Multimedia computer networks quality of service techniques evaluation and development.

DOGMAN, Aboagela A.

Available from the Sheffield Hallam University Research Archive (SHURA) at:

http://shura.shu.ac.uk/19569/

A Sheffield Hallam University thesis

This thesis is protected by copyright which belongs to the author.

The content must not be changed in any way or sold commercially in any format or medium
without the formal permission of the author.

When referring to this work, full bibliographic details including the author, title, awarding
institution and date of the thesis must be given.

Please visit http://shura.shu.ac.uk/19569/ and http://shura.shu.ac.uk/information.html for

further details about copyright and re-use permissions.
 G/VU is P!%j9L)S
snd 't;' A.fo 'sjjuoQ snaspy
S??'/v ti0,|L d^OjUf p u s 6UIUJB9-|
ti«s.:,»,un ojepH Piayjaijg .

REFERENCE
 ProQuest N um ber: 10694450

All rights reserved

INFORMATION TO ALL USERS

The quality of this reproduction is d e p e n d e n t upon the quality of the copy subm itted.

In the unlikely e v e n t that the a u thor did not send a c o m p le te m anuscript

and there are missing pages, these will be noted. Also, if m aterial had to be rem oved,
a n o te will ind ica te the deletion.

uest
ProQuest 10694450

Published by ProQuest LLC(2017). C opyright of the Dissertation is held by the Author.

All rights reserved.

This work is protected against unauthorized copying under Title 17, United States C o d e
M icroform Edition © ProQuest LLC.

ProQuest LLC.
789 East Eisenhower Parkway
P.O. Box 1346
Ann Arbor, Ml 4 8 1 0 6 - 1346
Multimedia Computer Networks Quality of Service
Techniques Evaluation and Development

Aboagela A. Abolghasm Dogman

A thesis submitted in partial fulfilment of the requirements of

Sheffield Hallam University
for the degree of
Doctor of Philosophy

February 2014
 ABSTRACT
Quality of Service Evaluation and Improvement of Multimedia Computer
Networks

The growth in the transmission of time-sensitive applications over computer networks means
that Quality of Service (QoS) needs to be managed in an efficient manner. Network QoS
management in this thesis refers to evaluation and improvement of QoS provided by integrated
wired and wireless computer networks. Evaluation of QoS aims to analyse and quantify network
performance with respect of meeting multimedia applications' transmission requirements. QoS
improvement involves the ability to take actions to change network performance toward
improved operation. Therefore, the main aims of this thesis are: (i) to develop techniques for
evaluation QoS in multimedia computer networks, (ii) to develop techniques that uses the
information from (i) to manage and improve network performance.

Multimedia traffic generates a large amount of data. Collecting this information poses a
challenge as it needs to be sufficiently fast and accurate. A contribution of this thesis is that
adaptive statistical sampling techniques to sample multimedia traffic were developed and their
effectiveness was evaluated. Three different adjustment mechanisms were incorporated into
statistical sampling techniques to adjust the traffic sampling rate: simple linear adjustment,
quarter adjustment, and Fuzzy Inference System (FIS). The findings indicated that the
developed methods outperformed the conventional non-adaptive sampling methods of
systematic, stratified and random.

The data collected included important QoS parameters, i.e. delay, jitter, throughput, and packet
loss that indicated network performance in delivering real-time applications. An issue is that
QoS needs evaluation in an informative manner. Therefore, the second contribution of this
thesis is that statistical and Artificial Intelligent (AI) techniques were developed to evaluate QoS
for multimedia applications. The application’s QoS parameters were initially analysed either by
Fuzzy C-Means (FCM) clustering algorithm or by Kohonen neural network. The analysed QoS
parameters were then used as inputs to a regression model or Multi-Layer Perceptron (MLP)
neural network in order to quantify the overall QoS. The proposed QoS evaluation system
differentiated the network’s QoS into a number of levels (Poor to Good QoS) and based on this
information, the overall network’s QoS was successfully quantified. In order to facilitate QoS
assessment, a portable hand-held device for assessing the QoS in multimedia networks was
designed, regression model was implemented on the microcontroller board and its performance
was successfully demonstrated.

Multimedia applications transmitted over computer networks require a large bandwidth that is a
critical issue especially in wireless networks. The challenge is to enable end-to-end QoS by
providing different treatments for different classes of traffic and efficient use of network
resources. In this thesis, a new QoS enhancement scheme for wireless-wired networks is
developed. This scheme consisted of an adaptive traffic allocation algorithm that is incorporated
into the network's wireless side to improve the performance of IEEE 802.1 le Enhanced
Distributed Channel Access (EDCA) protocol, and a Weighted Round Robin (WRR) queuing
scheduling mechanism that was incorporated into the wired side. The proposed scheme
improved the QoS for Multimedia applications. The average QoS for voice, and video
applications were increased from their original values by 72.5%, and 70.3% respectively.
 DEDICATION

I dedicate this work to:

o To the soul o f my mother, I still owe to her. Without her encouragement when
she was a live, I could not be able to reach this stage.
o My father, fo r his invaluable support throughout the years, who devoted his life
to the achievement o f this dream,
o My lovely wife whose unconditional love makes everything possible,
o My sons Yousf, Ahmed, and Abdul-Aziz who make every day new and precious.
o My brothers, sisters, and friends who shared with me my dream.
 ACKNOWLEDGMENTS

In the name of Allah, the Most Gracious, the Most Merciful.

All thanks to Almighty ALLAH for giving me the guidance and the strength to achieve
my aim of obtaining PhD degree.

First, I would like to express my deep thanks to my director of studies, Dr. Reza
Saatchi, for his guidance, support, encouragement, valuable suggestions and comments
on various aspects throughout my PhD study. The initial help I have received from Dr.
Reza in setting up my research focus and his subsequent professional support
throughout my research process were very valuable. I really appreciate your support!

Special thanks to my supervisor, Dr. Samir Al-Khayatt for being very helpful and
supportive. His valuable guidance, advice, and expertise throughout my research were
really helpful. Dr. Samir Al-Khayatt taught me how a supervisor can be a sincere friend.
I will never forget your help and support!

I also wish to thank researchers, who have been, or still are, PhD candidates within the
C3RI as well as many of the departmental staff, our friendly discussions helped me gain
confidence in my research.
 LIST OF PUBLICATIONS

• Dogman, A., Saatchi, R., and Al-Khayatt, S. (2014), Collaborative Scheme to

Improve QoS in WLAN-Wired Networks, The Mediterranean Journal of
Computers and Networks, Accepted for publication.
• Dogman, A., and Saatchi, R. (2014), Multimedia Traffic Quality of Service
Management Using Statistical and Artificial Intelligence Techniques, Institute of
Engineering and Technology Journal of Circuits, Devices, and Systems, Accepted
for publication.
• Dogman, A., Saatchi, R., and Al-Khayatt, S. (2013), Network Quality o f Service
Assessment Implementation Using a Microcontroller Board, Malaysian Journal of
Fundamental and Applied Sciences, Vol.9, No.2, pp 57 - 61.
• Dogman, A., Saatchi, R., and Al-Khayatt, S. (2012), Quality of Service Evaluation
using a Combination of Fuzzy C-Means and Regression Model, International
Journal of Electronics and Electrical Engineering, World Academy of Science
Engineering and Technology (WASET12), Vol. 6, pp. 58 - 65.
• Dogman, A., Saatchi, R., and Al-Khayatt, S. (2012), Evaluation o f Computer
Network Quality of Service Using Neural Networks, In Proceedings of IEEE
Symposium on Business, Engineering and Industrial Applications (ISBEIA12),
23rd - 26th September, Bandung, Indonesia, UK, IEEE Xplore, pp. 217 - 222.
• Dogman, A., Saatchi, R., and Al-Khayatt, S. (2012), Computer Network Quality of
Service Monitoring Using KEIL ARM Microcontroller, In Proceedings o f Regional
Annual Fundamental Science Symposium (RAFSS12), 10th - 13th December,
Persada Johor International Convention Centre, Johor Bahru, Malaysia.
• Dogman, A., Saatchi, R., and Al-Khayatt, S. (2012), Improving Quality of Service
in IEEE 802.1 le Enhanced Distributed Channel Access Protocol, In Proceedings of
the 8th IEEE International Symposium on Communication Systems Networks and
Digital Signal Processing (CSNDSP12), 18th - 20th July, Poznan University,
Poznan, Poland, UK, IEEE Xplore, pp. 1 - 6 .
• Dogman, A., Saatchi, R., Nwaizu, H., and Al-Khayatt, S. (2011), Adaptive
Statistical Sampling of VoIP Traffic in WLAN and Wired Networks using Fuzzy
Inference System, In Proceedings of the 7th IEEE International Wireless

v
Communications and Mobile Computing Conference (IWCMC11), 4th - 8th July,
Bahcesehir University, Istanbul, Turkey, UK, IEEE Xplore, pp. 1731 - 1736.
Dogman, A., Saatchi, R., and Al-Khayatt, S. (2010), An Adaptive Statistical
Sampling Technique for Computer Network Traffic, In Proceedings of the 7th
IEEE, DET International Symposium on Communication Systems Networks and
Digital Signal Processing (CSNDSP10), 21st -23rd July, University of
Northumbria, Newcastle, UK, IEEE Xplore, pp. 479 - 483.
Dogman, A., Saatchi, R., and Al-Khayatt, S. (2010), Evaluation of Adaptive
Statistical Sampling versus Random Sampling for Video Traffic, In Proceedings of
the 11th International Arab Conference on Information Technology (ACIT10),
14th-16th December, University of Benghazi, Benghazi, Libya.
 TABLE OF CONTENTS

ABSTRACT............................................................................................................................................... H

DEDICATION.......................................................................................................................................... HI

ACKNOWLEDGMENTS....................................................................................................................... IV

LIST OF PUBLICATIONS.......................................................................................................................V

TABLE OF CONTENTS........................................................................................................................VII

LIST OF FIGURES................................................................................................................................. XI

LIST OF TABLES........................................ XV

GLOSSARY TERMS...........................................................................................................................XVII

LIST OF VARIABLES...........................................................................................................................XX

CHAPTER 1 INTRODUCTION...............................................................................................................1

1. 1 R e se a r c h M o t iv a t io n s ....................................................................................................................................1

1.2 R e se a r c h A im and O b je c t iv e s .....................................................................................................................4

1.3 R e se a r c h C o n t r ib u t io n s .............................................................................................................................. 4

1.4 T hesis O r g a n isa t io n ........................................................................................................................................ 7

CHAPTER 2 RELEVANT THEORY AND BACKGROUND

2.1 I n t r o d u c t io n .............................................................................

2.2 Q ua l ity of S e r v ic e (Q o S): A n o v e r v ie w .......................

2.2.7 Definition of QoS:..................................................

2.2.2 QoS Parameters.................................................................................................................. 72
2.2.2.1 Throughput..........................................................................................................................................12
2.2.2.2 Delay ............................................................................................................................................. 13
22.23 Jitter......................................................................................................................................................13
2.2.2.4 Packet Loss R atio............................................................................................................................... 14
2.2.3 QoS Requirements o f Multimedia Applications................................................................... 14
2.2.4 Service Levels o f QoS.......................................................................................................... 75
2.2.5 QoS Components..................................................................................................................75
2.3 QoS in W ireless a n d W ired N e t w o r k s .................................................................................................. 16
2.3.1 QoS in Wireless Networks....................................................................................................16
2.3.1.1 IEEE 802.1 le Standard...................................................................................................................... 17
2.3.1.1.1 HCF Controlled Channel Access (HCCA)................................................................................17
2.3.1.1.2 Enhanced Distributed Channel Access (EDCA)....................................................................... 19
2.3.2 QoS in Wired Networks........................................................................................................22
2.3.2.1 Packet Scheduling Mechanisms........................................................................................................ 23
2.3.2.1.1 First-In, First-Out (FIFO) queuing mechanism......................................................................... 23
2.3.2.1.2 Priority Queuing Mechanism (PQ).............................................................................................24
vii
 2.3.2.1.3 Fair Queuing Mechanism (FQ)................................................... 25
2.3.2.1.4 Weighted Fair Queuing Mechanism (WFQ).............................................................................. 25
2.3.2.1.5 Weighted Round Robin queuing mechanism (WRR)...............................................................26
2.4 S t a t ist ic a l and A r t if ic ia l I n t e l l ig e n c e T e c h n iq u e s ..................................................................... 28

2.4.1 Regression Model................................................................................................................. 28

2.4.1.1 Multi-Linear Regression Model...................................................................................................... 29
2.4.2 Fuzzy logic...........................................................................................................................30
2.4.2.1 Fuzzy Inference System (F IS)........................................................................................................ 30
2.4.2.1.1 Fuzzification................................................................................. 31
2.4.2.1.2 Rule Base....................................................................................................................................... 31
2.4.2.1.3 Inference Engine............................................................................................................................32
2.4.2.1.4 Defuzzification..............................................................................................................................32

2.4.3 Fuzzy clustering......................................... 34

2.4.3.1 Fuzzy C-Means Clustering (FCM )................................................................................................. 34
2.4.4 Artificial Neural Network (ANN)..........................................................................................36
2.4.4.1 Multi-Layer Perceptron (MLP)....................................................................................................... 37
2.4.4.2 Kohonen Neural Network................................................................................................................ 39
2.5 S u m m a r y ............................................................................................................................................................... 4 0

CHAPTER 3 LITERATURE REVIEW.................................................................................................41

3.1 I n t r o d u c t io n .......................................................................................................................................................41

3.2 S a m pl in g A ppr o a c h e s fo r M ea su r in g Q o S P a r a m e t e r s .................................................................42

3.2.1 Non-adaptive Sampling Techniques..................................................................................... 42

3.2.2 Adaptive Sampling Techniques............................................................................................. 43
3.3 N e t w o r k Q u a l it y of S e r v ic e E v a l u a t io n ............................................................................................ 45

3.3.1 Quality o f Service Analysis Techniques............................................................................... 46

3.3.2 Quality o f Service Measurement Techniques.......................................................................47
3.4 I m pr o v in g Q u a l ity of S e r v ic e in C o m pu t e r N e t w o r k s .................................................................. 51

3.5 Q u a l ity of S e r v ic e M o n it o r in g T o o l s ....................................................................................................55

3.6 A ppl ic a t io n of S t a t ist ic a l and A r t if ic ia l I n t e l l ig e n t T ec h n iq u e s to C om puter

N e t w o r k .......................................................................................................................................................................... 56

3.7 S u m m a r y ............................................................................................................................................................... 58

CHAPTER 4 EXPERIMENTAL METHODOLOGY..........................................................................59

4.1 I n t r o d u c t io n ....................................................................................................................................................... 59
4.2 N e t w o r k E v a l u a t io n A p p r o a c h e s ............................................................................................................59

4.2.1 Network Simulation..............................................................................................................60

4.2.1.1 An overview of Evalvid Framework.............................................. 61
4.2.2 Network Topologies..............................................................................................................63
4.2.3 Physical Layer (PHY) Parameters.......................................................................................64
4.2.4 Medium Access Layer (MAC) Parameters...........................................................................64
4.2.5 Routing Protocols.................................................................................................................65
4.2.6 Queuing Mechanisms............................................................................................................ 66
4.2.7 Traffic Type and Traffic Characteristics............................................................................... 66
viii
 4.3 A n a l y sis of S im u l a tio n O u t p u t .............................................................................................................. 68

4.3.1 Transmission Requirements o f Applications........................................................................68

4.3.2 Calculation of QoS Parameters............................................................................................69
4.4 S u m m a r y ............................................................................................................................................................. 70

CHAPTER 5 DEVELOPMENT AND EVALUATION OF ADAPTIVE STATISTICAL

SAMPLING TECHNIQUES FOR MULTIMEDIA TRAFFIC............................................... 71

5.1 I n t r o d u c t io n .......................................................................................... 71

5.2 R e l a t e d W o r k .................................................................................................................................................. 72

5.3 A d a p t iv e S ta t ist ic a l S a m pl in g A p p r o a c h e s ......................................................................................72

5.3.1 Description o f Statistical Sampling Algorithm.....................................................................73

5.3.2 Adjustment of Inter-Sampling Section Interval (ISSI)..........................................................75
5.3.2.1 Linear Adjustment Mechanism of ISSI............................................................................................ 76
5.3.2.2 Quarter Adjustment Mechanism o f ISSI.......................................................................................... 76
5.3.2.3 ISSI Adjustment using Fuzzy Inference System............................................................................. 77
5.3.3 Implementations of Conventional Sampling Techniques......................................................80
5.3.4 Calculation of Sampling QoS Parameters and Sampling Analysis......................................80
5.3.5 Network Topology and Traffic Characteristics....................................................................82
5.4 R esu l t s and D is c u s s io n ................................................................................................................................83

5.4.1 Throughput........................................................................................................................... 83
5.4.2 Delay....................................................................................................................................90
5.4.3 Jitter.....................................................................................................................................95
5.4.4 Packet Loss Ratio............................................................................................................... 100
5.5 S u m m a r y ............................................................................................................................................................105

CHAPTER 6 TECHNIQUES TO EVALUATE NETWORK QUALITY OF SERVICE USING

STATISTICAL AND ARTIFICIAL INTELLIGENCE......................................................................106

6.1 I n t r o d u c t io n ................................................................................................................................................. 106

6.2 R e l a t e d W o r k s ...............................................................................................................................................106

6.3 D esc r ipt io n of the A p p r o a c h e s .............................................................................................................. 107

6.3.1 Analysis of QoS using Fuzzy C-means Clustering Algorithm............................... 109

6.3.2 Analysis o f QoS using Kohonen Neural Network...............................................................110
6.3.3 QoS Assessment using Regression Model.......................................................................... 112
6.3.4 QoS Assessment using Multi-Layer Perceptron................................................................. 114
6.3.5 Measuring Predication Accuracy....................................................................................... 116
6.3.6 Network Simulation and Traffic Models............................................................................. 117
6.4 R e su lts and D is c u s s io n .............................................................................................................................. 118

6.4.1 QoS Analysis using FCM Clustering Algorithm................................................................. 118

6.4.2 QoS Assessment using Regression Model:......................................................................... 123
6.4.3 QoS Analysis using Kohonen Neural Network................................................................... 127
6.4.4 QoS Assessment using Multi Layer Perceptron.................................................................. 130
6.5 S u m m a r y ............................................................................................................................................................ 134
C H A P T E R 7 IM P R O V E M E N T S IN Q U A L IT Y O F S E R V IC E IN C O M P U T E R N E T W O R K S 136

7.1 I n t r o d u c t io n ...................................................................................................................................................136

7.2 R e l a t e d W o r k ..............................................................................!..................................................................136

7.3 D e sc r ipt io n of Q o S E n h a n c e m e n t A p p r o a c h .................................................................................. 138

7.3.1 Adaptive Traffic Allocation Algorithm............................................................................... 138

7.3.1 Integration of Weighted Round Robin Queuing Mechanism.............................................. 140
7.1 N e t w o r k M o d e l l in g and S im u l a t io n .................................................................................................. 141

7.2 R e su l t s and D is c u s s io n ..............................................................................................................................143

7.2.1 Delay....................................................................................... 143

7.2.2 Jitter....................................................... 146
7.2.3 Packet loss ratio................................................................................... 149
7.2.4 Overall QoS.........................................................................................................................152
7.3 S u m m a r y ............................................................................................................................................................156

C H A P T E R 8 M IC R O C O N T R O L L E R B O A R D IM P L E M E N T A T IO N O F Q U A L IT Y O F

S E R V IC E A S S E S S M E N T S Y S T E M ....................................................................................................................... 157

8.1 I n t r o d u c t io n ................................................................................................................................................... 157

8.2 R e l a t e d W o r k .................................................................................................................................................158

8.3 K E IL A R M M C B 2300 E v a l u a t io n B o a r d ........................................................................................... 159

8.3.1 MCB2300 Hardware Components..................................................................................... 159

8.3.2 MDK-ARM Microcontroller Development Kit................................................................... 159
8.4 Q oS A sse ssm e n t I m p l e m e n t a t io n U sin g K E IL M C B 2300 A R M ................................................ 161

8.4.1 Proposed Regression Model to Assess QoS........................................................................ 162

8.4.2 Implementation o f QoS Assessment Technique using KEIL ARM Microcontroller 163
8.5 E x pe r im e n t a l p r o c e d u r e ........................................................................................................................... 165

8.5.1 Modelling and Simulation...................................................................................................165

8.5.2 Hardware and Software Setup.............................................................................................166
8.5.2.1 Hardware S etup.................................................................................................................................166
8.5.2.2 Software Setup ....................................................................................................................168
8.6 R e su lts and D is c u s s io n .............................................................................................................................. 170

8.6.1 VoIP Traffic......................................................................................................................... 170

8.6.2 Video Traffic........................................................................................................................172
8.7 S u m m a r y ............................................................................................................................................................174

C H A P T E R 9 C O N C L U S IO N S , A N D F U T U R E W O R K ................................................................................. 175

9.1 C o n c l u s io n s .....................................................................................................................................................175

9.2 F u t u r e W o r k ................................................................................................................................................... 178

R E F E R E N C E S .................................................................................................................................................................. 180

x
 LIST OF FIGURES

F ig u r e 1-1. T h e sc h e m a t ic o v e r v ie w o f t h e t h e s is ............................................................................................. 10

F ig u re 2-1. T h e c o n st r u c t io n of IE E E 802.1 1e H C C A ....................................................................................... 18

F ig u re 2-2. U pl in k a n d d o w n l in k tr a n sm issio n b e t w e e n AP a n d w ir e l e ss s t a t io n .......................... 18

F ig u re 2-3. T h e IE E E 802.1 1e ED C A m odel. ....................................................................................................... 19

F ig u r e 2-4. IE E E 802.1 1e E D C A o p e r a t io n ..............................................................................................................21

F ig u r e 2-5. T h e pro cess o f FIFO sc h e d u l in g m e c h a n ism ................................................................................. 24

F ig u r e 2-6. P r io r it y q u e u in g sc h e d u l in g m e c h a n ism ....................................................................................... 24

F ig u re 2-7. F a ir q u e u in g sc h e d u l in g m e c h a n ism ................................................................................................25

F ig u re 2-8. W F Q u sin g a w e ig h t e d b it - b y - b it ro u n d r o b in s c h e d u l e r .....................................................26

F ig u re 2-9. T h e o pe r a t io n o f W R R .............................................................................................................................27

F ig u r e 2-10. B in a ry lo g ic v e r su s f u zzy l o g ic ..................................................................................................... 30

F ig u r e 2-11. B l o c k d ia g r a m o f fu z z y in fe r e n c e sy st e m . ............................................................................... 30

F ig u r e 2-12. D eg r ees o f m e m b e r sh ip in G a u ssia n m e m b e r sh ip f u n c t io n ..................................................31

F ig u re 2-13. T h e pr o c e ss o f f u z z y in fe r e n c e sy st e m ........................................................................................ 33

F ig u r e 2-14. M u l t i -L a y e r P e r c e pt r o n : ( a ) T h e a r c h it e c t u r e , ( b ) M LP o p e r a t io n ............................. 38

F ig u r e 2-15. T h e st r u c t u r e o f k o h o n e n n e u r a l n e t w o r k ..............................................................................39

F ig u r e 3-1. C la ssific a t io n o f s a m pl in g t e c h n iq u e s : ( a ) Non a d a pt iv e sa m pl in g , ( b ) T he concept

OF ADAPTIVE SAMPLING............................................................................................................................................43

F ig u r e 3-2. I n teg ra ted W L A N - w ir e d n e t w o r k ................................................................................................... 51

F ig u r e 4-1. T h e sim u la tio n pr o c e ss o f N S -2.......................................................................................................... 61

F ig u r e 4-2. T h e pro cess o f tr a n s m it t in g v id e o u sin g E v a l v id fram ew ork and N S -2 .......................62

F ig u r e 4-3. T h e sim u la te d n e t w o r k to po l o g y ................... 63

F ig u re 4-4. G O P s e q u e n c e in M P E G -4 .......................................................................................................................67

F ig u r e 5-1. O pe r a t in g pa r a m e t e r s o f a d a pt iv e st a t ist ic a l s a m pl in g a l g o r it h m ............. 73

F ig u r e 5-2. T h e flo w c h a r t o f t h e a d a pt iv e s t a t ist ic a l sa m pl in g a l g o r it h m .................................... 75

F ig u r e 5-3. B lo c k d ia g r a m o f FIS fo r a d ju st in g ISSI: ( a ) F u z z y in pu t s , ( b ) Fuzzy o u t p u t ............. 78

F ig u r e 5-4. C o m pa r iso n o f t h r o u g h pu t w it h its sa m pl e d v e r sio n s u s in g : ( a ) A c t u a l , (b )

A d a pt iv e s a m pl in g b a se d o n FIS, (c ) A d a pt iv e s a m pl in g l in e a r a d ju st m e n t , ( d ) A d a pt iv e

SAMPLING QUARTER ADJUSTMENT, (E) SYSTEMATIC, (F) STRATIFIED, (G) RANDOM SAMPLING...........84

F ig u re 5-5. T h e len g th o f ISS I fo r a d a pt iv e s t a t ist ic a l sa m pl in g u s in g : ( a ) F uzzy in fe r e n c e

s y st e m , ( b ) L in ea r a d ju st m e n t m e c h a n ism , (c ) Q uarter a d ju st m e n t m e c h a n is m ................... 86

F ig u r e 5-6. C o m pa r iso n o f bia s o f sa m pl e d t h r o u g h pu t o b t a in e d fr o m n o n - a d a pt iv e sa m pl in g

WITH BIAS OBTAINED USING: (A) ADAPTIVE SAMPLING BASED ON FIS, (B) ADAPTIVE SAMPLING

BASED ON LINEAR ADJUSTMENT, (C) ADAPTIVE SAMPLING BASED ON QUARTER ADJUSTMENT 88

F ig u r e 5-7. R S E o f sa m ple d t h r o u g h pu t u s in g c o n v e n t io n a l s a m pl in g v e r s u s : ( a ) A d a p t iv e

SAMPLING BASED ON FIS, (B) ADAPTIVE SAMPLING BASED ON LINEAR ADJUSTMENT, (C) ADAPTIVE

SAMPLING BASED ON QUARTER ADJUSTMENT.............................................................................. 89

F ig u r e 5-8. C o m pa r iso n o f d e l a y w it h its sa m pl e d v e r sio n s u s in g : ( a ) A c t u a l , ( b ) A d a pt iv e

SAMPLING BASED ON FIS, (C) ADAPTIVE SAMPLING LINEAR ADJUSTMENT, (D) ADAPTIVE SAMPLING

QUARTER ADJUSTMENT, (E) SYSTEMATIC, (F) STRATIFIED, (G) RANDOM SAMPLING...............................91

F ig u r e 5-9. C o m pa r iso n o f bia s o f sa m pl e d d e l a y o b t a in e d fro m n o n - a d a pt iv e s a m pl in g w it h

BIAS OBTAINED USING: (A) ADAPTIVE SAMPLING BASED ON FIS, (B) ADAPTIVE SAMPLING BASED ON

LINEAR ADJUSTMENT, (C) ADAPTIVE SAMPLING BASED ON QUARTER ADJUSTMENT................................93

F ig u r e 5-10. R S E o f sa m pl e d d e l a y u s in g c o n v e n t io n a l sa m pl in g v e r s u s : ( a ) A d a p t iv e

SAMPLING BASED ON FIS, (B) ADAPTIVE SAMPLING BASED ON LINEAR ADJUSTMENT, (C) ADAPTIVE

SAMPLING BASED ON QUARTER ADJUSTMENT..................................................................................................... 94

F ig u r e 5-11. C o m pa r iso n o f jit t e r w it h its sa m pl e d v er sio n s u s in g : ( a ) A c t u a l , ( b ) A d a p t iv e

SAMPLING BASED ON F IS , (C) ADAPTIVE SAMPLING LINEAR ADJUSTMENT, (D) ADAPTIVE SAMPLING

QUARTER ADJUSTMENT, (E) SYSTEMATIC, (F) STRATIFIED, (G) RANDOM SAMPLING............................... 96

F ig u r e 5-12. C o m pa r iso n o f bia s o f sa m ple d jit t e r o b t a in e d fr o m n o n - a d a p t iv e s a m pl in g w it h

BIAS OBTAINED USING: (A) ADAPTIVE SAMPLING BASED ON FIS, (B) ADAPTIVE SAMPLING BASED ON

LINEAR ADJUSTMENT, (C) ADAPTIVE SAMPLING BASED ON QUARTER ADJUSTMENT................................ 98

F ig u r e 5-13. R S E o f sa m ple d jit t e r u s in g c o n v e n t io n a l s a m pl in g v e r s u s : ( a ) A d a p t iv e sa m pl in g

BASED ON FIS, (B) ADAPTIVE SAMPLING BASED ON LINEAR ADJUSTMENT, (C) ADAPTIVE SAMPLING

BASED ON QUARTER ADJUSTMENT..........................................................................................................................99

F ig u r e 5-14. P a c k e t loss r a t io w it h its sa m pl e d v er sio n s u s in g : ( a ) A c t u a l , ( b ) A d a p t iv e

SAMPLING BASED ON FIS, (C) ADAPTIVE SAMPLING LINEAR ADJUSTMENT, (D) ADAPTIVE SAMPLING

QUARTER ADJUSTMENT, (E) SYSTEMATIC, (F) STRATIFIED, (G) RANDOM SAMPLING............................. 101

F ig u r e 5-15. C o m pa r iso n o f bia s o f sa m pl e d p a c k e t lo sse s o b t a in e d f r o m n o n - a d a pt iv e

SAMPLING WITH BIAS OBTAINED FROM ADAPTIVE SAMPLING BASED ON: (A) FIS, (B) LINEAR

ADJUSTMENT APPROACH, (C) QUARTER ADJUSTMENT APPROACH............................................................... 103

F ig u r e 5-16. R S E o f sa m ple d p a c k e t loss r a t io u sin g c o n v e n t io n a l sa m pl in g v e r s u s : ( a )

A d a pt iv e sa m pl in g b a se d o n FIS, ( b ) A d a pt iv e sa m pl in g b a se d o n lin ea r a d ju st m e n t , ( c )

A d a pt iv e sa m pl in g b a se d o n q u a r t e r a d ju st m e n t ......................................................................... 104

F ig u r e 6-1. Q o S ev a lu a tio n sy st e m .........................................................................................................................108

F ig u r e 6-2. K o h o n e n n e u r a l n e t w o r k .................................................................................................................. 111

F ig u r e 6-3. S t r u c t u r e of an MLP........................................................................................................ 114

F ig u r e 6-4. C l u st e r in g QoS pa r a m eter s o f VoIP a ppl ic a t io n a t pr e d e fin e d t im e in t e r v a l 118

F ig u r e 6-5. C l u st e r in g QoS pa r a m e t e r s o f v id e o a p pl ic a t io n a t p r e d e fin e d t im e in t e r v a l . ... 119

F ig u r e 6 -6 . T h e d e g r e e o f m e m b e r s h ip b e t w e e n a s a m p l e o f d e l a y a n d t h e c l u s t e r ' s c e n t e r f o r :

( a ) V o IP , ( b ) V id e o a p p l ic a t io n ............................................................................................................................ 120

F ig u r e 6 -7 . T h e p r o g r e s s o f o b j e c t iv e f u n c t io n s d u r in g F C M a n a l y s is o f : ( a ) V o IP , ( b ) V id e o

APPLICATION......................................................................................................................................................................121

F ig u r e 6 -8 . C l u s t e r in g Q o S p a r a m e t e r s o f V o I P .................................................................................................. 122

F ig u r e 6 -9 . C l u s t e r in g Q o S p a r a m e t e r s o f v id e o ...............................................................................................123

F ig u r e 6 -1 0 . T h e Q o S o f V o IP u s in g r e g r e s s io n m o d e l ...................................................................................... 124

F ig u r e 6 -1 1 . T h e Q o S o f v id e o u s in g r e g r e s s io n m o d e l .............................................. 125

F ig u r e 6 -1 2 . C l a s s if y in g Q o S p a r a m e t e r s o f V o IP u s in g k o h o n e n n e t w o r k : ( a ) D e l a y , ( b ) J it t e r ,

(c ) P a c k e t l o s s r a t io , ( d ) Q o S p a r a m e t e r s l a b e l s , ( e ) S O M s a m p l e h i t s .....................................128

F ig u r e 6 -1 3 . C l a s s if y in g Q o S p a r a m e t e r s o f v id e o a p p l ic a t io n u s in g k o h o n e n n e t w o r k : ( a )

D e l a y , ( b ) J it t e r , ( c ) P a c k e t l o s s r a t io , ( d ) Q o S p a r a m e t e r s l a b e l s , ( e ) S O M s a m p l e h i t s .

