assigned Numbers Authority

ICMP Internet Control Message Protocol

IETF Internet Engineering Task Force

ICT Information and Communication Technologies

IEEE Institute of Electrical and Electronics Engineers

IntServ Integrated Services

IoT Internet of Things

IP Internet Protocol

IPv6 Internet Protocol version 6

ITS Intelligent Transportation Systems

ITS-G5 Intelligent Transportation Systems-G5

ITU-T International Telecommunication Union -

Telecommunication Standardization Sector

LAN Local Area Network

LOS Line-Of-Sight

LTE Long Term Evolution

LTE-A Long Term Evolution - Advanced

LTE-M Long Term Evolution - Machine-to-Machine

M2M Machine-to-Machine

MAC Medium Access Control

MBAC Measurement-Based Admission Control

MEC Mobile Edge Computing

MIMO Multiple Input Multiple Output

MOS Mean Opinion Score

NAT Network Address Translation

NFV Network Functions Virtualization

NGN Next Generation Networking

NLOS Non Line-Of-Sight

10

NS-2 Network Simulator 2

OBU On-Board Unit

OGC Open Geospatial Consortium

PBAC Parameter-Based Admission Control

PCF Point Coordination Function

POC Proof Of Concept

PQ Priority Queueing

PSQA Pseudo-Subjective Quality Assessment

QMCP QoS Measurement Control Protocol

QoE Quality of Experience

QoS Quality of Service

RA-WFQ Revenue-based Adaptive Weighted Fair Queueing

RA-WRR Revenue-based Adaptive Weighted Round Robin

RA-DRR Revenue-based Adaptive Deficit Round Robin

REAC Regressive Admission Control

RFID Radio Frequency Identification

RR Round Robin

RSU Road Side Unit

RSVP Resource Reservation Protocol

RTT Round Trip Time

SDN Software Defined Networking

SLA Service Level Agreement

SLS Service Level Specification

SWE Sensor Web Enablement

TCP Transmission Control Protocol

TOS Type Of Service

TTC Time To Collision

UDP User Datagram Protocol

UMTS Universal Mobile Telecommunications System

UWB Ultra Wide Band

V2I Vehicle-to-Infrastructure

V2V Vehicle-to-Vehicle

V2VRU Vehicle-to-Vulnerable Road User

V2X Vehicle-to-X

VM Virtual Machine

11

VRU Vulnerable Road User

WAVE Wireless Access in Vehicular Environments

WiMAX Wireless Interoperability for Microwave Access

WFQ Weighted Fair Queueing

Wi-Fi Wireless Fidelity

WLAN Wireless Local Area Network

12

List of original publications

This thesis is based on the following papers, which are referred to in the text by theirRoman numerals (I-V):

I Jutila M (2016) An Adaptive Edge Router Enabling Internet of Things. IEEE Internet ofThings Journal 3(6):1061-1069.

II Jutila M, Scholliers J, Valta M & Kujanpää K (2017) ITS-G5 performance improvementand evaluation for vulnerable road user safety services. IET Intelligent Transport Systems,in press.

III Jutila M, Prokkola J & Triantafyllidou D (2013) Regressive Admission Control Enabledby Real-Time QoS Measurements. International Journal of Computer Networks andCommunications (IJCNC 2013) 5(6):23-43.

IV Frantti T & Jutila M (2009) Embedded fuzzy expert system for Adaptive Weighted FairQueueing. Expert Systems with Applications (Elsevier) 36(8):11390-11397.

V Jutila M & Frantti T (2017) Cognitive Fuzzy Flow Control for Wireless Routers. Inter-national Journal of Autonomous and Adaptive Communications Systems (IJAACS), inpress.

13

14

Contents

AbstractTiivistelmäPreface 7Abbreviations 9List of original publications 13Contents 151 Introduction 17

1.1 Background and motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

1.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

1.3 Research approach and contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .20

1.4 Outline of the dissertation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2 Adaptive traffic management framework and techniques 272.1 Heterogeneous networking for the Internet of Things . . . . . . . . . . . . . . . . . . . . 28

2.1.1 Internet of Things communication challenges . . . . . . . . . . . . . . . . . . . . 30

2.1.2 Vehicular communications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.1.3 Vehicle to vulnerable road user (V2VRU) communicationrequirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2.2 Traffic management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

2.2.1 Queueing and scheduling mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

2.2.2 Network information management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2.2.3 Fuzzy logic based traffic management . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

2.3 QoS control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

2.3.1 Resource allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

2.3.2 Network admission control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3 Contributions of the thesis to adaptive traffic management 473.1 Functional requirements of the system architecture . . . . . . . . . . . . . . . . . . . . . . 47

3.2 Computing at the network edge. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .49

3.3 Regressive admission control (REAC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

3.3.1 REAC architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

3.3.2 REAC control logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

15

3.4 Fuzzy flow scheduler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 563.4.1 Fuzzy weighted queueing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573.4.2 Reasoning example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 604 Results and discussion 63

4.1 REAC measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634.2 Fuzzy flow scheduling results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

4.2.1 Fuzzy weighted queueing in LAN and WLANs . . . . . . . . . . . . . . . . . . . 674.2.2 Fuzzy weighted queueing in vehicular communications . . . . . . . . . . . 69

4.3 Vehicle to vulnerable road user (V2VRU) use case results . . . . . . . . . . . . . . . . 724.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

5 Conclusions 81References 83Original publications 91

16

1 Introduction

1.1 Background and motivation

The vision of future networking enhances the current view of connected people toconnect all devices, applications and media globally. Different wireless and wiredtechnologies link sensors and systems together utilizing wireless access points, gatewaysand routers that in turn connect to service-based intelligence with cloud computing. Inparticular, the Internet of Things (IoT) framework [1–4] aims to connect billions ofdevices ranging from various domains such as transportation, industry, smart home,smart city, medical services and energy systems. To transfer all the data into a cloud[5, 6] and the response back to the system, without any pre-processing and distributedcomputing, would consume not only the already scarce bandwidth resources but also thetime and money of different players in the IoT product chain.

In response to this problem, part of the network intelligence and data managementshould be distributed to gateways and routers in interconnected systems creating fogand edge computing operations. Edge computing, also known as fog computing[7–9], is an extension of the cloud computing paradigm. Computation tasks that arenormally executed in the cloud are brought closer to the end users at the network edge.Fog and cloud computing complement each other, providing mutually beneficial andinterdependent services to make computing, storage, control and networking possibleanywhere along the ubiquitous communications. Therefore, communication latenciescan be reduced as operations are executed closer to the user. However, this type ofnetworked world sets great demands on network control methods to efficiently managethe massive amounts of nodes and data with minimal human intervention.

In today’s Internet, there is a growing demand for high performance real-timeservices, which creates strong requirements for high bandwidth and the timely deliveryof information. There are various traffic management mechanisms developed for wiredand wireless networks to improve the Quality of Service (QoS) including: scheduling,queue management, QoS-based routing, traffic off-loading, mobility management, andadmission control (AC). Key components for the effectiveness of a QoS scheme includethe end-to-end traffic management and bandwidth estimations. The downside is thatmany of these traffic engineering mechanisms are rather complicated and static in theiroperations and often require extra signaling. In some cases, the solution is to use high

17

capacity links, but on the other hand provisioning resources for traffic bursts withoutproviding adaptive traffic engineering mechanisms eventually leads to over-provisioningand sub-optimal usage of the capacity.

For instance, in current fixed IP routers, many scheduling algorithms are staticand allocate resources according to predefined service models. The service modelsare based on estimated traffic loads and patterns from traffic history. However, theloads of different service classes already vary on a short timescale due to the trafficbursts. Therefore, the predefined resource allocation cannot adapt to dynamicallychanging traffic load conditions. The system should maintain an acceptable level offairness to other services at the same time as providing QoS for delay-sensitive real-timeapplications. The early QoS mechanisms based on prioritization, like DifferentiatedServices (DiffServ) [10] and Integrated Services (IntServ) [11] or IEEE 802.11e in aWLAN [12], can provide QoS for certain types of traffic flows. However, these methodsprovide only statistic QoS instead of guaranteed QoS, i.e., if too much high-prioritytraffic appear, the performance of all the users will collapse, similarly as with thebest-effort Internet.

The use of adaptive traffic management methods can overcome problems raisedby the emerging new applications and network congestion. However, adaptive flowscheduling and QoS control requires new functionalities to share the link resourcesbetween service classes. In order to support these requirements, the future Internetis expected to become more reactive and adaptive towards dynamically changingtraffic and network conditions. The new management approaches should enable theInternet to continuously optimize its usage and recover from faults and attacks, i.e., touse self-management principles, without disturbing the supported applications. Thisdissertation examines how adaptive traffic management techniques can be utilized toimprove network performance, provide QoS for traffic paths and share traffic in a fairway. The implemented methods include parts for traffic monitoring, estimating theoverall network conditions, and adapting to the current traffic conditions for resourceoptimization. Hence, applying cognition to process the information in communicationnetworks is essential for creating more efficient systems and to respond to demandingapplications’ requirements.

18

1.2 Problem statement

One problem in the Internet, and in any other bandwidth limited communicationsystem, is a shortage of resources to satisfy all traffic demands. The Internet has only abest-effort class of service, and no assurance of when and how many datagrams aredelivered, which is not sufficient, especially for the real-time applications. Duringnetwork congestion, datagrams are treated equally, even if a variety of applicationsrequire additional support from the network, which sets great demands on networkcontrol methods to efficiently manage the massive amounts of nodes and data. With theincreasing number of applications requiring real-time and delay sensitive operations, theneed for adaptive methods suitable for highly dynamic environments has increasedenormously. To overcome certain limitations of current traffic engineering operations,this thesis applies and embeds adaptive computing in the network devices. This enhancesthe systems’ abilities to analyze its environments and decide on appropriate actions aswell as to learn from previous decisions. The network performance measurements areutilized with network management operations for regressive admission control (REAC)and fuzzy weighted queueing (FWQ) to balance the traffic in heterogeneous networkenvironments. These techniques, together with an adequate amount of processingpower as a precondition, provide adaptive network knowledge to optimize the networkperformance and alleviate traffic overload situations giving response to time-of-dayeffects, channel quality variations and emerging new services.

The research question of this thesis is whether it is possible to develop adaptive

and cognitive management methods to monitor, balance, and share traffic that has

different capacity requirements in heterogeneous communication networks. As theresearch assumption, I state that efficient and fair traffic management requires adaptive

mechanisms for network admission control and flow scheduling to balance and control

the traffic in order to continuously reconfigure and fulfill the network management and

resource allocation objectives. My research hypothesis is that the presented methods

for regressive admission control and fuzzy flow scheduling are suitable for adaptive

traffic management in heterogeneous communication networks.Current methods in fixed networks rely on offline practices for resource management,

where a centralized computing entity is responsible for managing network performanceover long timescales. However, the Internet requires new traffic engineering methodsin order to adapt to network and traffic variations requiring more decentralized ap-proaches around and on the network edges with online traffic engineering capabilities.

19

Addressing this especially, in the IoT framework, part of the network intelligenceand data management should be distributed to gateways and routers creating fog andedge computing operations. These operations include, e.g., mobility support, resourcemanagement, low latency connectivity, location awareness, and QoS features. In orderto support the requirements of future connected systems, self-managed and autonomicfunctionalities have to be included in various domains extending the network operationand management.

1.3 Research approach and contributions

In order to allocate resources and to guarantee QoS for specified traffic for selectedIP connections in an adaptive way, routers have to monitor traffic, limit the access ofincoming flows and classify the incoming packets into appropriate traffic queues. Theparts for adaptive traffic management utilized in this thesis are presented in Figure 1.

Fig. 1. Adaptive network management and control.

In this thesis, the network traffic is being monitored and the traffic classified togain QoS awareness. The network admission of flows is controlled by the REACmethod working at the network edges. The fuzzy flow scheduler with the adaptiveFWQ algorithm calculates new bandwidth weight coefficients periodically on routers

20

preserving optimized path performance and application dependent QoS with flowfairness. Changes in the network conditions determine decision-making procedures atthe network edges for the possible flow rejection or marking and assigning the allowedbandwidth weight, thus bringing cognition to the network path. This way the networkpaths are controllable and the system is capable of adapting and self-managing thenetwork state and operations. For instance, if the traffic flows endeavoring into thenetwork, e.g., malicious traffic, are exceeding or likely to exceed the network capacity,they can be treated differently to alleviate and minimize the effects of bandwidthattacks. In this thesis, the network performance of different heterogeneous networktechnologies, e.g., IEEE 802.3, ITS-G5 (IEEE 802.11p) and IEEE 802.11b, is beingassessed, monitored and balanced in varying circumstances and conditions.

The evaluation of the developed solutions is based on simulations, test-bed mea-surements and real prototype implementations. The fuzzy scheduler presented in thisdissertation is implemented utilizing the Network Simulator NS-2 simulator environ-ment. The regressive admission control and IoT use cases are based on prototypeimplementations in test-beds and in real environments. The ITS-G5 measurementsare carried out in rural environments utilizing devices that test the state-of-the arttechnology.

This dissertation is based on five peer-reviewed original papers. The author’scontributions and the main areas of research regarding this dissertation are presented inthe following papers:

Paper I introduces adaptive solutions based on the REAC and FWQ methods workingat the network edge for IoT communications. The intelligent solutions are based onadaptive methods and algorithms controlling and optimizing the network path andresources. There are some distinguished challenges related to fog computing that ourwork addresses, such as QoS, network provisioning and resource management. Someparts of the cloud’s management tasks are distributed around the edges of networkedsystems including: real-time communication capabilities, and high quality messageexchange for applications with low latency and reliability requirements. The systemmonitors applications that collect data periodically from a multitude of data sources toprovide flow awareness and QoS. Adaptive traffic management and control methodsbased on REAC and FWQ are profitable solutions for the optimization of resources atthe network edges. This was a solo paper by the author discussing the problem space forefficient network traffic management utilizing edge computing in the IoT framework

21

and presenting adaptive solutions.

Paper II presents vehicular communications utilizing ITS-G5 (IEEE 802.11p) for timecritical communications allowing the delivery of safety information over a long rangewith low latency. However, obstacles in the link path and a large amount of vehiclessending at the same time may reduce performance. The purpose of the work was toresearch the performance of ITS-G5 and its suitability for demanding scenarios betweenvehicles and vulnerable road users (VRUs). We have tested various non-line-of-sight(NLOS) scenarios in urban environments and line-of-sight (LOS) simulations to supportC-ITS message prioritization and scalability with different amount of vehicles. Exampleuse cases with NLOS include pedestrians crossing streets from behind objects, andlow-visibility scenarios, e.g., when a VRU is behind a vehicle, behind a queue of vehicles,between vehicles, behind trees/bushes or behind a building. The LOS simulations utilizefuzzy weighted queueing (FWQ) mechanism for congestion control to overcome thepacket losses and for data prioritization. Based on the measured network performance,the suitability of cooperative ITS (C-ITS) messaging for VRU safety applications isassessed and performance improved. The author contributed to the implementationof measurement cases and conducting the measurements. The author was the maincontributor to the paper with input from co-authors: Johan Scholliers, Mikko Valta, andKaisa Kujanpää.

Paper III describes a regressive principle to Admission Control (AC) assisted byreal-time passive QoS monitoring. It makes regressive decisions on possible flowrejection and prioritization, according to which the network unconditionally acceptsnew flows, and evaluates their impact upon existing clients. Regression allows instantaccess of new flows without the need to wait for a decision thus bringing cognition tothe network path. The REAC system consists of traffic classifying, monitoring andflow control components. The developed AC logic monitors the QoS-level variationsand estimates when the QoS-level decrease affects the quality of the high-priorityapplications. Whenever the measurement indicates that the quality drops below a giventhreshold, a decision-making process is initiated to drop or mark the excessive lowpriority flows. The REAC method focuses on the temporal variation, e.g., of a delayparameter, averaged over a defined measurement window to make the resource controldecisions. The author participated in the design of the REAC module and in conductingthe test bed measurements. The author was the main author of the paper with input from

22

co-authors: Jarmo Prokkola and Despina Triantafyllidou.

Paper IV introduces an embedded fuzzy expert system for Adaptive Weighted FairQueueing (AWFQ). AWFQ is located in the network traffic router allowing AWFQ toacquire differentiated service for traffic classes according to QoS requirements. AWFQdesignates queue weights utilizing the fuzzy expert system’s two-input, single-outputcontrol strategy. The fuzzy scheduler fairly allocates resources for prioritized users forreal-time UDP traffic and best-effort TCP traffic. The network path is being monitoredand controlled by the fuzzy expert system which calculates new bandwidth weightcoefficients periodically on routers. It is shown through simulations that the adaptivemodel is more stable and reacts faster to different traffic states than the conventionalstatic method for Weighted Fair Queueing (WFQ). The author participated in thedesign phase of the QoS controlling strategy, the fuzzy expert system rule base tuning,implementing the simulation model and executing the simulations with the initial idea ofthe topic and input from the main author Tapio Frantti.

Paper V presents a cognitive fuzzy control system developed for enhancing the wirelessrouter’s throughput and delay. Developed Fuzzy Weighted Queueing (FWQ) schedulercontrols traffic flows by anticipating the required changes on weight coefficients. Thisway the devices operating in Wireless Local Area Networks (WLAN) can utilize analgorithm to adaptively and fairly share the available bandwidth for each service class.To provide continuous information about the network state, new functionalities to sharethe link resources between service classes were developed. In current Internet IP routers,the scheduling algorithms are static and allocate resources according to some predefinedservice models. However, the loads of different service classes already vary on ashort timescale due to the traffic bursts. Therefore, the predefined resource allocationcannot adapt to dynamically changing traffic load conditions, and thus adaptive methodsare profitable. The author contributed to the design and implementation of the FWQalgorithm for carrying out the required simulations. The author was the main contributorwith the assistance and input from the co-author Tapio Frantti.

