gprs

Traffic Engineering Concepts

for Cellular Packet Radio Networks

with Quality of Service Support

Presented by Yujing Wu

Based on Peter Stuckmann‘s Public PhD defense on20/06/2003

Peter Stuckmann‘s thesis work 20/06/2003

2

Outline

• Motivation and objectives

• Traffic models for existing and future applications• Simulation environment GPRSIM

• Analytical Traffic Engineering Approaches

• GPRS/EDGE performance analysis

• Performance of different applications

• Traffic engineering

• QoS support

• Mutual dependency of traffic engineering and traffic management


3

Motivation: Cost-effective Network Evolution• Traffic Engineering and Traffic Management

Design and upgrade the network in a cost-effective way Based on traffic-performance relation Service differentiation ensured by admission control and scheduling

->Influence on traffic-performance relation

term circuit-switched packet-switchedtraffic offered traffic in

Erlangamount of data per time in kbit/s

QoS parameter blocking probability throughput, delay,...

resources traffic channels packet data channels

tool simple formula or table

dimensioning graphs or tables

methodology Erlang-B formula simulation results, analytical/algorithmic techniques


4

Evolution from 2G to 3G

Requirements for 3G systems:

high data rate (144kbit/s outdoor and 2Mbit/s indoor) ; asymmetric traffic; packet switched; high spectrum efficiency.


5

Assignment of GSM Channels for GPRS

• Packet Data Channels (PDCHs) assigned out of pool of GSM physical channels

• Fixed PDCHs are permanently available

• On-demand PDCHs only available if not used for GSM circuit-switched traffic

pool of GSM physical channels GPRS packet data channels

x fixed PDCHs

y on-demand PDCHs


6

Dimensioning Approach• Dimensioning graphs for application-specific performance measures• Valid for the cell and load scenarios of interest• Applicability: only based on user number/ traffic volume in the busy hour• Accuracy: derived from realistic models for the protocol stacks, traffic patterns and radio channel

QoS

QoS limit

offered trafficacceptable traffic

QoS

QoS limit

offered trafficpredicted traffic

resource configuration 3




7

TE Methodology and Evaluation Scenarios

1. Analytical and algorithmic models: Lack of details of protocol stacks and realistic traffic model

(close-loop control of TCP and heavy tailed traffic)

2. Measurement: Lack of tunable traffic load and different protocol options

3. Simulation: Detailed implementations of GPRS and Internet protocols Traffic generator for common applications Models of the radio channel

Simulation Scenarios:

• Per cell: max PDCH no 8; max IP throughput 80kbits; 1-40 active stations;

• Traffic: Web browsing and email with small obj size; not much WAP traffic; no mobility model.


8

Traffic Management• Increase performance for best-effort services

Coupled RLC/MAC implementation considering urgency of RLC blocks for MAC scheduling

MAC scheduler considering link quality

• Support application-specific QoS (class differentiation on MAC level) Priority queuing Fairer scheduling algorithms introducing weights for traffic classes

QoS

QoS limit 2

QoS limit 1application 1

application 2

acceptable traffic offered traffic (aggregate)

QoS

QoS limit 2

QoS limit 1

acceptable traffic

capacity gain

application 1

application 2

offered traffic (aggregate)


9

Outline







• QoS support



10

Multimedia Traffic Modelling

• Aim definition of user profiles characterization of sessions

• Predicted applications for mobile users Internet (WWW, e-mail, FTP) Wireless Application Protocol (WAP) Streaming (Video & Audio) Video-Conferencing, VoIP

• Methodology Use measurement results for fixed Internet from literature Perform own measurements Use standardized models (e.g. UMTS 30.03) Use market prediction studies


11

WWW Session / Structure of Web Page

Sdfsadfsdasafsdfsafdsadfasfdsafsdfasfdsaf

Sdfsadfsda safsdfsafdSdfsadfsdasafsdfsafd Sdfsad fsdasafsdfsafdSdfs adfs a safs dfsafd Sdfsadfsda safsdfsafd Sdfsadfs da fgdfg dfg afsdfs afd gfdgs fgsdf sdfg sdg sdfg

page 1

Sdfsadfsdasafsdfsafdsadfasfdsafsdfasfdsaf

Sdfsadfsda safsdfsafdSdfsadfsdasafsdfsafd Sdfsad fsdasafsdfsafdSdfs adfs a safs dfsafd Sdfsadfsda safsdfsafd Sdfsadfs da fgdfg dfg afsdfs afd gfdgs fgsdf sdfg sdg sdfg

page n

object 1object 1 object 2object 2 object mobject mtobject page 2 size m

tread

picture

text

links


12

Mosaic Traffic Model [Arlitt and Williamson 1995]