................................................................................... 129

F ig u r e 6-1 4 . T h e p r o g r e s s o f t r a in in g M L P t o a s s e s s t h e Q o S f o r : ( a ) V o IP , ( b ) V id e o ..................130

F ig u r e 6-1 5 . C o m p a r is o n b e t w e e n n o r m a l is e d a c t u a l Q o S a n d c a l c u l a t e d o u t p u t s f r o m M L P

in c a s e o f : ( a ) V o IP Q o S a s s e s s m e n t , a n d ( b ) V id e o Q o S a s s e s s m e n t ............................................. 131

F ig u r e 6 -1 6 . T h e Q o S a s s e s s m e n t o f : ( a ) V o IP , a n d ( b ) V id e o a p p l ic a t io n u s in g M u l t i -L a y e r

P e r c e p t r o n .......................................................................................................... 132

F ig u r e 7 -1 . A d a p t iv e t r a f f ic a l l o c a t io n a l g o r it h m f l o w c h a r t ............................................................... 139

F ig u r e 7 -2 . B a r c h a r t r e p r e s e n t a t io n o f a v e r a g e d e l a y w it h a n d w it h o u t Q o S e n h a n c e m e n t

s c h e m e f o r : (a ) V o IP , ( b ) V id e o , ( c ) B e s t e f f o r t t r a f f ic , ( d ) B a c k g r o u n d t r a f f ic ................ 145

F ig u r e 7 -3 . B a r c h a r t r e p r e s e n t a t io n o f a v e r a g e j it t e r w it h a n d w it h o u t Q o S e n h a n c e m e n t

s c h e m e f o r : (a ) V o IP , ( b ) V id e o , ( c ) B e s t e f f o r t t r a f f ic , ( d ) B a c k g r o u n d t r a f f ic ................ 148

F ig u r e 7 -4 . B a r c h a r t r e p r e s e n t a t io n o f a v e r a g e p a c k e t l o s s w it h a n d w it h o u t Q o S

ENHANCEMENT SCHEME FOR: (A) V o IP , (B) VIDEO, (C) BEST EFFORT TRAFFIC, (D) BACKGROUND

TRAFFIC.......................................................................................... 151

F ig u r e 7 -5 . B a r c h a r t r e p r e s e n t a t io n o f o v e r a l l a s s e s s e d Q o S w it h a n d w it h o u t Q o S

ENHANCEMENT SCHEME FOR: (A) V o IP , (B) VIDEO, (C) BEST EFFORT TRAFFIC, (D) BACKGROUND

TRAFFIC.............................................................................................................................................................................. 154

F ig u r e 7 -6 . V is u a l c o m p a r is o n o f r e c o n s t r u c t e d F o r e m a n v id e o u s in g : ( a ) Q o S e n h a n c e m e n t

SCHEME, (B) LEGACY NETWORK SCHEME.................................................................................................................155

F ig u r e 8-1. K E IL M C B 2 3 0 0 E v a l u a t io n B o a r d ....................................................................................................... 161

F ig u r e 8 -2. M D K -A R M M ic r o c o n t r o l l e r D e v e l o p m e n t K it .......................................................................... 161

F ig u r e 8-3. Q o S m o n it o r in g S y s t e m .................................................................................................................... :........ 162

F ig u r e 8 -4. Q o S A s s e s s m e n t C o d e ...................................................................................................................................164

F ig u r e 8-5. T h e S im u l a t e d N e t w o r k ............................................................................................................................. 166

F ig u r e 8 -6 . T h e U L IN K -M E A d a p t e r ..............................................................................................................................167

xiii
F ig u r e 8-7. T h e c o n n e c t io n b e t w e e n t h e PC a n d t h e MCB2300 u sin g ULINK-ME JTAG A d a p t e r .
......................................................................................................................................................... 167
F ig u r e 8-8. ^V isio n IDE so ft w a r e t o o l ..................................................................................................................168

xiv
 LIST OF TABLES

T a b l e 2-1. T h e sen sitiv ity o f so m e c o m m o n a ppl ic a t io n s to t h e ir QoS pa r a m e t e r s .........................15

T a b l e 4-1. S im u l a tio n settin g s o f MAC and PHY p a r a m e t e r s in IE E E 802.1 1e ................................... 64

T a b l e 4-2. IE E E 802.1 1e a c c ess c a teg o ries p a r a m e t e r s ..................................................................................65

T a b l e 4-3. W R R P a r a m e t e r s .........................................................................................................................................66

T a b l e 4-4. T h e n u m b e r o f v id e o fra m e s a n d pa c k e t s o f t h e v id e o st r e a m in g s o u r c e s ................... 67

T a b l e 4-5. Q oS r eq u ir e m e n t s fo r v o ic e , v id e o , a n d d a t a as r e c o m m e n d e d b y IT U group (Z h a i e t

a l , 200 5 ).......................................................................................................................................................................68

T a b l e 5-1. M ea n a n d s t a n d a r d d e v ia t io n o f o v e r a l l st a t ist ic in pu t fu z z y m e m b e r sh ip

f u n c t io n s ....................................................................................................................................................................78

T a b l e 5-2. M e a n a n d st a n d a r d d ev ia tio n o f c u r r e n t sa m pl e in t e r v a l in pu t fu z z y m e m b e r s h ip

fu n c t io n s ....................................................................................................................................................................78

T a b l e 5-3. M ea n a n d s t a n d a r d d e v ia t io n o f sa m pl e in t e r v a l d iffe r e n c e o u t pu t fu z z y

MEMBERSHIP FUNCTIONS.......................................................................................................................................... 78

T a b l e 5-4. T h e fu zzy ru les u se d b y FIS t o a d ju st IS S I................. 79

T a b le 5-5. O pe r a t in g pa r a m e t e r s o f a d a pt iv e s t a t ist ic a l s a m pl in g a p pr o a c h e s ............................. 85

T a b le 5-6. T h r o u g h pu t m e a su r e m e n t r esu lts u s in g d iff e r e n t sa m pl in g m e t h o d s : ( a ) A d a p t iv e

SAMPLING BASED ON FIS, (B) ADAPTIVE SAMPLING BASED ON LINEAR ADJUSTMENT, (C) ADAPTIVE

SAMPLING BASED ON QUARTER ADJUSTMENT, (D) SYSTEMATIC, (E) STRATIFIED, (F) RANDOM

SAMPLING......................................................................................................................................................................87

T a b le 5-7. D el a y m e a su r e m e n t resu lts u s in g d iff e r e n t s a m pl in g m e t h o d s : ( a ) A d a p t iv e

SAMPLING BASED ON FIS, (B) ADAPTIVE SAMPLING BASED ON LINEAR ADJUSTMENT, (C) ADAPTIVE

SAMPLING BASED ON QUARTER ADJUSTMENT, (D) SYSTEMATIC, (E) STRATIFIED, (F) RANDOM

SAMPLING........................................................................... 92

T a b le 5-8. J itte r m e a su r e m e n t resu lts u s in g d iff e r e n t sa m pl in g m e t h o d s : ( a ) A d a p t iv e

SAMPLING BASED ON FIS, (B) ADAPTIVE SAMPLING BASED ON LINEAR ADJUSTMENT, (C) ADAPTIVE

SAMPLING BASED ON QUARTER ADJUSTMENT, (D) SYSTEMATIC, (E) STRATIFIED, (F) RANDOM

SAMPLING......................................................................................................................................................................97

T a b l e 5-9. P a c k e t lo ss r a t io m e a su r e m e n t r e su l t s u s in g d iff e r e n t s a m pl in g m e t h o d s : ( a )

A d a pt iv e s a m pl in g b a se d o n FIS, ( b ) A d a p t iv e s a m pl in g b a s e d o n l in e a r a d ju st m e n t , ( c )

A d a pt iv e s a m pl in g b a sed o n q u a rte r a d ju st m e n t , ( d ) S y s t e m a t ic , ( e ) S t r a t if ie d , ( f )

R andom sa m pl in g ................................................................................................................................................. 102

T a b le 6-1. Q oS pa r a m eter s o f V o IP a n d e x pe c t e d QoS u s in g : fu zzy in fe r e n c e sy s t e m , d ist a n c e

MEASUREMENT, AND REGRESSION MODEL..........................................................................................................126

T a b l e 6-2. Q oS pa r a m eter s o f v id e o a n d e x p e c t e d QoS u s in g : fu zzy in fe r e n c e s y s t e m , d ist a n c e

MEASUREMENT, AND REGRESSION MODEL..........................................................................................................126

XV
T a b le 6-3. Q oS param eters o f V o IP a n d e x pe c t e d Q o S u s in g : F u z z y I n fe r e n c e S y st e m (FIS),

D ista n c e M e a s u r e m e n t (DM), R e g r e s sio n M o d e l (RM), M u lti -L a y e r P e r c e pt r o n (MLP)

...................................................................................................................................................................................... 133

T a b le 6-4. Q oS pa r a m e t e r s o f v id e o a n d e x pe c t e d Q o S u s in g : F u z z y I n fe r e n c e S y st e m (FIS),

D ista n c e M e a s u r e m e n t (DM), R e g r e ssio n M o d e l (RM), M u lti -L a y e r P e r c e pt r o n (MLP)

...................................................................................................................................................................................... 134

T a b le 7-1. WRR q u e u e w e ig h t s ................................................................................................................................. 140

T a b le 7-2. T r a ffic c h a r a c t e r is t ic s ........................................................................................................................ 141

T a b le 7-3. WRR P a r a m e t e r s .......................................................... 142

T a b le 7-4. T h e A m o u n t o f s e n t , r e c e iv e d , a n d l o s t F orem an v id eo f r a m e s ....................................... 155

T a b l e 8-1. IEEE 802.1 1e A c cess C a t e g o r ie s ........................................................................................................ 166

T a b le 8-2. QoS P a r a m e t e r s , a n d t h e evaluated Q o S o f VoIP a ppl ic a t io n u sin g Q o S M o n it o r in g

D e v ic e (Q o S_M D ), F u z z y I n fe r e n c e S y st e m (FIS), a n d M u lti -L a y e r P e r c e pt r o n (MLP). 171

T a b le 8-3. QoS P a r a m e t e r s , a n d t h e evaluated Q o S o f v id e o a ppl ic a t io n u s in g QoS

M o n it o r in g D e v ic e (Q o S_M D ), F u zzy I n f e r e n c e S y st e m (FIS), a n d M u l t i -L a y e r

P e r c e pt r o n (MLP)................................................................................................................................................ 173

xvi
 GLOSSARY TERMS

AC Access Category
ACR Absolute Category Rating
ADC Analog to Digital Converter
AI Artificial Intelligent
AIFS Arbitration Inter Frame Space
AIFSN Arbitration Inter Frame Space Number
ANN Artificial Neural Network
AODV Ad hoc On-Demand Distance Vector
AP Access Point
CAC Call Admission Control
CAN Controller Area Network
CAP Controlled Access Phase
CBR Constant Bit Rate
CFP Contention Free Period
CF-Poll Contention Free Poll
cwmax Maximum Contention Window
CW m in Minimum Contention Window
DAC Digital to Analog Converter
DCF Distributed Coordination Function
DiffServ Differentiated Service
DIFS Distributed Inter Frame Space
DoS Denial of Service
DSDV Destination Sequenced Distance Vector
EDCA Enhanced Distributed Channel Access
FCM Fuzzy C- means
FIFO First In First Out
FIS Fuzzy Inference System
FLC Fuzzy Logic Controller
FQ Fair Queuing
FTP File Transfer Protocol
GloMoSIM Global Mobile Information System Simulator
GOP Group Of Picture
GUI Graphical User Interface
HC Hybrid Coordinator
HCCA HCF Controlled Channel Access
HCF Hybrid Coordination Function
HID Human Interface Devise
IC Integrated Circuit
IDE Integrated Development Environment
IDS Intrusion Detection System
IEEE Internet of Electrical and Electronic Engineering
IntServ Integrated Service
IP Internet Protocol
IPTV Internet Protocol Television
ISSI Inter-Sampling Section Interval
xvii
ITU International Telecommunication Union
KSOM Kernel SOM
LCD liquid crystal display
LED light-emitting diode
LSP Label Switched Path
MAC Medium Access Control
MANET Mobile Ad hoc Network
MCB Microcontroller Board
MDK Microcontroller Development Kit
MLP Multi- Layer Perceptron
MOS Mean Opinion Score
MSC Massage Storage Class
MSDU MAC Service Data Unit
MSE Mean Square Error
NAM Network Animator
NS-2 Network Simulator -2
OMNET++ Optical Micro Networks Plus Plus
OPNET Optimized Network Evaluation Tool
OSI Open System Interconnection
OTcl Object Tool Command Language
OWD One Way Delay
PAR Packet Arrival Rate
PCF Point Coordination Function
PCM Plus Code Modulation
PESQ Perceptual Evaluation of Speech Quality
PIFS PCF Inter Frame Space
PLCP Physical Layer Convergence Procedure
PLR Packet Loss Ratio
PQ Priority Queuing
PSNR Peak Signal to Noise Ratio
PSQM Perceptual Speech Quality Measure
QoS Quality of Service
RAM Random Access Memory
RED Random Early Detection
RNN Random Neural Network
ROM Read Only Memory
RSE Relative Standard Error
RST Request to Send
SD Secure Digital
SI Service Interval
SID Sample Interval Difference
SIFS Short Inter Frame Space
SOM Self Organising Map
TCL Tool Command Language
TCP Transmission Control Protocol
TD Transmission opportunity Duration
TORA Temporally Ordered Routing Algorithm
TSW Time Sliding Window
TXOP Transmission Opportunity
UDP User Datagram Protocol
xviii
USB Universal Serial Bus
VBR Variable Bit Rate
VoIP Voice over Internet Protocol
VVoIP Voice and Video over Internet Protocol
WFQ Weighted Fair Queuing
WLAN Wireless Local Area Network
WRR Weighted Round Robin

xix
 LIST OF VARIABLES

Symbol Definition Equation

n e t) Throughput in bit per second (bps) during ith time interval
No. of bits of all successfully received packets during i th 2.1
Pi ( 0 interval
D, Delay of the ith packet arrives at its final destination
Ri Sending timestamp of the ith packet. 2.2
Si Arrival timestamp of i th packet.
Ji Jitter of ith packet 2.4
Ph Packet Loss ratio in percentage (%) during ith timeinterval 2.6
w, Associated weight for queue (i) in weithed round robin
2.8
R Link capacity
y Dependent variable in regression model formula
No. n of Independent variables in regression model
••• •> formula
No. n of regression coefficients determined from 2.9
b0,b l f b2, ...bn
independent variables in regression model formula
A column vector of n error terms in regression model
e
formula
Mx Degrees of membership functions for fuzzy sets X
2.14
My Degrees of membership functions for fuzzy sets Y
In FCM, is matrix of size n X N representing a given set of
X
feature data
u Membership matrix of size n x C generated by FCM
V Matrix of clusters' centres generated by FCM
c No. of clusters generated by FCM
2.16
Controls the degree of fuzziness for the membership of the
m
cluster
Degree of membership between Xj to the centre v t of
N cluster i
D2
uu Euclidian distance between Xj to the centre of cluster i
s Resulting value of summation function of MLP
Xi Input i to MLP 2.21
Wj Associated neuron connection weight of input i
y The output of activation function <p (s) of MLP
d Desired output o f training example 2.22
e Calculated error between d and y
Xi Input i to SOM
2.23
wu Associated connection weight between input i and neuron j
Wij (n) Current weight of winning neuron
w if (n + 1 ) Updated weight of winning neuron 2.24
V Learning rate
Mean value of pre- sampling section for throughput of
meanl 5.1
traffic being sampled

XX
 Mean value of post- sampling section for throughput of
mean2
traffic being sampled
Median value of pre- sampling section for throughput of
medianl
traffic being sampled
Median value of post- sampling section for throughput of
median2 5.1
traffic being sampled
Standard deviation value of pre- sampling section for
std l
throughput of traffic being sampled
Standard deviation value of post- sampling section for
std2
throughput of traffic being sampled
Updated ISSI Updated length of Inter- Sampling Section Interval
5.2
Ai l Update the length of ISSI by increasing it linearly
p2 Update the length of ISSI by decreasing it linearly 5.3
V Update the length of ISSI in quarter adjustment mechanism 5.4 - 5.5
VMM Gaussian membership of value (x) of ith fuzzy set A1
Ct The mean of the ith fuzzy set A1 5.6
Standard deviation of the ith fuzzy set A1
Y Deffuzzification output using centroid method
Vi Centroid of fuzzy region i
5.7
Hi Output membership value
m Number of fuzzy sets after implication process
SID Sample Interval Difference 5.8
Shows the difference between mean of sampled version of
Bias
QoS parameter and the mean of its original population
Mt Mean of QoS parameters for the sampled version
M Mean of QoS parameters for original population 5.9
N Number of simulation runs
RES Relative Standard Error
SE Standard Error of sampled version 5.10
n Sample size

xxi
 Chapter 1 Introduction

Quality of Service (QoS) management is currently one of the principle technological

fields of development in computer networks. Computer networks are increasingly
integrated and carry a diverse set of traffic such as Voice over Internet Protocol (VoIP),
video streaming, video conferencing, and traditional data. The interconnection of wired
and wireless networks and the rapid growth of real-time and non real-time applications
transmitted over these networks have made QoS management an area that requires
further research and development. QoS management of these networks is important to
both users as well as the network service providers. Users are interested in determining
how well they receive applications. The network service providers and network
managers need QoS information to determine how well their networks are preforming.
Therefore, the main focus of the study is to develop mechanisms associated with QoS
management processes for multimedia computer networks.

The purpose of this chapter is to: clarify the rationales behind this research which are
presented in section 1.1, outline the aim and objectives in section 1.2, summarise the
contribution of this research in section 1.3, and present the outline and organisation of
this thesis in section 1.4.

1.1 Research Motivations

In this study, network QoS management refers to evaluation and improvement of QoS
in wired and wireless computer networks. Evaluation of QoS aims to analyse and
quantify a network performance with respect to meet the applications1 transmission
requirements. QoS improvement involves the ability to take actions to enhance network
performance. However, there are complexities associated with realising QoS
management and so the area is an important field of research. The following points
summarise the QoS management issues considered in this study:

(i) Computer networks carry different traffic types containing video, audio and data.
Analysis and interpretation of all traffic packets can be time consuming. Also,
collecting all packets poses a challenge as the process needs to be sufficiently fast,
should not load the network and interfere with the operation of the protocols
responsible for managing the network. Therefore, ways need to be found to
Chapter 1 Introduction

accurately sample packets and present them to the network management entities
responsible for their interpretation and decision making. Most of current sampling
techniques use a predetermined and fixed sampling rate irrespective of the extent
of traffic fluctuation with time. The sampling therefore is not the optimal as the
multimedia traffic is time varying. In a fixed rate sampling method, two situations
could occur: (a) If the traffic fluctuation (i.e. frequency of variation) is high, there
is a risk of losing important information (caused by under sampling), (b) If the
traffic fluctuation is low, resources would be under utilised (by over sampling)
(Giertl et al, 2006). In order to reduce the biasness presented in conventional fixed
rate sampling and to enhance the process o f gathering Quality of Service (QoS)
information, traffic should be sampled adaptively. In other words, a small sample
interval is required during the period of high activity, whereas a larger sample
interval is required during the period of low traffic fluctuation.

(ii) Gathered network information is indicative of its performance in delivering real­

time and non-real-time applications. The data collected include parameters related
to QoS such as delay, jitter, throughput, and packet loss. These parameters need
evaluating in an informative and effective manner. The evaluation of QoS is
currently carried out either by analysis or measurement techniques. The analysis
techniques are used to examine the characteristics of the traffic (Chen et al, 2009)
and (Timo et al, 2002), whereas the measurement techniques are applied to
determine how well the network treats the ongoing traffic (Palomar et al, 2008),
(Teyeb et al, 2006), (Cranley and Davis, 2005), and (Mishra and Sharma, 2003).
The current state-of-the-art of QoS analysis and measurement approaches have
several limitations. For example, they are not combined to form a mechanism to
evaluate QoS from an analysis and measurement point of view. Moreover, current
QoS measurement techniques can generate an excessive traffic load that affect the
operation of the network as in active measurement approach or they perform by
measuring the actual network traffic that requires collecting and processing a large
amount of recorded data packets in order to provide an indication of network
performance as in passive approaches (Brekne et al, 2002). Subjective and
objective QoS measurement approaches have also some limitations. The former
approach cannot be automated since it requires a controlled environment, and it is
time consuming to be repeated frequently due to its dependence on human
subjects (Palomar et al, 2008). Objective approaches are computationally

2
Chapter 1 Introduction

intensive because their operations are at the pixel level and they cannot take into
the account all the affected network parameters (Mohammed et al, 2001). These
limitations highlight the need for further development in evaluation of QoS. In
order to evaluate QoS effectively, the proposed methods need to combine analysis
and measurement techniques. The methods require evaluation of QoS in a manner
similar to human subjects and quantify the QoS without the necessity for complex
mathematical models, taking into the account the QoS requirements of every type
of multimedia applications. In addition, the methods should not add a significant
extra load to the network as is the case with active approaches, nor should it
depend on collection of all traffic packets.

(iii) Delivery of multimedia applications is still with limitations, especially for

wireless networks. A major bottleneck in transmission of real-time multimedia
applications is insufficient channel bandwidth. Accordingly, QoS could be
unpredictable. Therefore, deployment of network QoS enhancement techniques
can be beneficial for multimedia transmission. However, most previous studies
considered QoS support either in wireless local area networks (WLANs) or in
wired networks. There are few studies to enable end-to-end QoS in wired and
wireless networks (Skyrianoglou et al, 2002), (Park et al, 2003), and (Senkindu
and Chan2008). The limitations of these studies were the inclusion of an
intermediate layer between the MAC and IP layers in wireless stations, which in
turn added more complexity in managing the wireless side of the network, or the
low priority traffic was starved due to link congestions and QoS prioritisation.
The challenge to enable end-to-end QoS is the inclusion of both parts o f the
network (i.e. wired and wireless) to provide different treatments for different
classes of traffic and efficient use of network resources.

(iv) The existing QoS monitoring techniques are limited in scope. They obtain overall
network QoS indirectly and are not stand-alone. For example, the QoS monitoring
tool proposed by Graham et al (1998) was used to assess packet latency and loss
as an indication of network performance. The Surveyor tool proposed by Zseby
and Scheiner (2004) to assess end-to-end delay and packet loss required a further
analysis using a server to measure the network performance. The disadvantage of
these tools is that network managers have to do a variety of operations to assess
the overall network QoS. From these limitations, it can be obvious that the
process of monitoring QoS can be complicated, expensive, and time consuming.
3
Chapter 1 Introduction

Therefore, developing a portable hand-held device that accurately determines the

overall network QoS for multimedia applications can be very valuable. In this
study, a mechanism that assesses QoS by taking into the account the requirements
of multimedia applications is implemented on a portable microprocessor board.

1.2 Research Aim and Objectives

The focus in this research is on network QoS management which entails evaluation and
improvement of QoS. This raises the following research question: to what extent, the
QoS evaluation and improvement can be interrelated, and do they require different
mechanisms to be integrated.

The overall aim of this study is to develop techniques to evaluate QoS in multimedia
networks and to use this information as part of a network management process to
improve its performance. The study uses IEEE 802.1 le standard for the wireless side of
the network. The wireless is connected to a wired network. The objectives of the study
are to:

i. Develop approaches that allow QoS parameters of multimedia applications to be

collected efficiently and accurately.
ii. Develop techniques to analyse and interpret QoS parameters of multimedia
networks.
iii. Develop techniques to allow the QoS information to be used as part of network
management to improve its performance.
iv. Develop an electronic portable device that facilitates accurate QoS assessment of
multimedia networks.
v. Critically evaluate the developed techniques to determine their effectiveness and
accuracy.
This study will involve multimedia type networks that integrate both wired and wireless
parts.

1.3 Research Contributions

Improvement in managing QoS of multimedia applications is essential, particularly
when dealing with hybrid wired and wireless structures. The techniques proposed in this
study contribute to extending the knowledge in QoS management. This section outlines

4
Chapter 1 Introduction

the contributions of this research in line with its objectives. The research contributions
of this study are included in relevant chapters. These contributions are outlined below.

(i) Develop approaches that allow QoS parameters o f multimedia applications to

be collected efficiently and accurately. In order to facilitate effective traffic data
gathering, novel statistical adaptive sampling techniques are developed that
utilised traffic's statistical features. These techniques adjust the sampling interval
by considering the traffic's statistical features between two consecutive sampled
sections using: quarter adjustment approach (Dogman, et al., 2010a), simple
linear adjustment approach (Dogman, et al., 2010b) and fuzzy inference system
(Dogman, et al., 2011). Chapter 5 of this thesis introduces the adaptive statistical
sampling techniques which were based on three adjustment mechanisms: quarter
adjustment mechanisms, simple linear adjustment mechanisms, and fuzzy
inference system. Also, a comparison of the devised methods versus
conventional non-adaptive sampling techniques (i.e. systematic sampling,
stratified sampling, and random sampling) was carried out to validate the
effectiveness of proposed sampling techniques.

(ii) Develop techniques to analyse and assess QoS parameters o f multimedia

networks. The contribution of this study is development of mechanisms that
combines analysis and measurement techniques to evaluate QoS of multimedia
applications in an effective manner. Two innovative QoS evaluation systems are
proposed based on Artificial Intelligence (Al) and traditional techniques. The
latter combines Fuzzy C-Means (FCM) and the regression model to analyse and
assess the QoS of multimedia applications (Dogman, et al., 2012a). The former
system analyses and assesses the QoS of multimedia applications based on a
combination of supervised and unsupervised neural networks (Dogman, et al.,
2012b). The proposed QoS evaluation systems are introduced in Chapter 6. The
transmitted application’s QoS parameters are initially analysed either by the
FCM clustering algorithm or by the unsupervised learning Kohonen neural
network (i.e. Self Organising Map (SOM)). The analysed QoS parameters are
then used as inputs to a regression model or supervised learning Multi-Layer
Perceptron (MLP) neural network in order to quantify the overall QoS. The
proposed QoS evaluation systems provided relevant information about the
network’s QoS patterns and based on them, the overall network’s QoS was

5
Chapter 1 Introduction

successfully quantified. The process of evaluating QoS is explained in details in

Chapter 6.

(iii) Develop techniques to allow the QoS information to be used as part o f network
management to improve its performance. A scheme that improves QoS in both
wired and wireless domains of the network is proposed in (Dogman, et al.,
2012c). A description of the proposed QoS enhancement scheme is provided in
Chapter 7. The scheme uses a combination of MAC layer QoS control in the
wireless domains and network layer QoS control in the wired domains. A novel
aspect of the scheme is that an adaptive access category (AC) traffic allocation is
devised and incorporated into wireless access point (AP) in order to improve the
QoS of IEEE 802.l i e Enhanced Distributed Channel Access (EDCA) protocol,
and Weighted Round Robin (WRR) queuing scheduling mechanism is
implemented into congestion point (i.e. router) in wired networks to support fair
distribution of bandwidth among different traffic classes. The adaptive traffic
allocation algorithm determines the packet arrival rate (PAR) of the up- and
down-link traffic for each access category (AC). It then dynamically allocates
traffic of a lower priority AC to the next higher AC, when the higher AC is not
receiving traffic at the time. Whereas, WRR shares the network resources, based
on the traffic’s QoS requirements. The results explained in Chapter 7 show the
improvement in QoS provided by wired and wireless sides when the QoS
enhancement scheme was implemented.
(iv) Develop an electronic portable device that facilitates accurate QoS assessment
o f multimedia networks. A hardware QoS monitoring system is designed
(Dogman, et al., 2013). The system is used regression modelling, implemented
on the MCB2300 KEIL ARM microcontroller board. More details about QoS
monitoring system are provided in Chapter 8. The hardware QoS monitoring
system analysed the QoS requirements (i.e. delay, jitter and packet loss ratio) of
multimedia applications to determine their overall QoS. The evaluation o f QoS
monitoring system was carried out by comparing the results obtained with other
QoS assessment methods to indicate the effectiveness of the developed system
in monitoring multimedia QoS accurately.

6
Chapter 1 Introduction

1.4 Thesis Organisation

Figure 1-1 shows the schematic overview of the thesis. In addition to the introduction
chapter, which outlines the rationales for this study, study’s aim and objectives, and a
summary of the contribution of the research to QoS management area are explained.
There are eight further chapters in this thesis.

Chapter 2 covers the theoretical background essential for this research. This includes the
definitions of QoS and its parameters, QoS requirements of multimedia applications, the
service levels of QoS, and QoS components. Chapter 2 also provides a detailed
description of IEEE 802.1 le as an emerging WLANs standard to provide QoS, and
packet scheduling mechanisms as the most commonly mechanisms implemented in
wired network to support QoS. In addition to that, the theories o f statistical and A l
techniques used in this study are provided. Regression model as one of the most widely
employed statistical analysis methods is described. The fundamental principles of three
important paradigms in A l system: fuzzy logic, fuzzy clustering, and neural network
with their operational steps are also discussed.

In Chapter 3, the previous studies relevant to the aforementioned issues o f managing the
QoS in multimedia networks in section 1.1 are critically analysed and discussed. The
aim is to identify the potential limitations associated with these issues which in turn are
further developed. Chapter 3 is organised to review the sampling techniques used to
gather information from network traffic, and discuss the state-of-the-art o f QoS analysis
and assessment techniques used to evaluate network QoS. The relevant studies
considering QoS support in the context of wireless-wired networks are also analysed,
and the exiting monitoring tools used to assess the network performance are reviewed.
Moreover, the applications of statistical and A l techniques used in this study into the
field of network QoS management are reviewed in that chapter.

Chapter 4 provides an explanation of network evaluation approaches, network

simulation tools, and the general experimental procedure used throughout this study.
The description of Network Simulation 2 (NS-2) environment, transmission protocols,
queuing mechanisms, and traffic type, and its characteristics are included. The
measurement process, which includes a description o f the QoS metrics and
requirements, and the procedure for analysis simulation output are also provided.

7
Chapter 1 Introduction

The main results of this research are provided in Chapters 5-8. In Chapter 5, statistical
adaptive sampling techniques to adjust sampling rate based on traffic’s statistics are
introduced. A detailed description of the proposed adaptive statistical sampling
techniques which are based on three adjustment mechanisms: quarter adjustment
mechanisms, simple linear adjustment mechanisms, and fuzzy inference system is
provided. An implementation of conventional sampling techniques (i.e. systematic
sampling, stratified sampling, and random sampling) is explained. The experimental
results in Chapter 5 show the effectiveness of statistical sampling techniques by
carrying out a comparison of devised statistical sampling technique versus the
conventional sampling techniques using a simulated computer network.

Chapter 6 introduces the use of statistical and A l techniques in the area o f QoS
management. Two innovative QoS evaluation approaches are explained. The first
approach combines Fuzzy C-Means (FCM) and regression model to analyse and assess
QoS of multimedia applications in a simulated network, whereas the other approach
analyses and assesses QoS in multimedia applications using a combination of
supervised and unsupervised neural networks. The transmitted application’s QoS
parameters are initially analysed either by FCM clustering algorithm or by the
unsupervised learning Kohonen neural network (i.e. Self-Organising Maps (SOM)). The
analysed QoS parameters are then used as inputs to a regression model or supervised
learning Multi-Layer Perceptron (MLP) neural network in order to quantify the overall
QoS. Results in Chapter 6 show how the proposed QoS evaluation systems provide
information about the network’s QoS patterns and based on this information, how the
overall network’s QoS is quantified.

In Chapter 7, development of a new Quality of Service (QoS) enhancement scheme for

WLAN-wired networks is introduced and its performance is evaluated. The proposed
scheme consists of an adaptive Access Category (AC) traffic allocation algorithm that is
incorporated into the network's wireless side to improve the performance of IEEE
802.l i e Enhanced Distributed Channel Access (EDCA) protocol, and a Weighted
Round Robin (WRR) queuing scheduling mechanism that is incorporated into the wired
side of the network. A description of the proposed QoS enhancement scheme is
provided in this Chapter. The results discussed in Chapter 7 show the efficiency of
proposed enhancement scheme. The scheme provides an end-to-end QoS to be setup
which in turn provides an improved delivery of a variety of applications in the context
of wired-cum-wireless networks.
Chapter 1 Introduction

In Chapter 8, a network QoS monitoring system is designed and evaluated. The

proposed monitoring system incorporates the QoS assessment approach developed by
(Dogman et al, 2012a) that is based on regression model. The microcontroller board
MCB2300 KEIL ARM is used. Chapter 8 explains the MCB2300 KEIL ARM
microcontroller board, and outlines how the QoS assessment technique using regression
modelling is devised, and implemented on the MCB2300 KEIL ARM microcontroller
board. The performance of QoS monitoring system is compared with other QoS
assessment methods (e.g. QoS assessment using Fuzzy Inference System introduced by
(Al-Sbou et al, 2006), and Neural Network QoS monitoring approach proposed by
(Dogman et al, 2012b). The results indicated that the developed system is capable of
accurately assessing QoS.

Chapter 9 discusses the overall findings of this research, provides the conclusions, and
highlights future research directions.

9
Chapter 1 Introduction

QoS overview Research Motivations

Chapter 1:
IEEE 802.l ie QoS Introduction Aim and objectives
mechanisms

QoS in wired networks Research contributions

Chapter2:
Theory and Thesis organisation
Fuzzy logic background
Sampling techniques for
computer network traffic
Fuzzy clustering

Chapter3: QoS analysis and

Literature measurement techniques
Neural networks review
Improvement of QoS in
computer networks
Regression model
Hardware implementation
Chapter4: of QoS assessment
Simulation package Experimental techniaues
procedure
Quarter adjustment of
Network settings adaptive sampling scheme
Results chapters
Adaptive sampling scheme
Traffic types
Chapter5: based on linear adjustment
Adaptive sampling
Measurement process using FIS and Adaptive sampling using
linear techniques Fuzzy Inference System

FCM analysis technique Comparison between

Chapter6:
Evaluation of QoS adaptive and non adaptive
using Al sampling techniques
SOM analysis technique techniques
Adaptive allocation traffic
Chapter7: QoS algorithm in 802.1 le
RM assessment technique improvement in
WLAN-wired Implement WWR in wired
networks side of the network
MLP assessment technique
Chapter8: Comparison between QoS
Implementing enhancement scheme and
QoS assessment using keil QoS measurement network legacy scheme
MDK-ARM evaluation techniques on
board hardware
Conclusions

Comparison between Chapter9:

r
software and hardware QoS Conclusions, and
Further work
assessment techniques further work
V_ ^ —

Figure 1-1. The schematic overview of the thesis.

10
 Chapter 2 Relevant Theory and Background

2.1 Introduction
The aim of this chapter is to provide the background related to the main issues of this
study. Section 2.2 of this chapter gives an overview of QoS in multimedia networks.
This includes definitions of QoS and its parameters, QoS requirements of multimedia
applications, the service levels required of QoS, and QoS components. Section 2.3
discusses the QoS in wireless and wired networks. This section includes a detailed
description of IEEE 802.l i e as an emerging WLAN standard to provide QoS, and
packet scheduling mechanisms as the most commonly used mechanisms implemented
in wired network to support QoS. Section 2.4 introduces the theory of statistical and A l
techniques used in this study. This includes the basic concepts o f regression analysis as
one of the most widely employed statistical modelling methods and the fundamental
principles of three important paradigms in A l system: fuzzy logic, fuzzy clustering, and
neural network.

2.2 Quality of Service (QoS): An overview

Wired and wireless networks are becoming increasingly integrated. This is coupled with
the rapid growth of real-time and non-real-time applications transmitted over these
networks. These developments necessitated a greater emphasis on QoS of networks. As
its aim is to provide guaranteed services for different applications, QoS is currently one
of the principle research topics in computer network.

2.2.1 Definition of QoS:

In the world of telecommunications, each of the terms: “quality” and “service” has its
own definition. The latter term means the ability o f network as specified by the service
provider to transmit information to the end user, whereas the term “quality” is used to
assess the ability of the service whether it satisfies the stated and implied needs
(Gozdecki et al, 2003). The interrelation between the two terms and the importance of
their associations with three different parts: service provider, network manager, and the
end user have produced the concept of QoS. Therefore, QoS could have a number of
definitions. From the user point of view: QoS is “the collective effect o f service
Chapter 2 Relevant Theory and Background

performance, which determines the degree of satisfaction of the user o f the service”
(ITU, 1994). While from the technical aspects, QoS can be defined as the ability of
network to provide a better service to selected traffic over a variety of technologies.
QoS is thus "a collection of technologies, which allow network-aware applications to
request and receive predictable service levels in terms of data throughput capacity
(bandwidth), latency variations (jitter) or propagation latency (delay)" (Saliba et al,
2005). From service provider perspective, QoS is "a set of service requirements to be
met by the network, while transporting a flow." (Crawley et al, 1998). In this study, the
term QoS refers to the ability of network to assess and provide desired QoS
requirements in terms of delay, jitter, throughput, and packet loss for transmitted
multimedia applications.

2.2.2 QoS Parameters

In packet-switched networks, there are a number of QoS parameters that can be assessed
and measured to determine QoS. These parameters can express how well the network
treats packets during their journey from the source, throughout the network, and finally
to their destinations. However, as different applications require different QoS
parameters, a correct set of QoS parameters for a particular application being
transmitted should be determined in order to effectively evaluate the QoS provided by
the network for that particular application (Cheong and Lai, 1999). For instance, time
sensitive applications such as video and Voice over IP (VoIP) are susceptible to
transmission parameters such as packet delay, jitter, loss and throughput whereas time
insensitive operations such as a file transfer are more sensitive to packet loss, but can
tolerate transmission jitter (Kurose and Ross, 2005). In this study, multimedia
applications will be considered. Therefore, the most important QoS parameters would
be examined and quantified are throughput, delay, jitter, and packet loss ratio. The
following subsections explain these parameters.

2.2.2.1 Throughput

Throughput is used to assess the capability of network to transmit data over a given
period of time (Heckmann et al, 2002). It could be defined as the maximum
transmission rate of packets that can be sustained between two endpoints. Wang et al
(2000) defined the throughput as the amount of successfully received packets in a
predefined time. Equation (2.1) is used to calculate the throughput:

12
Chapter 2 Relevant Theory and Background

T, (t) = (2.1)
li

Where T* is the measured throughput in bit per second (bps) during the ith interval,
£ P i( t ) is the total bits of all successfully received packets during the ith interval, and
t t is the time duration of the ith interval.

2.2.2.2 Delay

Delay isdefined as the elapsed time for a packet to travel from its source to its
destination. It can be also defined as the total amount of time that the packet takes to be
sent from its source, through the network, until it reaches its destination (Heckmann et
al, 2002). However, as the packet travels from its source to its destination, it suffers
from various types of delay such as queuing delay, transmission delay, and processing
delay. A detailed explanation of these types of delay can be found in (Kurose and Ross,
2005). According to Wang et al (2000), the delay can be calculated using equation
(2 .2):

Di= Rt — St (2.2)

Where Dt is the delay of the ith packet arrived, Rt and St are the arrival and sending
timestamps of the ith packet. The average delay can be calculated using equation (2.3):

Average delay = —J]L i A' (2-3)

n

Most of real-time applications are delay sensitive. For example, video conferencing and
voice over IP (VoIP) have high sensitivity to delay because the transmitted packets need
to be replayed back at the receiver in real time (Nortal Networks, 2003).

2.2.2.3 Jitter

Jitter is defined as the variation of delay between two consecutive packets for a given
traffic flow (Nortal Networks, 2003). Jitter has a significant effect on real time
applications (video and audio) because they require packets to be received with a fixed
delay. For example, VoIP requires packets to be arrived at a constant rate, since the
received packets will be played back at real time. Therefore, a small amount of jitter
might be acceptable but when the jitter increases, its effect becomes obvious and might
lead to a stuttering communication with pops and clicks (Heckmann et al, 2002). Wang

13
i
Chapter 2 Relevant Theory and Background

et al (2000) pointed out that jitter can be calculated as the difference between the delay
of two consecutive packets using equation (2.4):

Ji = ID i~ D i-ll while i>0 (2.4)

Where Ji is the jitter of the ith packet, Dt is the delay of packet i, and is the delay
of the previous packet. Average jitter can be calculated as in equation (2.5):

l
Average jitter = —Y i=i Ji (2.5)

2.2.2.4 Packet Loss Ratio

Packets can be lost during their journey from source to destination. Packet loss occurs
due to many effects such as inadequate physical transmission medium, collisions
between packets, queuing overflow, and hardware failure (Nortal Networks, 2003). The
effect of packet loss leads to "distortion, which results in a stuttering and snatch
communication” (Heckmann et al, 2002). Packet loss can be defined as the percentage
of transmitted packets that failed to reach their destinations as given by (Wang et al,
2000).

PL K t) = 1 0 0 X ( l - § m ) (2.6)

Where PL\ is the loss ratio in percentage (%) during the ith interval, and 2 Ri (t ) and
2 St (t) are the total number of received and transmitted packets with the ith interval
respectively.