The overview of different traffic management levels with the positioning of thereference papers listed above is presented in Figure 2. The main focus of this thesis is onpresenting the management level adaptive mechanisms covered in all Papers I-V. PapersI, II and III present the importance of network monitoring in order to apply adaptive

23

Fig. 2. Adaptive traffic management levels and related papers.

management methods and assess the applicability of network technologies for certainapplications. Papers I, II and III also fall into data distribution and acquisition levelpresenting the edge router and road side unit mechanisms for data acquisition.

1.4 Outline of the dissertation

The dissertation is organized as follows. In the first chapter, the research problem andobjectives were stated including some background information. The first chapter createsthe basis for the contributions and the motivation of this thesis. In Chapter 2, workrelated to adaptive traffic management techniques is reviewed. Chapter 3 presents thekey technologies and methods relevant to this thesis, including the admission controlmechanism introduced in Paper III and fuzzy logic based management method appliedto heterogeneous network environments described in Papers IV and V. These methodsare applied to the IoT and ITS frameworks in Paper I and Paper II, respectively. Chapter

24

4 introduces the results presented in original Papers I-V for cognitive and adaptive trafficmanagement with discussions. Finally, Chapter 5 presents the conclusions of the thesis.

25

26

2 Adaptive traffic management framework andtechniques

The enormous growth, wide distribution and adaptation of information and communi-cation technologies (ICT) have heavily influenced the way consumers act, and howmanufacturers and operators have had to respond [13]. The new applications andservices have increasingly stringent requirements as the networks are becoming morecomplex. Networks are dynamic with precise QoS needs, even more heterogeneous withinteroperability requirements, less reliable with tight security demands and larger in size,which affects the system scalability. These changes and demands influence networksand systems, where different connectivity technologies, services and terminal equipmenthave to co-exist in a smooth and transparent way with the end users. The current Internetalready has some in-built autonomic features enabling recovery from different faults byutilizing methods to reroute traffic and roam between different network technologies.However, even more autonomic parts are required. The research efforts have focusedon studying autonomic and adaptive methods on different network elements that canadapt to various operational and contextual changes [14, 15]. The prerequisite is thatminimum external intervention is needed with the expectation that the systems arecapable of self-management. Large scale centrally managed and controlled systems areturning into distributed automated entities utilizing fog and edge computing [16–18].The decision-making processes are capable of reconfiguring the system and adapting tothe changed circumstances as well as learning from the previous operations.

Early Internet applications, like file transfer and email, were not bound by stringentperformance requirements and were well suited to the datagram delivery model basedon TCP/IP. However, the evolution of Internet with the growing amount of multimediaand other interactive and high-performance applications requires QoS guarantees.Various QoS methods have been developed to facilitate QoS provisioning including [19]congestion control, admission control, and traffic shaping and engineering.

Autonomic networking stems from autonomic computing [20, 21]. It aims tocontrol modern networks that are complex systems with heterogeneous nodes, linksand users operating over dynamic environments and contexts. In order to maintainperformance and service guarantees under changing circumstances, networks have toprovide mechanisms to adapt to these dynamic changes. Adaptive networking functions

27

can be implemented, e.g., with four basic mechanisms [22] such as constructinghop-by-hop connectivity, routing data, scheduling data transmission and controllingtransmission rate. In the next sections, the challenges and related work of consideringtraffic and resource management, QoS provisioning, admission control and schedulingare introduced.

In this chapter the related work is reviewed and the main functionalities for adaptivetraffic management and QoS control in heterogeneous networks are presented andsummarized. These principles work as a background for this thesis in researchingadaptive and cognitive methods for resource optimization and assessment.

2.1 Heterogeneous networking for the Internet of Things

The Internet of Things (IoT) presents an applied world of ICT to anytime, anyplaceand for anyone connectivity [2, 23]. Typical IoT architecture can be categorizedbroadly into four interconnected systems including ”things”, gateways/routers, networksand clouds. The connections are being multiplied creating an entirely new dynamicnetwork of networks. Heterogeneity becomes especially important in the IoT framework,which includes a number of sub-network deployments, where multiple heterogeneouswireless and wired communication solutions co-exist. A heterogeneous networkenvironment combines various different kinds of networking technologies and protocols.IoT architectures pose great demands on network control methods for the efficientmanagement of massive amounts of nodes and data.

The multiple access technologies include, e.g., cellular 2G/3G/4G technologies,wireless short-range technologies, e.g., Bluetooth technologies, ZigBee and Wi-Fi,which must all be effectively integrated to create a seamless interoperable platform. The3GPP technologies [24–26] are the dominant mobile network technologies, and thedevelopment of these systems has gone through several stages starting from the 2Gtechnologies (i.e., GSM/GPRS/EDGE) to 3G (i.e., UMTS/HSPA/HSPA+), and to 4G(i.e., LTE/LTE-Advanced/LTE-M [27]) and onwards to emerging standards for fifthgeneration wireless systems (5G). Local area networks (LAN) [28] based on IEEE 802.3Ethernet technology, are utilized in the case of fixed network access that supports datarates of 10 Mbps-10 Gbps with low latencies. The co-existence of different networktechnologies stems from different specializations and independently deployed systemsin different sub-domains. Managing these heterogeneous networking infrastructures

28

and the available resources, especially in dynamic environments, is a key technicalchallenge.

Wireless LAN (WLAN) and its numerous amendments can be used for short-rangecommunications in various environments providing hot spots for mobile and flexibleconnection to the Internet. The IEEE 802.11 WLAN [29] architecture supports ad-hocand infrastructure based operations. In ad-hoc mode the mobile nodes in a certain areacommunicate with one another in a peer-to-peer fashion, whereas the infrastructuremode supports centralized communication through the access point. The IEEE 802.11standard specifies both the MAC and physical layer. The MAC layer specifies twoprotocols, i.e., the Distributed Coordination Function (DCF) and Point CoordinationFunction (PCF). The DCF mechanism utilizes Carrier Sense Multiple Access withCollision Avoidance (CSMA/CA) protocol as the basic contention-based access scheme.PCF is only usable in an infrastructure based mode providing centralized channelpolling. Compared to wired communication systems, the wireless transmission mediumsuffers from unreliability due to the varying characteristics of the wireless air interface[30]. The typical impairments relates to temporal noise, multipath fading, multiuserinterference, obstacles on the link path and changing weather conditions. In WLAN,congestion typically arises when several nodes try to send at the same time due to ahidden node problem, in which the sender of one flow is out of the radio range of thesender of another flow resulting in throughput degradation, and hence unfairness, forsome of the flows.

The IEEE 802.11 architecture working at the 2.4 GHz band defines three modes ofsignal propagation namely, the FHSS (Frequency Hopping Spread Spectrum), DSSS(Direct Sequence Spread Spectrum) and the infrared system. The basic version of thestandard supports only 1 Mbps and 2 Mbps data rates, but it has been enhanced over theyears to support higher data rates with more advanced physical layer mechanisms. TheIEEE 802.11b amendment is backward compatible with the basic IEEE 802.11 schemeextending the data rates to 5.5 Mbps and 11 Mbps. IEEE 802.11a supports data rates upto 54 Mbps in the 5 GHz band; IEEE 802.11g supports data rates up to 54 Mbps in the2.4 GHz band; and IEEE 802.11n provides data rates of at least 100 Mbps in both the2.4 and 5 GHz bands using the Multiple Input Multiple Output (MIMO) technology.The IEEE 802.11e standard was introduced to improve the efficiency of IEEE 802.11with a new coordination function called Hybrid Coordination Function (HCF) and QoSfacilities. The HCF utilizes two access mechanisms, i.e., Enhanced Distributed ChannelAccess (EDCA) and HCF Controlled Channel Access (HCCA). The EDCA provides a

29

differentiated and distributed channel access method for different traffic priority classeswith admission control.

The existing IEEE 802.11 based WLAN solutions experience performance problemsin terms of providing QoS and bandwidth guarantees. For example, in IEEE 802.11a/b/gDCF-based access schemes, the medium access is given on a first come first served(FCFS) basis without any admission control mechanism. Performance degradationespecially occurs during heavy traffic loads [31]. The IEEE 802.11e scheme needsmore dynamicity for the scheduler operations to be able to tune bandwidth allocationsonline, and hence, support fairness [32]. When the network load increases, the backoffprocedure is facing problems that can lead to starvation of low priority flows [33]. TheQoS limitations related to misbehaving nodes should also be taken into consideration,where the EDCA parameters are used to manipulate low priority traffic to gain access tothe wireless medium using the high-priority parameters [34].

The IEEE 802.11 standards’ family also includes a MAC layer specification IEEE802.11p [35] allowing for the addition of wireless access in short-range vehicularenvironments (called WAVE in USA and ITS-G5 [36] in Europe). This includesmulti-channel data exchange between high-speed vehicles and between the vehicles andthe roadside infrastructure in the licensed ITS band of 5.9 GHz.

2.1.1 Internet of Things communication challenges

The IoT can be seen as an intersection between semantic-, internet- and things-orientedvisions [2]. The semantic oriented vision includes organizing, representing, storing,searching and interconnecting information. The IoT forms a greater Internet togetherwith the Internet of Energy (IoE), Internet of Media (IoM), Internet of Services (IoS) andthe Internet of People (IoP). Cisco [37] views the IoT as physical items, which togetherwith people, data and processes form a "network of networks" called the Internet ofEverything (IoE). The IoT challenges are very versatile concerning demands includingscalability, energy efficiency, intelligence, communication, integration, dependability,semantics, manufacturing and standards. The new techniques and concepts should alsobe easily integrated to enhance existing technologies. The system enablers especiallyfrom the networking perspective can provide the following technological characteristics[3]:

30

1) Reliability: Support for management of network mobility and heterogeneity inorder to guarantee reliable communications and system operations. Reliable energy-efficient communications must be configured to ensure dependability when billions ofheterogeneous devices are connected.

2) Scalability: The system should be robust, providing high performance andscalable algorithms and protocols capable of handling a varying number of devices inheterogeneous networks.

3) Security: It is necessary to provide security through the embedding and provi-sioning of new security solutions during the manufacturing of the device and when inoperation. The security solutions include the establishment of access control policies toheterogeneous networks and services as well as the development of processes for securesoftware development [38].

4) Intelligence: Efficient proliferation of intelligence is needed to save resourcesand energy including software and algorithms for distributed problem-solving anddecision-making to various management parts of the IoT systems.

5) Self-management: Systems with self-adaptive, self-configuration and self-healingfeatures targeted at systems to be more robust, intelligent and easier to manage.

6) Virtualization: Network virtualization techniques are among the importantenablers to ensure an evolutionary and modular path for the deployment of IoT applica-tions with assured QoS. Virtualization can perform various network functionalitiesfor hardware, device, network, cloud, or system. Network Function Virtualization(NFV) is an ETSI ISG initiative [39]. NFV is seen as highly complementary to, but notdependent on, Software Defined Networking (SDN). SDN can create virtual networksthat provide specific network services related to, e.g., admission and QoS control andresource management. Currently utilized methods for virtualizing different parts of thenetwork include, e.g., OpenFlow [40] for router and switch operations and OpenStack[41] for cloud virtualization functions.

Open standards are key enablers for IoT technologies and for any kind of Machine-to-Machine (M2M) [42] communication. Without globally recognized and interoperable

31

standards the expansion of the IoT and IoE (Internet of Everything) systems and servicescannot reach a global scale in many industrial sectors. In order to manage the largeamount of data coming from different data sources, a separate sensor middleware isoften required. Current IoT enabling standards include various middleware solutionsfor defining the requirements for a unique global identification, namely data fusion,scalability and interoperability to support all-IP based communications. For instance,the Sensor Web Enablement (SWE) initiative of the Open Geospatial Consortium(OGC) [43] standardizes web service interfaces and data encodings that can be usedas building blocks for a ”Sensor Web”. The suite of SWE standards enables sharing,finding and accessing networked sensors, transducers and sensor data repositories viathe web utilized for example in [44]. The current oneM2M standard [45] focuses onproviding an interoperable platform and technical specifications that can be readilyembedded with various hardware and software modules providing a common serviceand application development framework. 3GPP is standardizing LTE-M for M2Mapplications. LTE-M offers the benefits of wide spectrum and low cost cellular systemsfor M2M communications, which also enables a long battery life with enhanced coveragefor large numbers of devices [27].

2.1.2 Vehicular communications

Intelligent Transportation Systems (ITS) are an important part of the IoT framework,while an increasing amount of people are living in cities, using various ways ofcommuting and utilizing different Internet services. ITS and vehicle networking solutionsaim for traffic safety, fluency and informatics that require high quality connections forvehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communications, whichare known as a V2X hybrid solution. Vehicles can also communicate directly with theinfrastructure utilizing, e.g., IEEE 802.11p and/or cellular 3G/4G communications orthrough the Road Side Units (RSU) with their special on-board units (OBUs). RSUs actas an interface between the fixed and wireless networks including similar capabilities forintelligent traffic management and control as gateways, routers and other intermediatenodes. RSUs deliver up-to-date real-time service data to bypass vehicles but it canalso store and relay vehicle related data gathered through vehicle OBUs. Also, vehicleto vulnerable road user (V2VRU) communication is a rather new research topic withhigh performance requirements, although some solutions to detect VRU have beenproposed in the past years. VRUs include a wide range of road users, such as pedestrians,

32

Fig. 3. Heterogeneous wireless vehicle networking and communication types.

cyclists and powered two-wheelers. These different communication types for vehicularnetworking are presented in Figure 3.

The technologies operating at high-frequencies, like IEEE 802.11p, provide gooddelay performance but the operational range is not wide, and they are especially prone tolink failures. Therefore, the IEEE 802.11p requires careful network understanding andmeasuring in order to assess the suitability of cooperative ITS (C-ITS) applications withstringent timely demands and support for control algorithms. As vehicles and automotivesystems are moving towards higher automation levels and complex situations, therewill be limitations and needs for complementary communication systems with higherbandwidth, reliable range, latency, scalability and positioning accuracy. Therefore,other access methods are also required especially in rural areas, where IEEE 802.11pcoverage is difficult to attain. For instance, cellular 3G/4G/5G technologies can beutilized as alternative access methods in supporting seamless mobility and servicecontinuity. In particular, LTE with its support for mobile edge computing (MEC) [46] isa promising solution for vehicular networks, as the LTE performance according to rangeand throughput is larger than with IEEE 802.11p [47].

2.1.3 Vehicle to vulnerable road user (V2VRU) communicationrequirements

The current C-ITS messages have been specified for vehicle applications, and needmodifications to be able to support VRU applications. These applications include several

33

C-ITS systems, in which VRU and vehicle and/or infrastructure exchange informationwith each other using C-ITS messages. Safety applications either require periodicV2X data messaging, broadcasting based on CAM (Cooperative Awareness Message)[48], or event-triggered warning messages (e.g., traffic light status messages) based onDistributed Environmental Notification Message (DENM) [49]. The communicationbetween VRU and vehicle/infrastructure can be unidirectional (the VRU only broadcastsstatus messages, or the VRU can only receive warning messages), or bidirectional (theVRU both broadcasts status and warning messages and receives messages). Startingfrom the requirements for the different applications, a set of different communicationrequirements for C-ITS applications with VRUs are defined, which are related to: range,latency, scalability and position accuracy.

Range: One of the major benefits of cooperative safety services is the possibility todetect dangerous situations earlier than using on-board vehicle sensors. The possibleconflict between VRU and cars should be detected in time, in order to be able to warnroad users to take corrective action. The time of warning depends on the time to collision(TTC), which is composed of user reaction time, communication latency, the timeneeded to perform the maneuver, and a margin to take position inaccuracy into account.The phases for information, warning and intervention are system-specific. In order toavoid false warnings, collision avoidance warnings should only be issued to the driver atthe last moment to avoid the collision (ETSI TS 101 539-1). The range should hence besufficient to perform a risk assessment based on CAM information prior to issuing thewarning. As an initial guideline [50], a minimum range corresponding to a TTC of 5seconds is proposed. For oncoming traffic scenarios, this results in a range of about 100meters for pedestrian and cyclists in urban scenarios (car speed 50 km/h) and 160 metersin extra-urban scenarios (car speed 90 km/h).

Latency: Critical road safety and pre-crash applications require an estimated 300milliseconds (ms) end-to-end latency time as stated in ETSI TS 101 539-1 and ETSITS 101539-3. The end-to-end latency includes the communication network and thein-vehicle processing delays. The data transmission and receiving delays should be keptbelow 100 ms for time-critical systems with intervention. Safety systems aiming only atinforming and warning can work with slightly larger latencies. Non-critical applicationsfor comfort applications can cope with end-to-end latencies of 1-2 seconds.

34

Scalability: The system should work in an urban environment and perform well,e.g., in a use case where many road users are at the same intersection. This requiresthat both the communication network supports this as well as the developed services.This requirement can be challenging for communications systems based on cellularcommunications during heavy traffic.

Position Accuracy: For vehicle applications, ETSI TS 101 539-1 requires a positionaccuracy of 1 meter. For VRU applications the same requirement holds, however ahigher accuracy (0.5 meters) is desired, to be able to distinguish if the VRU is in a safearea (on sidewalk) or not (on the road). Additional sensor information, such as thatfrom vehicle sensors, is needed to increase the absolute position accuracy or the use ofsystems for relative positioning (e.g., radar, camera, time-of-flight, angle-of-arrival).High-accuracy maps that match the location accuracy level are required, which includeVRU specific elements (e.g. crosswalks).