Parameter Distribution Mean Variance

Pages per session Geometric 5 20.0

Reading time between pages [s] Exponential 12 s 144.0

Objects per page [byte] Geometric 2.5 3.75

Object size [byte] Log2-Erlang-k (k=17) 3700 1.36 x 10e6

Transformed Erlang 9.4 5.2


13

Choi’s Behavioral Model of Web Traffic

• Larger WWW pages with higher object sizes

• Not yet suitable for GPRS traffic engineering

• Important when performance of wireless Internet access will be comparable to today‘s fixed networks, e.g. with EGPRS or UMTS


14

E-mail Traffic Model


E-mail size (lower 80%) [byte] Log2-Normal 1700 5.2 x 10e6

Transformed Normal 10.0 2.13

E-mail size (upper 20%) [byte] Log2-Normal 15700 115 x 10e9


Base quota [byte] Constant 300 - - -

• Parameters derived by measurements made at the Lawrence Berkeley Laboratory (California, USA) by Paxson in 1994

• Fixed overhead of 300 byte

• Bimodal distribution of e-mail sizes Lower 80% can be interpreted as text-based mails Upper 20% represents mails with attached files

• Maximum size 100 kbyte


15

WAP Traffic Model


Decks per session Geometric 20.0 3800

Reading time between decks [s] Exponential 14.1 198.8

Packet size 'Get Request' [byte] Log2-Normal 96.1 3.75 x 10e3


Packet size 'Content' [byte] Log2-Normal 562.6 3.5 x 10e5


• Parameters are depending on the content

• Values derived by measurements performed at a WAP gateway in test operation Suitable for introduction scenarios Will change over the next years

(today: 1 kbyte for monochrome decks, 3 kbyte for colored decks)


16

Video Streaming Traffic Model

• Traffic model based on three real video sequences coded with the H.263 codec specified by the ITU-T (similar to MPEG)

• Sequences proposed by the Video Quality Expert Group each one representing a particular group of motion intensity

• Sequences are randomly concatenated producing a continuous video stream

Sequences

Q20 80-10-10 Mix

Claire 10.9 kbit/s

Carphone 26.7 kbit/s 14.39 kbit/ s

Foreman 31.7 kbit/s

Offered I P traffic

}


17

Outline







• QoS support



18

GPRS Architecture


19

GPRSIM

0.20.30.4

0.60.70.80.9

1

10002000 500070008000900010000

Blo

ckin

g R

ate

56 RA Slots24 RA Slots88 RA Slots

0.1

0.5

ChannelMgmt.

4000

LLC(SDL)

sessionmgmt.

6000Offered Load [byte/s]

FrameRelay

BSSGP

3000

BSSGSN

CAC

LLCRelay

SNDCP

BSSGP

FrameRelay

Circuit SwitchedGenerator

RLC/MAC

(SDL)

Transc.

SNDCP

LLC(SDL)

MS

(SDL)

RLC/MAC

Transc.

Channel ErrorModel

bG U m

UplinkbG

Downl.G b

0.10.20.30.40.50.60.70.80.9

0.20.40.60.81 1.21.41.61.82

Railway

Funet

Mobitex

Offered Load (G)

Thro

ug

hp

ut

(S)

GIST Statistical EvaluationWeb Interface

HTTP FTP WAP

WTP

Manager

IP

GPRSim Load Generator

TCP

SMTP Video

UDP

SessionArrivalProcess

SessionArrivalProcess

• Event-driven Simulator based on C++ and SDL

• Prototype implementations of protocol stacks at Mobile Station (MS) Base Station (BS) SGSN

• Stochastic traffic models to generate well-defined traffic load

• Channel and mobility models

• Evaluation and graphical representation

• Validation by measurement


20

Simulation Results

IP user/cell throughput

IP datagram delay

application response time

session blocking rate,

circuit switch call blocking rate

PDCH utilization

assigned PDCHs


21

Transmission time t for a file of size F:

Transition to steady state with the number of Round-tripperiods Nss:

Amount of data Bss transmitted in slow start:

Transmission time t for a file of size F:

Transition to steady state with the number of Round-tripperiods Nss:

Amount of data Bss transmitted in slow start:

Validation I (Analytical TCP Model, Meyer2001])

( )( ) ( ) SS

SS setup LCHTCP

F Bt F N RTT TBF D

R

( )log

log( )