2.2.3 QoS Requirements of Multimedia Applications

The QoS requirements of multimedia applications are significantly different from

traditional applications. The latter applications such as email, web browsing, and file
transfer can be elastic with some QoS parameters such as delay and jitter. However, the
former applications (i.e. multimedia applications) such as video conferencing and VoIP
have high sensitivity to QoS parameters and require a faster response from the network.
A large delay or jitter can seriously degrade their quality (Kurose and Ross, 2005). The
allocation of bandwidth usage for these applications can be also challenging to calculate.
This is because of the number of different variables, such as codec usage, resolution,
and transmission activity levels. In a computer network, some factors constrain realising
an acceptable QoS. For instance, network congestion in wired networks and interference
14
Chapter 2 Kelevant Theory and Background

problems in wireless networks are critical issues to provide service guarantee for
transmitted multimedia applications. Therefore, the acceptable range of QoS parameters
for multimedia applications must be identified. Table (2-1) indicates the sensitivity of
some common applications to QoS parameters (Zhai et al, 2005) and (ITU-T, 2001).

Table 2-1. The sensitivity of some common applications to their QoS parameters.

Typical Packet
Class Application Delay Jitter
bandwidth loss ratio

< 150 ms < 1ms <3%

VoIP 16-128 kbps
preferred preferred preferred
Real-time
< 150 ms < 30ms < 1%
Video 16-384kbps
preferred preferred preferred
Non real­ E-mail, file transfer, web
Minutes N/A Zero
time browsing

2.2.4 Service Levels of QoS

The service levels of QoS can be defined as the actual capability provided by a network
to deliver service required by specific application. Computer network could provide
three main levels of QoS agreements: best effort service (Floyd and Allman, 2008),
Differentiated Service (DiffServ) (Black et al, 1998), and Integrated service (IntServ)
(Braden et al, 1994). In the best effort service, the network QoS is unspecified. The
packet delivery rate varies depending on the current network load. In this case, the
network does not provide any guarantees or priority to the transmitted applications
although the diversity of their QoS requirements. The DiffServ architecture provides a
service differentiation mechanism to manage network resources based on QoS
requirements of transmitted applications. In DiffServ, the transmitted applications are
aggregated to a number of classes to receive a particular level of service. In IntServ
architecture, end-to-end QoS guarantee per flow is provided. A flow requiring QoS
guarantees must acquire sufficient network resources during its transmission to ensure
its QoS requirements are met. This is achieved by a reservation of network resources for
a specific flow (Shah, 2001) and (Kurose and Ross, 2005).

2.2.5 QoS Components

There are several components of QoS. These include: QoS mapping (Al-Kuwaiti et al,
2007), admission control (Georgoulas et al, 2004), traffic shaping (Lekcharoen, 2007),

15
Chapter 2 Relevant Theory and Background

policing (Koutsakis, 2009), packet queue scheduling (Yu and Meng, 2009), and
prioritisation (Senkindu and Chan, 2008).

The conversion of QoS representation between OSI model layers is referred to QoS
mapping. In admission control, the network take a set of actions during traffic
establishment phase to check if that particular traffic can be admitted or should be
rejected. A new traffic is admitted to the network when its desired QoS can be satisfied,
without causing any QoS violation to the already established services. The function of
traffic shaping mechanism is to ensure that there is a smooth data rate in order to meet
the specified QoS agreement. Traffic policing is used to monitor the transmitted traffic
by discarding or remarking traffic that exceed limits in order to protect the network
from malicious behaviour. Packet queue scheduling determines bandwidth allocation
among transmitted packets and the manner to service various applications with different
QoS requirements. In a prioritisation scheme, applications are classified based on their
QoS requirements and resources are assigned according to classes of priority depending
on network resource availability.

2.3 QoS in Wireless and Wired Networks

The QoS mechanisms in WLANs and wired networks are different. The QoS in
WLANs is enabled at MAC layer, whereas the most wired networks enabled QoS at IP
layer. This section discusses the QoS in WLANs and wired networks. A detailed
descriptions of IEEE 802.1 le as an emerging WLANs standard to provide QoS, and
packet scheduling mechanisms as the most common mechanisms implemented in wired
network to support QoS are provided.

2.3.1 QoS in Wireless Networks

WLANs have grown rapidly over the last decade, offering a range o f practical and
beneficial services for both home users and businesses. This growth relates to the fact
that WLANs allow users to communicate without using physical medium. Also, the cost
of installing WLANs is lower than wired networks because WLANs can be expanded
by just adding access points (AP) rather than installing cables as in wired networks.
Another influential factor in the growth of WLANs is the emergence the of IEEE
802.11 standard in 1997 and its subsequent amendments (1999 - 2005). Its cost
effectiveness, ease of deployment, and mobility support made IEEE 802.11 WLANs to
be used widely and became the dominating WLAN technology. IEEE 802.11 has
16
Chapter 2 Relevant Theory and Background

reached an unprecedented maturity in providing ever-growing bit rates (Lin et al, 2009).
However, with an emergence of multimedia communications over WLAN, the
provision of QoS in WLANs that is capable of guaranteeing QoS requirements of
multimedia applications becomes important. The demand for supporting QoS for
various applications with different QoS requirements has led to the development of a
WLAN standard so called IEEE 802.1 le.

2.3.1.1 IEEE 802.11e Standard

The original standard of IEEE 802.11 was not designed to support QoS (IEEE
Computer Society, 1999). In the legacy IEEE 802.11, applications with different QoS
requirements were treated the same, and the service differentiation in terms of
guaranteed bandwidth, bounded delay, and jitter for particular applications was
disregarded. Due to the growth transmission of time-sensitive and time-insensitive
applications over WLANs, the demand for supporting QoS for various applications has
increased. Thus, the enhancement version of IEEE 802.11 standard called IEEE 802.1 le
was proposed to provide QoS support for applications with different QoS requirements
(IEEE Computer Society, 2005). The IEEE 802.l i e introduces Hybrid Coordination
Function (HCF) which defines two medium access mechanisms: HCF Controlled
Channel Access (HCCA) and Enhanced Distributed Channel Access (EDCA).

23.1.1.1 HCF Controlled Channel Access (HCCA)

HCCA is an enhanced version of Point Coordination Function (PCF) in the legacy IEEE
802.11. Similar to PCF, HCCA has centralised access scheme which uses Hybrid
Coordinator (HC) implemented in the AP to access wireless medium, transmit at the
Contention Free Period (CFP), and the QoS Contention Free (CF-Poll) frame is used to
schedule the uplink traffic. However, HCCA has new operation parameters. These
include: Service Interval (SI), Controlled Access Phase (CAP), and Transmission
Opportunity (TXOP). SI is an interval between two successive TXOPs. CAP is a time
period when the HC maintains control over the medium after gaining medium access by
sensing the channel to be idle for the PCF Inter Frame Space (PIFS) period. TXOP is
time period where stations can transmit a number of MAC Service Data Unites
(MSDUs). Figure 2-1 shows the operation parameters of HCCA (Lee et al, 2011).

17
Chapter 2 Relevant Theory and Background

CFP ! CP
4------------------------------------------- >'<----------------------------------------------------------------------- N
SI [ SI
n ------------------------------------------------------- 1-------------- ►H----------------------------------------------------------------------- ►
i i

E HCCA HCCA HCCA ... EDCA HCCA HCCA HCCA ... EDCA
TXOP, TXOP2 TXOP3 TXOP TXOP, TXOP2 TXOP3 TXOP

! X f A P s ____________ [________ j____________ CAPs

MSDU, MSDU, ... MSDU,

Figure 2-1. The construction of IEEE 802.1 le HCCA.

In HCCA, each wireless station receives QoS CF-Poll frame from the HC before
transmitting data. While the medium is idle for PDFS, the HC can start a CAP by
sending the QoS CF-Poll frame to the station that is requesting to transmit data in order
to inform that particular station of the time allocated for its transmission. The station
should reply to this poll within a time interval equal to Short Inter Frame Space (SEFS).
For downlink transmission, the HC waits for PIFS and then start its transmission. In
HCCA, the HC has higher priority than other wireless stations. Thus, MSDUs for
downlink traffic can be transferred faster than uplink traffic (Villalon et al, 2007).
Figure 2-2 shows the operation of HCCA for uplink and downlink transmission (Lee et
al, 2011).

Start of a new TXOP Uplink Downlink

PIFS ISIFS PIFS

AP QoS CF-Poll Ac QoS data

ISIFS ISIFS SIFS

ST
QoS QoS Ack

TXOP

Figure 2-2. Uplink and downlink transmission between AP and wireless station.

A limitation of HCCA operation is its adaption to traffic rate. The allocation of time
resources to the flows must consider the calculation of scheduling parameters such as

18
Chapter 2 Relevant Theory and Background

SI, and TXOP duration (TD) as these parameters significantly affect the performance of
the scheduler. Another limitation of HCCA is the maximum TD which was set to 8160
pS. This value is an acceptable duration if the traffic rate is relatively small. However, a
sudden increase in traffic rate can exceed the limit of TXOP which in turn affects the SI
to be set to the optimal value. The computational complexity of HCCA and its overhead
can deteriorate the performance of high priority traffic in heavily loaded networks
(Rashid et al, 2007) and (He and Ma, 2011). Due to its limitations, the HCCA
mechanism has not been implemented widely. Thus, the investigation of HCCA
mechanism will be excluded from this study.

2.3.1.1.2 Enhanced Distributed Channel Access (EDCA)

The IEEE 802.1 le EDCA is designed to enhance Distributed Coordination Function

(DCF) in the legacy IEEE 802.11 standard. EDCA is based on a distributed access
scheme and supports service differentiation among different traffic classes (IEEE
Computer Society, 2005). The IEEE 802.l i e EDCA mechanism classifies the traffic
into four access categories (ACs) based on their QoS requirements as depicted in Figure
2-3 (Liang et al, 2006).

Classification of traffic to different access categories:

AC0 ACl AC2 AC3

Backoff Backoff Backoff Backoff

AIFS AIFS AIFS AIFS
CW CW CW CW

Virtual Collision Handler

Transmission attempt

Figure 2-3. The IEEE 802.1 le EDCA model.

Figure 2-3 shows that each AC forms an independent backoff entity with its own queue
and corresponds to a different level of transmission priority. To simplify the concept,
the four access categories (AC0 - AC3) assigned to different traffic priority. The traffic
with the highest priority is assigned to AC0, whereas AC3 is assigned to the traffic with
the lowest priority. The priority of an AC is determined by a set of parameters called
19
Chapter 2 Relevant Theory and Background

EDCA parameters. These parameters are: Arbitration Inter Frame Space (AIFS),
Minimum Contention Window (CWmin), Maximum Contention Window (CWmax),
and Transmission Opportunity (TXOP). The highest priority AC corresponds to the
smallest AIFS, CWmin, CWmax and largest TXOP (Lin et al, 2009).

The AIFS is a replacement of the Distributed Inter Frame Space (DIFS) in IEEE 802.11
DCF access method. The value of AIFS can be determined based on the Arbitration
Inter Frame Space Number (AIFSN) used to determine the length o f AIFS, the time unit
dictated by the physical layer characteristics SlotTime, and the Short Inter Frame Space
period (SIFS) used to manage and control frames. AIFS for particulate AC can be
calculated using equation 2.6 (Alahmadi and Madkour, 2008):

AIFS = AIFSN x SlotTime + SIFS (2.6)

The values of CWmin and CWmax determine the range of contention window for each
AC. These values determine the range of random backoff slots which in turn control the
waiting period for traffic before accessing the channel (Abeysekera et al, 2009). The
backoff value of a particular AC can be determined using equation 2.7:

Backoff [AC] = random [0, min (2K (CWmin [AC] + ! ) - ! , CWmax [A C ]] (2.7)

where K is the number of collisions for the currently transmitted frame, CWmin, and
CWmax are the minimum and maximum contention windows respectively. The high
priority traffic with small CW values have small waiting period before accessing the
medium whereas low priority traffic with a large CW values have a long waiting time to
access the channel.

The transmission opportunity (TXOP) defines the transmission holding time for each
AC. Each AC can transmit for a certain time interval whose length is determined by
TXOP Limit.

During the operation of IEEE 802.1 le EDCA as shown in Figure 2-4 (Lin et al, 2009),
each AC contends to access the channel as an individual virtual station and start its
backoff procedure after detecting the channel is idle for an AIFS period. AC with the
smallest AIFS has the highest priority and needs to defer for its corresponding AIFS
interval. When a particular station can initiate its transmission, it will be allowed to
transmit multiple data frames from the same AC continuously during time interval
defined by TXOP. The highest priority AC has a largest TXOP period, while the lowest

20
Chapter 2 Relevant 1 neory ana nacKgrouna

priority AC has a smallest TXOP period. Similar to a real packets collision, when an
internal collision occurs among the ACs within the same station, the higher priority AC
has the rights to transmit whereas the lower priority AC suffers from a virtual collision.

AIFS [AC2]
Backoff Counter
Window Frozen

AIFS [AC1]
Backoff
Immediate access when medium Counter
Window
is idle >= AIFS fACl Frozen
AIFS [ACO] Contention Window [0,CW[AC]]
AIFS [AC]
Busy Medium Backoff Next
Window Frame

Slot Time
Defer Access Select Slot and decrement backoff
as long as medium stays idle

Figure 2-4. IEEE 802.1 le EDCA operation.

IEEE 802.l i e EDCA aims to ensure better service to high priority traffic and offer a
reduced service for low priority traffic. The simplicity of EDCA operation as compared
with HCCA means it is implemented widely. However, although EDCA mechanism
improves the QoS of time sensitive applications, its performance is not optimal because
EDCA parameters are not adjustable according to the network conditions. EDCA
mechanism is reported to be unable to guarantee a good performance when the network
traffic load was high. The main reason was the excessive number of packet collisions,
which in turn was due to the fixed transmission parameters values assigned to the ACs
(Villalon et al, 2007) and (Lin et al, 2009).

The limitations of EDCA mechanism make it an area of research to improve its

performance. The study reported in this thesis, focused on the IEEE 802.l i e EDCA.
The allocation of traffic to the fixed AC will be investigated in order to enhance EDCA
performance.

21
Chapter 2 Relevant Theory and Background

2.3.2 QoS in Wired Networks

Most current wired networks that are based on IEEE 802.3 have error rates when
transmitting multimedia applications as compared with wireless networks. This is due to
a higher bandwidth which ranges between 10Mbps - 100 Mbps. However, with the
growth in transmission of real-time applications over wired networks, over-provisioning
which refers to enhancing the network capability by simply providing the network with
enough bandwidth, in order for all traffic to meet their QoS requirements might not be
an optimal solution. This is because over-provisioning approach can be difficult and
costly (Fraleigh et al, 2003). In addition, over provisioning bandwidth in the wired
network may not prove effective in dealing with QoS requirements of multimedia
applications (as bandwidth is a costly resource).

Multimedia applications must have priority over elastic applications because of their
higher sensitivity to QoS parameters. For example, high delay variation o f VoIP packets
would affect its quality which in turn leads to a stuttering communication with pops and
clicks.

An important issue that affects the QoS in wired networks is traffic prioritisation. When
multiple packets are serviced through a bottleneck such as a router in the same manner,
this would negatively affect the QoS, as it ignores the QoS requirements of transmitted
applications. The transmitted packets could for instance experience several types of
delay such as queuing delay, which occurs in the output buffer of a router. Also, when
the buffer of the congested router is overflowed, the transmitted packets could be
dropped regardless to their QoS requirements.

Therefore, an efficient utilisation of network resources must be considered during the

transmission of different applications in order for the packets to be serviced according
to their QoS requirements. This in turn improves network's QoS at the wired side of the
network.

There are several mechanisms to support QoS in wired networks. These include: Call
Admission Control (CAC), bandwidth reservation, congestion-management,
congestion-avoidance, traffic policing, and shaping (Szigeti and Hatting, 2005). In this
study, the network traffic prioritisation will be investigated as one of the most important
issues that affect the QoS in wired networks. Therefore, packet scheduling mechanisms
which are the most commonly congestion-management tools will be considered.

22
Chapter 2 Relevant Theory and Background

2.3.2.1 Packet Scheduling Mechanisms

Conceptually, the phrase “scheduling” refers to a set of rules that determine how a
frame or packet is treated when congestion or bottleneck occurs at the convergence
point. When bottleneck occurs, devices such as routers have buffers that allow packets
to be stored temporarily in order to be scheduled subsequently. This process is known as
“queuing”.

The two terms “scheduling” and “queuing” are complementary and their processes are
intertwined. Queuing process is engaged only when congestion occurs and deactivated
after the congestion is cleared. Similar to queuing, scheduling takes place when packets
experience congestion. However, the scheduler has to decide which packet should be
transmitted next, even when there is no congestion. Packet queue scheduling determines
bandwidth allocation among transmitted packets and the manner to service various
applications with respect to different QoS requirements (Szigeti and Hatting, 2005).

There are a number of queuing scheduling mechanisms. These include: First-In-First-

Out (FIFO), Priority Queuing (PQ), Fair Queuing (FQ), Weighted Fair Queuing (WFQ),
and Weighted Round Robin (WRR) (Semeria, 2001). These mechanisms are explained
in the following subsections.

2.3.2.1.1 First-In, First-Out (FIFO) queuing mechanism

FIFO queuing mechanism is the most basic queue scheduling algorithm. It is also
known as First-Come-First-Served (FCFS). The incoming packets are accepted in order
of arrivals. Figure 2-5 shows the process of FIFO scheduling mechanism.

FIFO serves the first packet in the queue regardless of any prioritisation or even fairness.
This feature makes it to be the simplest scheduling mechanism in terms of
implementation. However, FIFO is drop-tail based, when the buffer at the router
implementing FIFO mechanism becomes full, the arrived packets are dropped.
Therefore, FIFO can be insufficient mechanism in meeting QoS requirement for
particular applications such as real-time applications (Semeria, 2001) and (Hasegawa et
al, 2002).

23
Chapter 2 Relevant Theory and Background

Flow 1

Flow 2
FIFO

Flow 3 Port

Flow 4

Flow 5

Flow 6

Figure 2-5. The process of FIFO scheduling mechanism.

2.3.2.1.2 Priority Queuing Mechanism (PQ)

PQ is designed to provide a simple method of supporting service differentiation among

a variety of applications. The fundamental concept of PQ is to classify traffic in
different classes according to their QoS requirements and then place them into different
priority queues. Packets placed in the highest priority queue are served first, whereas the
packets located in the lowest priority queue are severed only when higher priority
queues are cleared. Lower classes could suffer from starving issue which in turn leads to
a significant rate of packet dropping (Miaji and Hassan, 2010). In PQ mechanism,
within each priority queues, packets are served by FIFO scheduling mechanism. Thus,
any packet arrives in the lower priority queue will be dropped without any consideration
if that particular queue is full. Figure 2-6 illustrates the process of PQ scheduling
mechanism (Semeria, 2001).

Packet classification Packets Reassemble

Scheduler

Flow 1
Highest
Flow 2 □HUD priority

Flow 3

Flow 4
mm Middle
priority

Flow 5
mm Lowest
priority

Flow 6

Figure 2-6. Priority queuing scheduling mechanism.

24
Chapter 2 Relevant Theory and Background

2.3.2.1.3 Fair Queuing Mechanism (FQ)

FQ scheduling mechanism was proposed by John Nagle in 1987 (Miaji and Hassan,
2010). It is designed to ensure that each flow has a fair distribution of the bandwidth
regardless of the traffic transmission rate. In FQ, packets are first classified into flows
and then each flow is assigned to a queue dedicated for that particular flow. During the
scheduling process, flows are serviced one packet at a time in round robin order and
empty queues are skipped. If a packet arrives at an empty queue after the scheduler is
visited, the packet has to wait in that queue until the next visit of the scheduler. Figure
2-7 shows the principle of FQ (Semeria, 2001).

Packet classification

\ Scheduler

\
Flow 1

Flow 2

Flow 3 Port

Flow 4
□U ED
Flow 5

Flow 6 1

Figure 2-7. Fair queuing scheduling mechanism.

2.3.2.1.4 Weighted Fair Queuing Mechanism (WFQ)

WFQ was proposed by Zhang, Demers, Keshav and Schenke in 1989 to address the
limitations of PQ, and FQ mechanisms. WFQ allocates bandwidth to different flows
according to their assigned weights in order to satisfy the QoS requirements for
different applications (Balogh and Medvecky, 2011). However, in WFQ, flows with
large packet size are not allocated more bandwidth than flows with small packet size.
Therefore, the distribution of bandwidth among variable length packets is carried out by
a weighted bit-by-bit round robin scheduling. This approach supports fair distribution of
bandwidth because it takes into the account the length of transmitted packets. Figure 2-8
shows a weighted bit-by-bit round robin scheduling serving three flows. 50% of the
bandwidth is assigned to flow 1 whereas the remaining bandwidth are allocated to flows
25
Chapter 2 Relevant Theory and Background

2 and 3 (i.e. 25% to each flow). During scheduling, 2 bits are transmitted from flow 1, 1
bit from flow 2, and 1 bit from flow 3. This causes the packet with 600 byte to be
transmitted before the packet with 350 byte, which in turn is transmitted before the
packet with 450 byte (Semeria, 2001).

However, WFQ implements complex algorithm requiring a significant amount per-

service class, and iterative check on each packet arrival and departure. The
computational complexity affects the performance of WFQ to support a large number of
flows on high speed network interfaces (Senkindu and Chan, 2008) and (Balogh and
Medvecky, 2011).

Packet scheduler using

a weighted bit-by-bit Packet Reassemble
round robin
\
Flow 3
25% bw 450

Flow 2
25% bw 350

Flow 1
50% bw 600

Figure 2-8. WFQ using a weighted bit-by-bit round robin scheduler.

2.3.2.1.5 Weighted Round Robin queuing mechanism (WRR)

WRR is also known as Class Based Queuing (CBQ). The operation of WRR is shown in
Figure 2-9. The packets sent throughout the outgoing port are first classified into
different classes and then assigned to a queue that is particularly dedicated to that class.
The queues are in turn serviced using the weights associated with them. The weights
indicate the number of packets to be sent for each class in a single service round. The
number of packets transmitted for queue (/) is calculated using equation (2.8)

Wi
Number o f serviced packets^ ——— X R (2.8)

Where W, is the associated weight for queue (/), n determines the number of the queues,
and R is the link capacity (Senkindu and Chan, 2008).

In WRR, capacity is allocated to different classes either by allowing high priority

queues to send more packets in a single service round, or allowing each queue to send a

26
Chapter 2 Relevant Theory and Background

single packet each time it is visited but high priority queues are visited multiple times
during a single service round (Semeria, 2001).

Queuel ,Weightl Queue scheduling

Packet classification

Queue2 ,Weight2

Port
Queue n-1,Weight n-1

Queue n ,Weight n
Packets sent through the port
Sent packets

Figure 2-9. The operation of WRR.

From the aforementioned packet scheduling queuing mechanisms, it can be observed

that the trade-off between these mechanisms is their complexity, control ability, and
level of fairness. Therefore, an aspect of this study is an appropriate utilisation of a
suitable queuing scheduling mechanism.

FIFO does not support QoS because it treats all traffic equally. It serves the first packet
in the queue regardless of any prioritisation or fairness. PQ provides premium service to
the high priority traffic at the expense of the low priority traffic. Low priority traffic are
denied access to the buffer space, whenever high priority traffic is transmitted, which in
turn causes the low priority traffic to experience excessive delay and high packet
dropping. FQ is not designed to support traffic with different QoS requirements, as it
allocates the same amount of bandwidth among multiple traffic. The computational
complexity of WFQ algorithm affects its scalability to support lager traffic with
different requirements at the edge of the network. WRR is designed to address the
limitations of FIFO, PQ, and FQ by classifying traffic based on their QoS requirements,
and ensuring that low priority traffic has access to buffer space and output port
bandwidth. The implementation of WRR is more popular and its operation is less
complex comparing with WFQ. Although WRR does not take packet size into account,
an accurate bandwidth allocation could provide an optimal algorithm for usage in
modem multimedia networks (Balogh and Medvecky, 2011). Therefore, in this study,
WRR was considered to provide traffic prioritisation because of its practicality and low
complexity.
27
Chapter 2 Relevant Theory and Background

2.4 Statistical and Artificial Intelligence Techniques

This section explains the principles of statistical and AI techniques used in this study.
The basic concepts of regression analysis as one of the most widely employed statistical
modelling tools is explained, and the fundamental principles of fuzzy logic, fuzzy
clustering, and neural network as the most popular AI techniques are also discussed.

The aforementioned techniques were used in this study to fulfil the main objectives of
this thesis. The use of these techniques is as follows:

i. Fuzzy Inference System (FIS) mechanism was employed to develop an approach

that allowed QoS parameters of multimedia applications to be collected
efficiently and accurately.
ii. Fuzzy C-Means (FCM) and Self Organizing Map (SOM) (i.e. Kohonen neural
network) were used to develop techniques to analyse QoS parameters of
multimedia networks.
iii. Regression model and Multi-Layer Perceptron (MLP) (i.e. a supervised neural
network) were employed to develop techniques to assess QoS parameters of
multimedia networks. The analysed QoS parameters by FCM and SOM were
combined and then a single value that represented the overall network's QoS was
produced.

2.4.1 Regression Model

Different regression models are used for prediction; they can be classified into linear
and nonlinear model. Linear models include Auto Regressive (Box and Jenkins, 1976),
Moving Average (Vandaele, 1983), and mixed of Auto Regressive and Moving Average
(Box and Jenkins, 1976) and (Ljung, 1999). The nonlinear models include Bilinear
Model, Threshold Auto Regressive Model, and Exponential Auto Regressive Model
(Priestly, 1988). A detailed description about linear and nonlinear regression models is
provided by (Box and Jenkins, 1976), (Vandaele, 1983), (Ljung, 1999, and (Priestly,
1988).

Due its simplicity and effectiveness, multi linear regression model is a commonly used
method for prediction purposes (Chatteijee and Hadi, 2006) and (Sweet and Grace-
Martin, 2010). Therefore, in this study, this type of regression is used.

28
Chapter 2 Relevant Theory and Background

2.4.1.1 Multi-Linear Regression Model

The multi-linear regression model is a widely employed statistical method due to its
effectiveness for creating functional relationships among variables (Jain, 1991). Its aim
is to analyse the relationship between several variables. One variable is considered to be
the dependent or response variable and the others are considered to be independent or
descriptive variables (Chatteijee and Hadi, 2006). In order for the regression model to
be valid and accurate predictor, there are some assumptions that need to be followed
(Jain, 1991) and (Chatteijee and Hadi, 2006). These are as follows: (i) dependent
variable and independent variables need to be linearly related, (ii) the independent
variables is non-stochastic and measured without error, and (iii) model errors are
independent and normally distributed.

The formula of multi linear regression model as shown in equation (2.9) defines the
relationship model between dependent variable ( y ) and independent variables
(xlt x 2, ...., xn) as follows (Chatteijee and Price, 2006) and (Jain, 1991):

y = b0 + + b2x 2 + ••• + bnxn + e (2.9)

In vector notation, the regression model can be written as in equation (2.10) or as in

equation (2.11).

yi T *11 *kl \b01 ■Cl'

Y2 —
1 *12 *k2 hi e2
+
y n. 1 *ln *kn- bn. .en.

Y=XB + e (2.11)

where Y is a (nX 1) vector of dependent variable, X is a matrix of independent variables

with size of (nX (k+1)), n is the number of observations, k is the number of independent
variables, B= {b0, blf b2, ... bn] are the regressioncoefficients determined from
recorded data,and e is a column vector of n error terms. Theregression coefficients are
calculated using equation (2.12).
B = (X TX )~1X TY (2.12)
where is X T is an inverse of matrix X. The vector of residual e (i.e. error terms) is given
by equation (2.13).

e = Y —XB (2.13)
29
Chapter 2 Relevant Theory and Background

2.4.2 Fuzzy logic

Fuzzy logic was introduced by Lotfi Zadeh in 1965 (Zadeh, 1965) as a methodology for
computing words rather than numbers. The concept of fuzzy logic is based on natural
human communication language because it has similarities with human knowledge and
reasoning (Klir and Yuan, 1995). The robustness of fuzzy logic due to the direct
expression of input/output relationships without a physical derivation of the rules, and
its flexibility to cope with imprecise and uncertain information and then draw definite
conclusions makes it an excellent and powerful mechanism (Jantzen, 1998) (Khoukhi
and Cherkaoui, 2008) and (Muyeen and Al-Durra, 2013). Unlike Binary logic (i.e.
Classical logic) which has a sharp boundary between true and false states, fuzzy logic
implements a gradient of possible states between true and false as shown in Figure (2-
10) (Cirstea et al, 2002).

Fuzzy logic is applied to many applications in various domains such as control, decision
making, optimisation, and evaluating systems (Klir and Yuan, 1995) (Naoum-Sawaya
and Ghaddar, 2005) (Saraireh et al, 2008) and (Muyeen and Al-Durra, 2013).

True False True False

Classical logic Fuzzy logic

Figure 2-10. Binary logic versus fuzzy logic.

2.4.2.1 Fuzzy Inference System (FIS)

Fuzzy Inference System (FIS) is built upon the theory of fuzzy logic. FIS includes four
main components: fuzzification, rules base, inference engine, and defuzzification as
show in Figure (2-11) (Jantzen, 1998).

Rule Base

Inputs ^ Fuzzification
I
Inference
Defuzzification ■ f Output

engine
i :

Figure 2-11. Block diagram of fuzzy inference system.

30
Chapter 2 Relevant Theory and Background

FIS is used to interpret (i.e fuzzify) the crisp values of inputs into linguistic terms, and
based on a set of predefined rules, it calculates linguistic output value which in turn is
converted (i.e. defuzzified) into real crisp output value (Naoum-Sawaya and Ghaddar,
2005) and (Saraireh et al, 2008). The following subsections outline each component of
FIS.

2.4.2.1.1 Fuzzification

This is a process of converting numerical input values into linguistic terms and
determining the degree of belonging to the appropriate fuzzy sets via membership
functions. In fuzzy sets, Cirstea et al (2002) reported that an element (jc,) in the universe
of discourse X is assigned a degree of membership p(;c,) obtained by a membership
function as shown in Figure (2-12). A membership function allows gradual transition
from full-belonging to a fuzzy set to not-belonging at all with intermediate values
presenting degrees of belonging (Al-Sbou et al, 2006). In fuzzification process, different
types of membership functions can be employed. These include Triangular,
Trapezoidal, Gaussian, Generalised, and bell Sigmoid (Mathworks, 2012(a)).

0 ll X
(X)

Figure 2-12. Degrees of membership in Gaussian membership function.

2.4.2.1.2 Rule Base

This component contains a set of IF-THEN rules represented in linguistic variables. The
set of IF-THEN rules is the bases of decision making process of FIS. The number of
rules depends on the number of inputs and outputs variables as well as the number of
membership functions associated with them (Jantzen, 1998). The common form of IF-
THEN rules as follows:

31
Chapter 2 Relevant Theory and Background

IF (Antecedent) AND (Antecedent) ...... THEN (Consequent)

where the Antecedent relates the linguistic term to a fuzzy set, and the Consequent
represents the conclusion from IF term. Each rule could have one or more connectives
(i.e. fuzzy operators). The most common fuzzy operations applied on IF-THEN rules are
Intersection, Union, and Complement which respectively defined by fuzzy operators
AND, OR, and NOT (Klir and Yuan, 1995) (Ross, 1995) and (Mathworks, 2012(a)). For
example, given that px and pYare the degrees of membership functions for fuzzy sets X
and Y respectively, the application of fuzzy operators AND, OR, and NOT can be
defined as given in equation (2.14) (Ross, 1995):

AND: Hxrw- ntin (Fx, F y)

OR: Fxuy —max (Fx >F y) (2.14)
NOT: f -*= I-Fx

2.4.2.1.3 Inference Engine

Fuzzy inference engine uses fuzzified inputs along with the rules to perform inferencing
(i.e. the process of implication and then aggregation) (Jantzen, 1998). The fuzzified
inputs can be related to more than one rule to specify how adequately each rule
describes the current situation by computing the degree of truth for IF condition. More
than one rule might be triggered at the same time describing the same situation. Each of
these rules produces Consequent or Conclusion to be taken in the THEN condition. This
process is performed by implication method which is defined as the shaping of output
membership functions. The input for the implication is a single number given by the
Antecedent of the rule, and the output is a fuzzy set. The truncated output fuzzy sets
from the implication process which describes the firing strength of the rules is then
processed by an aggregation method. In the aggregation process, the truncated output
fuzzy sets from the implication process are unified to produce one output fuzzy set
(Ross, 1995).

2.4.2.1.4 Defuzzification

This is the process that converts the output linguistic value (i.e. the aggregate output
fuzzy set) into a real numeric value. The input for the defuzzification process is the
aggregate output fuzzy set and the output is a single number. However, the aggregate of
a fuzzy set covers a range of output values which in turn must be defuzzified to produce
a single output value from the set. There are several methods can be used in

32
Chapter 2 Relevant Theory ana Background

defuzzification process such as centroid, bisector, middle of maximum, largest of

maximum, and smallest of maximum (Abdul Aziz and Parthiban, 2006).

Figure 2-13 shows the information flows through the process of fuzzy inference system:
commencing from fuzzifying inputs, through the process of applying fuzzy operator,
implication method, aggregation method, and terminating by defuzzification process
(Math work,2012(a)).
ZApoty 5frp
-AW
Aafon
1. e u z t f y f i p v t
tnemoo

fuzzy ru le 1

TV r>v etoponrjoncY

fuzzy ru le 2-
A .

1 3 a a 4. Appty
aggregation
fuzzy ru le 3 rm tfiod

input 1 input 2

/. * 6 D efu zztfy th e
H esu it OT a g g re g a te o u tp u t
a g g re g a tion (c*n trold>-

R esu /t o f o u tp u t
d efuzzification

Figure 2-13. The process of fuzzy inference system.

Fuzzy inference system has two methods: Mamdani and Sugeno inference methods. The
procedure of fuzzifying the inputs and applying the fuzzy operator during the fuzzy
inference process are similar in both methods. However, the main difference between
Mamdani and Sugeno is the manner the outputs are determined. Mamdani-type FIS is
based on defuzzification process to generate crisp output from output fuzzy set, while
Sugeno-type FIS uses weighted average to compute the crisp output. The advantage of

33
Chapter 2 Relevant Theory and Background

Mamdani FIS is that its outputs are expressed and interpreted. This feature is lost in the
Sugeno FIS since the consequents of the rules are not fuzzy (Arshdeep and Amrit, 2012).
The other difference is that Mamdani FIS has output membership functions whereas
Sugeno FIS has no output membership functions. Due to the interpretable and intuitive
nature of the rule base, Mamdani-type FIS is widely used particularly for decision
support application (Haman and Geogranas, 2008). Therefore, in this study, Mamdani-
type FIS will be used in this study.

2.4.3 Fuzzy clustering

Clustering techniques are used to partition data into correlated groups where different
data objects should belong to different clusters and similar data objects are assigned to
the same cluster. The aim of clustering is to reveal the underlying structures of data and
explore its nature (Rokach and Maimon, 2005). Clustering techniques can be classified
based on many criteria. For instance, clustering techniques can be divided into two main
groups: hierarchical and partitioning techniques (Farley and Raftery, 1998). Another
criterion divided clustering methods into: density-based methods, model-based
clustering, grid based methods, and soft computing methods (Han and Kamber, 2006).
In this study, the focus will be on Fuzzy C-means (FCM) as one o f the most widely
employed soft clustering methods. Unlike hard clustering methods where an object must
belong to only one particular cluster, in FCM, it can belong to several clusters with
different degrees of membership between 0 and 1 to indicate their partial membership
(Parker et al, 2012). Therefore, in this study, FCM was used to analyse network QoS
because the natural characteristics of network QoS patterns partly cover more than a
single cluster.

2.4.3.1 Fuzzy C-Means Clustering (FCM)

FCM clustering algorithms was originally introduced by Dunn in 1973 and improved by
Jim Bezdek in 1981 (Nascimento et al, 2000). This algorithm is one of the most widely
used clustering algorithms. FCM is defined as a mechanism to discover certain features
in a set of data and classify each element of data into a number of clusters with different
degrees of memberships (Wang, 2009) and (Chaabane et al, 2008). FCM can be applied
to partition a set of data with a form of matrix X of size n X AT as shown in equation
(2.15).