Through an analysis of the different solutions, four different ways of V2VRUcommunication can be distinguished. The first way is through the use of a tag by theVRU, transmitting only limited information, such as ID, and a reader in the vehicle,which localizes the VRU based on this signal. Several technologies, such as RFID(Radio-frequency Identification), radar systems, IEEE 802.15.4 and IEEE802.15.3aUWB (Ultra-wideband) have been evaluated for positioning and identification of theobject. The ADOSE project developed and assessed a system for VRU detectionbased on harmonic radar and passive transponders [51]. SafeWay2School developeda RFID based VRU unit for children that consisted of a standalone radio unit ableto communicate with intelligent bus stops, which warn drivers with flashing lightsabout the vicinity of VRUs [52]. The second approach is to use smartphones for VRUdetection, through exploiting the Bluetooth, Bluetooth Low Energy (BLE), Wi-Fi orcellular signals transmitted periodically by the phone. A receiver in the car should beable to position the VRU. The use of Wi-Fi Direct for pedestrian detection has beendemonstrated by General Motors [53]. The main challenges are large latencies (in theorder of seconds) due to the detection process, and the positioning of the smartphonebased on the radio signal. In the third approach, applications on a smartphone, whichalso transmit location data (possibly as standard cooperative ITS (C-ITS) messages) ashas been described by Engel et al. [54] and Liebner et al. [55] and Borroni-Bird [56] areused. According to Borroni-Bird, future phones can include ITS-G5 without additional

35

hardware. Challenges regarding technical performance include the automatic activationof the C-ITS application, battery management and interaction with other applicationsand position accuracy on the smartphone. The fourth solution is to use a separate device,transmitting location and other sensor data. The messages should be C-ITS compliantto communicate with standard in-vehicle and on-road C-ITS equipment. The use ofC-ITS devices on motorcycles has been demonstrated in the DRIVE-C2X project [57].Challenges are the miniaturization and battery management of the devices, especiallyfor pedestrians and cyclists.

2.2 Traffic management

A computer network is composed, e.g., of links with varying bandwidths, and routerswith varying buffer sizes that all have to be shared with different applications and users.Packet delays and losses occur if the network cannot handle all the traffic that is offeredto the system or due to variations in the traffic or in link quality. A network that supportsQoS should actively manage and monitor traffic and coordinate resource allocations.The network elements need to implement the following features: packet identificationand classification, traffic management and queueing, policing, and administration ofQoS policies and management.

Without appropriate traffic engineering and resource allocation mechanisms, thenetwork performance and service quality deteriorates under heavy traffic changes due todropped packets and congestion. Congestion can first be detected by utilizing networkperformance measurement and monitoring techniques. For congestion handling thereare two main approaches: congestion control and congestion prevention/avoidance.Congestion control is a reactive method and is activated when the network is overloaded.The intermediate nodes participate in congestion control so that a fast sender or manysenders cannot block the paths faster than a network can handle. Congestion preventionis a proactive approach hindering the network for becoming overloaded including manydifferent methods at application, transport, network and data link layers, such as packetsize optimization, rate control, routing algorithms, and admission control.

Congestion control involves the design issues to limit the offered traffic to match thecapacity constraints of the system. Especially for real-time traffic, it is important tounderstand how congestion arises and find efficient ways to keep the network operatingwithin its capacity. The basic design issues of congestion control are what providesfeedback to sources and decides how to react to the feedback. The network endpoints,

36

i.e., the source and destination, do not usually have the details of congestion point(s)and reason(s). Thus, the application-based adaptation works only when the networkendpoints support adaptation mechanisms, e.g., for adjusting the required data ratechanges. Intermediate nodes can use network layer techniques like ICMP (InternetControl Message Protocol) to inform hosts that congestion has occurred. However, theICMP feedback from routers can suffer from problems related to accuracy and reliability[58].

The fundamental TCP method did not include congestion control or avoidanceschemes and solely used Go-Back-N sliding window for end-to-end flow control. CurrentTCP implementation utilizes congestion control [59] and avoidance [60] methods thatadjust the source nodes transmission rates. The drawback of this is that the transmissionrate is decreased only after the detection of datagrams losses, which causes a time delay(due to round trip time, RTT) and results in buffer overflows in routers and further lossesof datagrams. Hence, the flow and congestion control schemes of TCP are not sufficientin terms of the network performance and overall service quality.

On the other hand, the real-time flows with stringent delay requirements make use ofUDP (User Datagram Protocol), which does not provide any mechanism to regulate theamount of data being transmitted. UDP does not return acknowledgements and cannotsignal congestion to the sender. The inability of UDP flows to regulate transmission rateat the transport layer makes it especially vulnerable to congestion. Therefore, for theUDP sessions, applications have to provide some form of integrated flow control.

2.2.1 Queueing and scheduling mechanisms

Queueing and scheduling algorithms participate in network resource sharing andallocation. Efficient queueing mechanisms allow the traffic to be split into multiplequeues after which the scheduler will decide the service order of the packets as shownin Figure 4. A packet scheduler has an important role in dequeueing the packets andkeeping track of the network resources. If there is a situation in which network resourcescannot serve all flows, queues will start to build up in the routers. Queueing methods arecommonly categorized as work-conserving and non-work-conserving ones [61]. Mostof the well-known schedulers are work-conserving and the principle is that they alwaysschedule packets when there are some waiting for a service. A non-work-conservingdiscipline proposes to reduce timely factors, like delay and jitter, by only schedulingpackets that are considered to be eligible.

37

Fig. 4. QoS traffic flowing and scheduling.

The most popular queueing algorithm is First-In-First-Out (FIFO), which determinesthe service order of packets based on their arrival order. In Priority Queueing (PQ)[62–64], traffic classes with the highest priority are forwarded with the least delay; inthis case the low priority packets are prone to starvation, particularly when there is asteady flow of high-priority packets. Class based queueing (CBQ) [65] provides anequal share of the bandwidth for each class. Round Robin (RR) algorithms processpackets in turn with an equal share and achieve very high accuracy and fairness in theoutput bandwidth sharing but cannot provide tight delay prioritization. These problemswere solved with Fair Queueing (FQ) techniques devised by Demers et al. in [66] ofwhich the WFQ is one of the most popular variations of FQ. A WFQ scheduler [67]can allocate resources or bandwidth to different flows by defining a weight parameterto each session flow. A WFQ scheduler selects packets from multiple queues basedon their arrival time, size and associated weight. A parameter called virtual time iscalculated for a packet every time it enters a router. Packets are inserted in a servicequeue sorted by the virtual times. The packet with the smallest virtual finishing time isscheduled first. This type of approach enables the sharing of resources between trafficaggregates in a fair and predictable way. Considering that B is the total throughput of anoutput link, and if all sessions of the WFQ scheduler are active, then each service classreceives a portion of the total bandwidth determined by its weight wi, which is equal towiB. The weights of different service classes should fulfill the following constraint:

m

∑i=1

wi = 1, 0.01≤ wi ≤ 1 (1)

In this equation m denotes the total number of service classes. Each weight wi indicatesthe portion of bandwidth that the service class will attain. For instance, if the weight w1

is 0.3, 30% of the available bandwidth will be scheduled to this service class and w2

38

attains (in this case) the rest (1-0.3=0.7) 70% of the bandwidth. In order to guaranteeminimum fairness to all service classes, the weight cannot be zero as can be seen in (1),in which the minimum attainable weight is 0.01.

However, conventional scheduling and queueing methods provide a rather weak formof resource reservation, e.g., in weighted bandwidth schemes, in which the weights areonly indirectly related to the bandwidth the flow receives. In addition in heterogeneousnetworking systems, scheduling and coordination of the resources is complex due tothe various QoS requirements of the different links and flows. In particular, allocatingresources fairly between different applications in a wireless environment is not atrivial task, where a higher amount of channel errors are present. Another problem ofconventional scheduling methods is that they are quite static in their operations. Thelatest development is directed at the dynamic adaptation of scheduling parameters,which gives better overall performance [68–73]. The adaptive approach to WFQ in [69]is a variation of the fair queue algorithm with dynamic priority scheduling. [70] presentsan adaptive approach to WFQ that uses a concept of revenue to adapt weights. This islater extended to a comparison and an analysis of several adaptive scheduling algorithms:Revenue-based Adaptive WFQ (RA-WFQ), Revenue-based Adaptive Weighted RoundRobin (RA-WRR) and Revenue-based Adaptive Deficit Round Robin (RA-DRR) [71].The Dynamic Weighted Fair Queueing (DWFQ) algorithm in [68] presents a new typeof fast packet scheduling algorithm. The adaptive WFQ algorithm is applied to IEEE802.16 (WiMAX) networks utilizing fuzzy logic in [72]. A rate-adaptive version ofWFQ algorithm with energy-aware scheduling is utilized in [73].

2.2.2 Network information management

Since, efficient QoS provisioning includes various mechanisms from the whole TCP/IPstack, cross-layer [74] solutions that collect information from multiple entities and layersare found to be useful. Especially in responding to common challenges in multimediadelivery and improving the user experience over the wireless Internet, QoS can beintroduced in multiple layers [75]. The mobility management mechanisms residingat the mobile device or at the network side are finding suitable ways to manage QoSbetween these entities.

To effectively manage the changes in resource adaptation, there needs to be a real-time monitoring mechanism for the changing network environment [19] and information.Network measurements can be divided in to active and passive measurements [76].

39

Active measurements are based on discovering the network performance by sendingspecial probe packets on network paths, and recording and analyzing the response.Passive measurements monitor the traffic at one or multiple points without interveningin the traffic itself. A single-point measurement can be utilized for monitoring theperformance of Local Area Networks (LANs) that are connected to larger networks.However, the variety of monitored performance metrics is rather limited when there is nopossibility of measuring time dependent values. Generally, throughput and distributionof different traffic types can be monitored with the single-point measurement. Withmulti-point measurement time-related parameters, such as packet loss, delay, and delayjitter, can also be measured, requiring that measurement point clocks are synchronized.Hence, the measurement accuracy increases but having multiple measurement pointsalso increases the complexity and amount of possible error points. The application QoSand overall quality of a certain link can be monitored well when setting the measurementpoints at traffic end-points. However, when having various heterogeneous links it can bedemanding to determine the actual traffic bottlenecks causing network congestion.

The Qosmet tool, developed by VTT [77], enables passive bi-directional mea-surements at multiple points of the traffic path for QoS metrics, such as delay, delayvariation (jitter), packet loss, throughput and number of sent/received packets. It alsoenables the estimation of the QoE with application-specific modes that are basedon real-time QoS measurements. Example models suitable for this are the (ITU-T)E-model [78] and models that perform pseudo-subjective analysis of the quality, suchas PSQA (Pseudo-Subjective Quality Assessment) [79]. Many QoE models providequantitative evaluation of the perceived mapped quality, for example, in the MeanOpinion Score (MOS) range from 1 to 5 where the numbers present a verbal counterpartof the perceived quality. We have used the Absolute Category Rating (ACR) where5 stands for ”excellent”, and 1 for ”unusable” quality. Typically, value 3 presents athreshold value, where the quality is on the average fair but impairments are alreadyslightly annoying and it is not suitable for long time use.

Formerly, the applications were easily associated with certain transport layer portsby Internet Assigned Numbers Authority (IANA) [80]. Nowadays, modern services likePeer-to-Peer (P2P) applications, utilize non-standard ports that makes the port-basedclassification unreliable [81]. To defeat this problem payload-based classifiers havebeen proposed that look into packet payloads for possible application signatures orother properties of packet content like in Deep Packet Inspection (DPI) [82], [83].Although having many advantages, DPI appears to be a computationally costly method

40

and cannot be utilized with encrypted traffic flows. Statistical classification mechanismsare considered to be more lightweight, while exploiting several flow-level measurementsinstead of accessing the packet payload. A two-phased traffic classification tooldeveloped by Hirvonen and Laulajainen [84] performs a robust traffic classificationutilizing statistical characteristics of the flows. Different service types are first trainedfor the classifier. In the first phase packet payload sizes and the direction of the trafficflow (uni- or bi-directional) are identified, the second phase considers more trafficcharacteristics such as average packet size, average packet inter-arrival time, and thepacket ratio between uplink and downlink. The K-means clustering method [85] isutilized for the clustering of traffic flows. The traffic classifier captures all packets, andnew flows endeavoring the network will go through the classification procedure.

2.2.3 Fuzzy logic based traffic management

Fuzzy logic has been widely applied to control theoretical problems, enabling the imple-mentation of advanced knowledge-based computing strategies for complex dynamicsystems. Control theory provides a wide amount of methodologies that are profitable toautonomic and adaptive network management systems, where the early interests havebeen applied, e.g., to flow control queueing theory [86], TCP’s operational optimization,and in multiple-input multiple-output (MIMO) control [87]. Control theory provides aframework to analyze and design closed loop systems [88] based on the properties ofstability, accuracy, settling time, rise time, undershoot and overshoot.

The control loops in systems’ management as described by Len Fehskens in 1989[89] already included the core parts namely for control, monitoring, policing andapplying knowledge. IBM introduced the autonomic computing principles in 2001which had strong comparisons to biological systems that also use control loops [20].Control loop monitors the state of the managed resources, analyzes the knowledge,e.g., about when and where the network congestion occurred, and acts according to thecontrol policies by adjusting the system accordingly. Autonomic computing leads toself-managed, i.e., self-configuration, self-healing, self-optimization and self-adaptivefunctionalities.

A control system is accurate if the measured output converges to the referenceinput [88], while ensuring that, e.g., throughput, delay or packet loss is maximizedwithout exceeding response time constraints. For dynamic systems the output may notconverge to a certain value but rather utilize operating regions, under which system

41

operation is acceptable. A control system has short settling times if it converges quicklyto its steady-state value, which is often important if there are time-varying workloads.Accurate systems are essential for guaranteeing that the control objectives are fulfilled,for instance differentiating between prioritized and best-effort service classes. Thesystem should also achieve its control objectives without undershoots or overshoots.

Lotfi Zadeh initially proposed fuzzy set theory [90]. Fuzzy logic was later developedfrom fuzzy set theory to present systems with uncertain information. Communicationsystems, and especially wireless systems, have nonlinear traffic characteristics thatcomplicate the designing of precise mathematical models due to transmission mediumuncertainties. The fuzzy control approach is very useful for systems where it is difficultto precisely quantify information.

In fuzzy control applications, a rule base (knowledge base) includes a control policythat is usually presented with linguistic conditional statements, i.e., if-then rules thatcan be converted into matrix equations. Fuzzy reasoning can be done either usingcomposition-based or individual-based inference. In the former all rules are combinedinto an explicit relation and then fired with fuzzy input whereas in the latter each of therules is individually fired with crisp input and then combined into one overall fuzzyset. In this thesis an individual based inference with Mamdani’s implication [91] wasused. The reason for the choice was its easier implementation because the results areequivalent for both methods with Mamdani’s implication.

2.3 QoS control

The Internet Engineering Task Force (IETF), the International TelecommunicationUnion - Telecommunication Standardization Sector (ITU-T) and the 3rd GenerationPartnership Project (3GPP) all have their own definitions for QoS. IETF determinesQoS [92] as follows: ”a set of service requirements to be met by the network whiletransporting a flow”. ITU-T approach [93] takes the user perceived quality aspectQuality of Experience (QoE) more into account by defining QoS as ”the collectiveeffect of service performance which determine the degree of satisfaction of a user of theservice”. The 3GPP defines the QoS concept and architecture for cellular systems.

The overprovisioning of network resources is not always possible or even feasiblein heterogeneous networks. Therefore, service differentiation becomes an essentialpart for many QoS-based methods. QoS denotes the properties of a system, affectingthe perceived quality of the service [94]. Service requirements are usually application

42

related, for example, real-time applications for multimedia, gaming and interactivecommunication are sensitive to round-trip delay; whereas, bulk file transfer is moredependent on average transmission rate. QoS metrics are typically expressed withparameters such as delay, jitter (variation of delay), throughput and packet loss rate [95].Generally, QoS mechanisms enable network operators to assign a certain order to howthe network traffic is treated and the capacity allocated. The network should providemultiple QoS strategies to support different services and accommodate the impact ofQoS metrics [32, 96]. In this thesis, the QoS metrics mainly focus on delay, throughput,packet loss rate and balanced bandwidth.

2.3.1 Resource allocation

The early QoS mechanisms included methods to manage the overall bandwidth andallocate network resources. Two broad categories, namely fine-grained and coarse-grained approaches, have been developed to provide a range of qualities of service.Fine-grained approaches provide QoS to individual applications or flows, whereascoarsegrained approaches provide QoS to large classes of data or aggregated traffic.Integrated Services (IntServ), developed by the IETF, and often associated with RSVP(Resource ReSerVation Protocol) [97] is an example of a fine-grained approach. TheIntServ architecture allocates resources to individual flows specified by a stream ofpackets with a common source address, destination address and port number. In IntServ,a packet scheduler is used to ensure the resource allocations with defining specificationsfor a number of service classes, such as guaranteed service and controlled load, aredesigned to meet the needs of certain application types. RSVP makes reservationsusing these service classes, e.g., guaranteed service class is designed for intolerantapplications, which require timely packet delivery, whereas controlled load service classis designed towards more tolerant applications, if the network is not too heavily loaded.Controlled load service emulates a lightly loaded network, even though the network maybe heavily loaded on a wider scale. Widely used mechanisms for isolating the controlledload traffic from other traffic are the use of a queueing mechanism, such as weighted fairqueueing (WFQ) [98] that is enhanced in this thesis.