TCP setup

initSS

ss

R RTT TBF

W MSSN

k

SSNinitSS

TCP

W MSSRTT k

R}

}

}1st RTT

2nd RTT

3rd RTT

TCP Client TCP Server

PSH + Data

STEA

DY S

TA

TE

SLO

W S

TA

RT

ACK

Model WWW (3700 byte) e-mail (1 kbyte)Analytical 14.9 kbit/s 22.7 kbit/sSimulation 17.2 kbit/s 22.9 kbit/s

1

1

SSNSS

SS initSS

kB W MSS

k


22

Validation II (Measurement)

0

5

10

15

20

25

30

1 1.5 2 2.5 3 3.5 4 4.5 5

Do

wn

link

IP t

hro

ug

hp

ut

[kb

it/s]

Number of mobile stations

Downlink IP throughput (FTP)

GPRSimMeasured

UmGPRS mobile

Notebook &

iNetwork

IP-Backbone

BSC

SGSNGb

GGSNG

Web

(WinDump)infrared

PPP

Server

BTS

External IP-Network

Internet

Measurement Point

Vodafone NL GPRS measurement settings

• CS-2

• 4 fixed PDCHs

• Multislot (dl/ul) 3/1


23

Outline







• QoS support



24

Fluid-flow Model Approach• Basic concept:

Traffic sources are water taps, being randomly turned on and off Regarded network element is a water reservoir with constant depletion

rate C

• Source model: Markov-modulated Rate Process (MMRP) Single source: behavior controlled by two-state Markov Chain Multiple sources: Superimposing N equal MMRP‘s again leads to an MMRP

• MMRP parameters: ON state probability (activity factor) Mean burst length ENB

Transmission rate during ON state h

10

100

1000

10000

100000

0 1 2 3 4 5 6 7 8 9 10

Me

an

IP

Da

tag

ram

De

lay

[ms]

Number of MS

alpha = 0.187, h = 3272 byte/s , EN_B = 9150 byte

GPRSim with ON/OFF sourcesGPRSim with WWW sources

Fluid-flow Analysis

OFF ONB

h

EN

1


25

MMAP/G/1 Queue [Vornefeld 2002]

• Arrival Process: analytically tractable representation of Choi‘s WWW model using Marked Markovian Arrival Process: Sojourn times in on and off phase approximated with PH-type

distributions (EM algorithm) Poisson arrivals of single IP datagrams during on phase

• Accounts for complicated stochastic nature of arrival process

• Traffic sources can have individual service time distributions

• No batch arrivals of IP datagrams

• Service Process: n-point distribution describing the number of time slots required for transmission of an IP datagram Link-level simulations, models of channel coding and radio channel Each IP packet (576 byte) leads to batch arrival of RLC blocks Size of batch determined by applied Coding Scheme (CS)

• Approximation of n-point distribution by cont. PH-dist. (EM alg.)


26

Result Comparison: System Capacity and CIRScenario parameters:• 1 MS• CS-2• MSC = #PDCHs

1e+01

1e+02

1e+03

1e+04

1e+05

1e+06

5 10 15 20 25 30

Mea

n IP

pac

ket

dela

y [m

s]

Mean C/I [dB]

Simulation 2 PDCHsAnalysis 2 PDCHs

Simulation 4 PDCHsAnalysis 4 PDCHs

Deviations caused by TCP protocol behavior: Batch arrivals on

IP level Slow start and

congestion avoidance (elastic traffic)


27

Outline







• QoS support



28

Performance and system measures

• Application response time for each received file (WAP deck, e-mail or WWW page) the difference

between the date of request from the client (GET request) and the date of reception at the client is calculated

• Downlink IP throughput per user during an ongoing transmission the downlink IP throughput for each user is

calculated for each TDMA frame

• Downlink IP datagram delay for each received IP packet the difference between the date of

transmission (IP data request) and the date of reception (IP data indication) is calculated

• Downlink IP system throughput per radio cell the quotient of the total amount of received IP bytes in one radio cell

divided by the regarded time period equals the offered IP traffic (loss-free system)

• Downlink PDCH utilization the quotient of the number of transmitted radio blocks containing data or

control information divided by the total number of transmitted radio blocks


29

General Simulation Parameter Settings

multislot cap. (DL/UL) 4/1coding scheme CS-2PDCHs fixed 8PDCHs on-demand 0, 8C/I [dB] 12 (BLER = 13.5 %)cluster size 3, 7cell radius [m] 300, 3000MS velocity [km/h] 7, 100TCP version RenoTCP MSS [byte] 512TCP maximum window size [kbyte] 8HTTP version 1.1Traffic mix WWW / email 30% / 70%Traffic mix WAP / WWW / email 60% / 12% / 28% Traffic mix Streaming / WWW / email 10% / 27% / 63%