34
Chapter 2 Relevant t heory and .Background

'* 1 1 *12 X1N'

*21 *22 *2 N
X = (2.15)
*nl *n2 *** x nN.

where each Xj e Rp, j = 1 , 2 is a given set of feature data representing by a

number of features p. FCM operates on the matrix X and minimises the FCM objective
function given in Equation (2.16) in order to partition matrix X into C clusters (Lei, et al
, 2012).

m n2 (2.16)
] QC'.u.V) = Zf=1Z?=1G»y)m Dfj

The value m controls the degree of fuzziness for the membership of the cluster where m
e [1, oo]. As the value of m decreased, the membership of the cluster becomes closer to
the binary clustering. U the membership matrix includes n x C values and can be
expressed as in formula (2.17):

V11 ^12 Vic

^21 ^22 1*2C
U= (2.17)
nl A*n2 A'nCJ

where each value of the matrixp ^ j = 1, ...,n and i = 1, ...,C indicates the degree of

membership between vector Xj and cluster Q and must meet the following criteria:

• Vij e [0,1],Vj= 1 ,...,C,Vy= 1, ...,n

• 2 f= iJ«y = 1. Vy=

During the clustering process, the objective function J (U, V) is minimized with the
following iterative steps (Parker et al, 2012) (Chen et al, 2009) and (Timo et al, 2002):

1. Membership matrix U is initialised with random values considering the

aforementioned membership degree criteria.
2. The clusters' centres V = { v lf v 2, —, v c] are calculated according to Equation (2.18).

{pij) XJ v .= 1 c (2.18)
1 1 .......

3. The distance Dfj which is the Euclidian distance between Xj to the centre of

cluster i which is calculated using equation (2.19).

35
Chapter 2 Relevant Theory and Background

2
(2.19)

4. The elements of matrix U are then updated using Equation (2.20).

i
(2 .20)

Equations (2.18) - (2.20) are repeated until the termination criteria of FCM are met. The
process of FCM can be terminated when the maximum number of iteration is reached or
the objective function improvement between two consecutive iterations is less than the
minimum amount of improvement (Timo et al, 2002).

2.4.4 Artificial Neural Network (ANN)

ANN is an adaptive parallel processing system capable of achieving results through the
process of learning. It provides a mean to model the human brain in a simplified form.
An ANN consists of a number of highly interconnected neurons that learn by interaction
with each other as each interconnection has an associated weight (Abraham, 2005). In
an ANN, the function of each neuron is to receive information, process it, and produce
an output (Haykin, 1999). During the training process, the output is used to adjust the
weight values to optimise the neural network performance. The mechanism to determine
the amount of change in the weights is the neural network’s learning algorithm (Cirstea
et al, 2002) and (Zaknich, 2003). ANNs can be classified based on the manner of
learning into supervised and unsupervised neural network. In supervised learning, they
are provided with training examples from known classes together with their desired
outcomes. In unsupervised learning, the neural network requires only training examples
to be trained (Nogueira et al, 2006).

The advantage of ANN is the ability to derive meaning from imprecise values with
highly parallel computing structure. This capability gives ANN the strength to model
complex systems in efficient and effective manner and then achieve desired results. In
this study, a combination of supervised and unsupervised neural networks was
considered to evaluate the QoS parameters for multimedia applications. The transmitted
application’s QoS parameters were initially analysed by the unsupervised learning
Kohonen neural network. The analysed QoS parameters were then used as inputs to a
supervised learning Multi-Layer Perceptron (MLP) neural network in order to quantify

36
Chapter 2 Relevant Theory and Background

the overall QoS. An explanation of MLP and Kohonen neural network (i.e. Self-
Organising Maps SOM) are in the next sections.

2.4.4.1 Multi-Layer Perceptron (MLP)

In this study, a multi-layer perceptron (MLP) neural network was employed to assess
the overall QoS due to its suitability and effectiveness. As one of the most popular
supervised ANN, MLP needs to be provided with representative examples from each
class, together with their corresponding class category. The architecture of MLP neural
networks is shown in Figure 2-14 (a) (Abraham, 2005).

MLP composes of an input layer, one or two hidden layers and an output layer. At each
layer, there are a number of neurons (i.e. processing elements). The function of each
neuron is to receive information (i.e. the inputs with their associated weights), process
them, and produce an output (Zaknich, 2003).

The operation of an MLP is shown in Figure 2-14 (b). During the training phase, a
known pattern (x,-) is applied to the input layer of the MLP, and its target (i.e. desired
value (d)) is applied to the output layer of the MLP. The elements of input (x,-) are
multiplied with their associated connection weights (w,) and the resulting value (s) is
obtained using a summation function as in equation (2.21) (Cirstea et al, 2002):

(2.21)

The output from the summation function (5 ) is processed by an activation

function tp (s ) to provide the output (y). The activation function ensures that the

neuron's output is limited to a predefined range (such as 0 to 1 or -1 to +1). There are

several activation function types. These include threshold function, sigmoid, signum,
and hyperbolic tangent (Cirstea et al, 2002) and (Mathworks, 2012(b)). Therefore, the
value of y depends on (5) and the type of activation function. Subsequently, the
calculated output (y) for each neuron at output layer is subtracted from the desired
output (d) to produce an error (e) as expressed in equation (2.22):

e — d —y (2 .22)

The calculated error is then used by the learning algorithm to adjust the weights in order
to reduce the error in the next iteration.

37
Chapter 2 Relevant Theory and Background

The process of learning and adapting are achieved by the learning algorithm. The
function of the learning algorithm is to use the calculated error (e) and the input data (x,-)
to adjust the values of the connections' weights (w,) in such a way as to reduce the
magnitude of the error in the following training iteration (Haykin, 1999). There are
several methods to realise the required learning. These include gradient descent back-
propagation, gradient descent with adaptive learning rate back-propagation, gradient
descent with momentum back-propagation, and gradient descent with momentum and
adaptive learning rate back-propagation (Abraham, 2005),(Cirstea et al, 2002).

Input layer Hidden layer Output layer

wl4 S~\ w48

(a)
Inputs Outputs

w37 w79 Processing

elements
i—
Inputs Weights

Summation and d (desired

transformation output)

(b) Output

Error

Learning algorithm

Figure 2-14. Multi-Layer Perceptron: (a) The architecture, (b) MLP operation.

The MLP training process is repeated until either the specified number of iterations is
reached or when there is insignificant error between the network output (y) and the
desired result (d) (Mathworks, 2012(b)). The termination criteria evaluate how effective
the MLP is trained.

After the completion of training phase, the trained MLP is examined during the test
phase. Unknown input patterns are fed to MLP input layer and outputs are produced by
the output layer.

38
Chapter 2 Relevant Theory and Background

2.4.4.2 Kohonen Neural Network

In contrast to the multilayer perceptron, the Kohonen network (i.e Self Organising Maps
SOM) is one of the most popular unsupervised neural networks require only training
examples to learn (i.e. no desired output is required). Self-Organizing Map was
introduced by Teuvo Kohonen as a data visualization technique (Kohonen, 1982). SOM
visualizes data by reducing its dimensions to a map, and represents similar data objects
into correlated clusters. An aspect of this study is that QoS parameters of multimedia
applications were intelligently classified using SOM. In situations such as network QoS,
where the natural characteristics of traffic cover multiple clusters, SOM could be an
effective clustering technique to analyse network QoS patterns.

The Kohonen network has a single layer of neurons known as a Kohonen map as shown
in Figure (2-15). The Kohonen map can be arranged in various topologies such as
rectangular and hexagonal topology.

A neighbourhood neuron

The winning neuron

Inputs
Xj
x2

Xi

Kohonen map A neighbourhood region

Figure 2-15. The structure of kohonen neural network.

As shown in Figure (2-15), the j th neuron in Kohonen map is connected to ith input
feature of a certain pattern (x) to the neural network. Each connection has an associated
weight(wij). The connections' weights are initially set to random values between 0 and

1. Then, the network learns by determining the Euclidean distance dj between the

features of normalised input pattern (xi) and the neuron’s weights. For each j th neuron
and N features for each example, the Euclidean distance is calculated using equation
(2.23).

39
Chapter 2 Relevant Theory and Background

di = pSLoKxi - wi] ) 2 (2.23)

The neuron that obtained the smallest Euclidean distance with the input patterns is
considered as the winning neuron. Its weights are adjusted using the learning algorithm
as expressed in equation (2.24).

Wij(n + 1) = w tj(ri) + 77 (ri)(pCi(ri) - wl7(n )) (2.24)

where, Wij(n + 1) is the updated weight, tj(n) is the learning rate 0 < a < 1, and the

term (** (n) — Wj; (n )) represents the error. The learning rate is usually a value between

0 and 1 which in turn controls the adaptation speed. The learning algorithm ensures that
the winning neuron's weights become iteratively closer to the input pattern category.
This in turn allocates the winning neuron to become representative o f that specific
category.
The weights associated with a number of neurons around the winning neuron are also
updated to a lesser extent. This enables specific regions of the Kohonen map to be
associated with different pattern categories. The neurons around the winning neuron
which their weights are updated are known as neighbourhood neurons, and the area of
the Kohonen map covered by them is referred to as the neighbourhood region as shown
in Figure (2-15) (Haykin, 1999).

2.5 Summary
This chapter provided a theoretical background related to the QoS of multimedia
networks. This includes the definitions, QoS parameters, QoS requirements of
multimedia applications, and QoS components. It also discussed the QoS in wireless
and wired networks. The IEEE 802.l i e as an emerging WLANs standard to provide
QoS was explained, and packet scheduling mechanisms, the most commonly
mechanisms implemented in wired network to support QoS were reviewed. This chapter
also provided the relevant theoretical background to statistical and Artificial Intelligent
(AI) techniques which were used in this study. The fundamental principles of regression
analysis which is one of the most widely employed statistical modelling methods were
discussed. The basic concepts of fuzzy logic, fuzzy C-means clustering algorithm,
Multilayer Perceptron Neural Network, and Kohonen neural network were explained.
The next chapter reviews the literature relevant to the process of managing QoS for
multimedia computer networks
40
 Chapter 3 Literature Review

3.1 Introduction
Managing Quality of Service (QoS) of multimedia applications is currently one of the
principle research topics in the field of computer networks. Two important factors make
the QoS management an issue of great importance, (i) Computer networks are
increasingly integrated (i.e. networks consist of heterogeneous networks: wired and
wireless), (ii) Computer networks carry diverse multimedia applications involving
video, audio, and data which require a large bandwidth and perceived quality
(Mohammed et al, 2001).

In this study, network QoS management refers to evaluation and improvement of QoS
provided by computer networks. Evaluation of QoS aims to analyse and quantify
network performance with respect to meeting the applications' transmission
requirements. The QoS improvement involves the ability to take actions to enhance
network QoS or change network performance toward a desired operation. However,
there are many issues related to the process of managing multimedia computer
networks. The main issues are: (i) Multimedia applications generate an extensive
amount of data in the form of information packets. The collection and processing of all
these packets in real time are not practically feasible, (ii) Gathered network information
which represents network performance in delivering diverse applications include a
multitude of parameters related to QoS. These parameters need evaluating in an
effective manner, (iii) Multimedia applications require a large bandwidth, that in turn
are considered to be as a critical issue because of the limitations of the physical
communication channels as well as the interfering noise, particularly in the wireless side
of the network. Accordingly, the QoS of the transmitted applications could be
unpredictable, (iv) The existing monitoring tools are unable to get directly the overall
network QoS. Network managers have to do an extra evaluation to assess the overall
network QoS. This process could be complicated, expensive, and time consuming.

In this chapter, the previous studies relevant to the aforementioned issues of managing
QoS of multimedia network are critically analysed and discussed. The aim of this

41
Chapter 3 Literature Review

chapter is to identify the potential gaps associated with these issues which in turn
require a further development and investigation.

This chapter is organised as follows: section 3.2 reviews the sampling techniques used
to gather information from network traffic. In section 3.3, the current QoS analysis and
assessment techniques used to evaluate network QoS are discussed. The relevant studies
considering QoS support in wireless and wired networks are evaluated in section 3.4.
Section 3.5 critically analyses the exiting monitoring tools used to assess the network
performance. Section 3.6 reviews the applications of statistical and artificial intelligent
techniques (used within this study) into the field of computer network management.
Finally, in section 3.7, a summary of this chapter is provided.

3.2 Sampling Approaches for Measuring QoS

Parameters
The growth in real-time transmission of multimedia applications over computer
networks means that the QoS parameters of these applications need to be recorded and
measured in an efficient manner. The quantification of QoS parameters allow QoS
provided by the network for the transmission of these applications to be assessed.
However, most real-time applications generate an extensive amount of data in the form
of information packets. The collection and processing of all these packets in real time
are not practically feasible and computationally intensive. Therefore, in order to reduce
the amount of collected data, sampling operation needs to be performed.

Sampling techniques are used to analyse the statistical characteristics of a population of

packets and to produce subset traffic with smaller number of packets that represents the
original traffic. Sampling techniques can be classified into two categories: adaptive
sampling and non-adaptive sampling (Hernandez et al, 2001).

3.2.1 Non-adaptive Sampling Techniques

In non-adaptive sampling techniques, the packet selection method can be time interval
based or packet number based. In the former method, the selection is based on
predefined time intervals, whereas in the latter, the packet selection decision is based on
a packet count (Zseby, 2004) and (Claffy et al, 1993).

42
Chapter 3 Literature Review

Non-adaptive sampling techniques can be classified into systematic, random, and

stratified. In a systematic sampling scheme, a packet is chosen at either a fixed time
interval or a fixed number packet count. In random sampling, a packet is chosen from
the parent population at a random time interval or in a random packet count number. In
stratified sampling, a fixed interval of time is chosen and a packet is randomly selected
from that interval (Claffy et al, 1993) and (Gan et al, 2009). Figure 3-1 (a) shows
systematic, random, and stratified sampling techniques respectively (Claffy et al, 1993).

1 d d d d
Systematic sampling V
oC3
b • • • 1 O
Random sampling

h 1• 1 I
Stratified sampling Sampling interval
(a) (b)

Figure 3-1. Classification of sampling techniques: (a) Non adaptive sampling, (b) The
concept of adaptive sampling.

The use of predefined sampling rate in non-adaptive sampling schemes to determine

how to sample packets may not be effective for sampling multimedia traffic. This is
because multimedia traffic is time varying. In case of sampling multimedia traffic using
non-adaptive sampling techniques, two situations might occur: (i) if the traffic rate is
too high, there is a risk of losing information caused by under sampling, (ii) If the traffic
rate is too low, resources would be underutilised by over sampling (Giertl et al, 2006).

Gathering traffic information from the network using conventional sampling techniques
poses a challenge as the process should reflect on the nature of ongoing network traffic
activity. Therefore, in order to reduce the biasness presented in conventional sampling
methods and to increase the effectiveness of measuring QoS parameters, traffic should
be sampled adaptively. This process is illustrated in Figure 3-1 (b).

3.2.2 Adaptive Sampling Techniques

Unlike non-adaptive sampling techniques, the sampling rate of adaptive sampling

techniques (i.e. the packet selection method) is adjusted during the sampling process
according to traffic characteristics. In other words, small sample interval are required

43
Chapter 3 Literature Review

during the period of high activity, whereas a larger sample interval is required during
the period of low traffic activity (Giertl et al, 2006).

A number of studies have been conducted using adaptive sampling to gather network
information. For example, in (Hernandez et al, 2001), two adaptive sampling schemes
have been proposed based on linear prediction and fuzzy logic. The experimental results
showed that both approaches achieved better results compared with systematic
sampling.

An adaptive sampling based on fuzzy logic was reported in (Giertl et al, 2006). The
contribution of the study was to assess the utilisation of bandwidth for real TCP/IP
activity of network interface, connecting a local network to the Internet using adaptive,
random and stratified sampling techniques. The study showed that adaptive sampling
was more effective than the non-adaptive sampling techniques.

In (Ma et al, 2003, and 2004), an adaptive sampling technique was devised using the
time-sliding window (TSW) algorithm to estimate traffic rate. A key element of the
devised technique was to predict future behaviour based on observed behaviour. The
experimental results showed that measuring QoS parameters of voice traffic using the
devised adaptive sampling technique was more accurate than the conventional sampling
techniques. Gan et al (2009) proposed another method to sample packets in an adaptive
manner. Their adaptive sampling algorithm automatically adjusted the sampling
granularity according to the packet arrival interval. Their results showed that the
proposed adaptive sampling method is more accurate and economical than static
sampling methods.

Adaptive sampling methods are not only used for traffic measurement, but they also can
be used for a number of other applications (i.e. network security application). For
instance, Zhang et al (2007) proposed a small packet threshold adaptive sampling
algorithm to capture malicious packets, which speeded up the sampling process and
achieved accurate results. Patcha and Park (2006) in their paper proposed an adaptive
sampling algorithm based on weighted least squares prediction. The proposed sampling
algorithm was tailored to enhance the capability of network based IDS at detecting
denial of-Service (DoS) attacks. The algorithm was not only proposed to reduce data
analysed by IDS but also, it maintained the intrinsic self-similar characteristic of
network traffic. This feature can be used by IDS to detect DoS attacks by using the fact

44
Chapter 3 Literature Review

that a change in the self-similarity of network traffic could be an indicator of a DoS

attack.

From the above discussion, most of the previous adaptive sampling schemes require
sophisticated computations as in the case o f (Hernandez et al, 2001), (Giertl et al, 2006),
and (Giertl et al, 2008). Other schemes consider one parameter as the reference-
coefficient of sampling granularity as the case in (Gan et al, 2009). None of the previous
studies considered the traffic’s statistics such as mean, median, and standard deviation
during the process of sampling multimedia traffic in an adaptive manner.

The contribution of this study is to propose a statistical adaptive sampling method by

considering the traffic's statistics. The method increases the interval between two
consecutive sampled sections, when the overall statistic of the traffic does not change
for those two sections, and the interval is decreased when the overall statistic o f the
traffic for the two sections significantly differs. The advantages of this sampling method
are the ease of its implementation and its fast response to the variation of the input
traffic.

3.3 Network Quality of Service Evaluation

The growth in transmission of critical real-time applications such as VoIP and video
applications over computer networks means that their QoS needs evaluating in an
effective manner. QoS evaluation of multimedia networks is very important for end
users, network managers, and service providers. Users are interested in determining how
well they are receiving services and whether the received services meet the agreed
levels of service between them and the service providers (Gozdecki et al, 2003).
Network managers need to evaluate QoS to determine how well their networks are
operating in order to identify network failures and to optimise network performance.
Service providers need to evaluate network QoS in order to comply with the level of
QoS that the customer expects (Molina-Jimenez et al, 2004).

The current evaluation of QoS is achieved either by analysis or measurement

techniques. The analysis techniques examine the characteristics of network traffic
(Timo et al, 20022), (Wang et al, 2009), and (Ting et al, 2010), whereas the
measurement techniques determine how well the network treats the ongoing traffic
(Palomar et al, 2008), (Teyeb et al, 2006), (Mishra and Sharma, 2003), (Pias and

45
Chapter 3 Literature Review

Wilbur, 2001), and (Malan and Jahanian, 1998). The following subsections discuss the
current QoS analysis and measurement techniques used to evaluate network QoS.

3.3.1 Quality of Service Analysis Techniques

Analysis of network QoS plays a crucial role in practicing effective management for
multimedia computer networks. The aim of QoS analysis is to obtain a comprehensive
view about the state of the network and simultaneously discover important details from
the transmitted traffic (Timo et al, 2002). The network traffic analysis has been
investigated by a number of studies using several mechanisms. The most explorative
techniques used to analyse the characteristics of network traffic are statistical analysis
techniques and AI techniques (Liu et al, 2012) and (Ting et al, 2010).

Statistical parameters such as mean, standard deviation, and mode that are used to
analyse network traffic may not be as effective as AI techniques. This is because
outliers of the analysed traffic using statistical parameters could affect the final
conclusion about traffic characteristics which in turn may give inaccurate results about
the analysed traffic (Jain, 1991). Therefore, statistical parameters will not be used in this
study to analyse QoS.

Due to the ability of AI techniques to derive meaning from imprecise values as

demonstrated by (Cirstea et al, 2002) and (Haykin, 1999), and because network traffic is
highly complex in natural as reported in (Wang et al, 2009), and (Ting et al, 2010), this
study will include AI methods to analyse the characteristics network traffic.

Fuzzy C-means (FCM) and Self-Organising Maps (SOM) (i.e. Kohonen neural
network) were previously used in a number of study to analyse network traffic..

FCM was used to cluster network traffic and produce application profiles which
contained significant information about the current status of the network in order to
manage network resources as reported in (Timo et al, 2002). A network administrator
assistance system was proposed based on FCM. The proposed system utilised a FCM
method to analyse users’ network behaviours and traffic-load patterns based on the
measured traffic data of an IP network. Analysed results were used to assist
administrators to determine efficient controlling and configuring parameters of the
network management (Chen et al, 2009). In wireless sensor networks, FCM algorithm
was used to create clusters which reduced the spatial distance between sensors nodes

46
Chapter 3 Literature Review

(Hoang et al, 2010). FCM was employed to detect routing attacks caused by abnormal
flows in a wireless sensor network. The study demonstrated that FCM can be a valuable
tool for intrusion detection (Wang et al, 2009).

SOM was also used by several studies to analyse network traffic. For example, Kemel-
SOM (KSOM) was proposed to introduce network traffic classification approach (Ting
et al, 2010). The experimental results showed that KSOM achieved high classification
accuracy and successfully categorised network traffic characteristics. In (Kiziloren and
Germen, 2007), network traffic types were analysed using SOM. The aim was to
distinguish between normal traffic and anomaly traffic, such as port scanning and
Denial of Service Attacks. The results demonstrated the usefulness of SOM to
distinguish three traffic types: port scanning, heavy-download, and other traffic.

However, none of the previous studies discussed in this section have utilised FCM or
SOM to analyse and classify QoS parameters (i.e. delay, jitter, and packet loss ratio). A
novel aspect of this study is that QoS parameters of multimedia applications are
intelligently classified using either FCM or SOM. In situations such as network QoS
parameters where the natural characteristics of traffic cover more than a single cluster,
FCM algorithm or SOM could be an effective QoS analysis technique.

3.3.2 Quality of Service Measurement Techniques

Measuring QoS of multimedia applications could be performed using different types of

approaches. These can be objective or subjective. An objective approaches measures the
QoS based on mathematical analysis that compares original and distorted multimedia
signals (CAKY et al, 2006). Examples of some objective methods are Mean Square
Error (MSE) and Peak Signal to Noise Ratio (PSNS) which measure the quality by a
simple difference between frames (Mohammed et al, 2001). Other examples of
objective methods are Perceptual Speech Quality Measure (PSQM) and Perceptual
Evaluation of Speech Quality (PESQ) which used particularly to measure the voice
quality (CAKY et al, 2006).

A number of studies quantified QoS based on objective approaches. For instance, Sun
and Ifeachor (2002) examined the impact of packet loss and gender of talker on
perceived speech quality using ITU PESQ measurement algorithm. In the study, they
found that the packet loss and gender of talker have an impact on perceived speech
quality. The quality for female talkers was worse than the male talkers for the same
47
Chapter 3 Literature Review

network conditions. Palomar et al (2008) assessed the quality o f audio streaming over
WLAN using EAQUAL which is a software tool based on ITU-R for objective
assessment. Nevertheless, the experimental results showed that the EAQUAL was not a
good approach to assess the audio quality in cases like the effect of packet loss.

Subjective approaches on the other hand measure the overall multimedia quality based
on human subjects (ITU-T, 2008) and (Mohammed et al, 2001). Subjective approaches
are carried out by having n human subjects viewing the distorted multimedia signals and
then rate their quality based on predefined scale. The most common scalar used by
subjective approaches is Mean Opinion Score (MOS) (ITU-T, 1998) (Patel et al, 2003).

Several studies have used subjective approaches to evaluate network performance. For
instance, Brauer et al (2008) assessed Voice and Video over IP (VVoIP) quality in IP
networks. They used Mean Opinion Score (MOS) Absolute Category Rating (ACR)
scalar to obtain the quality of VVoIP under certain conditions. Teyeb et al (2006)
evaluated the performance of heterogeneous networks in subjective manner. In their
study, the QoS of web browsing and video streaming services was subjectively
evaluated through usability test. The aim was to find out the effect of network
parameters on users' perception of quality of web browsing and video streaming.

However, objective and subjective approaches have some limitations. Subjective

approaches cannot be automated, they require controlled environment, and they are time
consuming to be repeated frequently due to their dependence on human subjects
(Palomar et al, 2008). While objective approaches require high calculation power
because their operation at the pixel level, and they cannot take into the account all the
affected network parameters (Mohammed et al, 2001).

Another type of QoS assessment is based on passive or active measurements. Active

measurements are carried out by generating probe packets and injecting them into the
network or a portion of the network (Brekne et al, 2002). The concept behind this
approach is that the performance experienced by probing packets gives an indication
about the performance experienced by real traffic.

A number of studies assessed QoS based on active measurement approaches. For

instance, Mishra and Sharma (2003) proposed active measurement approach to select an
appropriate Label Switched Path (LSP) which satisfies the QoS of the new connection.
The selection of LSP was based on the end-to-end delay of probing packets sent along

48
Chapter 3 Literature Review

each LSP. (Ma et al, 2003) and (Ma et al, 2004) used adaptive sampling and active
measurement to evaluate the performance of voice traffic in Multiprotocol Label
Switching MPLS-based IP networks.

In contrast with active measurements, passive measurements are carried out by

measuring the actual network traffic in order to provide an indication about the network
performance. This approach is non-intrusive as packet's information can be gathered
without adding probing packets which might disturb the operation of the network
(Brekne et al, 2002).

There are several studies based on passive approaches to quantify QoS. For instance,
Cranley and Davis (2005) investigated the effect that the background traffic has on
video streaming traffic. Their approach non-intrusively measured and recorded the
bandwidth utilisation of video streaming traffic. Al-Sbou et al (2008) passively
evaluated network performance of multimedia applications in mobile ad hoc networks
(MANET). Their proposed system indicated that the measured QoS can be used as an
indication of the network conditions and resource availability.

However, there are some drawbacks with both passive and active measurements.
Performing active measurements that give representative results is not a trivial task,
because excessive probing generates a significant load that might disturb the operation
of the network, whereas infrequent probing might not reveal the performance
characteristics of the network. The disadvantage of passive measurement however is the
requirement of collecting and processing a large amount of recorded packets which is
not practically feasible in real time (Brekne et al, 2002).

The limitations of the above QoS measurement approaches highlight the need for
further investigation and development. The contribution of this study is to develop QoS
assessment techniques that overcome some drawbacks in subjective and objective
approaches. The techniques should evaluate QoS in a manner similar to human subjects
and quantify the QoS without the necessity for complex mathematical models taking
into the account the QoS requirements of each type of multimedia application. In
addition, the QoS assessment techniques should not add extra load to the network as the
case of active approaches, nor depend on the whole collected packets like passive
approaches. The proposed assessment technique will be based on the analysed traffic
generated from the proposed analysis techniques in order to overcome some drawbacks
of both active and passive measurement methods. An aspect o f this study is that a
49
Chapter 3 Literature Review

regression model is developed and Multi-Layer Perceptron (MLP) neural network was
trained in order to combine the QoS parameters (i.e. delay, jitter, and packet loss ratio)
for each QoS class (or cluster) identified by SOM or FCM to estimate the overall QoS.
This is because a single QoS parameter could not reflect an application’s transmission
requirements. For instant, delay, jitter, and packet loss ratio could all have significant
effects on VoIP quality (Al-Sbou, 2010).

Regression model is a widely employed statistical method in networking domain.

Baldwin (1999) developed regression models from simulated data to predict network
behaviour in terms of throughput, mean delay, missed deadline ratio, and collision ratio.
An adaptive regression algorithm was proposed to monitor two arbitrary sensor nodes
and dynamically learn the linear relation among their measurements. The algorithm then
eliminated the redundant node, and estimated the deficient data without the need for
base station assistance (Olios and Vida, 2009). Regression model was used to predict
the collision ratio, collision rate variation, and queue status ratio in participant wireless
nodes in a mobile ad-hoc network and to subsequently adjust the Contention Window
(CW), Distributed Inter-Frame Space (DIFS) and transmission rate in order to improve
the network performance (Saraireh, 2006).

Neural networks were also used in the area of network QoS. For instance, in (Nogueira
et al, 2006), a modelling approach based on neural network was proposed to predict the
media access delay in a Wireless Local Area Network (WLAN). MLP was used to
predict packet loss in a real-time video transmission. The results showed that MLP
could predict packet loss rate with accuracy of 96% (Lavington et al, 1999). Random
Neural Networks (RNNs) were devised to automatically quantify the quality o f video
flows. The achieved results correlated well with human perception methods (Mohamed,
2002). In (Radhakrishnan and Larijani, 2011), three non-intrusive RNNs models (simple
feed-forward model, Multilayer feed forward model, and recurrent architecture) were
employed to measure and monitor voice quality transmission. The results indicated that
the feed forward architecture produced the best results as compared with the other two
architectures.

50
Chapter 3 Literature Review

3.4 Im proving Quality o f Service in Com puter

Networks

QoS improvement in this study involves the ability to take actions to improve QoS or
change network performance toward desired situation. Improving QoS has become
more pronounced in particular with the presence of multimedia applications and
integration of wired and wireless networks.

In many instances, users in WLAN exchange and access multimedia applications with
users in wired networks. Figure 3-2 shows how voice calls can take place between a
user connected to wireless network and user connected to wired network. Considering
that multimedia applications which are sensitive to transmission parameters, it is vital to
improve and sustain QoS in the integrated networks in order to enable the effective
delivery of multimedia services.

Voice
station

Figure 3-2. Integrated WLAN-wired network.

The integration of wired and wireless networks poses challenges to overcome QoS
issues in both network sides: wired and wireless. This is because the QoS issues are
different between wired and wireless networks. The QoS support in WLAN is enabled
at MAC layer whereas the most wired networks enable QoS at the IP layer (Senkindu
and Chan, 2008). Therefore, providing and improving network QoS for traffic being
exchanged between two different environments is considered to be a challenging task.

In terms of wireless networks, the demand for supporting QoS for various applications
has led to the development of a WLAN standard so called IEEE 802.1 le. This standard
provides two schemes to access the wireless channel: Enhanced Distributed Channel
Access (EDCA) and Hybrid Coordination Function (HCF) Controlled Channel Access

51
Chapter 3 Literature Review

(HCCA) (IEEE Computer Society, 2005). An explanation o f these schemes can be

found in Chapter 2. The simplicity of EDCA operation as compared with HCCA allows
it to be implemented widely. Therefore, this study focuses on the IEEE 802.1 le EDCA.

In contrast with the earlier legacy IEEE 802.11 DCF access method, IEEE 802.l i e
EDCA differentiates traffic according to their QoS requirements as reported in (Liang et
al, 2006), (Rauf et al, 2009), (Alahmadi et al, 2008), and (Peng et al, 2010). For
instance, in (Liang et al, 2006), stations operating the IEEE 802.1 le EDCA were
compared with IEEE 802.1 lb for transmitting a number of applications. Results showed
that IEEE 802.1 le EDCA mechanism provided a better QoS as compared with IEEE
802.11b DCF. An evaluation reported in (Rauf et al, 2009) showed that IEEE 802.l i e
EDCA introduced an effective service differentiation mechanism and provided QoS
support under light network load. Peng et al (2010) presented an analytical model to
study the performance of EDCA service differentiation schemes. The general
effectiveness of EDCA service differentiation was proved through analytical and
simulation results.

However, a number of studies have demonstrated the limitations of IEEE 802.1 le

EDCA for efficiently handling a variety of traffic types in congested networks. The
main reason was the static nature of resource allocation inherent in IEEE 802.1 le. For
instance, the inadequate QoS support for multimedia traffic in high load conditions
using IEEE 802.1 le EDCA wireless infrastructure was demonstrated in (Politis et al,
2011). The proposed mechanism so called X-EDCA was designed to cope with high
load traffic situations in IEEE 802.1 le. The scheme improved the QoS for multimedia
traffic in infrastructure IEEE 802.l i e networks. The transmission performance o f IEEE
802.1 le EDCA was evaluated and reported in (Villalon, et al, 2007). The study showed
that EDCA was unable to guarantee a good performance when the network traffic load
was high. The main reason was the excessive number of packet collisions, which in turn
was due to the fixed transmission parameters values assigned to the ACs. Lin et al
(2009) proposed an adaptive cross layer mapping algorithm was devised to improve the
quality of MPEG-4 video transmitted over IEEE 802.l i e EDCA. The proposed
algorithm outperformed the 802.l i e protocol by dynamically mapping MPEG-4 video
packets to appropriate access categories, according to the network’s traffic load and the
significance of video frames.

52
Chapter 3 Literature Review

Most previous studies either supported QoS provided by IEEE 802.11 EDCA for high
priority traffic, which may starve other transmitted traffic as in (Politis et al, 2011), or
required modification of all wireless stations, which in turn complicates the WLAN
operation as in (Lin et al, 2009). The contribution of this study is the development of an
adaptive allocation algorithm at the wireless Access Point (AP) to further improve the
QoS of IEEE 802.1 le EDCA. The algorithm determines the Packet Arrival Rate (PAR)
of the uplink and downlink traffic for each Access Category (AC). The algorithm then
dynamically allocates the traffic from a lower priority AC to the next higher AC, when
there is no traffic for the higher AC. The algorithm enables lower priority traffic to have
access to a higher priority AC ensuring more efficient use of network resources.

Most current wired networks which are based on IEEE 802.3 have low bit error rates in
transmitting multimedia applications as compared with wireless networks. However,
one of the most challenging tasks that affects the QoS in wired is traffic congestion.
Congestion occurs when multiple stations transmit traffic simultaneously using the
same router. Congestion has a negative impact on the network performance as increases
the probability of packet loss ratio and collision.

Several mechanisms and approaches were used to overcome the QoS issues and
enhance network performance (Farrel, 2008). One of these approaches is over­
provisioning which refers to enhancing the network capability by simply providing the
network with enough bandwidth in order for all traffic to meet their QoS requirements
(Fraleigh et al, 2003). Other approaches such as traffic class, resource reservation, and
queuing mechanism utilise available network resources according to the application’s
QoS (Farrel, 2008).

Over provisioning might be preferred in core networks, such as Internet backbones

(Fraleigh et al, 2003) (Menth et al, 2006). However, in the context of integrated wired
and wireless, it is preferred to guarantee QoS using service differentiation mechanisms
rather than over provisioning because the latter can be difficult and costly. For instance,
telecommunication providers prefer Admission Control (AC) to guarantee QoS in
packet-switched communication rather than Capacity Over provisioning (CO) which
can be complicated and expensive task (Menth et al, 2006). DAntonio (2003) pointed
out that committing resources is not sufficient since QoS degradation is often
unavoidable due to any fault in the behaviour of network elements or lack of using
sufficient resources.

53
Chapter 3 Literature Review

There are several mechanisms to support QoS in wired networks as demonstrated in

Section 2.3.2 in Chapter 2. In this study, the network traffic prioritisation will be
investigated as one of the most important issues that affect the QoS in wired networks.
Therefore, packet scheduling mechanisms which are the most commonly employed
congestion-management tools will be considered.

Several studies have used packet scheduling mechanisms to improve network QoS. For
instance, Frantti and Jutila (2009) proposed Adaptive Weighted Fair Queuing (AWFQ)
to differentiate service for traffic according to QoS requirements. The study showed that
AWFQ achieved improved results than traditional WFQ. Epiphaniou et al (2010)
discussed the performance of three different mechanisms: FIFO, RED and DiffServ, and
their effects on real-time voice traffic. The experimentation results proved that under
burst traffic conditions up to a congestion level, DiffServ seems to perform better on all
three categories examined, by mainly providing a better queue management and
throughput, reduced packet drop rates based on data transmitted, and One Way Delay
(OWD) within the acceptable levels for an acceptable voice call. Miaji and Hassan
(2010) investigated the performance of three scheduling mechanisms: First Come-First-
Serve, Priority Queuing, and Weighted Fair Queuing. According the their simulation
results, WFQ provided a better enhancement for multimedia applications and hence a
higher QoS. Balogh and Medvecky (2011) carried out a comparison between Weighted
Fair Queuing (WFQ), Worst-case Fair Weighted Fair Queuing+ (WF2Q+) and
Weighted Round Robin (WRR) from the point of their usage in modem converged
telecommunication networks. Simulation results showed that WFQ and WF2Q+ provide
fair distribution of bandwidth for variable length packets due to the calculation of
packet size at cost of high computing and memory usage requirements which limited
their usages in high speed backbone network. On the other hand, WRR was a quick
algorithm with low computing requirements which allows its usage in high speed
backbone networks though WRR does not take into the account the length o f packet
size.

From the above discussion, the trade-off between different queuing mechanisms is their
complexity, control ability and level of fairness. An aspect of this study is an
appropriate utilisation of a suitable queuing scheduling mechanism. FIFO does not
support QoS because it treats traffic equally. PQ provides premium service to the high
priority traffic at the expense of the lower priority traffic, causing the latter traffic to
experience excessive delay. FQ is not designed to support traffic with different QoS
54
Chapter 3 Literature Review

requirements, as it allocates the same amount of bandwidth among multiple traffic. The
computational complexity of WFQ algorithm affects its scalability to support larger
traffic with different requirements at the edge of the network (Semeria, 2001),
(Epiphaniou et al, 2010), and (Balogh and Medvecky, 2011).