Another widely used QoS mechanism Differentiated Services (DiffServ) can becategorized as a coarse-grained approach. Packets entering a router are marked as QoSservice classes to receive particular per-hop forwarding behavior. The QoS class of apacket is indicated by the DiffServ field of the IP header (Type of Service (TOS) octet

43

in IPv4 and Traffic Class octet in IPv6). According to DiffServ Code Points (DSCP)each router can mark, shape or drop the incoming packets. Many proposed DiffServapproaches simply divide traffic into two classes, for instance Expedited Forwarding(EF) and Assured Forwarding (AF) for selected IP packets. The mapping depends onbandwidth and provisioning of resources among the different DiffServ classes thatare controlled by the operators in order to satisfy their performance requirements andinteroperability with other operators based on Service Level Agreements (SLAs) andService Level Specifications (SLS). The simpler QoS mechanisms based on prioritization,like Differentiated Services (DiffServ) in the core network side, can provide QoS forcertain types of traffic flows. However, DiffServ can provide only statistic QoS, notguaranteed QoS, and when there too much high-priority traffic appears, the performanceof all users will collapse.

Other flow-based QoS schemes include the flow-aware networking (FAN) schemeproposed by France Telecom [99]. The Flow-state-aware (FSA) transport methodappeared as an ITU-T recommendation for Next Generation Networking (NGN) in 2008[100]. The FAN scheme blocks new flows to protect ongoing flows. Therefore, theend-to-end performance is being monitored at all times but actual admission controlis initiated when congestion is perceived. Two types of flows are defined: elastic(best-effort traffic) and streaming (delay-sensitive UDP traffic). The FSA method differsfrom FAN in a way that per-aggregate resource guarantees are provided alongsideper-flow QoS.

2.3.2 Network admission control

Load balancing with admission control is important when paving the way towardsinteroperable heterogeneous networking environments. Load balancing methods basedon IP addresses, usernames or other former methods are challenging and not scalabledue to the changes in the client’s perceived address resulting from Dynamic HostConfiguration Protocol (DHCP), Network Address Translation (NAT) and web proxies.On the other hand, flow-based configuration on switches, routers and gateways, createspossibilities for a wider realization of intelligent load balancing, traffic managementand Network Function Virtualization (NFV). Aiming at better resource utilization, thenetwork can be enhanced with some degree of adaptable admission control (AC) to itsflows.

44

Admission control limits the number of incoming flows in order to guarantee QoS foreach flow. When defining an admission control strategy, a tradeoff between the servicelevel and the network control overhead needs to be established [101]. Moreover, the typeand number of intermediate network nodes and entities involved in and controlled by anyAC algorithm affect the solution complexity. And fatally, a higher system complexityleads to lower scalability. Apart from that, controlling the admission of flows thatshare a service class in a network should follow several considerations, i.e., (i) avoidover-allocation of network resources, (ii) avoid new flows from impairing those alreadyaccepted, (iii) fulfill service level agreements and specifications (SLA/SLS) and, (iv)prevent instability and assure existing QoS [102].

Typically, AC methods can be coarsely classified as proactive (parameter-based)[103], [104] or reactive (measurement-based) [103], [105] ones. In parameter-basedadmission control (PBAC) schemes, the flow admittance or denial is based on someanalytical assessment or estimates of flow characteristics such as peak rate or bandwidthusage. Typically, many AC methods are coupled with DiffServ, which offers thescalability advantage of manipulating traffic aggregates, but cannot alone resolve thecongestion problem, as it does not have any control over the traffic load entering thenetwork. PBAC’s main drawback is that early flow characterization cannot always beaccurate, while the traffic models assumed are usually simplified and cannot capturethe flow dynamics, as in the case of a compressed video. Additionally, informationabout the system capabilities and, e.g., the probability of buffer overflow is required,so the relationship between the traffic load and the queue length distribution has tobe derived, which is not a trivial task. The good side of PBAC schemes is that theycan be analyzed and compared using analytical methods, whereas Measurement-basedAdmission Control (MBAC) can only be quantitatively compared through measurementson simulated or real networks.

For a long time, MBAC has been used to enhance the DiffServ architecture. Initially,measurements were performed at all nodes. Later, more centralized architecturesutilizing, e.g., bandwidth brokers (BB) [106] were proposed. A BB is a middlewareservice that controls and facilitates dynamic access to the shared resources of a particularadministrative domain. Independently of any QoS architecture at the access routers, aBB acts as a central agent, which allocates resources like bandwidth share. However, thetrend in MBAC schemes has gone in the direction of utilizing either passive [107], [108]or active measurements [109] to collect the feedback at the network edges. MBACtypically measures the actual traffic load, to dynamically adjust the AC parameters.

45

This is why the MBAC controllers are more robust with respect to the accuracy ofthe traffic model, although they are compromised by the statistical variation of themeasurements. When designing an MBAC scheme, the challenge is to calculate theacceptance region in a way that the network can tolerate traffic bursts. The peak rateshave to be known or assumed; otherwise token bucket filters are necessary. It is thenusual to identify maximal target utilization for resource reservation, whose apparenttrade-off is the under-utilization of the capacity. Another disadvantage is that it maygive an excellent performance under one scenario, but an inadequate performance inanother, since MBAC needs to be tuned depending on the network settings and the trafficscenario. To mitigate this problem a new MBAC solution is proposed by including aKnowledge Plane to the measurement algorithm [110], to maintain a broad view of thelink behavior, and predict the expected QoS to admit the flows to the network. Fairnessis an issue not considered among typical AC methods when traffic is managed intoaggregates. Schemes aiming for fairness in resource control within the traffic aggregate,e.g., [111] apply queue control functions and additional QoS router’s state calculation,which have to be controlled carefully not to increase the overhead considerably. In thework around fairness issue [112], an algorithm is developed that guarantees both fairoccupancy and optimal usage of the resources.

2.4 Summary

This chapter described the principles for adaptive traffic management techniques utilizedin this thesis. First, heterogeneous networking for the Internet of Things was presentedin Section 2.1 and various common technologies introduced. The IoT is a wide researchtopic, which includes different communication challenges that were also discussed inSection 2.1.1. Vehicular communications is one vastly growing research area relatedto the IoT that was described in Section 2.1.2. In addition, principles for vehicle tovulnerable road user (V2VRU) communication use case were introduced in Section2.1.3.

This chapter also discussed principles for traffic management in Section 2.2 includingqueueing and scheduling mechanisms in Section 2.2.1, network information managementin Section 2.2.2, and fuzzy logic based traffic management in Section 2.2.3. QoScontrol was introduced in Section 2.3 including resource allocation in Section 2.3.1 andadmission control methods in Section 2.3.2.

46

3 Contributions of the thesis to adaptivetraffic management

This chapter presents the solutions developed for adaptive traffic management inheterogeneous communication networks. It provides the contributions of original PapersI-V in detail and addresses the research question and assumption presented in Section1.2 along with references to the literature in Chapter 2. This chapter begins with adescription of the overall architecture and enhancements developed for adaptive trafficmanagement in Section 3.1. Section 3.2 describes the edge computing features proposedin this thesis. Section 3.3 presents the algorithm developed for the regressive admissioncontrol scheme. Similarly, Section 3.4 presents the model and principles for adaptivefuzzy control in flow scheduling. Section 3.5 provides a summary of the methodsdeveloped for adaptive traffic management.

3.1 Functional requirements of the system architecture

Cognitive traffic management applies the available information about the networkconditions to plan, decide, and act to improve the network QoS [113] in an adaptive way.This thesis proposes QoS-aware adaptive traffic management methods, i.e., cognitiveflow management schemes, based on regressive admission control (REAC) and fuzzyscheduling working on the network edges. Typically, when providing per-flow QoSguarantees, each flow’s traffic is separated in one queue. In per-flow QoS, the schedulerhas to know the available flow requirements. This is problematic in cases where thenumber of flows is great without providing any traffic aggregation, as a router has tohave a lot of resources to support classifier entries with separated queues for each flow.Also the existence of great amount of queues complicates the finding of an appropriatequeue and sharing the bandwidth. This thesis provides per-flow QoS support, whereevery flow is identified and classified, and the traffic is treated and monitored accordingto traffic aggregates. All the flows are admitted to the network, and the new flowsendeavoring the network are treated differently if they are ruining or likely to ruinthe performance. The task of the fuzzy scheduler is to provide a certain minimumamount of output bandwidth for each queue and give more priority to real-time UDPapplications. This kind of scheme controls network congestion and also prevents it by

47

Fig. 5. System architecture for adaptive QoS-aware networking in heterogeneous environ-ment.

making predictions about traffic behavior and reacting accordingly in a heterogeneousnetworking environment.

The overall system architecture is shown in Figure 5 including various elementsfrom different research projects. The architecture facilitates the overall framework fordifferent parts of the adaptive resource management:

- Application server delivers the up-to-date road information and C-ITS warningmessages.

- Core router runs the QoS multi-point monitoring part for passive end-to-end delay,packet loss and throughput measurements.

- Edge router runs QoS monitoring, traffic classifier (TC) part and the REAC controllerapplying AC intelligence.

- Fuzzy Controller applies the adaptive queueing and assignment of bandwidth weightsfor prioritized and best-effort service classes. The bandwidth weight tuning isbased on the prevailing traffic level to reach application dependent packet loss rateand end-to-end delay limits to stabilize the throughput.

- Access network is utilizing the IEEE 802.11p (ITS-G5) technology for short rangecommunication between vehicles, VRUs and RSUs.

This architectural model presents a high-level framework for information distributionand utilization regarding the status of mobile end users, available networks, bandwidth

48

and congestion, and prioritized users. Each of the Papers I-V in this thesis includes apart of the overall architecture that participates in defining the capacity and networkrequirements in order to be answered by adaptive and cognitive methods. Table 1shows the implemented functionalities in this thesis, and which part of the architecturalframework each relates to. More detailed explanations are provided in the followingsections.

Table 1. Implemented parts utilized in the adaptive network control (mapped to Figure 5).

Implemented functionality Part of the architectural framework

Network monitoring QoS-aware network paths (Papers I, II, III)

Assessing and optimizing the IEEE 802.11pperformance

Access network (Paper II)

Regressive admission control (REAC) Core and edge routers (Paper III)

Adaptive weighted queueing (AWFQ) for fixednetworks

Fuzzy controller (Paper IV)

Fuzzy weighted queueing (FWQ) for wirelessnetworks

Fuzzy controller (Paper V)

3.2 Computing at the network edge

Cloud computing is tightly connected to the IoT systems and greatly depending onservers and their computing capabilities. These server farms are available in a remotelocation, e.g., in a data center, which can result in slow response times and also scalabilityissues, especially when concerning delay-sensitive applications. However, efficientresponse times including system stability, accuracy and settling time issues alongwith system scalability take an important role when designing the IoT communicationsystems and services.

The terms edge computing and fog computing are often used interchangeably.The basic characteristics of fog computing are similar to cloud computing with datacomputing and storage but concentrating on features like mobility, proximity to end-users, low latency, location awareness, heterogeneity suitable for services, e.g., withconnected vehicles, smart grids and cities. In a typical fog computing model, thecloud retains its central role of analyzing data and orchestrating the operations andmanagement. When there are no resource constraints and the connectivity amongmultiple data sources is adequate, pure centralized cloud computing makes more sense.

49

Fig. 6. Overall system description. c© [2016] IEEE

However, the cloud can also delegate some tasks to the smart edge devices in order tolocalize part of the data analysis and decision-making. Typically, the mission of theintelligent edge devices is not to carry out in-depth data analysis, but to actively filterlocal data and selectively relay data to the cloud. The success of fog computing relies onthe ability of these intelligent edge solutions to speed up the deployment, cost-effectivescalability and ease of management with limited resources.

According to a recent survey on fog computing [114], the potential important issuesthat have to be considered include fog networking, QoS, interfacing and programmingmodel, computation offloading and load balancing, accounting, billing and monitoring,provisioning, resource management, and security and privacy. In this thesis, PapersI-V all consider possible solutions for fog networking, QoS, monitoring and resourcemanagement. The adaptive traffic management system in this thesis proposes computingfeatures shown in Figure 6, where the system description includes parts for admissioncontrol, flow monitoring and fuzzy scheduling. The functional steps from one to fiveinclude the following operations:

1) Network traffic is monitored. In addition, all flows are registered in the controllercomponent (data base) enabling the specification of the flows such as QoS requirements,scheduling and packet information. The traffic monitoring component is tested in variousheterogeneous environments. The latest work relates to real-time vehicle scenarios

50

utilizing ITS-G5 (IEEE 802.11p) presented in Papers I and II and cellular 3G/LTEnetworks [47] to assess the suitability of these technologies for time-sensitive vehicularcommunications.2) Flows are being identified and classified according to the method described in Section2.2.2. A flow entry contains information such as source/destination IP address/port, IPaddress of next hop, and the new destination IP address and status information whetherthe flow is active or inactive after a certain period of time.3) The controller exploits the REAC algorithm presented in Paper III to make the QoSlevel estimation and network admission decision for the flows. The REAC controllerutilizes the information conveyed by Qosmet and the traffic classifier to implement theadmission control logic. The purpose of the logic is to monitor the QoS-level variationby measuring the delay, and to estimate when the QoS-level decrease affects the qualityof the high-priority applications.4) The system first traces the suspect flow(s), and then either drops packets of these flows,or decreases or increases the flow priority using, e.g., DSCP markings and schedulingcapabilities depending on the network support for certain SLA/SLS.5) Flows are being scheduled and bandwidth weight assigned according to the adaptiveAWFQ and FWQ algorithms in Papers IV and V, respectively. The developed modelsprovide fast response times to different traffic states with application dependent delaytimes and packet loss rates.

3.3 Regressive admission control (REAC)

The REAC controlling method presented in Paper III is a cognitive flow managementsystem combining network monitoring, traffic identification, congestion control, resourcehandling, QoS, and intelligent decision-making. The aim is to limit the newest low-priority traffic flows in the network path in a way that the offered traffic load staysbelow the bounds that the network can handle. In this way, the users already enjoying aservice in the network will have QoS, and will be satisfied. If new traffic flows enteringinto the network are likely to exceed the network capacity, they are treated differentlythan the flows already in the network. Normally, these extra flows would often ruinthe performance of all the network users. When a flow is blocked its admission timeusually increases, but on the other hand, once it is accepted it receives very stable andreliable service from the network. Thus, what happens is that REAC does not even try toperform a complex end-to-end resource allocation, but instead the edge routers make the

51

(a) Traditional MBAC and PBAC decision.

(b) REAC decision.

Fig. 7. Admission control decision triggering.

decision for the flows individually. Restricting the number of flows, or even knowing themaximum number of flows, is not a trivial task in a packet switched network. This iswhere REAC shows its best advantage; it works in a regressive way. Regression is acommonly used statistical method to describe how the mean of the dependent variablechanges with changing conditions [115].

The REAC module accepts flows to the network a priori, which presents a newapproach among traditional MBAC and PBAC admission control methods. The concep-tual difference between REAC and traditional AC is shown in Figure 7. TraditionalAC enforces an admission decision upon each client arrival, as depicted in Figure 7a.This implies that the traditional methods perform an end-to-end negotiation for the ACdecision, also adding some extra delay. Figure 7b shows that the REAC decision is nottied to the arrival event. The AC logic monitors the QoS-level variations, and estimateswhen the QoS-level decrease affects the quality of the high- priority applications formaking the REAC decision as shown in 7b. For this, a continuous real-time QoSmonitoring is enabled in the network path between the end points. The monitoring ispassive, so it gives clear measures of how the application traffic is really performingover the network path. Another advantage of passive monitoring is that the controloverhead to the network is minimized. Whenever the measurement indicates that the

52

quality drops below a given threshold, a decision-making process is initiated to drop ormark the excessive low-priority flows. The REAC method focuses on the temporalvariation of the mean of the dependent variable, which is delay in our case.

3.3.1 REAC architecture

The REAC architecture consists of modules shown in Figure 8 namely flow monitoring,which consists of a traffic classifier and a QoS measurement modules, and an admissioncontrol module with a REAC controller.

Fig. 8. The REAC system’s architecture. c© [2016] IEEE

The traffic classifier is based on a two-phased traffic classification tool as presentedin Section 2.2.2. The traffic classifier provides information regarding the flow identityand status (active or inactive) updated to a flow repository. The flow repository update isdone in a way that for a new flow a registry is added, the terminated flow registries areremoved, and the rest gets a status and an age update. The Qosmet measurement tool isfully controllable remotely by the REAC module via its QMCP (QoS MeasurementControl Protocol) interface. Qosmet performs the QoS measurements between the

53

controlled-network path endpoints averaging the results over aggregated flows inrequested intervals. The QMCP message processor parses the measurement results andupdates the QoS repository. The repository holds sets of QoS values and updates themin a rolling manner, i.e., every new set of values replaces the oldest one in the repository.The REAC controller exploits the information conveyed by Qosmet and traffic classifierto implement the AC logic.

The purpose of the AC logic is to monitor the QoS-level variation and to estimatewhen the QoS-level decrease affects the quality of the high-priority applications. Uponsuch an indication, the system first traces the suspect flow(s), and then either dropspackets of these flows, or decreases the flow priority using, e.g., DSCP marks. Alternativescheduling methods, such as AWFQ and FWQ presented in the next section, to treat theflows may be also included, depending on the capabilities of the controlled networkpath.