30

GPRS with Fixed PDCHs

• Maximum user throughput of 22 kbit/s

• Maximum system throughput of 56 kbit/s for 8 fixed PDCHs


31

Effect of Multislot Capability and C/I

• Effect of multislot capability only visible in situations with low traffic load

• Low sensitivity of performance to mean C/I


32

GPRS with on-demand PDCHs

• Performance degradation only occurring with high coexisting speech traffic

• Effect of lower speech traffic visible in situations with medium GPRS traffic


33

Outline







• QoS support



34

WAP vs. Conventional Internet Applications (I)

• WAP and e-mail response times remain below 5 s for the whole load range for pure traffic scenarios, while WWW exceeds 30 s

• In the traffic mix scenario (60% WAP, 28% email and 12% WWW), WWW performance increases, while e-mail and WAP performance decreases slightly


35

WAP vs. Conventional Internet Applications (II)

• Low throughput performance for WAP because of small deck size

• E-mail performance remains stable in pure traffic scenario because of low offered traffic per session

• Similar behavior of WWW and e-mail in traffic mix scenario because of equal load conditions


36

Streaming vs. Background Applications (I)

• Streaming performance in traffic mix scenario stable over the whole load range for EGPRS, up to 20 MSs for pure Streaming

• For GPRS only 5 MSs (pure) and 15 MSs (mix) acceptable for Streaming applications

• WWW performance only affected in GPRS scenario


37

Outline







• QoS support



38

Dimensioning for Fixed and On-demand PDCHs

• Dimensioning graph for fixed PDCHs based on the performance for different resource configurations over the offered IP traffic

• Dimensioning graph for on-demand PDCHs based on the performance for different coexisting speech loads over the offered IP traffic


39

Traffic Engineering Rules

1)Define the QoS target

2)Estimate the number of users per cell

3)Define the offered IP traffic per user

4)Calculate the offered IP traffic per cell

5)Regard the operating point p defined by the QoS target on the y-axis and the offered traffic per cell on the x-axis and choose the next curve that lies above p

6)Result: Number of fixed PDCHs to be allocated or the acceptable coexisting speech traffic


40

Outline







• QoS support



41

Dimensioning Graphs without QoS support

• Streaming performance starts to decrease with an offered traffic of 20 kbit/s and 4 fixed PDCHs

• Streaming application can be seen as the critical application


42

Dimensioning Graphs with QoS support

• In using DWRR the performance of Streaming applications can be increased

• Depending on the QoS target for lower prioritized applications a resource configuration with 4 fixed PDCHs might be sufficient


43

Conclusions

• Traffic engineering rules for the cost-effective evolution of cellular packet radio networks Requirements: applicability and accuracy Approach: traffic models and prototype implementation (GPRSIM) Result: Dimensioning graphs for fixed and on-demand configurations

• Advanced traffic management techniques Proposed scheduling algorithms for best-effort services

• DPARR very effective and easy to implement Proposed scheduling algorithms for traffic class support

• Solution should be based on the operator´s strategy Connection admission control parameterization

• Mutual dependency of traffic engineering and traffic management Estimate the effects of QoS support and best-effort scheduling on

traffic engineering rules Stay inline with network evolution


44

Research Contributions

• Development of a comprehensive GPRS/EDGE emulation tool for radio interface performance analysis and capacity planning

• Identification and development of traffic models for existing and future mobile applications

• Comprehensive performance analysis for GPRS and EDGE networks considering a wide range of applications and system parameters

• Derivation of radio resources traffic engineering rules for the cost-effective evolution of cellular packet radio networks

• Development and performance evaluation of advanced QoS management algorithms for cellular packet radio networks

• Book publication “The GSM Evolution” (Wiley 2002)

• 2 journal publications

• More than 20 conference papers

• 1 patent


45

What can we learn from this work?

Thoughts on TE of CDMA Cellular Networks

• Can we borrow the TE methodology in this work?

• Survey of simulators of CDMA networks (not complete yet):

•NS-2, Glomosim, SSF, Telesim: not provide.

any other free network simulator?

•Several commercial products: e.g. Opnet wireless module, MACdma, Netplan (Motorola), CELLsim (Nomad Access) etc.

• At the initial stage, can we build a simple simulator (without implementation of full protocol stack) for a good enough evaluation? Must consider the key features of CDMA systems: interference-limited capacity.

• Theoretic analysis is always a good starting point.

gprs

Documents

wap traffic

traffic patterns

traffic engineering

objectives traffic models

traffic classes qos

traffic management design

active stations traffic

heavy tailed traffic