WRR addresses the limitations of FIFO, PQ, and FQ by classifying traffic based on
their QoS requirements, and ensuring that low priority traffic can access to buffer space
and output port bandwidth. The implementation of WRR is more popular and its
operation is less complex as compared with WFQ. Therefore, in this study, WRR was
considered to provide traffic prioritisation because of its practicality and low
complexity.

3.5 Quality of Service Monitoring Tools

With the provision of transmitting multimedia applications over computer networks, it
is crucial to monitor network performance to ensure that the QoS of these applications
is being sustained. Therefore, based on the need of monitoring QoS, many tools have
been proposed to monitor network performance. For example, the QoS monitoring tool
proposed by Graham et al (1998) was used to assess packet delay, jitter, and packet loss
as an indication of network performance. The Surveyor tool proposed by Zseby and
Schemer (2004) was used to assess end-to-end delay and packet loss to measure
network performance. The Surveyor tool included three main components: the
measurement machine, the database, and the analysis server. The network QoS matrices
were collected periodically by the measurement machine which in turn buffered them to
the database. Finally, the network QoS matrices were analysed by the analysis server. In
(Malhotra et al, 2011), a tool for QoS monitoring of multimedia sessions so called
StreamTrack was proposed. The tool was deployed in an IP infrastructure with
heterogeneous network access technologies. StreamTrack was used to calculate the QoS
parameters (i.e. bandwidth, delay, and jitter) of multimedia sessions such as
Voice/Video Call, and Internet Protocol Television (IPTV) sessions. Another QoS
monitoring tool that has been proposed and implemented for multiservice networks was
QMon (Carvalho et al, 2009). This tool was used to measure the QoS of distinct service
classes between network measurement points. The measurement results can be accessed
directly from the monitoring database or from a web interface available on the QServer.
QMon, as a multiplatform and generic tool, is a versatile and cost-effective QoS

55
Chapter 3 Literature Review

monitoring solution to be deployed in multiservice network environments, being useful

to assist traffic engineering tasks, service management and auditing.

However, the existing QoS monitoring tools have some limitations. For instance, they
are not determine directly the overall network QoS as in (Graham et al, 1998) and
(Malhotra et al, 2011). Network managers have to do a variety of operations to assess
the overall network QoS. Also, these tools were not developed to work as stand-alone
device as in (Zseby and Scheiner, 2004) and (Carvalho et al, 2009). From these
limitations, it can be concluded that the process of monitoring QoS can be complicated,
expensive, and time consuming. Therefore, developing a portable hand-held device that
accurately determines the overall network QoS for multimedia applications can be very
valuable. In this study, a mechanism that assesses QoS which in turn taking into the
account the QoS requirements of multimedia applications is implemented on a portable
microprocessor board to build QoS monitoring tool. The proposed tool could work on
its own to assess the QoS of multimedia applications based on their QoS requirements.

3.6 Application of Statistical and Artificial Intelligent

Techniques to Computer Network
Statistical and artificial intelligent techniques have been widely used in the area of
computer network for various tasks such as analysis, optimisation, and evaluation. A
number of studies have applied regression model as one of the most popular statistical
techniques in networking domain. For example, Akinaga et al (2005) proposed a
method based on regression analysis for forecasting network traffic using the user’s
properties and information about mobile network environment. The technique was used
to predict traffic fluctuations for a mobile network area. An enhancement of the Peak
Signal-to-Noise Ratio PSNR method to evaluate video streaming quality was introduced
by (Chan et al, 2010). The modified PSNR was based on linear regression technique
which in turn was used to derive two specific objective video quality metrics, PSNR-
based Objective MOS (POMOS) and Rates-based Objective MOS (ROMOS). Linear
regression prediction model was proposed to evaluate network security situation (Xia
and Wang, 2010).

Artificial intelligence techniques on the other hand were used widely in the area of
computer networks. For instance, Khoukhi and Cherkaoui (2008) proposed a fuzzy
logic system called FuzzyMARS for call admission control and service differentiation

56
Chapter 3 Literature Review

for wireless ad hoc networks. Their results showed that FuzzyMARS achieved service
differentiation delivery and reduced delay transmission of multimedia applications.
Frantti and Jutila (2009) embedded fuzzy expert system for Adaptive Weighted Fair
Queuing (AWFQ) to differentiate service for traffic according to QoS requirements. The
simulation results showed that AWFQ reacted faster to differentiate traffic classes than
traditional WFQ. Fuzzy logic was used to assess QoS and adjust the minimum
contention window ( CWmin) in mobile ad hoc networks (MANET) (Seriareh et al,
2008) and (Al-Sbou, 2010). The devised approach effectively assessed and improved
the QoS of multimedia applications in MANET.

Several studies have used FCM clustering in network traffic domain. For instance, a
network administrator assistance system was proposed based on FCM (Chen et al,
2009). The proposed system utilised a FCM method to analyse users’ network
behaviours and traffic-load patterns based on the measured traffic data o f an IP network.
Analysis results can be used to assist administrators to determine efficient controlling
and configuring parameters of the network management. In wireless sensor networks, a
FCM algorithm was used in order to create clusters which reduced the spatial distance
between sensors nodes (Hoang et al, 2010). A FCM clustering algorithm was also
developed to detect routing attacks caused by abnormal flows in a wireless sensor
network. The study demonstrated that FCM can be a valuable tool for intrusion
detection (Wang et al, 2009).

Neural network has been also used in the area of computer networks. In (Nogueira et al,
2006), a modelling approach based on a neural network was proposed to predict the
media access delay in WLAN. The prediction of the model was accurate even when the
number of active nodes was changed significantly. Multilayer perceptron MLP was
used to predict packet loss of real-time video transmission. The results showed that
MLP was capable of predicting the number of lost packets with 96% accuracy
(Lavington et al, 1999). Self-organising map was used to cluster network traffic types
and produce application profiles, which contained significant information about the
current status of the network, in order to manage network resources (Timo et al, 2002)
and (Kiziloren, and Germen, 2007). A network traffic classification approach based on
Kemel-SOM (KSOM) was proposed in (Ting et al, 2010). The experimental results
showed that the Kohonen SOM achieved high classification accuracy and successfully
categorised network traffic characteristics.

57
Chapter 3 Literature Review

The contribution of this research is to develop approaches using the aforementioned

statistical and A l techniques to intelligently manage QoS for integrated wired and
wireless network carrying multimedia applications. Fuzzy Inference System will be
used to sample multimedia traffic in adaptive manner, Fuzzy C-Means clustering
techniques and Self Organizing Map will be devised to analyse the QoS parameters in
order to provide valuable information about the network’s QoS patterns, and regression
model and Multi-Layer Perceptron will be employed to combine the analysed QoS
parameters and then produce a single value that represented the overall network’s QoS.

3.7 Summary
The main objective in this chapter was to provide an extensive literature review of
previous studies in the area of managing QoS of multimedia computer networks. The
main QoS management issues of multimedia networks that need further development
and investigations were discussed. Firstly, the state of the art of sampling techniques for
multimedia traffic was reviewed. Then, the current QoS analysis and assessment
techniques used to evaluate network QoS were reviewed. The relevant studies that
considered the QoS support in wireless and wired networks were discussed. The exiting
monitoring tools used to assess the network performance were critically analysed. The
applications of statistical and artificial intelligent techniques used within this study into
the field of computer network management were reviewed.

The next chapter focuses on a description of the experimental approaches that were
carried out to evaluate and validate the proposed techniques involved in the process of
managing QoS.

58
 Chapter 4 Experimental Methodology

4.1 Introduction
An explanation of network evaluation methodologies, tools, and general experimental
procedure used throughout this study is provided in this chapter. A more detailed
discussion of the procedures which are specific to individual studies are discussed later
in the relevant chapters. Section 4.2 of this chapter covers an overview of network
evaluation approaches, network simulation tools, description of simulation environment
and protocols, and traffic type and characteristics. The measurement processes which
include a description of QoS metrics and requirements, and analysis procedure of
simulation output are included in section 4.3. The main issues are summarised in section
4.4.

4.2 Network Evaluation Approaches

The aim of this section is to reveal the appropriate network evaluation approach that
suited the objectives of this study. Three main approaches were used to evaluate
network performance: analytical modelling, measuring physical networks, and
simulation (Jain, 1991). It was not practically feasible to use the analytical modelling
techniques in this study since computer networks, in particular for wireless networks,
have a dynamic behaviour. For instance, new traffic comes online, while others
terminate, the route that packets take can vary, and the bandwidth availability can vary
considerably over time. Evaluating network performance based on measuring physical
networks was also excluded because it is time consuming as well as the most expensive
approach among other approaches. Also, implementation of the devised methods in
realistic size and complexity networks could not be implemented in the duration o f this
study. Therefore, an appropriate approach that fitted to this research was based on
simulation. Simulation could provide a rich environment for experimentation at low
cost. In contrast with analytical modelling techniques, simulation techniques may
achieve more accurate results since they are often closer to the reality. Comparing with
measuring real networks, simulations could have more control over the network
conditions and allow changes to the network settings in more effective manner (Bajaj et
al, 1999).

59
Chapter 4 E x p e rim e n ta l M e tn o ao io g y

4.2.1 Network Simulation

Network simulation approach can be used to serve a variety of network engineering

needs. It allows engineers to simulate networks with realistic topologies in an effective
and inexpensive manner (Siraj et al, 2012). There are many simulation tools used in
network engineering research community (Jain, 2008), (Weingartner et al, 2009),
(Sarkar and Halim, 2011) and (Siraj et al, 2012). The most popular simulation tools are
Optimized Network Evaluation Tool OPNET (OPNET, 2012), Global Mobile
Information System Simulator (GloMoSim) (GloMoSim, 2012), Optical Micro-
Networks Plus Plus (OMNET++) (OMNET++, 2012), and NS-2 from Virtual
Internetwork Testbed project (VINT) (NS, 2012) and (Bajaj et al, 1999). The full
version of OPNET has a complete set of features with a well-developed Graphical User
Interface (GUI), but it is not open source, restricting the scope for customising its
operation. Although GloMoSim and OMNET++ are open source simulation tools, they
only support wireless networks (Jain, 2008) and (Sarkar and Halim, 2011). NS-2 on the
other hand is an open source and freeware simulation tool in nature. Hence, network
engineers tend to use NS-2 in order to test new protocols or modifying the existing ones
in a controlled, reliable and reproducible environment (Lucio et al, 2003). Moreover,
NS-2 can carry out trace-driven simulation using a record of events from a real system
(Caro, 2003). Therefore, this study will be based on NS-2.

NS-2 is a discrete event simulator based on two object oriented languages: C++ with an
Object Tool Command Language (OTcl) as shown in Figure 4-1. NS-2 has a rich set of
protocols such as TCP and UPD and traffic source behaviour such as FTP, Telnet, Web,
CBR and VBR (Chung and Claypool, 2004). NS-2 is also capable of transmitting video
streaming such as MPEG-4 and H.264 using Evalvid framework tool-set (Ke et al,
2008). Full details about the general architecture of the network components in NS-2
can be found in the documentation supported by NS-2 group (Fall and Varadhan, 2011).
In the simulation process of NS-2 as shown in Figure 4-1, users generate Tool
Command Language (TCL) script files to specify network topology, traffic applications,
and all the required settings. The TCL files are then handled by the C++ libraries and
OTcl interpreter. After the termination of simulation process, NS-2 produces simulation
results in two output files: NAM file and trace file. The NAM file is used to support
graphical tool called Network Animator (NAM) to visualize simulation traces. The trace
file contains information about packets (i.e. transmitted packets, received packets,

60
Chapter 4 experimental ivietnoaoiogy

dropped packet, packet types, packet ID, etc) (NS, 2012). NS-2 Users can extract
relevant information from the output trace file using script languages such as Awk, Perl,
or Grep and plot them using other script languages such as Xgraph, Gunplot, or Matlab
(Kumar, 2008).

Summary
Trace
files
OTcl interpreter
TCL Script
Simulation
File t NAM
C++ library
File

Simulation
r NAM Network
NS-2
Results Animator
Figure 4-1. The simulation process of NS-2.

4.2.1.1 An overview of Evalvid Framework

Evalvid framework tool set can be integrated within NS-2 simulator to facilitate the
transmission of real video applications over simulated networks. The formats of the
video applications supported by Evalvid framework which can be later transmitted
using NS-2 are YUV QCIF (176 x 144) or YUV CIF (352 x 288) (Zhou and Sik-Jang,
2008). Figure 4-2 shows the component of Evalvid framework and NS-2 simulator used
to transmit real video applications. The process of transmitting a video application using
Evalvid framework and NS-2 simulator is as follows (Ke et al, 2008) and (Abdel-Hady
and Ward 2007):

i. Video Encoder: the encoder is used to convert the source video file from YUV
format to a compressed H.264 or MPEG4 format at the sender side.
ii. Video Trace Generator: this component reads the compressed H.264 or MPEG4
file generated by video encoder and then fragments each video frame into
smaller segments of 1000 bytes for transmission, if the size of the video frame is
larger than the preset maximum packet size (Maximum packet length is 1028
bytes, including 20 bytes for IP header and 8bytes for UDP header). Video trace
generator produces a video trace file that contains information about every frame
in the real video file.
iii. During the simulation process of NS-2, there are three agents implemented
between Evalvid framework and NS-2. These are MyTrafficTrace, MyUDP, and
61
Chapter 4 experimental Metnoaoiogy

MyUDPSink agent. MyTrafficTrace agent reads the frame type and the frame
size from the video trace file, and sends these frame segments to the transport
layer according to the preset time settings specified in the simulation script file.
At the transport layer, MyUDP which is an extension of UDP agent allows users
to specify the output file name of the sender trace file and then records the
timestamp of each transmitted packet, the packet ID, and the packet size. At the
transport layer of the receiver side, MyUDPSink agent receives the fragmented
video frame packets sent by MyUDP and then records the timestamp for each
received packet, packet ID, and packet size in the receiver trace file specified by
user.
iv. Video Reconstruction: after simulation is terminated, the three trace files (i.e.
video trace file, sender trace file, and receiver trace file) are reconstructed by the
video reconstruction component to produce the received video file in a
compressed H.264 or MPEG4 format.

YUV M PEG-4

f Video Video Trace Video trace file

Source
Encoder Generator
video

NS-2 Environment

MyTrafficTrac Video trace file

Sender trace

Simulated
^^^ender^^" Network - ^ ^Receiv^^

Receiver trace file MyUDPSink

Video Video 1. Video trace file

Receive
d video
K
YUV
Decoder

MPEG-4
Reconstruction 2. Sender trace file
3. Received trace file

Figure 4-2. The process of transmitting video using Evalvid framework and NS-2.

62
^napier * e x p e r im e n ta l ivietnoaoiogy

4.2.2 Network Topologies

Wireless-cum-wired network topologies with different sizes (i.e. small, medium, and
large network) were simulated using the Network Simulator-2 (NS2) as shown in Figure
4-3. As the number of stations plays an important role in network performance, the
number of stations in the simulated network was varied from 8 to 64 stations according
to selected scenario. The connections between stations in wireless-cum-wired network
topology were unidirectional. The number of these connections was varied from 4 to 32
according to the number of transmitted stations. In most simulation scenarios, half of
the connections transmitted traffic from wireless to wired network, whereas the other
half transmitted traffic from wired to wireless. Each station transmitted one type of
traffic to its corresponding destination. At the wired side of network, all links had 5
Mbps bandwidth and 2 ms propagation delay. The WLAN side o f the network was
based on IEEE 802.1 le, and it used the Enhanced Distributed Channel Access (EDCA)
scheme. The main parameters that modelled the wireless channel were the default
settings for IEEE 802.1 le as shown in Table 4-1.

IEEE 802.1 le EDCA Network Wired Network

Figure 4-3. The simulated network topology.

The network topology covered an area of 500m X 500m and the stations were
positioned randomly within the specified area. This position was fixed during the
simulation time. Simulations were repeated 10 times. Each time a different initial seed
value was used to randomly position the stations and manage which node transmitted
first, as all nodes were requested to transmit at a given time. The randomness introduced
using different seeds avoided the bias of random number generation. The results o f the
10 simulations were then averaged. Simulation time was between 300 - 500 seconds.
These settings were considered appropriate to examine the long term behaviour o f the
IEEE 802.l i e protocol.

63
Chapter 4 Jkxpem iieiiiai iviem uuuiugj

4.2.3 Physical Layer (PHY) Parameters

At the wireless side of the network, the main physical layer parameter considered was
channel bit rate which includes basic rate for control frame transmission and data rate
for data transmission. Basic rate was set at 1 Mbps, while data rate was set to be 2 Mbps
or 11 Mbps for some selected simulation scenarios. The physical layer was modelled to
work as Lucent WaveLAN at a frequency of 914 MHz and DSSS radio interface card
(Fall and Varadhan, 2011). A summary of PHY layer parameters of this model is listed
in Table 4-1 (NS, 2012).

4.2.4 Medium Access Layer (MAC) Parameters

In this research study, the MAC layer was based on IEEE 802.1 le at the wireless side of
the network. This standard provides two schemes to access the wireless channel. These
are: Enhanced Distributed Channel Access (EDCA) and Hybrid Controlled Channel
Access (HCCA) (IEEE, 2005). This study focused on the IEEE 802.1 le EDCA due to
its simplicity as compared with HCCA. The main parameters that modelled EDCA were
the default settings for IEEE 802.1 le. A summary of MAC parameters of this model is
listed in Table 4-1 (NS, 2012).

Table 4-1. Simulation settings of MAC and PHY parameters in IEEE 802.1 le.

Parameter Value
Capture Threshold 10
Carrier Sense Threshold 1.559e-ll
Receiving Threshold 3.652e-ll
Power Transmission 0.28183815
Frequency Band 914e+6 i
Data Rate 2 .0 -1 1 .0 Mbps
Basic Rate 1.0 Mbps
I Modulation Technique DSSS
PHY Header 24 bytes
MAC Header 28 bytes I
SlotTime 20psecs
SIFS 1Op secs
Preamble Length 144 bits
PLCP Header Length 48 bits
RST Threshold 3000
ShortRetryLimit 7
LongRetryLimit 4

64
l^napier 4 niA peniiiem ai ivicuiuuuiugj

In this study, the transmitted traffic over IEEE 802.l i e EDCA were VoIP, video, best
effort traffic, and background traffic. These traffics were mapped into different Access
Categories (ACs) to represent different levels of priority as shown in Table 4-2 (IEEE,
2005).

Table 4-2. IEEE 802.1 le access categories parameters.

^ ^ T y p e of traffic
Best effort Background
VoIP Video
traffic traffic
Parameters
AIFS 2 2 3 7
c w min 7 15 31 31
c w max 15 31 1023 1023
TXOP 3.008 6.016 0 0

Classifying the traffic into different ACs was based on their QoS requirements. Due to
the high sensitivity of VoIP to QoS parameters, it was assigned to AC with the smallest
values of AIFS, CWmin, CWmax and largest value of TXOP. In contrast, background
traffic was assigned to AC with the largest values of AIFS, CWmin, CWmax and
smallest value of TXOP because of its tolerance to some QoS parameters such as delay.
Accordingly, VoIP had the highest priority, whereas the background traffic had the
lowest priority.

4.2.5 Routing Protocols

Several routing protocols are available in NS-2 environment. These include,

Destination-Sequenced Distance Vector (DSDV), Ad Hoc On-Demand Distance Vector
Routing (AODV), Temporally Ordered Routing Algorithm (TORA), and Dynamic
Source Routing (DSR) (Gupta and Saket, 2011) and (Said and Saatchi, 2009). DSDV
protocol is based on the Bellman-Ford routing algorithm. It uses the proactive table-
driven routing strategy. Whereas AODV, DSR, and TORA are reactive on-demand
routing protocols which initiate route discovery mechanism to establish a route between
the source and destination nodes (Shah et al, 2008). In this study, DSDV was chosen as
it maintains the routing information for all the nodes in the network and adds a new
route or update the existing routes periodically. This ensures the routes to any
destination are ready to use when needed (Elashheb, 2012).

65
s im p le r 4 l-/ApVl

4.2.6 Queuing Mechanisms

In the network, when multiple packets are serviced through a congestion point such as a
router, queuing mechanisms are required to determine the bandwidth allocation among
transmitted packets and the manner in which to service various applications with
different QoS requirements. In this study, two queuing scheduling mechanisms were
employed. These were First-In-First-Out (FIFO), and Weighted Round Robin (WRR).
Although no preference was given to the transmitted traffic regarding to its QoS
requirement in FIFO queuing scheme as its basis was first packet come first packet
served, FIFO was employed in most simulation scenarios due to its management
simplicity and implementation popularity. The default queue size of FIFO used in this
study was 50 packets. In some experimental scenarios in Chapter 7, W RR queue
scheduling mechanism was implemented between the router and the Access Point (AP)
at the wired side of the network to improve its QoS. WRR addresses the limitations of
FIFO by classifying traffic based on their QoS requirements, and ensuring that different
traffic priorities can access the buffer space and output port bandwidth (Semeria, 2001).
The operation of WRR is explained under Section 2.3.2 in Chapter 2. In this study,
time-sensitive applications had a higher priority than time-insensitive applications. This
was because the former had larger weights than the latter as shown in Table 4-3.

Table 4-3. WRR Parameters.

Preset No. of
Queue 1 Queue 2 Queue 3 Queue 4
queues in WRR
Video Best effort Background
Application type VoIP
streaming traffic traffic
WRR weights 3 3 2 2
Queue length 25 25 25 25

4.2.7 Traffic Type and Traffic Characteristics

In this study, different types of traffic were transmitted over the simulated networks.
These were: Voice over IP (VoIP), videoconferencing, video streaming, best effort
traffic, and background traffic. Constant Bit Rate (CBR) traffic was adapted to model
VoIP, videoconferencing, and best effort traffic. The VoIP packet size was 160 bytes
and its inter-packet interval was 20 ms, corresponding to G .711 voice encoding scheme
with 64 kbps Pulse Code Modulation (PCM) voice flows. The packet size of the video

66
Chapter 4 C A J J C l IIU C IIK U lT lC U I U U U lU g J

traffic was 512 bytes and the inter-packet interval was 10 ms. This generated a packet
transmission rate of 384 kbps (Markopoulou et al, 2003) and (Saraireh et al, 2007).

The video streaming sources were YUV QCIF (176 x 144) Foreman (400 frame) and
YUV QCIF (176 X 144) Highway (2000 frame) (YUV QCIF, 2012). Prior to the
transmission, each video frame was fragmented into packets, which in turn had a
maximum length of 1024 bytes. Video streaming applications were encoded using
MPEG-4 encoding scheme which defines three types of video frames: I (Intra-coded)
frame, P (Predictive-coded) frame, and B (Bidirectionally predictive-coded) frame as
shown in Table 4-4.

Table 4-4. The number of video frames and packets of the video streaming sources.

Number of frames Number of packets

Video format Total Total
I P B I P B
Foreman QCIF 45 89 266 400 237 149 273 659
Highway QCIF 223 445 1332 2000 461 451 1333 2245

The I frame is encoded and decoded independent of previous or successive frames. The
encoding of P frame requires information from preceding I or P frame in the video
sequence. The predictions from previous and successive I or P frames are also required
to encode the B frame. According to MPEG-4 scheme, the I frame is the most important
frame among other types of frame. Comparing P and B frames, MPEG-4 scheme
specifies that the former frame is more important than the latter (Lin et al, 2009).
During the decoding process, the video frames can be decompressed into Group Of
Pictures (GOP), which its pattern is defined by two parameters G (M, N), where N is the
I-to-I frame distance and M is the I-to-P frame distance as shown in Figure 4-4 (Zhou
and Sik-Jang, 2008).

Q
B B B B

Figure 4-4. GOP sequence in MPEG-4

B
0
The other types of traffic transmitted over the simulated networks were the best effort
traffic which modelled using CBR with different packet sizes and generation rates that
corresponded to non-videoconferencing or VoIP usage. The packet size of 200 bytes
with 12.5 ms inter-packet interval was used to generate 128 kbps data rate, while large

67
cnapier 4 UjAJIEI

packet size of 1000 bytes with 12.5 ms packet interval was to generate 500 kbps data
rate. File Transfer Protocol (FTP) application was also used as background traffic. FTP
was transmitted over TCP, whereas other traffics were transmitted using UDP transport
protocol.

4.3 Analysis o f Sim ulation O utput

Due to their importance in the process of QoS evaluation, the transmission requirements
of applications are considered during the analysis the simulation results. Throughout
this research, delay, jitter, and packet loss ratio are considered to be the main QoS
parameters. These parameters and the manner they were calculated are explained in the
following sections.

4.3.1 Transmission Requirements of Applications

According to the application type, the QoS requirements of time-sensitive applications

are significantly different from time-insensitive applications. The latter applications
such as email, web browsing, and file transfer can be tolerant with QoS parameters such
as delay and jitter. However, the former applications (i.e. multimedia applications) such
as VoIP and video applications are highly sensitive to QoS parameters, thus requiring a
faster response from the network. A large delay, jitter, or packet loss ratio can seriously
degrade their quality (Kurose and Ross, 2005). There are some factors that pose
challenges to prevent network providing sustained QoS for transmitted applications.
These include network congestion in wired networks and interference problems in
wireless networks. Therefore, the QoS requirements for traditional and multimedia
applications must be considered to provide QoS for these applications. Table 4-5
summaries the QoS requirements for time sensitive and insensitive applications as
recommended by ITU group (ITU-T, 2001) and (Zhai et al, 2005).

Table 4-5. QoS requirements for voice, video, and data as recommended by ITU group
(Zhai et al, 2005).

Class Application Delay Jitter Packet loss rate

< 150 ms < 1ms
VoIP <3 % preferred
Time- preferred preferred
sensitive < 150 ms < 1ms
Video < 1 % preferred
preferred preferred
Time- E-mail, file transfer,
Minutes N/A Zero
insensitive web browsing

68
i^ n a p ie r * XifApCl Ilildliai lUCUIVUUlVgJ

4.3.2 Calculation of QoS Parameters

This section explains the calculation method of QoS parameters (i.e. delay, jitter, and
packet loss ratio). After the TCL script file is simulated by NS-2, a detailed trace file is
generated which contains extensive information about transmitted traffic. This
information include: packet status (i.e. departed, arrived, and dropped), its timestamp,
packet ID, packet type, packet size, flow ID, sequence number, node ID, source and
destination addresses.

In this study, the QoS parameters were extracted from the data trace file using AWK
script language (Aho et al, 1988). Delay calculation process was associated with three
main fields: packet sent time, packet received time, and its unique ID. To calculate
delay, the sent time a packet was subtracted from the received time for the same packet.
Equation 4.1 illustrates how delay is calculated for particular packet:

D i= R i - Si (4.1)

Where Dt is the delay of the ith packet arrived, Rt and S* are the arrival and sending

timestamps of the ith packet.

Jitter was computed by calculating the absolute value of the difference between two
consecutive packets delays as shown in equation 4.2:

Ji = abs (fit —D j.i) (4.2)

Where Ji is the absolute jitter of the ith packet, Dt is the delay of packet i, and is
the delay of the previous packet.

The percentage of packet loss ratio during certain time interval was calculated based on
the total number of received packets with respect to the total number o f transmitted
packets during that time interval as in equation 4.3:

PLi(t) = 100 X ( l - 1 ^ | ) (4.3)

Where PL\ is the loss ratio in percentage (%) during the ith interval, and £ Ri (t) and
2 Si (t) are the total number of received and transmitted packets with the ith interval
respectively.

When the values of QoS parameters were calculated, they were fed to MATLAB for
further analysis (MATLAB, 2012). This analysis included averaging the values of QoS
69
v^napier h UAJJCl 'BJ

parameter for every n consecutive packets. The averaged values were then normalised
and limited in order to ensure that all values had the same contribution in QoS
evaluation process.

4.4 Summary
This chapter described the experimental procedure used to evaluate and validate the
techniques proposed throughout this study. The use of NS-2 to simulate computer
network scenarios (wireless-cum-wired network) was discussed. The network settings
including routing protocols, queuing mechanisms, PHY, and MAC layer parameters
were also explained. The characteristics and the QoS requirements of applications
transmitted over the simulated network were clarified. The chapter concluded by
describing the calculation of QoS parameters using AWK script language and then
MATLAB for further qualitative analysis. Chapters 5, 6 , 7, and 8 rely on the outlined
experimental approach discussed in this chapter in order to test, validate, and evaluate
the proposed techniques to manage QoS of multimedia computer networks.

70
Chapter 5 Development and Evaluation of Adaptive
Statistical Sampling Techniques for Multimedia
Traffic

5.1 Introduction
The rapid growth of real-time applications transmitted over multimedia networks,
makes measurement of their traffic increasingly important. These measurements allow
Quality of Service (QoS) for the transmission of the applications to be assessed.
However, most real-time applications such as VoIP and video-conferencing generate an
extensive amount of traffic data. Analysing these data in real-time is computationally
intensive. Therefore, in order to reduce the amount of collected data and their
processing, appropriate sampling techniques are required.

In fixed rate sampling techniques, the number of data packets processed remains
unchanged even when traffic characteristics change. However, in adaptive sampling, the
number of packets sampled varies in accordance with traffic fluctuations. This ensures
appropriate amounts of data are processed.

In this chapter, statistical adaptive sampling techniques to adjust sampling rate based on
traffic's statistics were developed based on a linear adjustment approach, quarter
adjustment approach, and fuzzy inference system. A comparison of the devised methods
versus conventional sampling techniques (i.e. systematic sampling, stratified sampling,
and random sampling) was also carried out using a simulated computer network.

The organisation of this chapter is as follows: Section 5.2 reviews the state-of-the-art of
adaptive sampling approaches which are used to reduce network QoS parameters.
Section 5.3 includes a description of the proposed adaptive statistical sampling
approaches, implementation of conventional sampling techniques, calculation of QoS
parameters from sampled versions, methods of sampling analysis, and demonstrates the
simulation set up. The experimental results are discussed in section 5.4. The summary
of this chapter is provided in section 5.5.

71
i^napier d /\uitpuvc ouiusuuti omupimg icuiiuqu cs

5.2 Related Work

There were a number of studies conducted using adaptive sampling approaches to
gather network information. These studies were reviewed in chapter 3. A number of
these studies used fuzzy logic to sample traffic in an adaptive manner, such as
(Hernandez, et al., 2001), (Giertl, et al., 2006), and (Giertl, et al., 2008). A further
improvement of adaptive sampling, based on fuzzy logic control (FLC), was proposed
by (Xin, et al., 2009). Modified FLC method can realize dynamic adaptive sampling
making it more suitable for high speed networks.

Other adaptive sampling approaches were used for a variety of applications. These
applications were: a control of resources allocation for network performance reported
by (Graci'a et al, 2008), an estimate of traffic rate proposed by (Ma, et al., 2004), and
capturing Denial of Service (DoS) attack packet as in (Zhang, et al., 2007).

In our work, novel statistical adaptive sampling methods were developed based on
traffic's statistics (Dogman, et al., 2010a), (Dogman, et al., 2010b), and (Dogman, et al.,
2011). The methods adjusted the sampling interval by considering the traffic's statistics
between two consecutive sampled sections.

In this chapter, simple linear adjustment mechanism, quarter adjustment mechanism,

and fuzzy inference system were devised considering traffic's statistics to adaptively
adjust the sampling rate. A comparison of the devised methods versus conventional
sampling techniques (i.e. systematic, stratified, and random sampling) was also carried
out using a simulated computer network.

5.3 Adaptive Statistical Sampling Approaches

The applications of intelligent and non-intelligent methods for sampling multimedia
traffic in an adaptive manner are described in this section. Three adaptive sampling
approaches are proposed. The traffic length o f the sampling interval for the three
devised sampling techniques was controlled using three different mechanisms: simple
linear adjustment mechanism, quarter adjustment mechanism, and fuzzy inference
system. The approach that used fuzzy inference system was devised based on the theory
of fuzzy logic, discussed in Chapter 2 (Section 2.4.2).

72
Chapter 5 /\uapuve oiauMicai sampling icum^ucs

5.3.1 Description of Statistical Sampling Algorithm

The algorithm samples packets by considering the statistics of QoS parameters of

transmitted traffic. Analysing QoS parameters of the traffic is important in order to
ensure sampling process becomes tuned to the traffic characteristics. In this study, the
statistics of throughput were considered during the sampling process. Multimedia
applications such as videoconferencing and VoIP have high sensitivity to QoS
parameters, such as throughput (Alkahtani, et al., 2003). Moreover, the statistics of
throughput can be easily computed during the sampling process which gives an
advantage to the algorithm to respond quickly to traffic changes. In this study,
throughput was calculated using equation (2.1) (Wang, et al., 2000).

The devised statistical adaptive sampling algorithm had a number of operating

parameters as shown in Figure 5-1. These parameters were:

• Pre- and post-sampling sections of the traffic: these traffic sections contained the
packets to be sampled. The number of packets in these sections was not changed
during the sampling operation.

• Inter-Sampling Section Interval (ISSI) of the traffic: the position of this traffic section
was between the pre- and post-sampling sections. The length of this section was
adaptively determined during the sampling process. A simple linear adjustment
mechanism, quarter adjustment mechanism, and fuzzy inference system (FIS) were
developed to adjust the length of ISSI. Section 5.3.2 demonstrates how the ISSI was
adjusted using these mechanisms.

• Threshold value p : this value was used to assess the variation between the statistics of

the pre- and post- sampling sections. The threshold value influenced when the ISSI
was increased or decreased.

Pre-Sampling Inter-Sampling Section Post-Sampling

Section Interval (ISSI) Section

Figure 5-1. Operating parameters of adaptive statistical sampling algorithm.

A flow chart of the statistical sampling algorithm’s operation is provided in Figure 5-2.
The user is required to initialise the length of the pre- and post- sampling sections,
length of Inter-Sampling Section Interval (ISSI), and threshold value p . The lengths

pre- and post-sampling sections are the same and are not changed during the algorithm’s

73
Chapter 5 Adaptive statistical sampling lecumquca

operation. An increase in the length of pre- and post- sampling sections enlarges the
sample size. As the goal of the algorithm is to produce smallest sample size with the
highest accuracy, the user needs to select a suitable value for the length of the pre- and
post-sampling sections, through experimenting with a number of different values. The
initial length of ISSI was used as the current sample interval for the first iteration of the
sampling process. However, the initial length of ISSI was not critical during the first
iteration of sampling process because it will be adjusted in the following iterations.

The algorithm determined the statistics (i.e. mean, median and standard deviation) of
the QoS parameter (i.e. throughput) for the pre- and post- sampling sections and then
quantified their overall statistic using equation (5.1).

Overall Statistic = (abs (meanl-mean2)/meanl) + (abs (medianl-median2)/medianl) +

(abs (stdl-std2)/stdl) (5.1)

Where abs represents the absolute value, meanl and mean2 are the mean values o f pre-
and post- sampling sections respectively for the throughput of traffic being sampled,
medianl and median2 are the median values of pre- and post-sampling sections
respectively for the traffic throughput, and std l and std2 are the standard deviation
values of pre- and post-sampling sections respectively for the traffic throughput. The
overall statistic assessed the discrepancy of statistics between pre- and post-sampling
sections. The algorithm updates the new length of ISSI based on the comparison
between the quantified overall statistic and threshold value p using simple linear

adjustment mechanism, quarter adjustment mechanism, and fuzzy inference system. An

explanation of these mechanisms is in section 5.3.2.

The above operation represents one update of the ISSI length. During the next step, the
updated ISSI is the current sample interval, the current post-sampling section becomes
the next pre-sampling section, and the location of next post-sampling section is
determined based on the updated value of ISSI. This process is repeated until the traffic
is fully sampled.

74
V/iiapici j ounnutai odiupiuig lcuuiiqucs

CL Start algorithm

Initialise: Pre- and post-sampling sections length, the length of Inter-sample

section interval (ISSI), threshold value p

Calculate: mean, median, and standard deviation of pre- and post-sampling

sections

Quantify the overall statistic of the pre- and post-sampling sections using
Equation (5.1)

Current sample interval^ ISSI

Quantify the new length of ISSI based on the comparison between the
quantified overall statistic and threshold value p using different adjustment
mechanisms:
1. Linear adjustment mechanism.
2. Quarter adjustment mechanism.
3. Fuzzy inference system.

Current pre-sampling section = previous stage post-sampling section

Determine location of post-sampling section using the new length of ISSI

No

Is traffic fully
sa m n led ?

Yes

CL End of aleorithm

Figure 5-2. The flow chart of the adaptive statistical sampling algorithm.

5.3.2 Adjustment of Inter-Sampling Section Interval (ISSI)

Three adjustment mechanisms were proposed and incorporated into statistical sampling
algorithm to adaptively adjust the length of Inter-Sampling Section Interval (ISSI).