REAC does not assume knowledge of the controlled network path capabilities, butinstead, it tracks one of the most important QoS metrics: delay. The focus is on thetemporal variation of the delay, averaged over the last updating interval also called ameasurement window. There is a strong relation between delay characteristics, networkutilization [116] and application dependence [117] when the network reaches the pointof congestion. Delay is a metric, which reflects the status of the network well: steadyand low delay is an indication of good network conditions. Furthermore, an increasingdelay can be an indication of upcoming network congestion, thus allowing trafficpredictions to be made. Whereas, high and often highly variable delay is a sign ofcongestion indicating, e.g., that the network is operating at the edges of the reliableradio coverage or there are other obstacles on the wireless air interface [30]. The typicalimpairments relates to noise, multipath fading, multiuser interference, obstacles onthe link path or changing weather conditions. Paper II researches the effect of variousphysical non-line-of-sight (NLOS) obstacles by doing delay and other performancemeasurements over IEEE 802.11p in a real outdoor environment.

3.3.2 REAC control logic

A brief overview of the REAC control algorithm is given next, and more details can befound in Paper I. The REAC logic is implemented by a Finite State Machine (FSM) thathas five states: SETUP, NORMAL, ALERT, PREDICT, and ACTION shown in Figure9.

54

Fig. 9. Operational work flow of REAC algorithm. c© [2013] AIRCC

The FSM operation is synchronized to the arrival and processing of the QoS samples.The system operation starts with a SETUP phase by filling its repositories to collect asufficient amount of QoS history values. In its NORMAL state the system calculates themean delay value and compares it to the previous value. We allow for up to exponentialgrowth before entering the ALERT state denoted by the system’s increased alerts counter.When a number of maximum alerts (max_alerts) is reached, the system resets thecounter and enters a new state, namely PREDICT. The system transits to the PREDICTstate because the mean delay has continued increasing for at least a max_alerts interval.Instead of just comparing the mean delay to a system-dependent maximum value for adelay tolerance, REAC makes a projection into the future of its current behavior. Theprediction assumes that, since the mean delay has been increasing in an exponentialway during the last updating intervals, it will maintain the same tendency. The purposeof such a prediction is to give the system some self-knowledge, and the possibilityto diagnose whether or not its current state is progressive. The system activates afalse alarms counter if the mean delay stops increasing exponentially. It allows for themaximum number of predictions to occur before considering that the shock period isfinished and reverting back to the NORMAL state. On the other hand, if the predictionshows that in the next interval the system will have surpassed the ceiling, it immediatelyenters the ACTION state to drop or mark excessive flows.

55

3.4 Fuzzy flow scheduler

Fuzzy logic is used to include human reasoning aspects, for example, in a controllingalgorithm. The controller design incorporates membership functions, rules and rulesinterpretation. In this thesis the fuzzy control model was applied to an edge router toprovide an adaptive solution for updating the queue weights of the scheduling algorithm.The fuzzy scheduler calculates an adaptive weight coefficient that determines thebandwidth share, e.g., for delay-sensitive real-time services and C-ITS applicationswith UDP traffic and best-effort TCP flows. Other flow sharing and prioritizationprimitives such as gold, silver and bronze user group labeling can also be specified, butthe fuzzy rule base has to be tuned accordingly. Another solution to treat certain classesinside a flow classification is to utilize a cascaded fuzzy system model. The developedadaptive scheduler achieves differentiated quality levels, fairness and improved networkperformance. By using multiple inputs, the model can tune the assigned outputbandwidth weights.

The structure of the applied fuzzy control model to a queueing system is presented inFigure 10. The fuzzy controller has four modules: generation of membership functions,fuzzification module, reasoning module and defuzzification module. Input variables arerepresented in linguistic form after (normalization and) fuzzification of physical valuesinto linguistic form. Defuzzification operates to calculate a crisp numerical outputweight value.

Fig. 10. Fuzzy queueing control model.

This section introduces, with references to Papers IV and V, an adaptive fuzzy expertsystem applying concepts of fuzzy logic to solve flow-scheduling problems in WLANsand wired IEEE 802.3 LAN networks. The fuzzy model is also applied in the Internet ofThings use case scenario for vehicular and V2VRU networking utilizing IEEE 802.11ptechnology in Papers I and II.

56

3.4.1 Fuzzy weighted queueing

Adaptive traffic management by a FWQ algorithm has to have special properties suchas diversified bandwidth and buffer management. The FWQ capable router deploysservice classes according to an adaptive fuzzy expert system’s reasoning method andcalculates new queue weights periodically for each class. The algorithm assigns aweight to real-time UDP traffic queues and to TCP based best-effort type traffic queuesand classifies incoming traffic accordingly as shown in Figure 11. The fuzzy modelreacts to traffic states and controls the weights of output queues. The UDP serviceclass is controlled in a way that guarantees more reliability to UDP traffic. The fuzzycontroller has modules: fuzzification module, reasoning module and defuzzificationmodule as shown in Figure 11.

Fig. 11. Adaptive fuzzy weighted scheduling.

The weight update requires two inputs. The first is the share of the UDP and TCPinput traffic data rate (S), which is calculated in the following way initially presented inPaper IV:

S =QUDP

QTCP +QUDP(2)

where QUDP is queue length of the UDP traffic queue and QTCP is queue length of theTCP traffic queue. The other input is the changed share of the input traffic data rate (∆S)calculated as follows:

∆S =CQUDP

CQTCP +CQUDP(3)

57

where CQUDP is the change of received UDP packets and CQTCP is the change ofreceived TCP packets.

3.4.2 Reasoning example

A linguistic model of a system was described by a group of linguistic relations (rules)that can be converted into numerical equations. The linguistic relations form a rule baseof the system that can be converted into matrix equations. The logic or rule base of themodel was composed by analyzing the QoS requirements, the dynamics of input trafficto routers, i.e., traffic density (number of incoming datagrams/time unit), delays and jitterof incoming datagrams as well as transient responses and steady-state properties of thesystem. The mapping of linguistic relations to linguistic equations for this application isdescribed in Figure 12, where the size of the rule base was 25 rules.

Fig. 12. The structure of the rule base.

Suppose, as an example, that Xi j, i=1,2; j = 1, 3,..., m (j is uneven number), is a lin-guistic level (e.g. negative big, negative small, zero, positive small, and positive big) fora variable Xi. The linguistic levels are replaced by integers −( j−1)

2 , ...,−2,−1,0,1,2, ...,( j−1)

2 . The direction of the interaction between fuzzy sets is presented by coefficientsAi j={−1,0,1}, i=1,2; j = 1,..., m. This means that the directions of the changes in theoutput variable decrease or increase depending on the directions of the changes in theinput variables [118]. Thus a compact equation for the output Zij is:

m

∑j=1

m

∑i=1

Ai jXi j = Zi, j. (4)

58

share of UDP and TCP received packets

gra

deo

f m

ember

ship NB NS ZE PS PB

0 10.05 0.20 0.40 0.65 0.75 0.80 0.90 0.95

0

1

0.70

0.30

0.92

Fig. 13. Labels of the input variable share of UDP and TCP packets. c© [2017] Inderscience

change of share of received packets

gra

deo

f m

ember

ship NB NS ZE PS PB

-1 1-0.95

-0.90 0.000.50

0.600.90

0.950

1

0.70

0.30

-0.10

-0.07

Fig. 14. Labels of the input variable change of the share of received packets. c© [2017]Inderscience

weight value = 0.803

gra

deo

f m

ember

ship NB NS ZE PS PB

0 1

0

1

-0.07

0.500.40 0.60 0.750.80

0.850.90

0.95

Fig. 15. Labels of the output variable weight value. c© [2017] Inderscience

Considering an example presented in Paper V, that the share of UDP and TCP inputtraffic S is 0.92 (= TCP dominates over UDP), which is after fuzzification in linguisticform positive small at the grade of membership 0.70 and positive big at the grade ofmembership 0.30 (see Fig. 13). Suppose that the change of share of received UDP andTCP packets ∆S is -0.07 (share of TCP has decreased compared to UDP), which is afterfuzzification in linguistic form negative small at the grade of membership 0.70 and zero

59

at the grade of membership 0.30 (see Fig. 14). Now we can read from Figure 12 that

IF S is positive small AND ∆S IS negative small, THEN the weight value is posi-

tive small at the grade of membership 0.70 andIF S is positive small AND ∆S IS zero, THEN the weight value is zero at the grade of

membership 0.30 andIF S is positive big AND ∆S IS negative small, THEN the weight value is zero at the

grade of membership 0.30 andIF S is positive big AND ∆S IS zero, THEN the weight value is negative small at the

grade of membership 0.30.

In linguistic equations this can be presented as follows (first value in each multiplicationdetermines the interaction index):

d 1∗1+−1∗−12 e= 1 at the grade of min(0.70,0.70)

d−1∗1+1∗02 e= 0 at the grade of min(0.70,0.30)

d−1∗2+−1∗−12 e= 0 at the grade of min(0.30,0.70)

d−1∗2+1∗02 e=−1 at the grade of min(0.30,0.30)

where de returns the next highest integer value by rounding up the value if necessary.Using individual based inference with Mamdani’s implication the weight value isnegative small at the grade of membership 0.30, zero at the grade of membership 0.30and positive small at the grade of membership 0.70. The linguistic weight value is thentransformed back into the physical domain to find the crisp output value for the weight

value using the center of area method (CoA). Therefore, the defuzzified crisp outputvalue is 0.803 (see Fig. 15), which is used in the router’s output queue as a share of UDPtraffic. The share of TCP traffic is 1.0 - (the share of UDP traffic) = 0.197. The modelredefines weight values in every circle, i.e., after every 50 received packets, which wasdetected to be a suitable round [119].

3.5 Summary

This chapter described the methods and algorithms for adaptive network managementin heterogeneous networks including methods to facilitate QoS provisioning withadmission control and flow scheduling. Paper I provides a combined view of the adaptive

60

management capabilities in this thesis for distributing the computing to the networkedges. These edge computing functions include:

1) real-time communication and high quality message exchange for applications thatrequire low latencies and reliable delivery of information.

2) admission control solution that monitors, measures and collects data periodicallyfrom a multitude of data sources to provide flow awareness and Quality of Service(QoS).

3) cognitive fuzzy flow scheduling for traffic prioritization and bandwidth optimiza-tion and management.

The regressive method for admission control (REAC) presented in Paper III isa cognitive system combining network monitoring, traffic identification, congestioncontrol, dynamic resource handling, QoS, and reasoning. When comparing REACwith respect to the previous paradigms: REAC is a distributed, measurement-basedsystem and belongs to the MBAC category. The main difference between REAC and theother reactive schemes presented in Section 2.3.2 is that congestion feedback is notgenerated by the traffic itself via, e.g., lost packets count or missing ACKs, but renderedby an external entity, i.e., the passive measurements tool for QoS-level estimationpurposes. As REAC performs no reservation of resources, there is no wasted capacitywith predefined traffic classes. The AC policy keeps track of the flows’ QoS, whetheror not the network path actually supports QoS, providing cognition to the networkpath. For this, a continuous real-time QoS monitoring is enabled to the network pathbetween the end points. The monitoring is passive, so it gives clear measures of how theapplication traffic is really performing, while at the same time the control overhead isminimized. Depending on the network capabilities and support for prioritization (e.g.

DiffServ), the REAC’s QoS classification is either mapped to the network’s QoS classes,or dropping (or selective dropping) is performed.

It is common to use fixed resource allocations for certain application types inconventional scheduling and queueing methods. This provides only statistic QoS insteadof guaranteed QoS, and is missing the capability to anticipate real-time traffic changesand adapt to them. In dynamically changing environments, the fixed allocations arenot always sufficient to satisfy all traffic flows and their demands. Therefore, QoSprovisioning and the adjustment of traffic prioritizations need to be carried out adaptivelyaccording to the service requirements. Therefore, the latest development is goingtowards the cross-layer mechanisms as explained in Section 2.2.2 and also combiningadmission control and scheduling for flow-based QoS methods described in Section

61

2.3.1. For example, in weighted scheduling schemes, the current development is directedtowards adaptation of scheduling parameters, e.g., assigning dynamic bandwidth weightsas described in Papers II, IV and V.

62

4 Results and discussion

This chapter describes the results of the original publications in detail and addressesthe research question presented in Section 1.2. This chapter begins with the resultsconcerning regressive admission control (REAC) proof-of-concept measurements inSection 4.1. Section 4.2 presents the results attained with fuzzy flow scheduling utilizedin LAN, WLANs and in vehicular communications. Vehicle to vulnerable road user(V2VRU) communication use case results are presented in Section 4.3.

4.1 REAC measurements

The REAC method was evaluated in a test-bed in a laboratory environment and focusedon validating the algorithm operation with two scenarios: the proof-of-concept (POC)scenario and high-priority scenario. The POC scenario and its results are presented here.Results of the high-priority scenario, where the network was pushed to its extreme byusing many low- and a high-rate videos can be found in Paper III. The POC scenariowas implemented by measuring a number of high-priority video clients along withartificially generated high-rate low-priority traffic, which was inserted to congest thenetwork. The first video streaming started after a 30 second idle period. Every newvideo was streamed for 60 seconds before entering the high-rate burst. The bursts haddurations of 30 second over the 10 Mbps bottleneck link. After the burst was ended,there was a 60 second wait before entering a new video. Inserted videos were enteredperiodically in turns having both low-quality (500 kbps-1 Mbps) and high-quality (2Mbps-4 Mbps) videos. This scenario imitated a case, where even a single high-rate flowdominates the load of the network path. Measurements were carried out for differentnetwork path setups: 1) pure BE, 2) DiffServ with DSCP marks, 3) REAC over BE(flow dropping) and 4) REAC over DiffServ (DSCP marks). In the REAC case, the ACmodule ensures that the first videos will continue to enjoy good quality even though thetotal offered traffic load exceeds the capacity. When exceeding the capacity, the Qosmetmonitoring solution quickly notices the QoS degradation, and the AC module startsthe decision-making procedure causing no notable damage to the existing users in thenetwork.

Measurement runs of the POC scenario with different path setups are presentedin Figures 16-19. The BE path in Figure 16 suffers from long delay periods, when

63

Fig. 16. Proof-of-concept with pure BE. c© [2013] AIRCC

Fig. 17. Proof-of-concept with DiffServ. c© [2013] AIRCC

the high-rate traffic burst enters the network. All five multimedia streams are runningwhen the time is around 600 seconds. At this point, the overall bit rate of the videosbegins to exceed the network path capacity and there is a lot of fluctuation observed inthe delay behavior. The DiffServ configuration gives a better performance than thepure BE since, by its nature, it is able to cope with the traffic bursts of the BE class asthe videos belong to the high-priority class. However, it fails when there is too muchhigh-priority traffic, as perceived in Figure 17: the delay behavior becomes very similarto that of the BE case (Figure 16). As the network path is congested by high-priority

64

Fig. 18. Proof-of-concept with REAC BE. c© [2013] AIRCC

Fig. 19. Proof-of-concept with REAC marking. c© [2013] AIRCC

traffic, prioritization does not help any longer, and all the flows start to suffer from badquality. This demonstrates the problem of statistical QoS. In the REAC case in Figures18 and 19, the AC module ensures that the first videos will continue enjoying goodquality even though the total offered traffic load exceeds the capacity. When exceedingthe capacity, the monitoring solution quickly notices the QoS degradation, and theAC module starts the decision-making procedure causing no notable damage to theexisting users in the network. The REAC performance was evaluated by QoS metricsand the subjective quality MOS values presented in Section 2.2.2. REAC is able to cope

65

with the BE traffic bursts, using the AC policy, even though the network path itselfwould not support prioritization. Table 2 shows the average results related to MOSand delay values. The overall average delay for the BE path is 23 milliseconds, whichis twice the delay compared to the REAC cases as noticed in Table 2. Even thoughthe overall subjective score (MOS) is 4.3, which is acceptable; the subjective qualityis poor approximately 16% of the time, which is a very high value. This shows thatthe congestion, observed as high delay values, reflects the subjective quality as well.Considering the DiffServ case, the subjective quality is poor approximately 8% of thetime (Table 2).

Table 2. Averaged subjective quality and delay values in POC scenario. c© [2013] AIRCC

pureBE DiffServ REAC drop REAC mark

Subjective score (MOS) 4.3 4.6 4.8 4.8

% of time good quality 73.4 87.3 91.6 91.5

% of time poor quality 15.9 8.0 3.1 1.9

Delay [ms] 23.2 11.1 10.7 12.2

The REAC performance with measured delay values is shown in Figures 18 and 19,with drop and marking policies, respectively. The graphs show that during the high-rateburst periods, there is a short delay spike at the beginning, which represents the time ittakes for REAC to make decisions considering the newly entered flow. REAC is ableto cope with the BE traffic bursts, using the AC policy, even though the network pathitself would not support prioritization. The drawback with the REAC marking case isthat normal DiffServ allows direct prioritization, and there is no delay spike when theBE traffic enters the network path. However, while DiffServ fails in the presence oftoo many high-priority flows the REAC does not: good quality is guaranteed to themajority of the flows. The limit, for how many flows the quality can be guaranteed,comes directly from the relation between the controlled network path capacity andload of the traffic flows. Roughly, the QoS can be guaranteed for the flows that are ofhigh-priority class, and as an aggregate do not exceed the network path capacity. Thesubjective quality is poor only about 2-3% of the time, depending on the REAC method,being clearly better than with DiffServ, and considerably better when compared with thepure BE path. The difference between REAC dropping and marking performancesis not very notable in the POC scenario, as observed from Table 2. REAC is capable

66

of enabling guaranteed QoS for a subset of the aggregate of traffic regardless of thenetwork path’s support for prioritization.

4.2 Fuzzy flow scheduling results

Fuzzy flow scheduling was applied to IEEE 802.3 namely with adaptive weighted fairqueueing (AWFQ) in Paper IV and to IEEE 802.11b with fuzzy weighted queueing(FWQ) in Paper V. The model was further applied to an environment with mobile nodeswith IEEE 802.11p in Papers I and II. The research was carried out by simulationsutilizing Network Simulator 2 (NS-2) [120], [121]. In order that the adaptive controlmodel can be applied, it should converge to its steady-state value, which is important ifthere are time-varying workloads. The system should also achieve its control objectiveswithout undershoots or overshoots. The primary aim of the developed AWFQ and FWQcontrol methods was to update router’s weights for output queues, i.e., to determine theamount of allowed bandwidth for prioritized users with delay-sensitive UDP servicesand for the best-effort users with general purpose TCP services.

4.2.1 Fuzzy weighted queueing in LAN and WLANs

The adaptive fuzzy scheduling model for IEEE 802.3 model was first introduced andresearched for two different scenarios: when the input traffic data rate was significantlyhigher than the output line capacity, and when the input traffic data rate was significantlylower than the output line capacity. The weight coefficients of the comparative WFQmodel were set to 0.2 and 0.8 for TCP and UDP output traffic on the router in order togive more priority to real-time UDP traffic. The fuzzy model was tuned to producethe same output traffic relation. From Tables 3 - 4 it can be noted that the AWFQmodel reacts faster (shorter fall, rise and settling times) than the WFQ model. For TCPtraffic, fall times were 67 seconds for the WFQ model and 63 seconds for the AWFQmodel, whereas the UDP traffic rise times were 62 seconds and 59 seconds, respectively.Settling times for the WFQ model were 97 seconds and 76 seconds for TCP and UDPtraffic. For the AWFQ model they were 75 seconds and 59 seconds for TCP and UDPtraffic. Undershoot and overshoot for the WFQ model were about 20 kbps for both TCPand UDP traffic. For the AWFQ model, there was a 10 kbps undershoot for TCP trafficbut no overshoot for UDP traffic. Similar results were achieved with other input datarates when they exceeded the allocated bandwidth on the output link. In the second

67

scenario, when the input traffic rate was significantly lower than the output capacity,both WFQ and AWFQ models behaved the same way, while the router was capable oftransmitting all the input traffic.

Table 3. Rise, fall and settling times in 802.3 when input >> output.

Traffic Rise time Settling time

WFQ AWQ WFQ AWQ

TCP 67 s (FT) 63 s (FT) 97 s 75 sUDP 62 s 59 s 76 s 59 s

Table 4. Overshoots, undershoots and throughput in 802.3 when input >> output.

Traffic Overshoot WFQ Overshoot AWQ Throughput WFQ Throughput AWQUndershoot Undershoot

TCP US 20 kbps US 10 kbps 520 kbps 570 kbpsUDP OS 20 kbps - 1 Mbps 1 Mbps

In Paper V, the fuzzy FWQ model was applied to IEEE 802.11b, and the simulationmodel included 10 wireless hosts distributed randomly around the 400 meter x 400 meterarea. Two source nodes (one TCP and one UDP source), a router node, a destinationnode, and 6 wireless background nodes were observed. The FWQ model was designedto reach the target delay of 100 milliseconds. The transmission interval i was varied fromi ∈ [0.005 s,0.0065 s] to i ∈ [0.005 s,0.015 s] in 10 milliseconds intervals for backgroundtraffic. The FWQ model satisfied the end-to-end delay and packet loss requirements andresponded well to the throughput requirements. The simulation results of drop-tail,WFQ and FWQ for delay are presented in Table 5.

Table 5. Delays. c© [2017] Inderscience

BG traffic Average Delay for UDP Average delay for TCP

interval [s] DropTail FWQ WFQ DropTail FWQ WFQ[ms] [ms] [ms] [ms] [ms] [ms]

0.005-0.0065 77.6 63.0 89.4 119.0 92.4 94.80.005-0.0075 78.6 29.4 40.0 88.5 83.1 84.10.005-0.0085 73.7 21.2 24.0 73.8 75.0 78.20.005-0.015 3.7 4.1 4.1 4.1 3.7 3.7

68

Fig. 20. ITS architectural framework.

In the first case with a 6.5 millisecond average background traffic interval, thenetwork was pushed to the limits (delay above 100 ms) considering the basic drop-tailscheduling. With drop-tail the buffer is filled to its maximum capacity before it dropsthe newly arrived packets until the buffer has room to accept incoming traffic. Hence,it is less sensitive to the amount of background traffic than WFQ and FWQ. FWQobtains better delay performance with different amounts of background traffic comparedto the drop-tail mechanism. Still, FWQ is capable of sustaining TCP delay at thetarget level (below 100 ms). Basic WFQ can reach rather good performance when thedelay-sensitive traffic is prioritized (weight now 0.8 for UDP and 0.2 for TCP). With thelongest average background traffic interval of 5-15 milliseconds, all the methods canobtain equally low delay values for TCP and UDP.

4.2.2 Fuzzy weighted queueing in vehicular communications

In Paper I, The FWQ model was also applied to IEEE 802.11p technology withsimulation topology shown in Fig. 20. It is considered that six best-effort TCP (packetsize of 256 bytes) traffic sources with link bandwidth 200 kbps (1200 kbps altogether)and six delay-sensitive UDP (packet size of 512 bytes) traffic sources with link bandwidth350 kbps (2100 kbps altogether) are connected to the RSU through a LAN in the corenetwork.

The RSU including the FWQ logic is connected to the mobile vehicles. In thefuzzy scheduler, the flows are treated as aggregates respectively for UDP and TCP. By

69

increasing the Pareto distributed 1 datagram bursts randomly to the simulated link theburst nature of the traffic was enhanced. The adaptive FWQ scheduler is operating at thenetwork edge between fixed infrastructure and wireless IEEE 802.11p network. Thewireless receiving vehicle and four other vehicles are moving around the 600 meter x600 meter area with a speed of 0-60 km/h (urban area). The four other vehicles are alsosending background Constant Bit Rate (CBR) traffic at random intervals [0.01 0.05].The effect of increasing the number of vehicles is not considered in this paper. In fact,even with only one vehicle, by increasing the source data rate, we can analyze the FWQcontrol algorithm functionality that can be reached in IEEE 802.11 networks in similarconditions. Here we are operating at the capacity limits for IEEE 802.11p with one RSUdeployment [122] testing a congested situation when the input data rate is bigger thanthe output capacity.

Utilizing the FWQ mechanism shorter settling, rise and fall times as well as lowerovershoots and undershoots were attained compared to the traditional WFQ algorithm asshown in Figures 21-24.

Time [s]0 50 100 150 200 250 300 350 400 450 500

Th

rou

gh

pu

t [k

bit

/s]

0

500

1000

1500Throughput of TCP as a function of time. WFQ

settling time 220s

average throughput 640 kbit/s

rise time 26s

overshoot 50 kbit/s

Fig. 21. Throughput of TCP when WFQ was used. Bandwidth of the output line << inputdata rate with 802.11p. c© [2016] IEEE

The developed FWQ scheduler was designed to prioritize real-time UDP trafficbut it can be applied to prioritize, e.g., cooperative vehicle applications that are alsosensitive to end-to-end delay. The tuned rule base for the fuzzy system anticipates theupcoming traffic and makes it possible to react more smoothly and faster to prevailingtraffic conditions increasing QoS as shown in Tables 6 and 7. For TCP traffic, rise timeswere 26 s for the WFQ model and 17 s for the FWQ model, whereas for the UDP trafficrise times were 23 s and 23 s, respectively. Settling times for the WFQ model were 220s and 180 s for TCP and UDP traffic. For the FWQ model they were 70 s and 65 s for1Data network traffic has self-similar and long-range dependent nature, which is known to obey Paretodistribution with Pareto distributed interval times.

70

Time [s]0 50 100 150 200 250 300 350 400 450 500

Th

rou

gh

pu

t [k

bit

/s]

0

500

1000

1500Throughput of TCP as a function of time. FWQ

settling time 70 s

rise time 17sundershoot 45 kbit/s average throughput 645 kbit/s

Fig. 22. Throughput of TCP when FWQ was used. Bandwidth of the output line << inputdata rate with 802.11p. c© [2016] IEEE

Time [s]0 50 100 150 200 250 300 350 400 450 500

Th

rou

gh

pu

t [k

bit

/s]

0

500

1000

1500Throughput of UDP as a function of time. WFQ

average throughput 940 kbit/srise time 23s

undershoot 90 kbit/s settling time 180s

Fig. 23. Throughput of UDP when WFQ was used. Bandwidth of the output line << inputdata rate with 802.11p. c© [2016] IEEE

Time [s]0 50 100 150 200 250 300 350 400 450 500

Th

rou

gh

pu

t [k

bit

/s]

0

500

1000

1500Throughput of UDP as a function of time. FWQ

average throughput 940 kbit/s

undershoot 85 kbit/s

settling time 65s

rise time 23 s

Fig. 24. Throughput of UDP when FWQ was used. Bandwidth of the output line << inputdata rate with 802.11p. c© [2016] IEEE

TCP and UDP traffic. For the WFQ model with TCP and UDP traffic, there was 50kbits/s overshoot and 90 kbits/s undershoot, respectively. For the FWQ model, there was45 kbits/s undershoot for TCP and 85 kbits/s undershoot for UDP.

The reason for the shorter settling and rise times of the FWQ model may also bethat the rule base anticipates the upcoming traffic and makes it possible to react moresmoothly and faster to prevailing traffic conditions. The rule base lets the UDP burst to

71

Table 6. Rise and settling times in 802.11p when output << input. c© [2016] IEEE

Traffic Rise time Settling time

WFQ FWQ WFQ FWQ

TCP 26 s 17 s 220 s 70 sUDP 23 s 23 s 180 s 65 s

Table 7. Overshoots, undershoots and throughput in 802.11p when output << input. c©[2016] IEEE

Traffic Overshoot WFQ Overshoot FWQ Throughput WFQ Throughput FWQ

Undershoot Undershoot

TCP OS 50 kbps US 45 kbps 640 kbps 645 kbps

UDP US 90 kbps US 85 kbps 940 kbps 940 kbps

utilize breaks in TCP flows and vice versa. The rule base has a significant role for therise and settling times as explained in detail in Section 3.4.2. Hence, the rule base has tobe tuned for the overall aim in order to take care of tradeoffs between contradictorysubtargets. A large normal area, i.e., several rule combinations, produces normal output,stabilizes the behavior of the control system, but increases the rise and fall times. Anarrow normal area may lead to decreased rise and fall times but increased settling timedue to oscillations.

4.3 Vehicle to vulnerable road user (V2VRU) use case results

In Paper II, the performance of ITS-G5 for V2VRU communication in time-criticalsafety conflict scenarios was assessed and optimized. The tests included variousnon-line-of-sight (NLOS) scenarios in urban environments and line-of-sight (LOS)simulations to support C-ITS message prioritization and scalability with differentamounts of vehicles. A suitable test system was developed for V2VRU communicationshown in Figure 25, which monitors network performance in real-time. In order to mirrorthe performance of VRU safety applications, we have to be aware of the underlyingnetwork performance metrics. The test system uses the Qosmet tool running in bothVRU and vehicle terminals to be able to perform bi-directional measurements overthe IEEE 802.11p radio link. Current smartphones do not support the IEEE 802.11penhancement, but we emulated a complementary system utilizing NXP’s Cohda MK4

72

Fig. 25. V2VRU test setup. c© [2017] IET

[123] connected to a laptop running Qosmet and receiving the real-time applicationdata enabled by GPS. In the vehicle terminal, there is a similar 802.11p setup sendingapplication data and Qosmet measuring the real-time application performance.

Example use cases with NLOS communication included pedestrians crossing streetsfrom behind objects, and low-visibility scenarios, e.g., when VRU is behind a vehicle,behind a van, behind a queue of vehicles, between vehicles, behind trees/bushes orbehind a building. The LOS simulations with different amounts of vehicles utilized thefuzzy weighted queueing (FWQ) mechanism presented in Section 4.2 for critical dataprioritization.

Details of V2VRU NLOS scenarios are presented in Table 8 including the obstacle,VRU’s place with respect to the obstacle, and distance from VRU to roadside (inmeters). The tests were performed for three different VRU antenna transmitting powers,corresponding to different uses: vehicle grade antenna (20 dBm transmission power),less performant smartphone antenna (10 dBm), and a smartphone close to the humanbody (0 dBm). For emulating the impact of the human body, the power was reducedfrom the smartphone antenna power with 10 dBm [124]. In each of the scenarios, thetest vehicle was first driven outside the radio range, after which it approached the VRU,except in scenario 1 where the effect of different vehicle speeds having LOS connectivity

73

Table 8. Details of V2VRU performance monitoring scenarios.

Obstacles VRU’s place Distance from VRU to roadside

Scenario 1 line of sight side of the road 1 m

Scenario 2 1 car behind a car 1 m

Scenario 3 2 vans between vans 1 m

Scenario 4 1 car and 1 van between a car and a van 1 m

Scenario 5 bushes side alley 5 m

Scenario 6 building behind a building 8.4 m

building behind a building 60 m

building behind a building 112 m

Scenario 7 queue of vehicles behind a queue 1 m

was tested. Table 9 presents the performance values for delay, jitter, packet loss, reliablerange and time to collision (TTC) concerning the measured scenarios.

Starting from the measured values, and the scenario description, the TTC (explainedin Section 2.1.3) is calculated, i.e., when both VRU and vehicle have reached thecrossing point for urban scenarios, where cars move at moderate speed (50 km/h) andfor extra-urban scenarios with higher car speeds (90 km/h); and for different VRU types,i.e., for pedestrians (5 km/h), cyclists and e-bikes (25 km/h). Based on this analysis,a VRU with a device transmitting only 0 dBm (e.g. a smartphone close to the body),will not be perceived by fast moving cars in time when they are entering the road frombehind an obstacle or coming from an obstructed side road. For cars at moderate speed(50 km/h), the time of warning prior to the potential conflict is in the range of 3.5-5seconds, which is still sufficient for these cars to come to a full stop prior to the conflict.However, a user in front of a queue of cars (scenario 7) is not noticed in time. Hence,the scenario in which the VRU has a smartphone close to the body is not considered tobe suitable for cooperative traffic applications. A smartphone or other device, withtransmitting power of 10 dBm, which is, for example, fixed to a VRU or not close to theperson’s body (e.g. in a backpack), is detected in time by oncoming cars at moderatespeeds. However, for cars moving at high speeds (extra-urban scenarios), VRUs comingout from side roads obstructed by bushes and vehicles are challenging as the time tocollision is in the range 3.4-5 seconds, so the car may not be able to come to a stop intime. For motorcyclists, who have higher relative speeds towards vehicles, a vehicle unit

74

Table 9. Performance values for the measured scenarios and TTC for two conflict types:urban/pedestrian (car speed 50 km/h; pedestrian 5 km/h) and extra-urban/cyclist (car speed90 km/h; cyclist 25 km/h). c© [2017] IET

Power Delay Jitter Pkt loss Range TTCurban [s]2

TTC extra-urban [s]2

[dBm] [ms] [ms] [%] [m]1 pedestrian cyclist

Scen 1(LOS) 20 1.01 0.17 4 1002 72.1 40.1

10 1.05 0.24 5 327 23.6 13.10 1.41 0.63 18 121 8.7 4.9

Scen 2(1 car) 10 1.14 0.34 8 152 10.9 6.1

0 1.90 1.17 25 60 4.4 2.4Scen 3

(2 vans) 10 1.09 0.30 11 123 8.9 4.90 1.80 1.10 19 54 3.9 2.1

Scen 4(car&van) 10 1.14 0.29 7 157 11.3 6.3

0 2.13 1.39 14 54 3.9 2.1Scen 5

(bushes) 10 1.38 0.68 21 101 7.3 4.00 1.81 1.03 28 0 3.6 0.7

Scen 6(8.4m) 10 1.03 0.20 1 85 6.1 6.1

0 1.60 1.01 35 61 6.1 6.1

(60m) 20 1.05 0.19 6 90 42.8 42.810 1.10 0.33 9 11 42.8 42.80 1.71 1.02 47 1.0 42.8 42.8

(112m) 20 1.25 0.55 9 18 80.6 80.610 0.34 0.10 28 2.7 80.6 80.6

Scen 7(queue) 20 0.99 0.19 0 203 14.6 8.1

10 1.02 0.20 2 84 6.1 3.40 1.90 1.36 13 38 2.7 1.5

1 Values less than 160 meters in italics, less than 100 meters in bold italics2 Values less than 5 seconds in bold italics

with the same transmitting power as vehicle stations (20 dBm) should be used. Moredetailed results and figures of the scenarios can be found in Paper II.

75

The LOS simulations were tested with commonly supported WFQ and adaptiveFWQ scheduling methods for UDP traffic with C-ITS message prioritization andscalability. In the simulations, the amount of mobile background vehicles was changedfrom 6 to 12 with the overall number of vehicles changing from 8 to 14 (including thetest vehicle, receiving VRU, and background vehicles) and a RSU connected to theapplication server. All the vehicles were moving around the 100 meter x 100 meter areafor 600 seconds, while considering the minimum range corresponding TTC of 5 secondsfor pedestrian and cyclists in urban scenarios described in Section 2.1.3. The test vehiclewith the adaptive logic prioritizes the more critical UDPprio1 traffic over the UDPprio2

traffic, and sends it to the VRU node. Background vehicles send CBR UDP trafficin random intervals [0.01, 0.05] with packet a size of 128 Bytes. The measurementswere conducted by NS-2 (Network Simulator) to simulate V2VRU communicationswith transmit powers of 10 and 0 dBm (to emulate the VRU antenna powers) for IEEE802.11p. The LOS packet loss for 0 dBm, already started to suffer high packet loss ratesat a distance of 100-200 meters as seen in Table 9.