75
Chapter 5 Aaapuve aiausucai sampling leumiqucs

These mechanisms are simple linear adjustment mechanism, quarter adjustment

mechanism, and Fuzzy Inference System (FIS). These mechanisms are explained in the
following subsections.

5.3.2.1 Linear Adjustment Mechanism of ISSI

This scheme examines the overall statistic along with predefined threshold value p and

then linearly increases or decreases the length of ISSI. If the overall statistic value is
less than the threshold, then the length of ISSI is updated using equation (5.2).
Updated ISSI = current sample interval + //, (5.2)

If the overall statistics value is more than or equal to the threshold value, then the ISSI
is updated using equation (5.3).
Updated ISSI = current sample interval - p 2 (5.3)

The terms//, and p 2in equations (5.2) and (5.3) control the update magnitude. The initial

value for both p xand ju2 was 1. During the sampling operation, a further increase to the

value of //, and p 2 is applied which in turn increases or decreases the length of ISSI

linearly. However, in case of the overall statistic value is equal or more than threshold
value p , the value of p 2 was less than the length of current sample interval, at least by

one.

5.3.2.2 Quarter Adjustment Mechanism of ISSI

In this approach, the calculated overall statistic was compared with the user specified
threshold value p . If the overall statistic value is less than the threshold, then the Inter-

Sampling Section Interval (ISSI) was updated using equation (5.4)

Updated ISSI = current sample interval + round (ISSI/p) (5.4)
If the overall statistic value was more than or equal to the threshold value, then the ISSI
was updated by equation (5.5)
Updated ISSI = current sample interval - round (ISSI/p) (5.5)
The round function ensured that fractional values were rounded to the nearest integer
value. The term p in equations (5.4) and (5.5) updates the length of ISSI. The value of

p was determined though experimenting with different values. Small value o f p may

significantly change the length of ISSI, whereas as larger values o f p make finer

changes when ISSI length was updated. As the goal of the sampling algorithm is to
produce smallest sample size with the highest accuracy, the user needs to select a
76
 A u a p u v c o u i u a u u t i o a iiip iiu g ic u u i^ u ra

suitable value of f i , through experimenting with a number of different values. In this

study, the value of // was determined though experimenting with different values and a

value of 4 was chosen for this parameter. Therefore, the amount of ISSI change was a
quarter of its previous length.

5.3.2.3 ISSI Adjustment using Fuzzy Inference System

In this approach, the Mamdani type of Fuzzy Inference System (FIS) was used to adjust
the length of ISSI. Two inputs were fed into the FIS: the length of current sample
interval and the overall statistic, which was used to measure the discrepancy of statistics
between pre- and post-sampling sections. The overall statistic was calculated using
equation (5.1).

Each fuzzy input variable was represented by five fuzzy sets to create input membership
functions. The amount of overlap and the range of each variable were determined by
experimenting with a number of suitable values and selecting the ones that gave best
outcomes. The locations, the degree of overlap between the generated membership
functions, and their corresponding fuzzy linguistics variables are shown in Figure 5-3
(a).

The fuzzy input variables were used to produce a single fuzzy output called Sample
Interval Difference (SID) which was then used to determine whether ISSI should be
increased or decreased or remain unchanged. The fuzzy output variable was also
partitioned into five membership functions as shown in Figure 5-3 (b).

The fuzzy inputs and the output were fuzzified using the Gaussian membership
function. This membership function is smooth and has concise notation. The
mathematical formula of Gaussian membership function is expressed in equation (5.6):

\iAi(x)= exp ( ~ ^ ~ f ) (5.6)

where ct and cr* are the mean and standard deviation of the ith fuzzy set A1, respectively
(Saraireh et al., 2007).

Tables (5.1) - (5.3) show respectively the values of membership function parameters for
fuzzy inputs (i.e. the length of current sample interval and the overall statistic) and
fuzzy output (i.e. Sample Interval Difference SID).

77
C hapter 5 A daptive sta tistica l sa m p lin g 1 ecm uques

Very low Low High Veiyi Small Lagre Very large

0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 350
Corrent sample interval
Threshold values

Decrease Hi; High

■100 ■80 ■40 •20 0 20 40 60 80 100

Sample Interval Difference (SID)

(b)
Figure 5-3. Graphs of FIS for adjusting ISSI: (a) Fuzzy inputs, (b) Fuzzy output.

Table 5-1. Mean and standard Table 5-2. Mean and standard deviation
deviation of overall statistic input fuzzy of current sample interval input fuzzy
_______ membership functions.________ membership functions.________
Current sample
Membership Overall statistics Membership
interval
functions functions
Mean St. dev. Mean St. dev.
Very low 0 0.005 Very small 0 53.09
Low 0.012 0.005 Small 125 53.09
Medium 0.025 0.005 Medium 250 53.09
High 0.037 0.005 Large 375 53.09
Very high 0.05 0.005 Very large 500 53.09

Table 5-3. Mean and standard deviation of sample interval difference output fuzzy
membershi p functions.
Sample interval difference
Membership functions
Mean St. dev.
Decrease low (DL) -100 21.24
Decrease High (DH) -50 21.24
No change (NC) 0 21.24
Increase low (IL) 50 21.24
Increase high (IH) 100 21.24

78
Uftapter 5 g-vuajjiive cjuimuLai aam p m ig k u u u iju m

The relationship between the inputs and the output was defined by a set of fuzzy rules.
The number of fuzzy rules was set according to the number of inputs and their
associated fuzzy sets. Examples of the rules are provides in Table 5-4.

Table 5-4. The fuzzy rules used by FIS to adjust ISSI.

Overall Sample interval difference

Current sample interval
statistic (SID value)
Very small Very low j Increase high (IH)
Small Very low Increase high (IH)
Medium Very low Increase low (IL)
Large Very low Increase low (IL)
Very large Very low No change (NC)
Very small Low Increase high (IH)
Small Low Increase low (IL)
Medium Low Increase low (IL)
Large Low No change (NC)
Very large Low Decrease low (DL)
Very small Medium Increase low (IL)
Small Medium Increase low (IL)
Medium Medium No change (NC)
Large Medium Decrease low (DL)
Very large Medium Decrease low (DL)
Very small High No change (NC)
Small High Decrease low (DL)
Medium High Decrease low (DL)
Large H igh Decrease High (DH)
Very large High Decrease High (DH)
Very small Very high No change (NC)
Small Very high Decrease low (DL)
1 Medium Very high Decrease low (DL)
Large Very high Decrease High (DH)
Very large Very high Decrease High (DH)

Fuzzy reasoning (i.e. the process of implication and then aggregation) is based on
(minimum-maximum) inference method. Each rule is applied to the corresponding
membership function and the minimum is mapped into associated output membership
function. The output fuzzy set from the implication process for each rule is combined
together via aggregation process to produce one fuzzy set. In this study, the FIS output
was generated from aggregated fuzzy set (i.e. defuzzification) using the centriod
scheme. The centriod method returns the centre of area under the curve of the
aggregated output values using equation (Al-Sbou et al, 2006) (5.7).

r = I ly M J L n , ^

79
 Chapter 5 AUitJJUVC I3UIUSI1UII o a i l l J J l l u g ic v iu m ju v j

where m is the number of fuzzy sets obtained after implication, y t is the centriod of
fuzzy region i, and p* is the output membership value.
The adaptive sampling algorithm used the generated fuzzy output (i.e. Sample Interval
Difference SID) along with the current sample interval to update the length of ISSI.
Equation (5.8) is used to calculate the new length of ISSI.
Updated ISSI = round ((SID/current sample interval)* 100+current sample interval)
(5.8)

5.3.3 Implementations of Conventional Sampling Techniques

In this study, conventional sampling techniques (i.e. systematic, stratified, and random
sampling) were implemented using the count based approach where the packet selection
decision was based on the packet count. This approach was chosen due to its simplicity
(Zseby, 2002). The implementations of conventional sampling techniques are as
follows:
i. Systematic sampling: in this approach, for every n packets, the nth packet is
selected. In implementation of systematic sampling a counter is set initially to n and
it is decreasing it by 1 on receiving each packet. When the counter is zero, the
packet is selected. This operation represents one packet selection. During the next
step, the counter is reset and the process is repeated. For several experimental runs,
the starting point of packet selection is chosen randomly to get different sets of
samples for the same size.
ii. Random sampling: in this approach, for a sample size of n taken from a population
of N, n random numbers need to be generated for a range 1 to N, and then the
packet selection is performed according to the position of the packets in the flow.
For each experiment, a new n random numbers should be generated in order to get
different sets of samples of the same size.
iii. Stratified sampling: is similar to the implementation of random sampling. For
every strata which has a size of N packets, random numbers n are generated in the
range [1, IV], and the packets are selected according to their position. For every run,
a new n random numbers are generated for the same sample size.

5.3.4 Calculation of Sampling QoS Parameters and Sampling

Analysis
This section explains how the QoS parameters can be calculated from sampled versions
obtained using adaptive statistical sampling techniques and non-adaptive sampling
80
 z iu ap iiT c; j u i i i i s u v a i uani[iiiiA g iv v iu u « ju v a

techniques. It also demonstrates the methods used to compare the sampled versions
versus original populations.
i. Throughput calculation: throughput was calculated by multiplying the number of
received packets (N) by the packet size (PS) and then dividing it by the difference
between receiving time of two successive sampled packets.
ii. Delay calculation: end-to-end delay was calculated by the difference of arrival and
sending times of packets which were selected during the sampling process.
iii. Jitter calculation: the difference between the delays of two consecutive sampled
packets for specific flow was used in order to calculate the jitter.
iv. Packet loss calculation: the percentage of loss ratio during the ith time interval was
calculated based on the total number of received and transmitted packets during that
th
time interval. The number of received packets during the i time interval was
obtained from the sampled version, whereas the number of sent packets was
computed from the difference between the sequence numbers of two successive
sampled packets.

After calculating the QoS parameters (i.e. throughput, delay, jitter, and packet loss ratio)
of the sampled versions, obtained using adaptive statistical sampling and non-adaptive
sampling approaches, a comparison between the QoS parameters of sampled versions
and the QoS parameters of original population was carried out. The aim of the
comparison was to determine which sampling approach could be used to produce
sampled version that effectively represented the whole population. The comparison was
carried out by calculating the mean, and standard deviation of the original population
and its sampled versions using adaptive and non-adaptive sampling techniques.

In order to compare the sampled version versus original populations, the mean and
standard deviation of the sampled version may not be sufficient to assess the accuracy
of sampled version in terms of representing the original population as they are affected
by the outliers (Brase, 2010). Therefore, additional criteria were used to assess the
discrepancy between the original population and its sampled version. These were:

i. Bias: the bias shows how far the mean of the sampled version lies from the mean of
the original traffic (Zseby, 2004). Bias is the averaged difference of all samples of
the same size. Equation (5.9) was used to calculate the bias:

B ias = i Ef=i Mt - M (5.9)

81
^napier z> / i U i t p u v c o u i u d u u t i o a n i p i i i i g 1CU11U4UM

where N is the number of simulation runs, Mi and M are the means of the QoS
parameters for the sampled version and its original population respectively.
ii. Relative Standard Error (RSE): RSE measures the reliability of sampled version.
RSE is expressed as a percentage and can be defined as the standard error of the
sample (SE) divided by the sample size (n) as show in equation (5.10)

RSE = — X 100 (5.10)

n

iii. Curve fitting (i.e. data trend): another criterion to examine the behaviour of
sampled version in terms of representing the original population is evaluating the
trend of sampled data versus its original counterpart by applying the curve fitting.
Curve fitting is a useful tool for representing a data set in a linear, quadratic or
polynomial fashion. Curve fitting could be based on two functions, polynomial
curve fitting function, and polynomial evaluation function which can quickly and
easily fit a polynomial to a set of data points. The general formula for a polynomial
is given in equation (5.11):
f ( x ) = a 0x N + + a 2x N~2 + ...... 4- a N- i x + aN (5.11)
The degree of a polynomial is equal to the largest value of the exponents (N), (x) is
a set of data, and (a) is a set of polynomial coefficients. Polynomial curve fitting
function computes a least squares polynomial for a given set o f data (x) and
generates the coefficients of the polynomial which can be used to simulate a curve
to fit the data according to the specified degree (N). Whereas, polynomial
evaluation function evaluates a polynomial for a given set of (jc) values and then
generates a curve to fit the data based on the coefficients found using curve fitting
function (Lindfield and Penny, 2012).

5.3.5 Network Topology and Traffic Characteristics

To validate the performance of adaptive statistical sampling schemes, wireless-cum-

wired network topology with a size of 16 unidirectional connections as shown in Figure
4-3 were simulated using NS2. Half of the connections transmitted traffic from wireless
to wired network whereas the rest transmitted traffic from wired to wireless. The WLAN
was based on IEEE 802.lie , and it used the Enhanced Distributed Channel Access
(EDCA) scheme. The main parameters that modelled the wireless channel were the
default settings for IEEE 802.1 le. These parameters are shown in Table 4-1.

82
v^napier d ziuapuvc juiiiduuti oaiiipimg i^uuu\juvo

The traffic transmitted over the simulated network were: VoIP, video streaming, best
effort traffic, and background traffic. Constant Bit Rate (CBR) traffic was adapted to
model VoIP. VoIP was modelled as G.711 voice encoding scheme. The packet size of
VoIP was 160 bytes and the transmission rate was 64 kbps. The video streaming source
was YUV QCIF (176 X 144) Foreman (400 frame) (YUV QCIF, 2012). Prior to its
transmission, each video frame was fragmented into packets, which in turn had a
maximum length of 1024 bytes. The best-effort traffic had a fixed packet size of 1000
bytes and 125 kbps transmission rate. File Transfer Protocol (FTP) was used for the
background traffic. FTP was transmitted over TCP, whereas other traffics were
transmitted using the UDP transport protocol. The transmitted traffic over IEEE 802.1 le
EDCA (i.e. VoIP, video, best effort traffic, and background traffic) were mapped into
different access categories to represent different levels of priorities as shown in Table 4-
2. VoIP had the highest priority, whereas the background traffic had the lowest priority.
Each simulation lasted 500 seconds. Simulations were repeated 10 times for each
experiment. Each time, a different initial seed was used in order to randomly manage the
node’s transmission period as all the nodes were requested to transmit and stop at a given
time. The randomness introduced for different seeds avoided the bias of random number
generation.

5.4 Results and Discussion

In this section, the main QoS parameters of VoIP (i.e. throughput, delay, jitter, and
packet loss ratio) with their sampled versions using adaptive and non-adaptive sampling
techniques are presented in the following subsections.

5.4.1 Throughput

Figures 5-4 (a) - (g) show respectively the actual throughput with its sampled versions
using adaptive statistical sampling based on fuzzy inference system (FIS), adaptive
statistical sampling based on linear adjustment mechanism, adaptive statistical sampling
based on quarter adjustment mechanism, systematic sampling, stratified sampling, and
random sampling. The results obtained using adaptive statistical sampling approaches
were based on the initial parameters settings shown in Table 5-5. These initial
parameters settings were chosen experimentally, i.e. different values were tested to
monitor the response of the adaptive statistical sampling approaches and the most
suitable settings were chosen.

83
^ n a p ie r a n u ajiu y c k
3 uai.iai.ivai aaiiipnug i t m m iju w

I I — Original throughput Sampled throughput using fum sampling approach

I I
— Data trend_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
I I 1 1 I l 1 I 1

I I 1 1 1 1 1 1 1

I I 1 1 1 1 1 1 1
1 1114 1 li ill 1 1 1 1 1 1 1 -L 1- 1 4
| I t U
'H I HI i | | in li 1 1 1 i i 1 T 1 1 T

y y k a | |

ipiii y pm
/ U T O lI " l i
f i i 1 1 1 1 1 1 1
.1___ ii___ ii___ i___ i___ 1i____ .1............
i----- 11----- . 1----- 11----------
_l---------- ----
0 50 100 ISO 200 250 300 350 400 450 500
Simulation Time (Sec)

Sampled throughput using adaptive linear sampling approach Sampled throughput using adaptive quarter sampling approach
“ Data trend Data trend

50 100 150 200 250 304 350 400 450 500

Simulation time (see)
Simulation time (sec)

Sampled throughput using systematic sampling approach — Sampled throughput using stratified sampling approach 1
Data trend _________ Pd a trend _ _ _ _ _ _ J
1
. — 1--------- 1 -------- 1------------- i------------ --------- 1
l i l t l i

Simulation time (sec) Simulation time (sec)

(e)

450

(g)
Figure 5-4. Comparison of throughput with its sampled versions using: (a) Actual, (b)
Adaptive sampling based on FIS, (c) Adaptive sampling linear adjustment, (d) Adaptive
sampling quarter adjustment, (e) Systematic, (f) Stratified, (g) Random sampling.

84
Chapter 5 /vuapuve oiauM icai aam pu ug i cuuuijura

Table 5-5. Operating parameters of adaptive statistical sampling approaches.

■------- Operating parameters

Pre- and
Initial value post­ Threshold
of ISSI sampling values
Adaptive statistical *— ___
sections
sampling approaches
Adaptive sampling based on FIS 100 2 0.1
Adaptive sampling based on linear
100 5 1
adjustment mechanism
Adaptive sampling based on quarter
10 10 0.6
adjustment mechanism

In Figures 5-4 (b) - (g), the sample size for each sampled version was 240 packets (i.e.
sample fraction was 4.8 % of the actual traffic). The data trends shown in Figures 5-4
(a)-(g) were used to describe the behaviour of the observed sampled versions of
throughput and to illustrate the degree of discrepancy for each sampling method from
the original population (i.e. actual throughput).

It can be noticed from the trend of data in Figures 5-4 (a) - (g), how different sampling
approaches represent the actual throughput. As each sampling approach follows certain
procedure, the degree of discrepancy from the actual throughput is different for each
sampling method. However, it is observed that the sampled versions of throughput
obtained using adaptive statistical sampling approaches is closer to the actual
throughput as compared with the non-adaptive sampling approaches (i.e. systematic,
stratified, and random sampling). This is because the adaptive statistical sampling
approaches based on FIS, linear adjustment mechanism, and quarter adjustment
mechanism varied the length of Inter Sampling Section Interval (ISSI) according to the
statistical variations of throughput.

Figures 5-5 (a) - (c) show respectively the response of adaptive statistical sampling
approaches based on FIS, linear adjustment mechanism, and quarter adjustment
mechanism. It can be seen from Figures 5.5 (a) - (c) that during the sampling process of
adaptive statistical sampling approaches based on linear adjustment mechanism, quarter
adjustment mechanism, and FIS, whenever the calculated overall statistic of throughput
was less than pre-defined threshold values indicated in Table 5-5, the ISSI was
increased according to equation (5.2), (5.4), and (5.8) respectively. Otherwise, the ISSI
was decreased using equations (5.3), (5.5), and (5.8).

85
i^napier a ziuap uvc ou iusu i.ai oaiu pu ug i clihiiijucs

3fl 40 50
Iteration number

40 50 60 15 20 25
Iteration number Iteration number

(b) (c)

Figure 5-5. The length of ISSI for adaptive statistical sampling using: (a) Fuzzy
inference system, (b) Linear adjustment mechanism, (c) Quarter adjustment mechanism.

Tables 5-6 (a) - (f) summarise respectively the throughput measurement for different
sample fractions using adaptive statistical sampling based on (FIS), linear adjustment
mechanism, quarter adjustment mechanism, and non-adaptive sampling methods (i.e.
systematic, stratified, and random). It is established from Tables 5-6 (a) - (f) that the
variations of sampling versions using all sampling methods from actual mean and actual
standard deviation are increased as the sample size is decreased. The calculated absolute
error is also increased for all sampling methods as the sample fraction is decreased and
vice versa.

The summary statistics provided in Tables 5.6 (a) - (f) also indicate that sampled
versions of throughput for different sampling fractions obtained using the three adaptive
sampling approaches are closer to the original throughput as compared with the versions
obtained using non adaptive sampling approaches. This indicates that the proposed
adaptive sampling techniques outperform conventional sampling techniques.

86
v^napier 3 n u a p u v c o u tiu u ta i aaiiipiiug ic u im q u c a

Table 5-6. Throughput measurement results using different sampling methods: (a)
Adaptive sampling based on FIS, (b) Adaptive sampling based on linear adjustment, (c)
Adaptive sampling based on quarter adjustment, (d) Systematic, (e) Stratified, (f)
Random sampling.
__________________________________(a)__________________________________
Actual values Sample fraction (%)
Units: (kbps)
15.7 8.18 4.8 3.34
Mean throughput 68.35 68.37 68.45 67.57 69.33
Standard deviation 19.73 19.58 19.46 19.15 20.5
Absolute error 0.02 0.1 0.78 0.98

(b)

Units: (kbps) Actual values Sample fraction (%)

15.7 8.18 4.8 3.34
Mean throughput 68.35 68.1 67.9 67.5 69.3
Standard deviation 19.73 19.31 20.32 19.06 22.8
Absolute error 0.25 0.45 0.85 0.95

(c)
Units: (kbps) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean throughput 68.35 67.74 67.59 69.37 67.15
Standard deviation 19.73 21.4 21.87 22.41 23.27
Absolute error 0.61 0.76 1.02 1.2

(d)

Units: (kbps) Actual values Sample fraction (%)

15.7 8.18 4.8 3.34
Mean throughput 68.35 67.58 69.25 67.17 69.69
Standard deviation 19.73 22.32 22.71 22.85 23.8
Absolute error 0.77 0.9 1.18 1.34

(e)
Units: (kbps) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean throughput 68.35 69.1 67.5 67.04 66.17
Standard deviation 19.73 24.19 22.33 23.8 24.6
Absolute error 0.75 0.85 1.31 2.18

(f)

Units: (kbps) Actual values Sample fraction (%)

15.7 8.18 4.8 3.34
Mean throughput 68.35 69.12 67.38 69.5 66.59
Standard deviation 19.73 23.09 23.16 23.42 24.82
Absolute error 0.77 0.97 1.15 1.76

87
Chapter 5 AUajJUVC kJUtll»U(.ai OaiII|JlllIg Iiwuui|uvi}

The mean and standard deviation were not only the criteria used to access the accuracy
of throughput sampled version for representing the actual population. Bias and RSE
calculated using equations 5.9 and 5.10 respectively were also used to assess the
discrepancy between the original throughput and its sampled versions.
Figures 5-6 (a) - (c) illustrate respectively the comparison of the bias of sampled
throughput versions from the mean of actual throughput for different sample fractions
using conventional sampling approaches versus adaptive sampling based on FIS, linear
adjustment mechanism, and quarter adjustment mechanism. The results illustrate that
the bias was decreased and became closer to zero for all sampling approaches when the
sample size was increased. However, it can be observed from the figures that the
adaptive statistical sampling approaches have a lower bias as compared with systematic,
stratified, and random sampling techniques. For example, the bias of sampled
throughput for 8.18% sample fraction (i.e. sample size 409 packets) using adaptive
statistical sampling based on FIS was 0.1, whereas the bias values when using
systematic, stratified, and random sampling were 0.9, -0.85, and -0.97 respectively.

S flJ

Sample Fraction I%)

(a)
1 "
* i 4 Original Throughput * 1 4 Original Throughput
i 1
+ •

• ■ Adaptive sampling using linear adjustment ■ Adaptive sampling using quarter adjustment
1

V r * % Systematic sampling Systematic sampling

i 1 •
i * Stratified sampling 1 ★ Stratified sampling
i 1 Random sampling 1 • Random sampling
I 1 l 1 1
f
J J 1 f
_ J_ _ _ _ _ _ _ _ _ -------------- -
+
- f - 1 1 1 1
i 1 1 I ■ 1 1
9U i
L
I 1 l
9U 1
1 1 I * 1 ' *
■ ■ t
i I i ■ I 1
: * i i I I I 1
i i 1 i i r
• i i i i i i i
i i I i i i
i i i *
i i
★ | i i I i i
i i i i i i
* it II
Sample fraction!^) Sample fractioii (S»)

(b) (c)
Figure 5-6. Comparison of bias of sampled throughput obtained from non-adaptive
sampling with bias obtained using: (a) Adaptive sampling based on FIS, (b) Adaptive
sampling based on linear adjustment, (c) Adaptive sampling based on quarter
adjustment.
Chapter 5 A daptive sta tistica l sa m p lin g i eu im q u cs

RSE calculated from throughput sampled versions obtained from conventional sampling
approaches (i.e. systematic, stratified, and random) compared with the RSE obtained
using adaptive statistical approaches based on FIS, linear adjustment, and quarter
adjustment as shown in Figures 5-7 (a) - (c) respectively. It is established that the results
obtained from the three adaptive sampling approaches are more accurate as compared
with systematic, stratified, and random sampling. For example, the overall reductions of
RSE were respectively 13.46%, 16.67% and 16.67% when using adaptive sampling
based on FIS as compared with systematic, stratified, and random sampling.

1.2
A daptive sam pling using F IS
1.1 ISystematic sam pling
IStratified sam pling
IR andom sam pling
0.9

0.7

(a)
0.5

0.4

0.3

0.2
0.1

Sam ple fraction (% )

1 -------- r - A daptive sam pling using lin ear a d ju stm en t

1 i _
_1 i H System atic sam pling
1 j _ H i S tratified sam pling
1 I
H R andom sam pling
1 I i I 1
_i j _L 1
1 i l 1 1

1 I I 1 1

(b) 1 I i 1 1

' I i 1 1

! i i 1 1

1 I 1 1
IIH i _ _ _ _ _ _ _ : ____
1 I 1 1
I I j
1
_L_ H I m l
4 6 8 10 12 14 16 18
Sam ple fraction (% )

1
l A daptive sam pling using q u a r te r ad ju stm e n t
I H System atic sam pling
i
H i Stratified sam pling
I
i H R andom sam pling
i i i
i i i
i i i
i i i
r I i
,i i i
(c) i i i

i i i

!i i i

i
nIII i

i
t i

i
i II 8 10 12
mi
Sam ple fraction (% )

Figure 5-7. RSE of sampled throughput using conventional sampling versus: (a)
Adaptive sampling based on FIS, (b) Adaptive sampling based on linear adjustment, (c)
Adaptive sampling based on quarter adjustment.

89
unapter s /lUUpUVC iJUlUaUMll Oailipmig aW1UU1|UVU

5.4.2 Delay

Delay was another QoS parameter measured from sampled versions obtained using
adaptive statistical sampling based on FIS, linear adjustment mechanism, and quarter
adjustment mechanism, along with non-adaptive sampling techniques (i.e. systematic,
stratified, and random sampling) as depicted in Figure 5-8 (b) - (g).

Figures 5-8 (a) - (g) show the comparison between the actual delay and its sampled
versions using the above mentioned sampling techniques respectively. The mean and
standard deviation of data trend for original delay were 33.5 msec and 7 msec
respectively, whereas the sampled versions o f delay obtained from the adaptive
sampling based on FIS, linear adjustment mechanism, and quarter adjustment
mechanism had the mean of 32.67 msec, 33.99 msec, and 32.33 msec and standard
deviation of 7.7 msec, 5.7 msec, and 7.5 msec respectively. However, the mean and
standard deviation of data trend for sampled delay using systematic, stratified, and
random sampling were (35.82 msec, 34.95 msec, and 35.85 msec) and (10.73 msec,
5.42 msec, and 5.3 msec) respectively. This indicates that sampled versions of delay
using adaptive statistical sampling approaches represented the original delay more
accurately and effectively.

Tables 5-7 (a) - (f) statically demonstrate how different sample fractions obtained using
adaptive and non-adaptive sampling approaches represent the actual delay. For all
sampling methods, as the sample size was increased, the variation of sampled mean,
standard deviation from the actual mean and standard deviation decreased accordingly.
This is because a large sample size includes more packets that in turn increase the
probability of obtaining more details from the actual delay. However, the delay sampled
versions obtained from the proposed adaptive sampling techniques represent the actual
delay more closely than the conventional sampling techniques. For example, the
absolute error of 8.18% sample fraction of systematic, stratified, and random sampling
was increased by 32.2%, 31.03%, and 33.33%, as compared with the absolute error
obtained using the adaptive sampling using linear adjustment mechanism.

The improvement of adaptive sampling approaches over conventional sampling

techniques is due to the selection of packets in the former approaches depending on the
statistical variation of the traffic whereas the packet selection in the latter techniques
depended either on a fixed or random sample rate.

90
cnapier 3 nua|Jiivc ju iiisu i.ai uaiii[iuiig iwun»|tavu

— Sampled delay using adaptive fuzzy sampling approach

Data trend
__ 1 I— — Data trend

200 250 300 50 100 150 200 250 300 350 400 450 500
Simulation Time (set) Simulation time (set)

(b)
Sampled delay using adaptive linear sampling approach ~ Sampled delay using adaptive quarter sampling approach
"Data trend ■j 1--------- [--------- -j-. Data trend ______ ______ ______ _ _ _ _ _ _

50 100 150 200 250 300 350 400 450 500

Simulation time (sec) 0 50 100 150 200 250 300 350 400 450 500
Simulation time (set)

— Sampled delay using systematic sampling approach Sampled delay using stratified sampling approach
_ _ 1 I____ Data trend Dati trend

200 250 300

Simulation Time (sec) Simulation time (sec)

(e) — Sampled delay using random sampling approach

— Data trend_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

o 80

400

(g)
Figure 5-8. Comparison of delay with its sampled versions using: (a) Actual, (b)
Adaptive sampling based on FIS, (c) Adaptive sampling linear adjustment, (d) Adaptive
sampling quarter adjustment, (e) Systematic, (f) Stratified, (g) Random sampling.
91
cnapter 5 Adaptive statistical sampling i ecnniques

Table 5-7. Delay measurement results using different sampling methods: (a) Adaptive
sampling based on FIS, (b) Adaptive sampling based on linear adjustment, (c) Adaptive
sampling based on quarter adjustment, (d) Systematic, (e) Stratified, (f) Random
sampling.

(a)

Units: (msec) Actual values Sample fraction (%)

15.7 8.18 4.8 3.34
Mean delay 33.5 33.61 32.64 32.6 34.56
Standard deviation 22.63 22.09 22.01 21.44 20.47
Absolute error 0.11 0.86 0.9 1.06

(b)
Units: (msec) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34 !
Mean delay 33.5 33.98 34.3 34.48 32
Standard deviation 22.63 22.7 22.2 21.93 21.75
Absolute error 0.48 0.8 0.98 1.5

(c)
Units: (msec) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean delay 33.5 32.85 34.45 34.94 35.26
Standard deviation 22.63 22.88 23.08 23.81 24.05
Absolute error 0.65 0.95 1.44 1.76

(d)
Units: (msec) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean delay 33.5 32.64 34.68 35.24 31.57
Standard deviation 22.63 24.19 24.81 24.54 24.82
Absolute error 0.86 1.18 1.74 | 1.93

(e)
Units: (msec) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean delay 33.5 34.46 34.66 35.03 31.73 |
Standard deviation 22.63 24.19 25.64 24.32 24.17
Absolute error 0.96 1.16 1.53 1.77

(f)
Units: (msec) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean delay 33.5 32.52 34.7 35.85 35.88
Standard deviation 22.63 24.19 24.94 25.68 24.61 !
Absolute error 0.98 1.2 2.35 2.38

92
^napier s A daptive d ia u su ca i sa m p lin g i ecnm ques

Bias and RSE were also used to assess the accuracy of sampled delay. The bias for
adaptive sampling based on FIS, linear adjustment mechanism, quarter adjustment
mechanism were closer to zero as compared with the non-adaptive sampling approaches
as shown in Figures 5-9 (a) - (c) respectively. For example, the bias of 3.34% sample
fraction was reduced by 55.46%, 36.97%, and 26.05% in case of adaptive sampling
based on FIS, linear adjustment mechanism, quarter adjustment mechanism respectively
comparing with random sampling approach.

♦ Original Delay
-— 1 - - =-

•
Adaptive sampling using FIS

1
1
1
1
1
1
1
1

1
1
1
1
1
% Systematic sampling

1
1
1
1
j
^ Stratified sampling
Random sampling

4
-------- 1--------
____ 1 ____

-------- 1--------
1

im
1
L

1
1
1

1
1
1
1
1
1
1
1
1
1
1

i
i
i
i
i
i
i
r

i
i
i
i i i i i
i i i i i
---------
--- + -!- + - - f ------------> --------- r -------------
1 1 1 1 1
t i l l 1
1 1 1 1
l ■ L lB i - — — 4 .
t i l l i
t i l l i
J L J L L
1 1 1 1 1
★ 1 1 1 1 1
% I I I I l
Sample Traction (%)

(a)
------ 1 ■ |
• 1 f Original Delay 4 Original Ddai
i i
i i 1 Adaptive sampling using inear adjustment I Adaftii e sampling using quarter adjustnwnt
£ Systematic sampling i Systematic sampling
i i ★ Stratified sampling
* ♦ Stratified sampling
i i
i Randomsampling I Random sanpline ______
i i i
1
------------- _ _ « --------
----------- r
>1 i i
i i i
J 1 L
1 1 1
1 1 1
♦ it 7 i — ♦
i i i
i i i
i t r
i i i
i i i
i
i i i
i i i
------- I - --------------
i i i
★ i i i
* i i i........ -
8
Sample fraction Irr)
18 8 II
Sample fraction IIt)

(b) (c)

Figure 5-9. Comparison of bias of sampled delay obtained from non-adaptive sampling
with bias obtained using: (a) Adaptive sampling based on FIS, (b) Adaptive sampling
based on linear adjustment, (c) Adaptive sampling based on quarter adjustment.

Relative Standard Error RSE as shown in Figures 5-10 (a) - (c) proves that adaptive
statistical approaches based on FIS, linear adjustment, and quarter adjustment
respectively outperformed conventional sampling approaches.

93
cnapier 3 A daptive a ia u su c a i sa m p lin g 1 ecnniques

Although the value of RSE decreased and become closer to zero for all sampling
methods when the sample size was increased, the RSE values obtained from the three
adaptive sampling approaches were less than RSE calculated from sampled delay using
systematic, stratified, and random sampling. The overall RSE was increased by 3.7%,
1.85% and 3.7% when using systematic, stratified, and random sampling respectively
compared with adaptive statistical sampling based on quarter adjustment.

1.2
A daptive sam pling using F IS
1.1 I System atic sam pling

1 I S tratified sam pling

I R andom sam pling__________
0.9

0.8

0.7

S 0.6
(a) as
0.5

0.4

0.3

in
0.2

0.1

10 12
Sam ple fraction (% )

1 1
1 1 A daptive sam pling using lin e a r ad ju stm en t
1 H System atic sam pling
1
1
Hi Stratified sam pling
1 H R andom sam pling
r I n --------r
i i i i
- - -
I i i i

;1 i i i
L
(b) 1
1
i
i J
i i
i
;1 i i i
11 i i i
1 i i i
1 i i i
r i n t
s i i i i

i i i i

i
r _ J ' '
10 12
Sam ple fractio n (% )

1 1
1 1 A daptive sam pling using q u a r te r ad ju stm en t
1 1 H S ystem atic sam pling
1 1 H S tratified sam pling
r i
i i H R andom sam pling
r r
i i i
i i i
(C) - - -
! i i 1
i 1
i i 1
i i 1
i i 1
i i 1
n i 1
i i 1
i
i 1
H i
i 1

i
•I. j m i
8 10 12
S am ple fractio n (% )

Figure 5-10. RSE of sampled delay using conventional sampling versus: (a) Adaptive
sampling based on FIS, (b) Adaptive sampling based on linear adjustment, (c) Adaptive
sampling based on quarter adjustment.
94
Chapter 5 Adaptive Statistical Sampling Techniques

5.4.3 Jitter

Jitter was another important QoS parameter to be measured from sampled versions
obtained using the adaptive and non-adaptive sampling approaches. The results shown
in Figures 5-11 (a) - (g) highlighted the potential of sampling the actual jitter using
different sampling approaches.

It can be perceived from Figures 5-11 (b) - (d) how adaptive statistical sampling based
on FIS, linear adjustment, and quarter adjustment represent the actual jitter shown in
Figure 5-11 (a). The statistic of data trend obtained from jitter sampled versions using
adaptive statistical sampling based on FIS, linear adjustment, and quarter adjustment
had respectively; 12.99 msec, 12.63 msec, and 12.34 msec for their means and 0.76
msec, 1.18 msec, and 1.57 msec for their standard deviations. These values were very
close to the means and standard deviations of the trend of actual jitter which were 12.92
msec, and 0.81 msec. However, the mean and standard deviation of data trend obtained
using the non-adaptive sampling techniques as shown in Figures 5-11 (e) - (g) were
respectively 13.19 msec, 2.14 msec for systematic, 13.27 msec, 3.17 msec for stratified,
and 13.7 msec, 1.81 msec for random sampling. This indicates that the three proposed
adaptive statistical sampling approaches provided results closer to the actual jitter as
compared with the non-adaptive sampling techniques. The statistic results of jitter for
different sample fractions using adaptive statistical sampling based on FIS, linear
adjustment, quarter adjustment approaches and conventional sampling techniques are
summarised in Tables 5-8 (a) - (f) respectively.

From the Tables 5-8 (a) - (f), the percentage of difference between the values of mean
and standard deviation for 3.34% sample fraction obtained from adaptive sampling
approaches and the actual mean and standard deviation were respectively 4.01%, 2.33%
for fuzzy adjustment approach, 7.45%, 5.5% for linear adjustment approach, and 3.37%,
10.53% for quarter adjustment approach. Whereas the percentage of difference between
the values of the actual mean and standard deviation, and the values of mean and
standard deviation obtained from conventional sampling techniques were respectively
8.24%, 16.56% for systematic sampling, 9.31%, 20.42%, for stratified sampling, and
11.57%, 29.01% for random sampling. This analysis verified that the proposed adaptive
statistical sampling with fuzzy adjustment, linear adjustment, and quarter adjustment
mechanisms generated sampled versions which are very close to the original jitter as
compared with non-adaptive sampling approaches.

95
Chapter 5 Adaptive statistical sampling 1 ecnmques

— Sampled jtter using adaptive fuzn sampling approach

— Data trend

500
Simulation Time (sec) Simulation Time (see)

Sampled jitter using adaptive linear sampling approach Sampled jhter using adaptive quarter sampling approach

— Data trend Data trend

70-

500

(c) (d)
— bampted jitter using svstemanc sampling approacn Sampled jitter using stratified sampling approach
Data trend___________________________ — Data trend
50-

30-

500
Simulation time (set) Simulation time (set)

(e)
— Sampledjitter using random sampling approacb
Data trend _________________

(g)

Figure 5-11. Comparison of jitter with its sampled versions using: (a) Actual, (b)
Adaptive sampling based on FIS, (c) Adaptive sampling linear adjustment, (d) Adaptive
sampling quarter adjustment, (e) Systematic, (f) Stratified, (g) Random sampling.
96
Chapter s Adaptive statistical sam pling i ecmuqucs

Table 5-8. Jitter measurement results using different sampling methods: (a) Adaptive
sampling based on FIS, (b) Adaptive sampling based on linear adjustment, (c) Adaptive
sampling based on quarter adjustment, (d) Systematic, (e) Stratified, (f) Random
sampling.

__________________________________ (a)__________________________________
Units: (msc) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean jitter 12.92 12.99 13.02 13.09 13.46
Standard deviation 7.56 7.57 7.47 7.67 7.74
Absolute error 0.07 0.1 0.17 0.54

(b)
Units: (msc) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean jitter 12.92 13.1 13.18 13.42 13.96
Standard deviation 7.56 7.39 7.77 7.11 8
Absolute error 0.18 0.26 0.5 1.04

(c)
Units: (msc) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean jitter 12.92 13.01 13.1 12.64 13.37
Standard deviation 7.56 7.9 8.01 8.23 8.45
Absolute error 0.09 0.18 0.28 0.45

(d)

Units: (msc) Actual values Sample fraction (%)

15.7 8.18 4.8 3.34
Mean jitter 12.92 12.72 12.63 13.71 14.08
Standard deviation 7.56 9.89 9.92 8.9 9.06
Absolute error 0.2 0.29 0.79 1.16

(e)
Units: (msc) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean jitter 12.92 12.7 13.23 12.22 11.82
Standard deviation 7.56 9.68 9.1 9.67 9.5
Absolute error 0.22 0.31 0.7 1.1

(f)
Units: (msc) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean jitter 12.92 13.16 13.27 13.7 11.58
Standard deviation 7.56 10.01 8.72 10.04 10.65
Absolute error 0.24 0.35 0.78 1.34

97
Chapter 5 Adaptive statistical sampling i ecnniques

The bias values shown in Figures 5-12 (a) - (c) demonstrate that the bias obtained from
adaptive and non-adaptive sampling approaches decreased accordingly as the sample
size increased.

The bias values obtained from adaptive sampling based on fuzzy approach, linear
adjustment approach, and quarter adjustment approach were closer to zero as compared
with non-adaptive sampling approaches. For instance, the biases of sample size of 409
packets (i.e. 8.18% sample fraction) obtained using adaptive sampling were 0.1 for
fuzzy approach, 0.26 for linear adjustment approach, and 0.18 for quarter adjustment
approach whereas the biases obtained from the same sample size using conventional
sampling techniques were 0.29 for systematic sampling, 0.31 stratified sampling, and
0.35 for random sampling. This indicates that actual jitter could be represented
effectively using adaptive sampling approaches rather than conventional sampling
approaches.

ft Original Jitter
I Adaptive sampling using FIS
^ Systematic sampling
★ Stratified sampling
ft Random sampling_______

Sample fraction (%)

(a)
1 t Original Jitter
ui— ♦ Original JiUer
1 1 Adaptive sampling using linear adjustment I Adaptive sampling using quarter adjustment
* 1 ^ Systematic sampling ^ Systematic sampling
1 1 f Stratified sampling * Stratified sampling
1
ft Random sampling i Random sampling_ _ _ _ _ _ _ _ _ _
1 ft 1
1 1
------- - 4 --------------+ -------
1
t
ff
----------- +
- ~ . L - ♦ —
i
i * ! *
i I

i I
i ★ I
i I
i I
“1
★ r I
i I
• i I
i i._
4 4 8 II 14 16 1
Sam ple fra c tio n I rt) Sample fraction (%)

(b) (c)

Figure 5-12. Comparison of bias of sampled jitter obtained from non-adaptive sampling
with bias obtained using: (a) Adaptive sampling based on FIS, (b) Adaptive sampling
based on linear adjustment, (c) Adaptive sampling based on quarter adjustment.
98
Chapter 5 Adaptive statistical sampling i ecnmques

Relative Standard Error RSE as shown in Figures 5-13 (a) - (c) also illustrated how the
sampling errors for all sample fractions obtained from jitter sampled versions using
adaptive statistical sampling approaches were closer to zero as compared with non-
adaptive sampling approaches.

0.5
A daptive sam pling using F IS
0.45 ! System atic sam pling
I Stratified sam pling
0.4 i R andom sam pling__________

0.35

0.3

| 0.25
(a)
0.2

0.15

0.1

0.05

1 J i Il
10 12
Sam ple fraction (% )

0.5 1
1 1
1 1 f c : A daptive sam pling using lin ear a d ju stm en t
0.45 - - - H S ystem atic sam pling
1 1
1 1 H S tratified sam pling
0.4 1 “1 R £ R andom sam pling
1 1 1 " i " ............. i
1 1 J 1 1
0.35 1 1 1 1
1
1 1 1 1 1
0.3
1 1 1 1 1
1 1 1 1 1

(b) gu s 1 1 1 1 1
. I_ 1 1 1 1
0.2 1 1 1 1 1
1 1 1 1 1
0.15
1 1 1 1 1
11
0.1 ------- - 11
1
11 l i j
0.05 1
1
_1_ II H l i b
10 12
Sam ple fraction (%)

0.5
Adaptive sam pling using q u a r te r ad ju stm en t
0.45 I S ystem atic sam pling
I S tratified sam pling
0.4 I R andom sam pling_________________________

0.35

0J ------- 4 ---
(c) u
£ 0 .2 5

0.2

0.15

0.1

0.05

0 10 12
Sam ple fraction (% )

Figure 5-13. RSE of sampled jitter using conventional sampling versus: (a) Adaptive
sampling based on FIS, (b) Adaptive sampling based on linear adjustment, (c) Adaptive
sampling based on quarter adjustment.

99
Chapter 5 A d a p tiv e S ta tistic a l S am p lin g T e c h n iq u e s

As an example of the accuracy of adaptive statistical sampling, the RSE value for 4.8%
sample fraction as shown Figure 5-13 (a) was reduced by 16.67%, 23.08%, and 25.93%
respectively when fuzzy adjustment approach was applied as compared with systematic,
stratified, and random sampling.

5.4.4 Packet Loss Ratio

Figures 5-14 (a) - (g) illustrate the actual packet loss ratio in addition to the sampled
versions of loss ratio using adaptive and non-adaptive sampling approaches with a
sampling fraction of 3.34%. These figures show how adaptive and non-adaptive
sampling approaches tracked and sampled the actual traffic losses.

It can be observed from Figures 5-14 (a) - (g) that the accurate estimation of packet
losses were obtained using adaptive sampling techniques based on fuzzy adjustment
approach, linear adjustment approach, and quarter adjustment approach. The mean and
standard deviation of sampled packet losses obtained from adaptive sampling
techniques were 0.27, and 0.82 for fuzzy adjustment approach, 0.26, and 0.77 for linear
adjustment approach, and 0.38, 0.91 for quarter adjustment approach. These results
were very close and comparable with the mean and standard deviation of actual packet
loss ratio which were respectively 0.31, and 0.81. The data trends shown in Figures 5.14
(a) - (g) also demonstrate that sampled versions of packet loss ratio obtained using the
adaptive statistical sampling techniques with the three different approaches closely
represented the actual packet loss ratio as compared with non adaptive sampling
approaches.

Tables 5-9 (a) - (f) demonstrate how the actual packet loss ratio was represented by
different sample fractions obtained using adaptive and non adaptive sampling
approaches. For all sampling approaches, as the sample size was decreased, the
variation of sampled mean, standard deviation from the actual mean and actual standard
deviation increased accordingly. However, the packet loss ratio calculated from
sampled versions obtained using the proposed adaptive sampling techniques represented
the actual loss ratio better than the conventional sampling techniques. For example, the
absolute error of sample fraction of 4.8% was increased by 15%, 25%, and 60% in case
of systematic, stratified, and random sampling comparing with the absolute error
obtained using adaptive sampling based on quarter adjustment which was 0.2.

100
Chapter 5 Adaptive Statistical sampling 1 ecnniques

— Original packet loss ratio — Sampled packet loss ratio using adaptive fuzzy sampling approach
— Data trend — Data trend ___________________

200
Simulation rune (sec) Simulation Time (sec)

(a)
l — Sampled packet loss ratio using adaptive linear sampling approach — Sampled packet loss ratio using adaptive quarter sampling approach
I — Data trend t t t — Data trend_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
L _ _ J I I 1 ____ I I- ----------------- 5----- 1----- 1------ i----- 1------1------ i------ 1----- 1------ t------
_

I I I I I I I
I I I I I I I
I I I I I I I

150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Simulation Time (sec) Simulation Time (sec)

(c)
• I I I
Sampled packet loss ratio using systematic sampling approach 1 1 I Sampled packet loss ratio using stratified sampling approach
— Datatrend ____________________ 1 1 1 — Datatrend
J 1 ' L -t I

0 50 100 150 200 250 300 350 400 450 500

Simulation Time (sec) Simulation Time (sec)

(e) — Sampled packet loss ratio using random sampling approach

Datatrend_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

0 50 100 200 250 300

(g)

Figure 5-14. Packet loss ratio with its sampled versions using: (a) Actual, (b) Adaptive
sampling based on FIS, (c) Adaptive sampling linear adjustment, (d) Adaptive sampling
quarter adjustment, (e) Systematic, (f) Stratified, (g) Random sampling.
101
cnapter :> Adaptive siau su cai sam pling i ecnniques

Table 5-9. Packet loss ratio measurement results using different sampling methods: (a)
Adaptive sampling based on FIS, (b) Adaptive sampling based on linear adjustment, (c)
Adaptive sampling based on quarter adjustment, (d) Systematic, (e) Stratified, (f)
Random sampling.

(a)
Units: (%) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean packet loss ratio 0.31 0.31 0.31 0.29 0.27
Standard deviation 0.81 0.8 0.82 0.8 0.82
Absolute error 0.005 0.01 0.07 0.11

(b)
Units: (%) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean packet loss ratio 0.31 0.3 0.32 0.29 0.26
Standard deviation 0.81 0.8 0.83 0.76 0.77
Absolute error 0.02 0.03 0.07 0.16

(c)
Units: (%) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean packet loss ratio 0.31 0.3 0.29 0.37 0.38
Standard deviation 0.81 0.81 0.76 0.9 0.91
Absolute error 0.015 0.05 0.2 0.21

(d)
Units: (%) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean packet loss ratio 0.31 0.3 0.33 0.27 0.38
Standard deviation 0.81 0.8 0.83 0.75 0.95
Absolute error 0.018 0.07 0.23 0.23

(e)
Units: (%) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean packet loss ratio 0.31 0.33 0.37 0.23 0.39
Standard deviation 0.81 0.83 0.82 0.63 0.93
Absolute error 0.08 0.18 0.25 0.26

(1D
Units: (%) Actual values Sample fraction (%)
15.7 8.18 4.8 3.34
Mean packet loss ratio 0.31 0.37 0.38 0.41 0.47
Standard deviation 0.81 0.88 0.9 0.96 0.99
Absolute error 0.21 0.23 0.32 0.54

102
Chapter 5 Adaptive statistical sampling i ecnniques

The accuracy of sampled packet loss ratio was also measured using bias as shown in
Figures 5-15. As shown in the Figures, the bias of sampled packet loss ratio using
adaptive sampling based on FIS, linear adjustment mechanism, quarter adjustment
mechanism were closer to zero for all sampling fractions as compared with the non-
adaptive sampling approaches. For example, the bias of 4.8% sample fraction was
reduced by 75%, 75%, and 25% in case of adaptive sampling based on FIS, linear
adjustment mechanism, quarter adjustment mechanism respectively comparing with
stratified sampling approach.
0 .2 5
* O rig in a l P a c k e t loss ratio
■ A d ap tiv e sam p lin g u sin g F IS
0.2 * Sy stem atic sam p lin g
* S tra tified sam p lin g

0 .1 5 * R a n d o m sam p lin g___________

(a)
£ 0 .0 5

♦- - 1~
-0 .0 5

-0 . 1.

Sam ple fractio n

0 .2 5
* O rig in a l P a c k e t loss ratio
■ A d a p tiv e s a m p lin g u sin g lin e a r a d ju s tm e n t
0.2 * S y stem atic sa m p lin g
* S tra tified sa m p lin g
0 .1 5 * R a n d o m sa m p lin g _________ _______________

0.1
(b)

r
-0 .0 5

-0.1,
S am p le fra c tio n ( % )

0 .2 5
* O rig in a l P a c k e t loss ratio
■ A d ap tiv e s a m p lin g u sin g q u a rte r a d ju s tm e n t
0.2
* System atic sa m p lin g
* S tra tified sa m p lin g
0 .1 5 * R an d o m sam p lin g __________________________

0.1
(c)
a 0 .0 5

-
t
-0 .0 5

- 0 . 1.

Sam p le fra c tio n (% )

Figure 5-15. Comparison of bias of sampled packet losses obtained from non-adaptive
sampling with bias obtained from adaptive sampling based on: (a) FIS, (b) linear
adjustment approach, (c) quarter adjustment approach.
103
Chapter 5 Adaptive Statistical sampling l ecnniques

The comparisons between RSE of sampled packet losses obtained using conventional
sampling approaches, and the RSE obtained using adaptive statistical approaches based
on FIS, linear and quarter adjustments are shown in Figure 5-16 (a) - (c) respectively.
The results obtained from the three adaptive sampling approaches were more accurate
as compared with systematic, stratified, and random sampling. The overall reductions of
RSE were respectively 10.5%, 15% and 22.7% when using adaptive sampling based on
FIS compared with systematic, stratified, and random sampling.

0.05 1 1
1 1 ■ j A daptive sam pling using F IS
0.045 H System atic sam pling
1 1
H I S tratified sam pling
1 1
0.04 1 1 ■ R andom sam pling
1 1
0.035
1 1
1 1
0.03 1 1
1 1
0.025 - - - 1-
(a) DC 1 1
L 1
0.02 i 1
l 1
0.015
I 1
L
0.01

0.005
:i
1

l; r
1
1 1
III
Ih III 1
0.
4 6 8 10 12 14 16 18
Sam ple fractio n (% )

0.05
i A daptive sam pling using lin e a r ad ju stm e n t
0.045 — u --------------- H System atic sam pling
I
I H i S tratified sam pling
0.04 I ■ R andom sam pling
I l I
0.035
i i i
i i i
0.03 i i I
i I I

(b) 2060-025 i I i
i I I
0.02 I i
I I
1
0.015
1
i J
0.01 i I
i i
|
0.005

! 1 1 II I m i
0.
10 12
Sam ple fractio n (% )

0.05
A daptive sam pling using q u a rte r ad ju stm en t
0.045 I S ystem atic sam pling
I S tratified sam pling
0.04 I R andom sam pling_________________________

0.035

0.03

0.025
(C) DC

0.02

0.015

0.01

0.005

0 llll
10 12
Sam ple fraction (%)

Figure 5-16. RSE of sampled packet loss ratio using conventional sampling versus: (a)
Adaptive sampling based on FIS, (b) Adaptive sampling based on linear adjustment, (c)
Adaptive sampling based on quarter adjustment.

104
tn ap ier a Adaptive diausucai sampling lecnmques

5.5 Summary
In this chapter, three adaptive statistical sampling techniques to adjust the sampling rate
of multimedia traffic were developed and evaluated. A novel aspect of the proposed
techniques was an adjustment of sampling rate based on traffic's statistics. The sampling
rate of the three devised sampling techniques was controlled using three different
mechanisms: a simple linear adjustment mechanism, a quarter adjustment mechanism,
and a fuzzy inference system. The proposed techniques decreased the sampling rate
when the statistics of the traffic did not significantly change and increased the sampling
rate when the statistics of the traffic significantly changed.

A comparison of the three proposed adaptive statistical sampling techniques with

conventional sampling techniques (i.e. systematic, stratified, and random sampling) was
also carried out in this chapter. The findings indicated that the developed adaptive
sampling methods were more effective than conventional sampling methods. The
sampled versions produced using adaptive statistical sampling techniques were more
representative to the original population than the sampled versions produced using
conventional sampling techniques. The significant difference between adaptive
statistical sampling techniques and conventional sampling techniques was the manner of
sampling traffic. The sample interval was adjusted during the sampling process
according to traffic's statistics in case of adaptive statistical sampling approaches based
on linear adjustment mechanism, quarter adjustment mechanism, and FIS. In other
words, the sampling rate was decreased when the statistics of the traffic did not
significantly change and increased when the statistics of the traffic significantly
changed.

Conversely, the sampling rate of conventional sampling techniques was either constant
as in systematic sampling or changed randomly as in stratified and random sampling.
The fixed and random sampling rates may result in a significant discrepancy between
the actual data and its sampled version.

Advantages of the proposed adaptive statistical sampling techniques were the ease of
implementation and the quick response to traffic changes.

105
Chapter 6 Techniques to Evaluate Network Quality
of Service Using Statistical and Artificial Intelligence

6.1 Introduction
Evaluation of QoS is an important task in managing computer networks. This is
currently carried out by analysis or measurement techniques. Analysis techniques are
used to examine the characteristics of the traffic, whereas measurement techniques are
applied to determine how well the network treats an ongoing traffic. The contribution of
this study is to propose mechanisms that combine analysis and measurement techniques
to evaluate QoS in multimedia applications in an effective manner. In this chapter, two
innovative %QoS evaluation approaches are proposed. The first approach combined
Fuzzy C-Means (FCM) and regression model to analyse and assess QoS o f multimedia
applications in a simulated network, whereas the other analysed and assessed QoS in
multimedia applications using a combination of supervised and unsupervised neural
networks. The transmitted application’s QoS parameters were initially analysed either
by FCM clustering algorithm or by the unsupervised learning Kohonen neural network
(i.e. Self-Organising Maps (SOM)). The analysed QoS parameters were then used as
inputs to a regression model or supervised learning Multi-Layer Perceptron (MLP)
neural network in order to quantify the overall QoS. The proposed QoS evaluation
system provided information about the network’s QoS patterns and based on this
information, the overall network’s QoS was successfully quantified.

Section 6.2 of this chapter presents a discussion of the related studies. The proposed
FCM clustering algorithm, regression model, Kohonen neural network, and Multi-Layer
Perceptron (MLP) neural network in addition to the simulation of experiments and
traffic models are discussed in section 6.3. The findings of these studies are presented in
section 6.4. Finally, the summary of this chapter is provided in section 6.5.

6.2 Related Works

In section 3.3 (see Chapter 3), taxonomy of quality of service evaluation with the state
of the art of recent related studies were discussed in details. Previous studies generally
aim to evaluate network quality of service either by analysing the characterises of

106
c n a p te r o E v a lu a tio n 01 iNetworK Q u a lity 01 s e rv ic e

network traffic as reported in (Chen et al, 2009), (Timo et al, 2002), (Hoang et al, 2010),
(Wang et al, 2009), (Ting et al, 2010), and (Kiziloren and Germen, 2007) or by
measuring the network performance to determine how well the network treats the
ongoing traffic as reported in (Palomar et al, 2008), (Brauer et al, 2008), (Mishra and
Sharma, 2003), (Al-Sbou et al, 2008), (Brekne et al, 2002),. and (Mohammed et al,
2001). A novel aspect of this study is to propose an evaluation system that combines
analysis and measurement techniques to effectively evaluate network QoS. The first
contribution of this study is to analyse and classify network QoS parameters (i.e. delay,
jitter, and packet loss ratio) using, either Fuzzy C-means (FCM) or Self Organizing
Maps (SOM). Due to the ability of FCM and SOM to derive meaning from imprecise
values as reported in (Cirstea et al, 2002) and because of the natural characteristics of
network QoS parameters, where a single cluster could not be clearly identified, FCM
algorithm and SOM are suitable for QoS analysis.

Another contribution of this study is to develop QoS assessment techniques. The

proposed techniques should evaluate QoS in a manner similar to human subjects and
quantify the QoS without the necessity for complex mathematical models as in
objective approaches taking into the account the QoS requirements of each type of
multimedia application. Also, the proposed QoS assessment techniques should not add
extra load to the network as the case of active approaches, nor depend on the whole
collected packets, like passive approaches. The proposed assessment techniques are
based on the analysed traffic generated from the proposed analysis techniques in order
to overcome some drawbacks of both active and passive measurement methods. A
regression model was developed and multi-layer perceptron (MLP) was trained to
combine the QoS parameters (i.e. delay, jitter, and packet loss ratio) for each QoS class
identified by SOM or FCM to estimate the overall QoS.

6.3 Description of the Approaches

A schematic diagram of the proposed network QoS evaluation systems is shown in
Figure 6-1. The evaluation system combined analysis and measurement techniques in
order to effectively evaluate network QoS. The evaluation system was a combination of
FCM algorithm and regression model or a combination of supervised neural network
(i.e. MLP architecture) and unsupervised neural networks (i.e. Kohonen network). As
shown in Figure 6-1, the QoS parameters (i.e. delay, jitter, and packet loss ratio) were
obtained from the simulated network. The extracted QoS parameters were then used as

107
Chapter 6 Evaluation of INetworK quality 01 service

inputs to the FCM clustering algorithm or to the Kohonen network to be analysed and
organised into correlated groups. The regression model and MLP relied on the identified
groups of QoS parameters generated by FCM or Kohonen network to assess the overall
QoS. The following subsections explain how the aforementioned mechanisms were
developed to perform the QoS evaluation process.

< s>
<£>

Delay Jitter Packet loss ratio

Analyse and Classify QoS parameters using:

1. Fuzzy C-mean clustering algorithm (FCM)
2. Kohonen neural network

Classes of QoS
parameters:
Class 1
Class 2

Class i

Calculate the overall QoS assessment for class 1, class2 ,.. and class i using
1.Regression Model
2. Multi-Layer Perceptron (MLP)

Outcomes:
Assessed QoS for class 1, class2 ... and class i

Figure 6-1. QoS evaluation system.

108
Chapter 6 Evaluation of Network Quality oi service

6.3.1 Analysis of QoS using Fuzzy C-means Clustering

Algorithm

FCM was developed to partition the QoS parameters (i.e. delay, jitter, and packet loss
ratio) of critical real-time applications (i.e. VoIP and video applications) into clusters.
This classification provided an informative view about multimedia traffic behaviour and
consequently discovered valuable information from ongoing traffic.

The values of delay, jitter, and packet loss ratio were measured respectively using
equations (2.2), (2.4), and (2.6) which can be found in section 2.2.2, chapter 2. The
measured QoS parameters values were then represented by a matrix ( Q o S _ P ) to be used
as inputs to FCM algorithm as shown in equation (6.1)

D 1 J i P L R -l

D 2 J2 P L R 2
QoSP = (6 . 1)
P n Jn P L R n .

where Q o S _ P is the matrix of QoS parameters, DjJj, and P L R j , j = 1,2, ...,n are the

measured delay, jitter, and packet loss ratio respectively. In this study, FCM was
employed at predefined time interval. FCM operated on the matrix ( Q o S _ P ) and
minimised the FCM objective function given in equation (6.2) in order to partition
(Q o S _ P ) matrix into ( C ) clusters, generate membership matrix (U), and produce clusters
centres (V ).

c n

J (QoS.P; U , V ) = 2 , 2 , f r y ) " D‘i (6-2)

i=l j =1

The exponent value for partition (i.e. m) controls the degree of fuzziness for the
membership of the clusters (Chen et al, 2009). This is commonly set to 2 as other values
either introduce insufficient or too much fuzziness.

During the clustering process as explained in section 2.4.3 (chapter 2), the elements of
U were updated using Equation (2.17), the clusters centres V = {v lf v 2, —, v c) were
calculated according to Equation (2.18), and the Euclidian distance Dfj between

D j,J j, and P L R j to the centre v t of cluster i were calculated using equation (2.19). The

clustering process was terminated when the maximum number of iteration was
performed or the objective function improvement between two consecutive iterations
was less than the minimum set amount of improvement.
109
Chapter 6 Evaluation of Network Quality of Service

In this study, the maximum number of iterations was 200, and the minimum set amount
of improvement was 10'5. These parameters were chosen experimentally, i.e. different
values were experimented to monitor the FCM clustering response and the best values
were selected.

The generated membership matrix U contained values indicating the degree of

membership between vectors Dj J j, PLRj and cluster Q . The generated membership

matrix U expressed as

Fdi M71 Fplr i

Pd2 P/2 V-PLR2
U= (6.3)
Pdn P/'n \lPLRn

The number of clusters C was chosen based on the Xie-Beni index cluster validity
method (Xie and Beni, 1991). The function of Xie-Beni index method is defined by

nmintjWvi-Vjf

In this study, the optimal number of clusters for VoIP and video traffic was three,
associated with small Xie-Beni index S=0.0002 for VoIP, and S=0.0004 for video
traffic. The FCM algorithm produced three clusters classifying the QoS parameters into
three categories: Low, Medium, and High. These clusters were represented by clusters
centres matrix V as expressed as

D ei J c i P L R ci

Clusters c ent res (P) = D C2 Jc 2 P L R c2 (6.5)

D C3 Jc 3 P L R C3 .

where Dci,Jci, PLRci , i = 1,2,3 are the cluster centres of delay, jitter, and packet loss
ratio respectively.

6.3.2 Analysis of QoS using Kohonen Neural Network

Kohonen neural network (i.e. Self Organising Map SOM) with size of 100 neurons (i.e.
a 10 by 10 neuron grid) as shown in Figure 6-2 was trained to partition the QoS
parameters of multimedia applications (i.e. VoIP and video) into their correlated groups.
The aim of partition using SOM was to provide information about multimedia traffics
behaviour and their inherent groups of QoS.

110
C h a p te r 6 ii valuation 01 ivetworK quality 01 service

OOOOOOOOOO — A neighbourhood neuron

oooooeeooo
OOOOOOOOOO —- The winning neuron
0000000000 Wij

OOOOOOOOOO X Dj Jj PLRj
OOOOOOOOOO
OOOOOOOOOO D 2 J2 P L R 2

OOOOOOOOOO
OOOOOOOOOO D i h P L R i

OOOOOOOOOO
Kohonen map A neighbourhood region

Figure 6-2. Kohonen neural network.

Prior to the training process, the values of QoS parameters for VoIP or video application
were pre-processed and arranged to form examples to train the Kohonen neural
network. The values of QoS parameters (delay, jitter, and packet loss ratio) were
measured using equations (2.2), (2.4), and (2.6). The training examples considered the
classification of QoS parameters of VoIP and video applications listed in Table 2-1 (see
chapter 2). The QoS parameters of VoIP (i.e. delay, jitter, and packet loss ratio) were
classified into: Low (i.e. delay < 1 5 0 ms, jitter < 1 ms, and packet loss ratio < 2%),
Medium (i.e. 150 < delay < 400 ms, 1 < jitter < 3 ms, and 2% < packet loss ratio < 4%),
and High (i.e. delay > 400 ms, jitter > 3 ms, and packet loss ratio > 4%), whereas the
QoS parameters of video application were classified into: Low (i.e. delay < 1 5 0 ms,
jitter < 10 ms, and packet loss ratio < 1%), Medium (i.e. 150 < delay < 400 ms, 10 <
jitter < 20 ms, and 1% < packet loss ratio < 2%), and High (i.e. delay > 400 ms, jitter >
20 ms, and packet loss ratio > 2%). After considering the classification o f QoS
parameters and during the arrangement phase, the QoS parameters (delay, jitter, and
packet loss ratio) were labelled to provide visual differentiation between the generated
QoS classes. Label (L) indicated Low range for delay, jitter, and packet loss ratio. Label
(M) indicated Medium range for delay, jitter, and packet loss ratio, whereas label (H)
indicated High range for delay, jitter, or packet loss ratio. The values of QoS parameters
and their labels were represented in a matrix notation as shown equation (6.6) in order
to train the Kohonen network.

111
Chapter 6 E v a lu a tio n o t INetworK q u a l ity 01 s e rv ic e

D1 J-l PLRi Li
^ 2 J2 P L R 2 L>2
QoS p a r a m e t r e s = (6.6)
f i n Jn PLR n Ln .
where Di,Jit PLRi Lit i = 1,2,..., n were respectively the measured delay, jitter, packet
loss ratio, and their labels (i.e. L,M, or H).

During the training process, the ith input feature of delay, jitter, or packet loss ratio was
connected to j th neuron in Kohonen map. Each connection was associated with weight
Wij. The network learned by determining the Euclidean distance dj between input

patterns with N elements and the connection weights vtfy for j th neuron as explained in

Equation (2.23) (see chapter 2). The neuron where its associated weights provided the
smallest Euclidean distance to the input pattern was considered as the winning neuron.
The weights associated with the winning neuron were then updated according to the
Kohonen learning rule represented by Equation (2.24). This ensured that the winning
neuron's weights iteratively moved closer to the specific input pattern categoiy.
However, the weights associated with a number of neurons around the winning neuron
were also updated to a lesser extent. This allowed improved training.

In this study, the maximum number of training iterations was 1000. According to
(Vesanto et al, 1999), the number of training steps should be at least 10 times of the
number of neurons. The initial learning rate of learning Kohonen algorithm rj was set to

(0.5). This value was decreased inversely proportional to the number of iterations
(Haykin, 1999). This allowed the training to be course to start with and then it became
gradually finer. The analysis process of Kohonen network produced three distinct groups
that correlated with low, medium, and high QoS.

During the test phase, the trained Kohonen neural network was fed with other values of
QoS parameters in order to classify the QoS parameters into their correlated groups. The
groups produced by the Kohonen network provided information about the relationships
between different QoS parameters of transmitted VoIP or video traffics and subsequently
discovered relevant information about network operation.

6.3.3 QoS Assessment using Regression Model

Due to the relevance between the quality of multimedia applications and their QoS
parameters (delay, jitter, and packet loss ratio) as shown in Table 2-1 (see Chapter 2),
regression model was devised in this study to combine the QoS parameters of VoIP, and
C h a p te r 6 E v a lu a tio n o r INetworK q u a l ity or s e rv ic e

video applications in order to quantify the overall QoS.

The regression model was developed by using the theory of regression analysis
discussed in Section 2.4.1, Chapter 2. The values of the independent variables
(xl t x2, x 3) in the regression model were represented by delay, jitter, and packet loss
ratio respectively, whereas the values of dependent variable (y) were represented by the
overall QoS. The regression expression was determined by considering the QoS
requirements listed in Table 2-1 in order to provide the outputs that reflect the overall
QoS. The QoS parameters shown in Table 2-1 were categorised as: Low, Medium, and
High. The overall QoS on the other hand was classified as Good, Average, and Poor
quality corresponding to the categories of QoS parameters. In this study, the overall
QoS spanned between (0-100%). QoS=0% represented the worst case of network
performance, whereas the best network performance was for QoS=100%.

The regression formula for VoIP application was calculated by arranging the values of
independent variables (i.e. delay, jitter, and packet loss ratio) and the values of
dependent variable (i.e. overall QoS) into matrices as follows: Low QoS parameters (i.e.
delay < 1 5 0 ms, jitter < 1 ms, and packet loss ratio < 2%) corresponded to good
overall QoS which ranged between (67-100%), medium QoS parameters (i.e. 150 <
delay < 400 ms, 1 < jitter < 3 ms, and 2% < packet loss ratio < 4%) corresponded to
average QoS (i.e.33% < QoS < 67%), whereas high QoS parameters (i.e. delay > 400
ms, jitter > 3 ms, and packet loss ratio > 4%) corresponded to poor QoS (i.e. QoS <
33%).

The regression formula for video application was calculated using the same mapping
used in VoIP application. However, good overall QoS which ranged between (67-
100%) corresponded to low value of delay < 150 ms, jitter < 10 ms, and packet loss
ratio < 1%), average QoS (i.e.33% < QoS < 67%) corresponded to medium QoS
parameters (i.e. 150 < delay < 400 ms, 10 < jitter < 20 ms, and 1% < packet loss ratio
< 2%), whereas high QoS parameters (i.e. delay > 400 ms, jitter > 20 ms, and packet
loss ratio > 2%) corresponded to poor QoS (i.e. QoS < 33%).

After the mapping process of QoS parameters to the overall QoS, the matrices of QoS
parameters and the overall QoS were then organised to form the regression formula as
expressed by

113
Chapter 6 e v a lu a tio n 01 JNeiworK q u a l ity 01 service

Q0 S 1 1 D± A PLR± rfcoi ei
QoS2 ~ 1 D2 J2 PLR2 bi e2
+ (6.7)
b2
jQoSn. Dn Jn P L R nm -b^. ? n.

where DiJi,PLRi,QoSi , i = 1,2, ...,n are delay, jitter, packet loss ratio, and overall
QoS of VoIP or video applications. The regression coefficients b0, b 1, b 2, b 3 were
determined from the recorded data using equation (2.12). The vector o f residual (i.e.
error terms) was then calculated using equation (2.13). In this study, the calculated
errors produced from regression formula for VoIP and video traffics were normally
distributed. This indicated that the mean of error terms ei f i = 1,2, ...,n was zero. This
implied that the estimated regression model determined was:

QoSi = b0 + b1 * Di + b2 * Ji + b3 * PLRt (6 .8)

where Q0 S1 ,Di,Ji,PLRi t i = 1 ,2 ,...,n are the overall QoS, delay, jitter, packet loss
ratio for ithpacket respectively.

6.3.4 QoS Assessment using Multi-Layer Perceptron

A Multi-Layer Perceptron (MLP) neural network was chosen to assess the overall QoS
due to its suitability and effectiveness (Abraham, 2005). The proposed MLP neural
network model in this study composed of an input layer with three neurons, a hidden
layer with three neurons, and an output layer with one neuron as shown in Figure 6-3.

The inputs A , and P L R i fed into the input layer of MLP were delay, jitter, and packet
loss ratio respectively, whereas the values of desired output (QoSi) were represented by
the overall QoS.

Input layer Hidden layer Output layer

Inputs w !4 w48 Outputs

QoS]
Q 0 S2

Q o S i

w37 w67

Figure 6-3. Structure of an MLP.

114
Chapter 6 Evaluation of Network Quality of Service

The inputs (i.e. delay, jitter, and packet loss ratio) and their desired output(i.e. QoS) were
arranged in a matrix by considering the QoS requirements listed in Table 2-1 in order to
provide outputs that reflected the overall QoS. The QoS parameters of VoIP and video
application shown in Table 2-1 were categorised into three classes: Low, Medium, and
High. The overall QoS on the other hand was classified as Good, Average, and Poor
quality corresponding to the categories of QoS parameters. The overall QoS spanned
between (0-100%).

Low QoS parameters of VoIP application (i.e. delay < 1 5 0 ms, jitter < 1 ms, and packet
loss ratio < 2%) corresponded to good overall QoS which ranged between (67-100%),
medium QoS parameters (i.e. 150 < delay < 400 ms, 1 < jitter < 3 ms, and 2% <
packet loss ratio < 4%) corresponded to average QoS (i.e.33% < QoS < 67%), and
high QoS parameters (i.e. delay > 400 ms, jitter > 3 ms, and packet loss ratio > 4%)
corresponded to poor QoS (i.e. QoS < 33%). Whereas in video application, good
overall QoS which ranged between (67%-100%) corresponded to low value of delay <
150 ms, jitter < 10 ms, and packet loss ratio < 1%), average QoS (i.e.33% < QoS <
67%) corresponded to medium QoS parameters (i.e. 150 < delay < 400 ms, 10 < jitter
< 20 ms, and 1% < packet loss ratio < 2%), and high QoS parameters (i.e. delay > 400
ms, jitter > 20 ms, and packet loss ratio > 2%) corresponded to poor QoS (i.e. QoS <
33%).