The static weight coefficients of the WFQ model were assigned to 0.6 and 0.4 forUDPprio1 and UDPprio2 output traffic respectively, in order to give more priority toUDPprio1 traffic. The FWQ model dynamically changes the weight value giving morepriority to UDPprio1, as it is capable of monitoring the traffic states and utilizing the idleperiods of UDPprio2 for UDPprio1, and vice versa. The bottleneck link from the testvehicle with WFQ and FWQ to the VRU was assigned to 1 Mbits/s bandwidth. The datarates for UDPprio1 and UDPprio2 flows instantiated between the test vehicle and theVRU were first set to 250 kbits/s for three UDPprio2 flows (750 kbits/s altogether) and100 kbits/s for two UDPprio2 flows (200 kbits/s altogether) that is slightly below thebottleneck limit.

The 20 dBm transmit power was also ignored in the simulations (area 100 meter x100 meter) because 200 meters is adequate to detect VRU more than 5 seconds beforeTTC for extra-urban pedestrian and cyclist scenarios. Figure 26 presents packet loss as afunction of number of vehicles for 10 dBm and Figure 27 for 0 dBm. For less than 10vehicles, the packet loss is less than 3% for both the 10 and 0 dBm cases. However,the FWQ packet losses for UDPprio1 and UDPprio2 are less than for WFQ. With morethan 12 vehicles, the packet loss starts to grow rapidly. For example at 10 dBm with 14vehicles, the FWQ algorithm is able to provide approximately 19% lower packet lossfor UDPprio1 and 21% lower packet loss for UDPprio2 than the static WFQ algorithmas shown in Figure 26. With 0 dBm transmit power and 14 vehicles in Figure 27, the

76

Number of vehicles8 9 10 11 12 13 14

Pac

ket

loss

[%

]

0

5

10

15

20

25

30Packet loss as a function of number of vehicles with 10 dBm

FWQ priority 1FWQ priority 2WFQ priority 1WFQ priority 2

Fig. 26. LOS packet loss as a function of number of vehicles with 10 dBm transmit power. c©[2017] IET

Number of vehicles8 9 10 11 12 13 14

Pac

ket

loss

[%

]

0

5

10

15

20

25

30Packet loss as a function of number of vehicles with 0 dBm

FWQ priority 1FWQ priority 2WFQ priority 1WFQ priority 2

Fig. 27. LOS packet loss as a function of number of vehicles with 0dBm transmit power. c©[2017] IET

FWQ algorithm provides approximately 29% lower packet loss for UDPprio1 and 10%lower packet loss for UDPprio2. These results also give indications for the NLOS usecase scenarios that suffer from packet losses, in order that the QoS of the critical C-ITSsafety messages in V2VRU communication can be guaranteed. NLOS use cases canbenefit from prioritized scheduling but at the same time also preserve fairness towardsother information by providing dynamic weighted bandwidth sharing by FWQ.

4.4 Discussion

The primary target of this thesis was to research and apply adaptive traffic managementmethods in heterogeneous networking by providing means for traffic adaptation andpriority management inside a router. This chapter provided an overview of the resultsobtained in Papers I-V. The beginning of the Chapter 3 introduced the overall architecturaltopology for information dissemination from core network through the access networkto the end-users including the adaptive mechanisms.

77

The developed methods for adaptive traffic management included the following:

- Referring to the needs for adaptive traffic management at the network edge- Need for traffic classification, monitoring and measuring along with adaptive methods- Developing regressive admission control- Developing fuzzy weighted scheduling mechanism- Applying adaptive methods in V2VRU communications use case

These results answered the research question of whether it is possible to develop

adaptive and cognitive management methods to monitor, balance, and share traffic

that has different capacity requirements in heterogeneous communication networks

presented in Chapter 1. The need for adaptive traffic management distributed to thenetwork edges is starting to be evident in highly connected environments transmittingand processing massive amounts of data. These edge and fog computing features inclose proximity to mobile subscribers offer ultra-low latency, high bandwidth, trafficoffloading, data analysis, aggregation and QoS-aware networking capabilities. Thedeployment place of edge computing depends, e.g., on scalability, physical deploymentconstraints, performance requirements (e.g., latency) and the network information thatneeds to be exposed as discussed in Paper I. Part of the edge computing can be deployed,e.g., at the RSU as shown in Paper II being an aggregation point possibly located at theedge of the core network. In order to utilize available networks in the most efficient way,and to deliver, e.g., accurate real-time C-ITS messages, it is necessary that the networksare monitored, performance indicators and QoS measured and mobility is controlled.Mobility control is not considered in this thesis more widely but part of the adaptivemethods developed in this thesis are also important when providing seamless servicecontinuity. A piloting use case for vehicular networking with intelligent handover ispresented in Paper I.

The developed method for regressive admission control in Paper III monitored theQoS-level variation by measuring the delay, and estimated when the QoS-level decreaseaffected the quality of the high-priority applications. The system first traced the suspectflow(s), and then either dropped packets of these flows, or decreased the flow priority.Roughly, the QoS can be guaranteed for the flows that are of high-priority class, and asan aggregate, do not exceed the network path capacity. The overall average delay for abest-effort path was 23 milliseconds, which was twice the delay compared to REAC forflow dropping and marking cases.

78

Fuzzy flow scheduling was applied to IEEE 802.3 namely with adaptive weightedfair queueing (AWFQ) in Paper IV and to IEEE 802.11b with fuzzy weighted queueing(FWQ) in Paper V. The model was further applied to an environment with mobile nodeswith IEEE 802.11p in Papers I and II. A fuzzy model was first introduced and researchedfor two different scenarios: when the input traffic data rate was significantly higher thanthe output line capacity, and when the input traffic data rate was significantly lower thanthe output line capacity. Utilizing AWFQ and FWQ mechanisms shorter settling, riseand fall times as well as lower overshoots and undershoots were attained compared to thetraditional WFQ and drop-tail algorithms. FWQ obtained better delay performance withdifferent amounts of background traffic compared to drop-tail and WFQ mechanisms.

The developed FWQ scheduler was applied to prioritize cooperative vehicle applica-tions for V2VRU communication that are also sensitive to end-to-end delay as shownin Paper II. The tests included various NLOS scenarios in urban environments andLOS simulations to support C-ITS message prioritization and scalability with differentamount of vehicles. A suitable test system was developed for V2VRU communicationthat monitored the network performance in real-time. The LOS simulations with FWQmechanism utilized the tuned rule base which anticipated the upcoming traffic and madeit possible to react more smoothly and faster to prevailing traffic conditions increasingQoS.

By utilizing REAC and FWQ scheduling, the system does not need to guaranteecertain SLAs or classes of service, but instead the QoS can be introduced incrementally.However, if the system does not support any service classes, and REAC does the brutalflow dropping, this is not necessarily a negative action. The system can keep most of theusers satisfied with a guaranteed QoS and the blocked flows can, e.g., be redirected to analternative path. Improved bandwidth capacity and incremental deployment imply costefficiency and savings compared to many QoS architectures.

79

80

5 Conclusions

The primary aim of this thesis was to efficiently manage network resources in hetero-geneous communication networks utilizing adaptive traffic management methods. Inorder that the different architectural parts and available networks are utilized in the mostefficient way, it is necessary that the networks are monitored, performance indicators aremeasured and traffic load is optimized.

Regressive admission control (REAC) presented in this thesis, provides flowmanagement and prioritization capabilities towards real-time traffic as well as fairnesstowards other best-effort flows. The management system adjusts the transceiver’straffic flow(s) for prevailing network conditions to achieve optimal network capacity.REAC is able to cope with the best-effort traffic bursts, using the regressive admissioncontrol policy, even though the network path itself would not support prioritization. Thesubjective quality is poor only about 2-3% of the time utilizing REAC methods. This isclearly better than with DiffServ, which suffers from poor quality approximately 8% ofthe time, and considerably better when compared to the pure BE path that suffers frompoor quality 16% of the time. With REAC methods the delay is halved from over 20 msfor the pure BE traffic case around 10 ms.

The primary aim of the developed fuzzy scheduling methods was to satisfy andfairly allocate resources to UDP and TCP queues through adaptive queue weights. Thedeveloped AWFQ and FWQ mechanisms are able to react faster to traffic changes andguarantee better quality for delay-sensitive UDP traffic and lower packet loss rates forTCP traffic. These methods decreased TCP and UDP traffic settling, rise and fall times,decreased overshoot/undershoot and stabilized throughput in the output connections.Therefore, the fuzzy scheduling model enables an increase in the amount of usersand still provides adequate QoS. More stable throughput decreases the need for flowcontrol between the end nodes and makes the capacity utilization easier on the outputconnections.

The FWQ method was applied to prioritize cooperative vehicle applications forV2VRU communications that are also sensitive to end-to-end delay. The simulationswith the FWQ mechanism utilized the tuned rule base, which anticipated the upcomingnetwork traffic. In situations where the transmission medium is congested, V2VRUcommunication benefits from adaptive flow prioritization, especially when operating

81

around the system limits. In a congested situation where a vulnerable road user has thedevice close to the body (0 dBm antenna transmit power), the FWQ algorithm providesapproximately 30% lower packet loss for the prioritized UDP flows and 10% lowerpacket loss for other UDP flows than the static flow scheduling.

The large scale IoT and ITS systems, in particular, have to be made intelligent andcapable of self-management with adaptive features, due to the challenging managementof the overall system architecture. This thesis proposes to complement traditional cloudcomputing by running adaptive traffic management components at the edge of thenetwork closer to the end users to guarantee better performance, e.g., for delay-sensitiveapplications. The use of adaptive traffic management methods can overcome problemsraised by emerging new applications and network congestion, hence creating cognitionin the network path.

82

References

1. Ashton K (2009) That ’Internet of things’ thing. RFiD Journal 22(7): 97–114.2. Atzori L, Iera A & Morabito G (2010) The Internet of things: A survey. Computer Networks

54(15): 2787–2805.3. Vermesan O, Friess P, Guillemin P, Gusmeroli S, Sundmaeker H, Bassi A, Jubert IS, Mazura

M, Harrison M, Eisenhauer M et al. (2011) Internet of things strategic research roadmap.Internet of Things: Global Technological and Societal Trends 1: 9–52.

4. Miorandi D, Sicari S, De Pellegrini F & Chlamtac I (2012) Internet of things: Vision,applications and research challenges. Ad Hoc Networks 10(7): 1497–1516.

5. Armbrust M, Fox A, Griffith R, Joseph AD, Katz RH, Konwinski A, Lee G, Patterson DA,Rabkin A, Stoica I et al. (2009) Above the clouds: A berkeley view of cloud computing.Technical Report UCB/EECS-2009-28, EECS Department, University of California,Berkeley.

6. Zhang Q, Cheng L & Boutaba R (2010) Cloud computing: state-of-the-art and researchchallenges. Journal of Internet services and applications 1(1): 7–18.

7. Cisco (2013). Fog Computing, Ecosystem, Architecture and Applications. Category:Experimental.

8. Chiang M & Zhang T (2016) Fog and IoT: An overview of research opportunities. IEEEInternet of Things Journal 3(6): 854–864.

9. Bonomi F, Milito R, Zhu J & Addepalli S (2012) Fog computing and its role in the internetof things. In: Proceedings of the first edition of the MCC workshop on Mobile cloudcomputing, MCC ’12, pp. 13–16. ACM.

10. Blake S, Black D, Carlson M, Davies E, Wang Z & Weiss W (1998). An Architecture forDifferentiated Services. IETF Request for Comments: 2475.

11. Braden R, Clark D & Shenker S (1994). Integrated Services in the Internet Architecture anOverview. IETF Request for Comments: 1633.

12. IEEE Std 802.11e-2005 (Amendment to IEEE Std 802.11). IEEE Standard for Infor-mation technology– Amendment 8: Medium Access Control (MAC) Quality of ServiceEnhancements pp. 1–212.

13. Index: CVN (2016) Global mobile data traffic forecast update 2014–2019. white paperc11-520862 .

14. Parashar M & Hariri S (2005) Autonomic computing: An overview. In: UnconventionalProgramming Paradigms, pp. 257–269. Springer.

15. Singh S & Chana I (2016) Qos-aware autonomic resource management in cloud computing:a systematic review. ACM Computing Surveys (CSUR) 48(3): 42.

16. Vaquero LM & Rodero-Merino L (2014) Finding your way in the fog: Towards a compre-hensive definition of fog computing. ACM SIGCOMM Computer Communication Review44(5): 27–32.

17. Varghese B, Wang N, Barbhuiya S, Kilpatrick P & Nikolopoulos DS (2016) Challenges andopportunities in edge computing. In: IEEE International Conference on Smart Cloud, pp.20–26.

18. Mahmud R & Buyya R (2016) Fog computing: A taxonomy, survey and future directions.

83

19. Malik A, Qadir J, Ahmad B, Yau KLA & Ullah U (2015) QoS in IEEE 802.11-basedwireless networks: A contemporary review. Journal of Network and Computer Applications55: 24 – 46.

20. Horn P (2001) Autonomic computing: IBM’s perspective on the state of informationtechnology.

21. IBM (2003). An architectural blueprint for autonomic computing.22. Mahmoud Q (2007) Cognitive Networks, Towards Self-Aware Networks. Wiley.23. ITU (2005) The Internet of Things. Technical report.24. Halonen T, Romero J & Melero J (2002) GSM, GPRS, and EDGE performance: Evolution

towards 3G/UMTS. John Wiley & Sons, Ltd.25. Holma H & Toskala A (2010) WCDMA for UMTS: HSPA Evolution and LTE. John Wiley

& Sons, Ltd., 5th edition.26. Holma H & Toskala A (2011) LTE for UMTS: Evolution to LTE-Advanced. John Wiley &

Sons, Ltd., 2nd edition.27. LTE-M - Optimizing LTE for the Internet of Things (2015).28. IEEE Std 802.3 (2008) - Carrier sense multiple access with collision detection (CSMA/CD)

access method and physical layer specifications - section one.29. IEEE Std 802.11 (1997) - Wireless LAN medium access control (MAC) and physical layer

(PHY) specifications. Institute of Electrical and Electronics Engineers Inc., New York, USA.30. Pahlavan K & Krishnamurthy P (2002) Principles of Wireless Networking: A Unified

Approach. Prentice Hall PTR, New Jersey, USA.31. Mohapatra P, Li J & Gui C (2003) QoS in mobile ad hoc networks. IEEE Wireless

Communications 10(3): 44–53.32. Ksentini A, Nafaa A, Gueroui A & Naimi M (2007) ETXOP: A resource allocation protocol

for QoS-sensitive services provisioning in 802.11 networks. Performance Evaluation 64(5):419–443.

33. Lenagala R & Zeng Q (2006) Starvation prevention scheme for the IEEE 802.11e EDCFusing dynamic MAC layer parameters. Wireless Telecommunications Symposium pp. 1–7.

34. Szott S, Natkaniec M, Canonico R & Pach A (2008) Impact of contention window cheatingon single-hop IEEE 802.11e MANETs. Wireless Communications and NetworkingConference, WCNC pp. 1356–1361.

35. IEEE 802.11p (2010) - Specific requirements– Part 11: Wireless LAN Medium AccessControl (MAC) and Physical Layer (PHY) Specifications Amendment 6: Wireless Access inVehicular Environments. Institute of Electrical and Electronics Engineers Inc., New York,USA.

36. ETSI EN 302 663 (2012). Intelligent Transport Systems (ITS); Access layer specificationfor ITS operating in the 5 GHz frequency band.

37. Evans D (2012) The internet of everything.38. Keoh S, Kumar S & Tschofenig H (2014) Securing the Internet of Things: A standardization

perspective. Internet of Things Journal, IEEE 1(3): 265–275.39. ETSI NFV white paper. URI: https://portal.etsi.org/NFV/NFV_White_Paper.pdf, accessed:

2017-02-13.40. McKeown N, Anderson T, Balakrishnan H, Parulkar G, Peterson L, Rexford J, Shenker S &

Turner J (2008) OpenFlow: enabling innovation in campus networks. ACM SIGCOMMComputer Communication Review 38(2): 69–74.

41. OpenStack homepage. URI: http://www.openstack.org, accessed: 2017-02-13.

84

42. Boswarthick D, Elloumi O & Hersent O (2012) M2M Communications: A SystemsApproach. Wiley, 1st edition.

43. Bröring A, Echtefhoff J, Jirka S, Simonis I, Everding T, Stasch C, Liang S & Lemmens R(2011) New generation sensor web enablement. Journal of Sensors 11(3): 219–234.

44. Jutila M, Karhula P, Rivas H & Pantsar-Syväniemi S (2015) End-to-end safety solution forchildren enabled by a wearable sensor vest. Journal of Ubiquitous Systems & PervasiveNetworks 6(1): 33–39.

45. OneM2M homepage, uri: http://www.onem2m.org, accessed: 2017-02-13.46. ETSI Mobile-Edge Computing introductory technical white paper (2014) Mobile-Edge

Computing Industry Initiative.47. Valta M, Jutila M & Jämsä J (2015) IEEE 802.11p and LTE as enablers of cognitive vehicle-

to-infrastructure communication. 6th IEEE Conference on Cognitive Infocommunicationspp. 71–76.

48. ETSI EN 302 637-2 (v1.3.1): Intelligent transport systems (ITS); vehicular communications;basic set of applications; part 2: Specification of cooperative awareness basic service,september 2014.

49. ETSI EN 302 637-3 (v1.2.1): Intelligent transport systems (ITS); vehicular communications;basic set of applications; part 3: Specifications of decentralized environmental notificationbasic service, september 2014.

50. Scholliers J, Bell D, Morris A & García-Meléndez A (2014) Potential of ITS to improvesafety and mobility of VRUs. In: 10th ITS European Congress.

51. Saebboe J, Viikari V, Varpula T, Seppa H, Cheng S, Al-Nuaimi M, Hallbrjorner P & RydbergA (2009) Harmonic automotive radar for VRU classification. International Radar Conferencepp. 1–5.

52. Anund A, Dukic T, Borsbo B & Falkmer T (2010) Piloting smart safe school bus: explorationof security gains from implementation of a driver support system, additional technicalequipment and intelligent bus stops. European Transport Research Review 2(3): 157–163.

53. General motors - developing wireless pedestrian detection technol-ogy. URI: http://media.gm.com/media/us/en/gm/news.detail.html /con-tent/Pages/news/us/en/2012/Jul/0726_pedestrian.html, accessed: 2017-02-13.

54. Engel S, Kratzsch C & David K (2013) Car2pedestrian-communication: Protection ofvulnerable road users using smart phones. Advanced Microsystems for AutomotiveApplications, Smart Systems for Safe, Sustainable and Networked Vehicles pp. 31–41.

55. Liebner M, Klanner F & Stiller C (2013) Active safety for vulnerable road users based onsmartphone position data. IEEE Intelligent Vehicles Symposium pp. 256–261.

56. Enabling connected and electric vehicles, car managementbriefing seminars, traverse city, michigan, 5-8 august, 2013.www.cargroup.org/assets/speakers/presentations/117/080513_car_conference_.pdf,accessed: 2017-02-13.

57. Maruyama K, Chiba T, Kizaki T & Horozovic A (2014) Vehicle-To-X functions forimproved motorcycle safety. Auto Tech Review 3: 50–55.

58. Ravaioli R, Urvoy-Keller G & Barakat C (2015) Characterizing ICMP rate limitation onrouters. In: IEEE International Conference on Communications (ICC), pp. 6043–6049.

59. Allman M, Paxson V & Stewens W (2001). TCP congestion control. IETF Request forComments: 2581.

85

60. Stewens W (1997). TCP Slow Start, Congestion Avoidance, Fast Retransmit, and FastRecovery algorithms. IETF Request for Comments: 2001.

61. Wang Z (2001) Internet QoS: Architectures and Mechanisms for Quality of Service. MorganKaufmann Publishers Inc., San Francisco, CA, USA, 1st edition.

62. Nagle J (1985) On packet switches with infinite storage. Technical report, IETF Request forComments: 970.

63. Sanjay K & Hassan M (2002) Engineering Internet QoS. Artech House, Inc., Norwood,MA, USA.

64. Huitema C (2000) Routing in the Internet. Prentice Hall PTR, Upper Saddle River, NJ,USA, 2nd edition.

65. Floyd S & Jacobson V (1995) Link-sharing and resource management models for packetnetworks. ACM Transactions on Networking 3(4): 365–386.

66. Demers A, Keshav S & Shenker S (1989) Analysis and simulation of a fair queueingalgorithm. In: SIGCOMM ’89: Proceedings in Communications Architectures & Protocols,pp. 1–12.

67. Parekh AK & Gallagher RG (1994) A generalized processor sharing approach to flowcontrol in integrated services networks: the multiple node case. ACM Transactions onNetworking 2(2): 137–150.

68. Crawford L & Marshall A (2001) A dynamic and fast packet scheduling algorithm for openand programmable networks. In: Seventeenth UK Teletraffic Theory Symposium. Dublin,Ireland.

69. Horng M, Lee W, Lee K & Kuo Y (2001) An adaptive approach to weighted fair queuewith QoS enhanced on IP. In: IEEE International Conference on Electrical and ElectronicTechnology, volume 1, pp. 181–186.

70. Sayenko, A, Hämäläinen, T, Joutsensalo, J & Siltanen, J (2003) An Adaptive Approachto WFQ with the Revenue Criterion. In: ISCC ’03: Proceedings of the Eighth IEEEInternational Symposium on Computers and Communications.

71. Sayenko, A, Hämäläinen, T, Joutsensalo, J & Kannisto, L (2006) Comparison and analysisof the revenue-based adaptive queuing models. Computer Networks 50(8): 1040–1058.

72. Akashdeep & Kahlon K (2014) An embedded fuzzy expert system for adaptive WFQscheduling of IEEE 802.16 networks. Expert Systems with Applications 41(16): 7621–7629.

73. Andrews M & Zhang L (2014) Rate-adaptive weighted fair queueing for energy-awarescheduling. Information Processing Letters 114(5): 247–251.

74. Srivastava V & Motani M (2005) Cross-layer design: A survey and the road ahead. IEEECommunications Magazine 43(12): 112–119.

75. Toumpis S & Goldsmith A (2003) Optimization, and cross-layer design of media accessprotocols for wireless ad hoc networks. IEEE International Conference on Communications3: 2234–2240.

76. Paxson V (1996) Towards a framework for defining Internet performance metrics. In:Proceedings of INET, volume 96, pp. 50–56. Citeseer.

77. Prokkola J, Hanski M, Jurvansuu M & Immonen M (2007) Measuring WCDMA and HSDPAdelay characteristics with Qosmet. In: IEEE International Conference on CommunicationsICC’07, pp. 492–498.

78. ITU-T Recommendation G.107 (2011) - The E-model: A computational model for use intransmission planning. URI: http://www.itu.int/, accessed on 2017-02-13.

86

79. Varela M (2005) Pseudo-subjective quality assessment of multimedia streams and itsapplications in control. Ph.d. thesis, University, Rennes, France.

80. IANA (2013) Service name and transport protocol port number registry. IANA Registry.URI: http://www.iana.org/assignments/port-numbers, accessed on 2017-02-13.

81. Karagiannis T, Broido A, Brownlee N, Claffy K & Faloutsos M (2004) Is P2P dyingor just hiding? P2P traffic measurement. In: Global Telecommunications Conference,GLOBECOM ’04, volume 3, pp. 1532–1538.

82. Moore AW & Papagiannaki K (2005) Toward the accurate identification of networkapplications. In: Proceedings of the 6th International Conference on Passive and ActiveNetwork Measurement, PAM’05, pp. 41–54.

83. Lin PC, Lin YD, Lai YC & Lee TH (2008) Using string matching for deep packet inspection.Computer 41(4): 23–28.

84. Hirvonen M & Laulajainen J (2009) Two-phased network traffic classification method forquality of service management. In: 13th IEEE International Symposium on ConsumerElectronics, ISCE ’09, pp. 962–966.

85. Duda R, Hart P & Stork D (2000) Pattern Classification. Wiley-Interscience, Second Edition.86. Keshav S (1991) A control-theoretic approach to flow control. In: Proceedings of ACM

SIGCOMM, volume 21, pp. 3–15.87. Diao Y, Gandhi N, Hellerstein JL, Parekh S & Tilbury D (2002) Using MIMO feedback

control to enforce policies for interrelated metrics with application to the Apache Webserver. IEEE Network Operations and Management, NOMS pp. 219–234.

88. Hellerstein J, Diao Y, Parekh S & Tilbury D (2004) Feedback Control of Computing Systems.Wiley-Interscience.

89. Fehskens L (1989) An Architectural Strategy for Enterprise Management. In: Proceedingsof the IFIP TC6/WG6.6 Symposium on Integrated Network Management, pp. 41–60.

90. Zadeh L (1965) Fuzzy Sets. Information and Control 8: 338 – 353.91. Mamdani E (1976) Advances in linguistic synthesis of fuzzy controllers. International

Journal of Man-Machine Studies 7(1): 1–13.92. Crawley E, Nair R, Rajagoplan B & Sandick H (1998). A Framework for QoS-based

Routing in the Internet. IETF Request for Comments: 2386.93. ITU-T Recommendation E.800. (1993) Terms and definitions related to Quality of Service

and network performance including dependability.94. Chalmers D & Sloman M (1999) A survey of Quality of Service in mobile computing

environments. IEEE Communication Surveys 2(2): 2–10.95. Tanenbaum A (1996) Computer Networks. Prentice Hall, New Jersey, USA, 3rd edition.96. Jha S & Hassan M (2002) Engineering Internet QoS. Norwood, Massachusetts, USA;

Artech House.97. Braden R, Zhang L, Berson S, Herzog S & Jamin S (1997). Resource ReSerVation Protocol

(RSVP). IETF Request for Comments: 2205.98. Peterson LL & Davie BS (2007) Computer networks: a systems approach. Morgan

Kaufmann Publishers Inc., San Francisco, CA, USA, 4th edition.99. Oueslati S & Roberts J (2005) A New Direction for Quality of Service: Flow-Aware

Networking. In: Conference on Next Generation Internet Networks (NGI), pp. 226–232.100. ITU-T recommendation Y.2121 (2008) - Requirements for the support of flow-state-aware

transport technology in NGN.

87

101. Lima S, Carvalho P & Freitas V (2007) Admission control in multiservice IP networks:Architectural issues and trends. Communications Magazine, IEEE 45(4): 114–121.

102. Lima S, Carvalho P & Freitas V (2004) Distributed admission control for QoS and SLSmanagement. Journal of Network and Systems Management 12(3): 397–426.

103. Jamin S, Shenker P & Danzig P (1997) Comparison of measurement-based admission controlalgorithms for controlled-load service. IEEE International Conference on Communications(ICC) pp. 973–80.

104. Kim J & Jamalipour A (2001) Traffic management and QoS provisioning in future wirelessIP networks. Personal Communications, IEEE 8(5): 46–55.

105. Floyd S (1996) Comments on measurement-based admissions control for controlled-loadservices. Technical report.

106. Zhang ZL, Duan Z, Gao L & Hou YT (2000) Decoupling QoS control from core routers:A novel bandwidth broker architecture for scalable support of guaranteed services. In:Proceedings of the Conference on Applications, Technologies, Architectures, and Protocolsfor Computer Communication, SIGCOMM ’00, pp. 71–83.

107. Ping Y, Schlembach J, Skoe A & Knightly E (2001) Design and implementation of scalableedge-based admission control. Computer Networks 37(5): 507–518.

108. Dabrowski M & Strohmeier F (2003) Measurement-based admission control in the AQUILAnetwork and improvements by passive measurements. In: Architectures for Quality ofService in the Internet, pp. 189–202. Springer.

109. Melander B, Bjorkman M & Gunningberg P (2000) A new end-to-end probing and analysismethod for estimating bandwidth bottlenecks. In: Global Telecommunications Conference,GLOBECOM, volume 1, pp. 415–420.

110. Ammar D, Begin T, Guerin-Lassous I & Noirie L (2012) KBAC: Knowledge-basedadmission control. In: IEEE 37th Conference on Local Computer Networks (LCN), pp.537–544.

111. Li M & Hoang DB (2005) FIAC: A resource discovery-based two-level admission controlfor differentiated service networks. Computer Communications 28(18): 2094–2104.

112. Casagrande D, Montessoro P & Blanchini F (2013) Fair and optimal dynamic admissioncontrol of elastic flows. Journal of Computer Networks 57(5): 1277–1288.

113. Frantti, T and Majanen, M (2011) Fuzzy Based Flow Management of Real-Time Traffic forQuality of Service in WLANs. Intech., Croatia, first edition.

114. Yi S, Li C & Li Q (2015) A survey of fog computing: Concepts, applications and issues. In:Proceedings of the 2015 Workshop on Mobile Big Data, Mobidata’15, pp. 37–42. ACM.

115. Rawlings JO, Pantula SG & Dickey DA (1998) Applied regression analysis: a research tool.Springer Science & Business Media.

116. Jacobsen V & Karels M (1998) Congestion avoidance and control. ACM SIGGCOM pp.314–329.

117. Yan C, Farley T & Ye N (2004) Qos requirements of network applications on the internet.Journal on Information, Knowledge and Systems Management 4(1): 55–76.

118. Juuso EK (1993) Linguistic Simulation in Production Control. In: Proceedings of theUKSS’93 Conference of the United Kingdom Simulation Society, pp. 34–38.

119. Frantti T & Jutila M (2009) Fuzzy service classifier for QoS improvement. In: 3rd Interna-tional Conference on Anti-counterfeiting, Security, and Identification in Communication(ASID), pp. 10–15.

120. The network simulator - ns2. URI: http://www.isi.edu/nsnam/ns/, accessed: 2017-02-13.

88

121. Issariyakul T & Hossain E (2008) Introduction to Network Simulator NS2. SpringerPublishing Company, Incorporated, 1st edition.

122. Msadaa I, Cataldi P & Filali F (2010) A Comparative Study between 802.11p and MobileWiMAX-based V2I Communication Networks. In 4th International Conference on NextGeneration Mobile Applications, Services and Technologies (NGMAST) pp. 186–191.

123. NXP press release, NXP and Cohda unveil new V2X device for connected vehicles.URI: http://www.nxp.com/news/press-releases/2013/07/nxp-and-cohda-unveil-new-v2x-device-for-connected-vehicles.html, accessed: 2017-2-13.

124. Pedersen G, Olesen K & Larsten S (1999) Bodyloss for handheld phones. In: IEEEVehicular Technology Conference Proceedings, volume 2, pp. 1580–1584.

89

90

Original publications

I Jutila M (2016) An Adaptive Edge Router Enabling Internet of Things. IEEE Internet ofThings Journal 3(6):1061-1069.

II Jutila M, Scholliers J, Valta M & Kujanpää K (2017) ITS-G5 performance improvementand evaluation for vulnerable road user safety services. IET Intelligent Transport Systems,in press.

III Jutila M, Prokkola J & Triantafyllidou D (2013) Regressive Admission Control Enabledby Real-Time QoS Measurements. International Journal of Computer Networks andCommunications (IJCNC 2013) 5(6):23-43.

IV Frantti T & Jutila M (2009) Embedded fuzzy expert system for Adaptive Weighted FairQueueing. Expert Systems with Applications (Elsevier) 36(8):11390-11397.

V Jutila M & Frantti T (2017) Cognitive Fuzzy Flow Control for Wireless Routers. Inter-national Journal of Autonomous and Adaptive Communications Systems (IJAACS), inpress.

Reprinted with kind permissions from IEEE (I), IET (II), AIRCC (III), Elsevier (IV) andInderscience Publishers (V)

Original publications are not included in the electronic version of the dissertation.

91

92

A C T A U N I V E R S I T A T I S O U L U E N S I S

Book orders:Granum: Virtual book storehttp://granum.uta.fi/granum/

S E R I E S C T E C H N I C A

594. Darif, Bouchra (2016) Synthesis and characterization of catalysts used for thecatalytic oxidation of sulfur-containing volatile organic compounds : focus onsulfur-induced deactivation

595. Juholin, Piia (2016) Hybrid membrane processes in industrial water treatment :separation and recovery of inorganic compounds

596. Augustine, Bobins (2016) Efficiency and stability studies for organic bulkheterojunction solar cells

597. Ylioinas, Juha (2016) Towards optimal local binary patterns in texture and facedescription

598. Mohammadighavam, Shahram (2017) Hydrological and hydraulic design ofpeatland drainage and water treatment systems for optimal control of diffusepollution

599. Louis, Jean-Nicolas (2016) Dynamic environmental indicators for smart homes :assessing the role of home energy management systems in achieving decarbo-nisation goals in the residential sector

600. Mustamo, Pirkko (2017) Greenhouse gas fluxes from drained peat soils : acomparison of different land use types and hydrological site characteristics

601. Upola, Heikki (2017) Disintegration of packaging material : an experimental studyof approaches to lower energy consumption

602. Eskelinen, Riku (2017) Runoff generation and load estimation in drained peatlandareas

603. Kokkoniemi, Joonas (2017) Nanoscale sensor networks : the THz band as acommunication channel

604. Luoto, Petri (2017) Co-primary multi-operator resource sharing for small cellnetworks

605. Yrjölä, Seppo (2017) Analysis of technology and business antecedents forspectrum sharing in mobile broadband networks

606. Suikkanen, Essi (2017) Detection algorithms and ASIC designs for MIMO–OFDMdownlink receivers

607. Niemelä, Ville (2017) Evaluations and analysis of IR-UWB receivers for personalmedical communications

608. Keränen, Anni (2017) Water treatment by quaternized lignocellulose

C609etukansi.kesken.fm Page 2 Wednesday, February 22, 2017 11:26 AM

UNIVERSITY OF OULU P .O. Box 8000 F I -90014 UNIVERSITY OF OULU FINLAND

A C T A U N I V E R S I T A T I S O U L U E N S I S

University Lecturer Tuomo Glumoff

University Lecturer Santeri Palviainen

Postdoctoral research fellow Sanna Taskila

Professor Olli Vuolteenaho

University Lecturer Veli-Matti Ulvinen

Planning Director Pertti Tikkanen

Professor Jari Juga

University Lecturer Anu Soikkeli

Professor Olli Vuolteenaho

Publications Editor Kirsti Nurkkala

ISBN 978-952-62-1520-4 (Paperback)ISBN 978-952-62-1521-1 (PDF)ISSN 0355-3213 (Print)ISSN 1796-2226 (Online)

U N I V E R S I TAT I S O U L U E N S I SACTAC

TECHNICA

U N I V E R S I TAT I S O U L U E N S I SACTAC

TECHNICA

OULU 2017

C 609

Mirjami Jutila

ADAPTIVE TRAFFIC MANAGEMENT IN HETEROGENEOUS COMMUNICATION NETWORKS

UNIVERSITY OF OULU GRADUATE SCHOOL;UNIVERSITY OF OULU,FACULTY OF INFORMATION TECHNOLOGY AND ELECTRICAL ENGINEERING,DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

C 609

ACTA

Mirjam

i JutilaC609etukansi.kesken.fm Page 1 Wednesday, February 22, 2017 11:26 AM