The matrix expressed in equation (6.9) which included training examples was fed into
MLP for its training. The matrix represented the QoS parameters and the overall QoS
where D iJ ifPLRi,QoSi , i = 1,2, ...,n were delay, jitter, packet loss ratio, and the
overall QoS respectively:

D± Ji PLRi QoSi
^2 J2 PLR2 Q0 S 2
Training examples = (6.9)
&n Jn PLRn QoSn.

During the training phase of MLP, the inputs (Di,Jit PLRi) were multiplied with their
associated weights (Wi) and the resulting values are summed by the summation function
using equation (2.21) (see Chapter 2). The output from summation function (5 ) was then
processed by the activation function to produce the output (y). In this study, the
hyperbolic tangent activation function was used as it gives continuous output between -
1 and +1 and thus it gave a larger range than sigmoid activation function which provides

115
Chapter 6 Evaluation of Network Quality of Service

a range between 0 and 1 (Karlik and Olgac, 2010). Equation (6.10) formulates the
hyperbolic tangent function:

(6. 10)

The calculated output (p(s) (i.e. y) was subtracted from the desired output (QoSi) as *n
equation (2.22) (See Chapter 2) to produce an error (e) which in turn was used by the
learning algorithm in order to reduce the magnitude of the error in the next iteration.

In this study, the commonly used a gradient descent with momentum learning algorithm
was employed. The function of this algorithm was to use the calculated error (e) and the
input data (DiJit PLRi) to the processing neuron to adjust the values of the connections'
weights (W/) which in turn reduced the magnitude of the error in the following training
iteration. Gradient descent with momentum learning algorithm updated the weights using
equation (6.11) (Eberhart and Dobbins, 1990):

Wnew ” W0id “h V 00* + ft [A VP0^] (6 . 11)

where Wnew is the new updated weight, W0id is the previous weight, the term rj is the

learning rate parameter, (e) is the calculated error, x is the input to the processing
element, the momentum term ( a ) ensures that the learning algorithm does not get stuck
in a local minimum and finds the desired global minimum, and the term A W0id
represents the amount of change in the weights from the previous iteration.

The learning rate parameter rj is a positive constant limited to the range 0 < i} < 1

whereas the momentum factor a can take values between 0 and 1. In this study, the
learning rate parameter rj and the momentum factor a were set to their default values

which were 0.01 and 0.9 respectively. These values provided the most reliable results.

The training process of MLP was terminated when the maximum number of iterations
reached (in this study was 1000) or when there was insignificant error (i.e. 0.001)
between network output (y) and desired result (QoSi). During the test phase, the trained
MLP was fed with other values of QoS parameters (D*,/*, PLRi) in order to assess the
overall QoS. The actual output values used to train MLP and the MLP output values
during test phase were then correlated to ensure that MLP had been trained effectively.

6.3.5 Measuring Predication Accuracy

The prediction accuracy of proposed QoS assessment methods was measured using a
correlation coefficient. The correlation coefficient (R) is widely used to evaluate the

116
C h a p te r 6 E v a lu a tio n o t INetworK q u a lity 01 s e rv ic e

prediction accuracy (Chatteijee and Hadi, 2006). The magnitude of R is between 0 and
1. The magnitude closest to 1 indicates a perfect correlation, whereas a correlation less
than 0.5 would be described as weak correlation. The correlation between the actual
values (yA) and predicted values (y P) is calculated using equation (6.12).

n _ Z C y A i-W Cy pi -y p)
y A ' yp V S (y /ii-5M)2 Z ( y p i- y F )2

where y^is the actual value, Ya ls the mean of the actual values, y Pis the predicted
value, and y k is the mean of the predicted values.

6.3.6 Network Simulation and Traffic Models

A wireless-cum-wired network topology illustrated in Figure 4-3 was simulated using

Network Simulator- 2 (NS-2). The network topology consisted of 10 wireless nodes, 10
wired nodes, and 1 base station to form 10 unidirectional connections. At the wired side
of the network, all the links had 5 Mbps bandwidth and 2 ms propagation delay. The
queue management mechanism was Drop-Tail and the queue size was 50 packets. The
WLAN side of the network was based on IEEE 802.1 le, and it used the Enhanced
Distributed Channel Access (EDCA) scheme. The main parameters that modelled the
wireless channel were the default settings for IEEE 802.1 le as shown in Table 4-1.

The types of traffic transmitted over the simulated network were: VoIP, video streaming,
best effort and background. VoIP was modelled as G.711 voice encoding scheme by
adapting Constant Bit Rate (CBR) traffic. The packet size for VoIP was 160 bytes and
the transmission rate was 64 kbps. The video streaming source was YUV QCIF (176 X
144) Highway (2000 frame) (YUV QCIF, 2012). Prior to its transmission, each video
frame was fragmented into packets that in turn had a maximum length of 1024 bytes.
The best-effort traffic had a fixed packet size of 1000 bytes and 125 kbps transmission
rate. File Transfer Protocol (FTP) application was used for the background traffic. FTP
was transmitted over TCP, whereas other traffics were transmitted using UDP transport
protocol. The transmitted traffic over IEEE 802.l i e EDCA (i.e. VoIP, video, best effort
traffic, and background traffic) were mapped into different access categories to represent
different priority levels, as shown in Table 4-2. VoIP had highest priority, whereas the
priority of the background traffic was lowest.

The simulation time was 500 seconds. During the first third of the simulation, two VoIP
applications, and two best effort traffics were transmitted. During the period (170s-
117
Chapter 6 n,valuation 01 i\eiw orK i^uamy 01 service

340s), the number of transmitted applications increased to three VoIP, two video, two
best effort traffics, and background traffic making the network load heavy and
approximately 100% of channel capacity was used. However, after the transmission of
video traffic was completed, the network load became light during the last third of
simulation time.

6.4 Results and Discussion

This section is divided into four subsections: two sections provide QoS analysis results
for Fuzzy C-Means and Kohonen Neural Network. The remaining two sections provide
QoS assessment results using Regression Model and Multi-Layer Perceptron.

6.4.1 QoS Analysis using FCM Clustering Algorithm

In this study, FCM clustering algorithm was applied at predefined regular time intervals
to analyse VoIP and video traffic. FCM analysis results for VoIP traffic during the
simulation time interval 350 - 400 seconds are shown in Figures 6-4. The values of QoS
parameters (i.e. delay, jitter, and packet loss ratio) of VoIP were grouped into three
clusters, representing Low, Medium, and High values. Each cluster was represented by
its own centre. Figure 6-4 shows that the packet loss ratio of VoIP is zero. This is
because VoIP was mapped to the access category that had the highest priority.
However, due to the heavy network load, a number of VoIP packets experienced
fluctuated delay that in turn resulted in high jitter values.

_□ _ _
5

i*
-
ts 3 Centers of clustefc

2 G Clusterl
n Cluster2
0 Cluster3
1 © Centerl
□ Centerl
© © Q Centerl

°10 15 20 25 30 35 40 45 50 55
Delay

Figure 6-4. Clustering QoS parameters of VoIP application at predefined time interval.
118
cnapier o evaluation 01 i>eiworK ^mamy 01 service

FCM analysis result for video traffic during the simulation time interval 200 - 250s is
presented is presented in Figure 6-5. Due to the mapping of video application to the
Access Category (AC) that had a lower priority as compared with the (AC) assigned to
VoIP, and the heavy load on the network, the QoS parameters (i.e. delay, jitter, and
packet loss ratio) varied extensively. These variations resulted in three clusters
representing Low, Medium, and High QoS parameters. During this interval, the video
quality was affected by the high values of QoS parameters in cluster 3. The centre of
cluster 3 characterised high delay, high jitter, and medium packet loss ratio which were
respectively 540.6 ms, 24.4 ms, and 1.2%.

2.5 O Cluster!
D Cluster2
0 Cluster3
, Centers of dusters 0 Centerl
1 Center 2
^ Center 3

-<r
-O ' '

600
500
400
-o- 300
200
100

Jitter (ms) Delay (ms)

Figure 6-5. Clustering QoS parameters of video application at predefined time interval.

The fuzzy partition matrix (produced by FCM algorithm) indicated the degree of
membership of each QoS parameter to each cluster. Figures 6-6 (a) - (b) show
respectively the degree of membership for a sample of delay for VoIP, video
application, and their clusters’ centres. As indicated in the Figures, each value had
different degree of membership to the three clusters between 0 and 1 with the total
being 1. In this study, the fuzzy partition feature of FCM made it a valuable clustering
tool because the characteristics of QoS parameters did not allow crisp (binary) partition.

119
Chapter h evaluation 01 iveiworK i^uaiuy ui service

H C l u s te r l
E S I C lu s te r2
I i C lu s te r3

S a m p le d v a lu e s o f d e la y

(a)
WMC l u s te r l
H W C la s te r 2
1 IC lu s te r 3

S a m p le d v a lu e s o f d e la y

(b)

Figure 6-6. The degree of membership between a sample of delay and the cluster's
center for: (a) VoIP, and (b) Video application.

During the clustering process of VoIP and video application, the objective function
indicated the progress of FCM clustering algorithm over the number of iterations
performed. This is shown inFigures 6-7 (a) - (b). The clustering process of FCM
stopped either when it reached the maximum number of iterations orwhen the objective
function improvement between two consecutive iterations was less than the predefined
120
L^napier o u vaiuiiuu ii ui n ciw u iK v^uamy ui cjci vice

minimum amount of improvement. From Figures 6-7 (a) - (b), it can be noticed that the
clustering process of VoIP and video application terminated when the objective function
improvement between two consecutive iterations was less than le-5 (predefined
minimum improvement) although the maximum number of iterations was set as 200.

,x i ol

c
1s
£
Ic

Iteration Count

(a)

Iteration Count

(b)
Figure 6-7. The progress of objective functions during FCM analysis of: (a) VoIP, (b)
Video application.

Figure 6-8 illustrates the cluster centres for 50 seconds time intervals during
transmission of VoIP. The FCM algorithm classified the QoS parameters of VoIP into
three levels at each time interval. The Figure shows that VoIP started its transmission
approximately at 170s of the simulation time. During that time, there were two other
VoIP applications using the channel since their transmission started during the first third
of the simulation (i.e. Os-170s). Therefore, during the time interval 170s-185s, the

121
v^iiapicr o LvYdiuciuuii u i n c m u i i\ y u a m j u i j c i v it c

values of delay, jitter, and packet loss ratio of VoIP were at medium range of QoS
parameters. As the number of transmitted applications increased to three VoIP, two
video applications, two best effort traffics, and background traffic during the time
interval 215s - 275s, the load became heavy. This in turn made the network incapable of
meeting minimum QoS requirements for a VoIP application. The values of QoS
parameters for cluster 3 were in the high range, 500 ms for delay, 5 ms for jitter, and 6%
for packet loss ratio, indicating a poor quality VoIP. However, once the two video
applications were transmitted during 240s - 260s, the network load became light. This
allowed the network to meet the QoS requirements of the VoIP application during the
time interval 275s - 500s. The values of delay, jitter, and packet loss ratio were in the
low range of QoS parameters.

Simulation Time (Sec)

C enters or C lusters
Figure 6-8. Clustering QoS parameters of VoIP.

The FCM algorithm classified the QoS parameters of video application into three levels
for each time interval (i.e. 5 seconds) as shown in Figure 6-9. Video application
transmission started approximately at 170s and terminated at around 240s. Figure 6-9
shows that the majority of QoS parameters values for video application were in the high
range, particularly in cluster 3, around 550 ms for delay, 30 ms for jitter, and 3% for
packet loss ratio during the simulation time 170s - 230s. The reasons for this were: (i)
122
c iid p ie r u £ iV iiiu aiiu ii u i i i c m u i i v y u a m j u i i j t i y iv t

the allocation of the video application to an access category which had a lower priority
than the access category that VoIP was assigned (ii) the network load was heavy due to
the transmission of multiple applications during the same simulation time interval.
Consequently, the network was incapable of meeting the QoS requirements for the
video application.

600

500
o

100

170
175
180
185
190
195
200
205
210 Lc3
215 Jc3
D c3
220 Lc2
225 Jc2
230 Jcl
Lcl D e l
Simulation Time (Sec) D el
Centers of Clusters

Figure 6-9. Clustering QoS parameters of video.

6.4.2 QoS Assessment using Regression Model:

In this section, the developed regression model was used to combine the QoS
parameters (i.e. delay, jitter, and packet loss ratio) for each centre of the generated
clusters by FCM in order to estimate the overall QoS. Figure 6-10 shows the results
from the devised regression model. The results show the predicted QoS of VoIP for the
centres of generated clusters at each 50 seconds time interval. Figure 6-10 shows that
the QoS values reflected the corresponding QoS parameters indicated in Figure 6-8. In
other words, as the values of QoS parameters decreased, the values of overall QoS
increased accordingly. When VoIP started its transmission at 170s, its overall quality
was poor (i.e. below 30%). This is because of its contention to access the channel which
carried other VoIP applications with the same priority. As the number of transmitted

123
L^napier o j ^ v a i u a u u u u i i i c i w u i iv v £ u a ii i j u i o c i t iv c

applications increased to three VoIP, two videos, two best effort traffics, and
background traffic during the time interval 215s - 335s, the VoIP quality degraded by
overall 51.83% as compared with its quality of 55.97% during the period 1 8 5 -2 1 5 . The
heavy load on the network made its performance incapable of meeting minimum QoS
requirements for VoIP application. However, during the last third of simulation time,
the network load became light due to the termination of two video applications. The
VoIP quality increased sharply at certain time interval. For example, VoIP quality
reached 87.9% at time interval 365s - 395s.

H Q oS of center 1
I IQoS of center 2
1 IQoS of center 3

Simulation Time (Sec)

Figure 6-10. The QoS of VoIP using regression model

The developed regression model was also used to assess the QoS of video application
for each centre of the FCM generated clusters shown in Figure 6-9 by combining the
QoS parameters: delay, jitter, and packet loss ratio. Figure 6-11 shows the predicted
QoS of the video application for the centres of generated clusters at each 5 seconds time
interval. The Figure shows that the video quality was good at 170s. This is because the
transmission of video started prior to other traffic such as VoIP, other video application,
and background traffic during the second third of simulation time (170s - 340s).
However, due to the low priority of video as compared with VoIP and the load of the
network became heavy, the quality of video degraded sharply and its QoS range
remained between poor and average level (i.e. 3.04% - 73.05%).

124
L.napter o CiVillUilllUH u i l i c m u i i v y u a m j u i u v i > iw

f 50

170 175 180 185 190 195 200 205 220 225 230
Simulation Time (Sec)

Figure 6-11. The QoS of video using regression model

The results obtained from the devised regression model were compared using other QoS
assessment methods. These were: fuzzy inference system mechanism, and distance
measurement evaluation system reported in (Al-Sbou, 2010(a)) and (Al-Sbou, 2006(b)).
Tables 6-1 and 6-2 show respectively the comparison between the aforementioned
assessment methods used to quantify the overall QoS of VoIP, and video application.
From the sampled values of evaluated QoS provided in Tables 6-1 and 6-2, it is
indicated that the three assessment methods provided results which were comparable.
The values of the correlation coefficients between the QoS determined using regression
model, and the other QoS assessment methods (i.e. fuzzy inference system mechanism,
and distance measurement evaluation system) were respectively: 0.95 and, 0.97 for
VoIP application, and 0.89, and 0.93 for video application. Although some outputs were
slightly different, they were still in the same region (i.e. poor, average, or good). The
discrepancies were due to the fact that each method followed a different operation.

However, the values of QoS obtained from devised regression model spanned between
(0%-100%), whereas the range of QoS values produced by FIS was between (10%-
90%). This indicates that the devised regression model provides more accurate results.

125
cnapier o H r V itiu a u u ii u i m i n u i i\ y u n u i y u i o c i * n -c

Table 6-1. QoS parameters of VoIP and expected QoS using: fuzzy inference system,
distance measurement, and regression model.
QoS parameters Assessment of overall QoS using:

Fuzzy
Delay Jitter Packet loss Distance Regression
inference
(ms) (ms) ratio (%) measurement model
system
36.69 4.62 0.00 9.64 29.69 29.78
108.84 5.00 0.00 9.28 24.13 21.92
599.94 5.00 6.00 9.28 0.01 4.31
24.21 2.87 0.00 71.50 55.79 64.14

54.76 4.89 0.18 9.31 25.71 23.68

101.20 4.83 0.00 9.34 26.49 25.29
16.14 1.66 0.00 81.97 95.40 87.90 |
24.01 3.60 0.00 36.49 44.45 49.84
46.32 4.97 0.00 9.28 24.72 22.87
17.48 1.87 0.00 82.96 93.93 83.77
25.38 3.96 0.00 18.60 39.18 42.78
46.24 4.65 0.00 9.57 29.20 29.14

Table 6-2. QoS parameters of video and expected QoS using: fuzzy inference system,
distance measurement, and regression model.
QoS parameters Assessment of overall QoS using:
Fuzzy
Delay Jitter Packet loss Distance Regression
inference
(ms) (ms) ratio (%) measurement model
system
57.15 13.03 0.00 80.42 92.69 82.29
85.37 24.52 2.61 13.81 24.90 40.72
289.30 25.64 1.38 12.22 32.19 33.59
534.18 24.00 1.06 10.11 20.25 15.69
600.00 29.55 1.16 9.29 8.30 3.04

599.95 9.84 0.87 9.39 17.88 25.44

599.98 29.10 0.78 9.32 8.87 7.51
367.18 17.90 0.34 42.56 43.11 44.92
492.40 24.09 0.59 11.91 23.97 24.49
597.01 16.90 0.93 9.40 17.70 18.16
600.00 5.89 j 1.65 9.39 17.54 21.04
600.00 17.27 0.96 9.39 17.28 17.20

126
Chapter o n .v u iu a ii u n u i ix c i w u i k y u a u i j ' u i a c i v itc

6.4.3 QoS Analysis using Kohonen Neural Network

Kohonen neural network (i.e. Self-Organising Map SOM) processed the QoS
parameters (i.e. delay, jitter, and packet loss ratio) of VoIP, and video applications into
correlated groups. The values of delay, jitter, and packet loss ratio of transmitted VoIP
and video were grouped into three classes representing Low, Medium, and High values
as shown in Figures 6-12 and 6-13. Low values of delay, jitter, and packet loss ratio are
represented in blue, Medium values are represented in green, and High values are
represented in red.

Figures 6-12 (a), (c) show that low delay, and packet loss ratio activate most neurons in
SOM. This was due to VoIP having the highest priority to access the channel amongst
other transmitted traffic such as video, and best effort traffic. However, the heavy load
on the network and the contention with other VoIP applications having the same priority
caused high fluctuation to the delay. This in turn generated high values of jitter as
shown in Figure 6-12 (b). The high values of jitter would have a negative effect on the
overall VoIP quality.

The labels of the QoS parameters in Figure 6-12 (d) provided a differentiation between
the three generated QoS classes. Label L identified the active neurons which represent
Low values of QoS parameters; label M covered the region where Medium values of
QoS parameters were located; whereas High QoS parameters were represented by other
neurons with label H. The groupings produced by the Kohonen network provided
information about the relationships between the different QoS parameters of transmitted
VoIP traffic and subsequently discovered how VoIP was treated as high priority traffic
transmitted over heavy loaded network.

Figure 6-12 (e) shows the response of a Kohonen neural network to the QoS parameters
of VoIP. SOM differentiated delay, jitter, and packet loss ratio into almost three
clusters. The Figure shows that Low and Medium QoS parameters have well defined
regions. The winning neurons for these regions were characterised by small number of
QoS values. In contrast, the neighbourhood region of high QoS parameters was large.
This was because most jitter values for VoIP were high .This in turn had an impact on
VoIP quality, although other values such as packet loss ratio were 0% during most of
VoIP transmission period. The winning neuron for high QoS parameters region was
activated by 69 values.

127
VyliapiVl A VI T1VV

Delay (ms) Jitter (ms)

593

306 2 .7 3

1 8 .6 0 .4 5 6

(a) (b)
Packet loss ratio (%)
5 .8 1

2 .9

(c) (d)
Medium QoS parameters

High QoS ° ■ Low QoS

param eters [32T5To [ 4114 If 3 ifllU g 0 0 param eters

m m f* ° o
0 0 0 0 0 3
0 0 0 0 0 0 0
o 0 0 0
5}(7) ® (§) <2 0 (g)

(e)
Figure 6-12. Classifying QoS parameters of VoIP using kohonen network: (a) Delay,
(b) Jitter, (c) Packet loss ratio, (d) QoS parameters labels, (e) SOM sample hits.

The analysis of SOM for QoS parameters of video application is shown in Figure 6-13.
During its transmission, the video application experienced high delay, jitter, and packet
loss ratio as indicated in Figures 6-13 (a) - (c). The network was incapable of meeting
the QoS requirements for video application. This was because of the heavy load on the
network as multiple applications were being transmitted at the same simulation time
interval. Moreover, the video application was allocated to a lower priority access
category compared with VoIP allocation.

Although some individual values of delay, jitter, and packet loss ratio were within an
acceptable level of quality, high values of one the aforementioned parameters would
have a negative effect on the overall quality of video. The labels of QoS parameters and
the response of SOM to the QoS parameters of video application shown in Figures 6-13

128
(d) and (e) respectively provided a differentiation between the generated QoS classes.
Both figures indicated that Low QoS parameters identified by label L covered a very
small region of SOM. Two neurons were activated by low QoS values and the winning
neuron contained 16 values. The region of Medium values of QoS parameters which
had label M was large as compared with Low QoS parameters region. The
neighbourhood region of high QoS parameters represented by label H covered most of
the Kohonen map. Figure 6-13 (e) shows that high QoS parameters had four sub-
regions. The number of values activated the winning neurons for these regions were
relatively large (i.e. 121, 56, 19, and 8). It can be concluded from Figure 6-13 that high
value of QoS parameters of video application: delay, jitter, or packet loss ratio could
have negative impact on its overall quality. The analysis of SOM demonstrated how the
simulated network could manage and treat different applications with different priorities
during light and heavy loads.

D ela y (m s) Jitter (ms) 2 9 .9

599

(a) 342 (b) 1 6 .4

8 5 .1
2 .9 3

Packet loss ratio ( % )

(c) (d)

L o w Q oS p aram eters

M ed iu m ^ ^
Q oS ^ m
p aram eters W
0 O’
(e) @0
® o
0 CD
ZlSD@®

H ig h Q oS p aram eters

Figure 6-13. Classifying QoS parameters of video application using kohonen network:
(a) Delay, (b) Jitter, (c) Packet loss ratio, (d) QoS parameters labels, (e) SOM sample
hits.
t^napier o j&vaiuauuii ui i^ eiw o m v^uauiy ui o erv icc

6.4.4 QoS Assessment using Multi Layer Perceptron

After the QoS parameters of VoIP and video application were processed by Kohonen
neural network, the average of QoS parameters (i.e. delay, jitter, and packet loss ratio)
activated a winning neuron and a number of neurons in the neighbourhood region as
shown in Figures 6-12 and 6-13 were combined by MLP to assess the overall QoS. The
MLP was required to be trained correctly in order to assess the QoS in an effective
manner. Figures 6-14 (a) and (b) show respectively the training process of the MLP to
assess the QoS of VoIP and video application. The training process of MLP was
terminated either when the maximum number of iterations was reached or when there
was insignificant error (i.e. 0.001) between MLP outputs and desired results. In this
study, the training process of MLP in case of QoS assessment of VoIP or video
application was terminated when the maximum number of training iterations (i.e. 1000)
was reached. The value of maximum number of training iterations was chosen
experimentally, i.e. different values were tested and the MLP response was monitored.
It is also indicated from the Figures that the Mean Squared Error (MSE) values for MLP
when the training process terminated to assess the QoS of VoIP and video were
respectively 0.003 and 0.005. These values indicated an insignificant error between
MLP outputs and desired results.

Train I
— Best |

•gio'

£i<f

0 100 200 300 400 500 (00 700 800 900 1000
1000 Epochs

— Train
Best

ElO

0 100 200 300 400 500 600 700 800 900 1000
1000 Epochs
Figure 6-14. The progress of training MLP to assess the QoS for: (a) VoIP, (b) Video.
130
^ iia p ier o r.vaiuauoii ui n eiw orK v£u a ,,lJ U1 vice

Figures 6-15 (a) and (b) show respectively the comparison between normalised actual
QoS and calculated outputs from MLP during its training to assess QoS of VoIP and
video. It can be perceived from the Figures that the actual output values used to train
MLP and the output values following the termination of its training were strongly
correlated. The correlation coefficients were 0.997 when MLP trained to assess QoS of
VoIP and 0.996 when MLP trained to assess QoS of video application. The correlation
coefficient values indicated that the MLP had been trained effectively.

Training: R=0.99673

O.
O.
o.
« o.

-o.
-o.
-o.

Target

(a)
Training: R=0.99535

Target

(b)
Figure 6-15. Comparison between normalised actual QoS and calculated outputs from
MLP in case of: (a) VoIP QoS assessment, and (b) Video QoS assessment.

131
so ap ier o

The results in Figures 6-16 (a) and (b) show respectively the predicted QoS of VoIP and
video applications obtained from MLP during the test phase. From the Figures, it can be
observed that the range of QoS values reflected the QoS parameters regions shown in
Figures 6-12 (e) and 6-13 (e). In other words, Good QoS reflects the region of Low QoS
parameters, whereas High QoS parameters region produced Poor QoS. Due to its
priority over video application, the average estimated QoS of VoIP for Low, Medium,
and High QoS parameters regions was elevated by 56.01% as compared with the
average estimated QoS of video application. However, the heavy load on the network at
certain time intervals made its performance incapable of meeting the minimum QoS
requirements for VoIP and video applications.

90
• Good QoS
■ Average QoS
80 ♦ Poor QoS

70

60

50
40

30

20

lO

8 lO 12 14 16 18 20
No. of Neuron

100
• P oor Q oS
■ A v era g e Q oS
90 ♦ G ood Q oS

80

70

60

50
40

30

20 10 11 12 14
N o . o f N eu ro n

Figure 6-16. The QoS assessment of: (a) VoIP, and (b) Video application using Multi-
Layer Perceptron.

The results of expected QoS for VoIP and video obtained using MLP were compared
with other QoS assessment methods such as the developed fuzzy inference system FIS
(Al-Sbou, 2 0 1 0 ( a ) ) , Euclidean distance measurement system (Al-Sbou, 2 0 0 6 ( b ) ) , and

132
^napier n jj>yaiuauuu ui n c in u m ^ u a n tj v i uvi »iw

regression model (Dogman et al, 2012)) as shown in Tables 6-3 and 6-4 respectively.
The values of correlation coefficients between the QoS of VoIP determined using multi­
layer perceptron neural networks, and the other QoS assessment methods were: 0.90 for
FIS, 0.99 for distance measure, and 0.98 for the regression model. Whereas correlation
coefficients between the QoS of video determined using MLP and other QoS
assessment techniques were: 0.97 for FIS, 0.97 for distance measure, and 0.99 for
regression model. From the values of correlation coefficient and the values of evaluated
QoS provided in Table 6-3 and 6-4, it was concluded that the aforementioned QoS
assessment techniques for VoIP and video application provided results which were
closely comparable. Although some outputs were slightly different, they were in the
same QoS region. The discrepancies were due to the fact that each method followed a
different process. However, the values of QoS obtained from the MLP architecture
ranged from (1% - 100%), whereas the range of QoS values produced by the FIS was
between (10%-90%). This indicates that the MLP was more effective.

Table 6-3. QoS parameters of VoIP and expected QoS using: Fuzzy Inference System
(FIS), Distance Measurement (DM), Regression Model (RM), Multi-Layer Perceptron

QoS parameters Assessment of overall QoS using:

Packet Loss
Delay (ms) Jitter (ms) FIS DM RM MLP
ratio (%)
55.44 4.59 0.00 9.73 29.98 29.29 27.41
| 35.36 4.58 0.00 9.76 30.26 30.63 29.16 j
180.85 5.00 1.25 9.28 23.94 21.90 26.12
35.23 2.26 0.00 81.98 67.77 69.95 72.98
37.96 3.18 0.00 60.30 50.42 54.20 51.68
20.28 2.41 0.00 80.39 65.44 68.27 66.29
71.67 2.99 0.00 67.51 52.66 55.46 52.89
24.24 2.76 0.00 74.18 57.69 62.11 56.44

52.51 2.78 0.00 73.58 56.26 60.13 55.92

77.99 4.06 0.00 16.91 37.29 36.97 29.67
27.64 0.81 0.00 89.33 95.29 94.96 97.05
18.77 1.70 0.00 82.21 94.94 80.39 96.54
21.63 0.24 0.25 90.31 96.11 98.53 97.02

24.57 1.56 0.00 83.64 94.95 82.43 96.76

133
 i ^ v a i u a u u u u i n c m u i iv y u a i u j u i ljv i ▼i w

Table 6-4. QoS parameters of video and expected QoS using: Fuzzy Inference System
(FIS), Distance Measurement (DM), Regression Model (RM), Multi-Layer Perceptron
(MLP)
QoS parameters Assessment of overall QoS using:

Packet loss
Delay (ms) Jitter (ms) FIS DM RM MLP
ratio (%)

55.37 4.92 0.00 88.41 93.82 81.95 85.58

142.73 4.02 | 0.00 87.94 93.64 73.94 83.10
264.02 12.10 0.00 54.99 90.51 57.03 64.93
320.64 14.77 0.00 48.92 83.58 49.82 54.23
249.58 11.04 0.00 57.66 91.14 59.10 68.02
402.18 10.48 0.00 42.80 46.27 44.48 50.17
304.88 8.28 0.00 62.54 87.78 55.39 65.02
600.00 29.04 0.00 9.32 10.01 13.56 11.79
336.61 14.00 3.00 10.47 21.77 27.37 31.97
380.12 16.05 3.00 10.47 20.27 21.83 19.13
516.10 5.56 3.00 10.22 13.98 14.99 21.14
595.00 2.49 3.00 9.41 8.57 9.16 14.39
600.00 7.79 3.00 9.39 7.75 5.38 7.34
599.61 29.66 3.00 9.29 0.51 8.14 5.70

6.5 Sum mary

This chapter introduced two novel Quality of Service (QoS) assessment systems. The
first system included a combination of fuzzy C-means (FCM) and regression model to
analyse and measure the QoS of VoIP and video traffic transmitted over a simulated
network. Whereas the other system used a combination of supervised and unsupervised
neural networks (i.e. Kohonen network and multi-layer perceptron (MLP)) to evaluate
the QoS of VoIP and video applications.

The QoS parameters of VoIP and video traffic were analysed by FCM and Kohonen
network. The capability and robustness of these techniques to cope with imprecise QoS
patterns made them effective clustering mechanisms for QoS analysis. FCM and
Kohonen network classified the values of QoS parameters of transmitted VoIP and
video into clusters representing Low, Medium, and High values of QoS. The regression

134
V /U c tJ J tC l U i^Ydiuaiiuu u i i i c i n u i *v y u a u i j u i a c t t i v t

model and MLP in turn combined the QoS parameters (i.e. delay, jitter, and packet loss
ratio) for each centre of generated clusters and produced a single value that represented
the overall QoS. The overall QoS was a good indication of network performance. The
overall QoS can be used to monitor the network and to avoid congestion. The values of
assessed QoS were strongly correlated to a number of previously studied QoS
assessment methods.

135
 Chapter 7 Improvements in Quality of Service in

Computer Networks

7.1 Introduction
In this chapter, a new Quality of Service (QoS) enhancement scheme for WLAN-wired
networks is developed and its performance is evaluated. The proposed scheme consists
of an adaptive Access Category (AC) traffic allocation algorithm that is incorporated
into the network’s wireless side to improve the performance o f IEEE 802.1 le Enhanced
Distributed Channel Access (EDCA) protocol, and a Weighted Round Robin (WRR)
queuing scheduling mechanism that is incorporated into the wired side of the network.
The adaptive traffic allocation algorithm determines the Packet Arrival Rate (PAR) of
up-link and down-link traffic for each AC. It then dynamically allocates the traffic of
the lower priority AC to the next higher AC, when the higher AC is not receiving traffic
at the time. On the wired side of the network, the aim of WRR is to share the network
resources based on the traffic’s quality o f service (QoS) requirements. The performance
of the proposed scheme was compared with the standard IEEE 802.1 le EDCA and
FIFO queuing mechanisms (i.e. WLAN-wired network legacy scheme). The
incorporation of the scheme improved the performance of the WLAN-wired network,
thus enhancing the QoS for transmitted applications. The scheme allowed an end-to-end
QoS to be set up which in turn provided improved delivery of a variety of applications
in the context of wired-cum-wireless networks.

This chapter is organised as follows: the relevant studies are discussed in section 7.2.
Section 7.3 introduces a detailed description of the proposed QoS enhancement scheme.
The simulated network and traffic models are presented in section 7.4. The results are
presented and discussed in section 7.5. The conclusions are provided in section 7.6.

7.2 Related Work

The growth in the transmission of multimedia applications with different transmission
time sensitivities has raised the challenge of facilitating their QoS. Timersensitive
applications such as videoconferencing and Voice over IP (VoIP) are susceptible to

136
Lnapter / xmprovemenu* 111 ^uaui}' ui aervite in ^uinpuier i^eiworiu»

communication parameters such as packet delay, jitter and throughput. On the other
hand, some time-insensitive applications, such as File Transfer, are more affected by
packet loss and reliability (this includes bits error rate), but are unaffected by jitter
(Kurose and Ross, 2005). The interconnection between wired and wireless networks also
requires that QoS of traffic being exchanged to be appropriately realised.

Most previous studies explored QoS support either in wireless local area networks
(WLANs) or in wired networks as discussed in section 3.4 of chapter 3. However, there
were only few studies to enable end-to-end QoS for wired and wireless networks. For
instance, Skyrianoglou et al. (2002) proposed a Wireless Adaption Layer (WAL) to
provide an integrated QoS between WLAN and IP infrastructure. The WAL was located
between the MAC layer and the IP layer. The function of WAL was to provide service
differentiation to the IP traffic transmitted between WLAN and a fixed IP network to
improve the performance in a wireless IP networks. An architecture to map IP layer
Differentiated Services Code Point (DSCP) QoS to MAC layer EDCA Access categories
(ACs) was proposed by Park et al. (2003). The integrated scheme used DSCP of the IP
packets to allocate packets to an appropriate AC. Another mapping scheme of Enhanced
Distributed Channel Access (EDCA) access categories to IP traffic class in wired
network was proposed by Senkindu and Chan (2008). Their mapping aim was to ensure
that end-to-end service guarantees were provided for multimedia applications.

However, most previous studies either required modifications of wireless stations, which
in turn complicated the WLAN operation, or supported QoS of high priority traffic
which in turn starved other transmitted traffic. The limitation of WAL proposed by
Skyrianoglou et al. (2002) is that an intermediate layer was introduced between the MAC
and IP layers in wireless stations. This in turn resulted in more complexity to WLAN
management. Also, the modification of wireless stations to support DiffServ
functionality was a disadvantage of the method introduced by Park et al. (2003).

The scheme by Senkindu and Chan in (2008) was simple to implement, but the low
priority traffic’s packet drop rate was high due to link congestions and QoS prioritisation.
Therefore, the challenge in end-to-end QoS is that both wired and wireless parts of the
network provide suitable treatment for each class of traffic and to efficiently use network
resources.

The main contribution of this study is the development of a new QoS enhancement
scheme that improves delivery of a variety of applications in both wired and wireless
sides of the network. The scheme provides an integrated MAC layer QoS for the
137
L im p ier / uiijji uvciiiciiio in \ £ u a n i y ui oci vice 111 cu iu p u in licinuiivd

wireless side of the network and network layer QoS controls in the wired side. A novel
aspect of the approach is that an adaptive AC traffic allocation algorithm was devised
and incorporated into wireless access point (AP) in order to improve QoS in the IEEE
802.l i e EDCA protocol. Also, Weighted Round Robin (WRR) queuing scheduling
mechanism was implemented into the congestion point (i.e. router) in wired networks to
support a fair distribution of bandwidth among different traffic types.

7.3 Description of QoS Enhancement Approach

The aim was to introduce an enhanced integrated QoS in the wired and wireless sides of
the network. The approach consists of: (i) an adaptive traffic allocation algorithm at the
wireless side of the Access Point (AP), and (ii) a Weighted Round Robin (WRR)
queuing scheduling mechanism implemented at the congestion points of the network’s
wired side. The proposed QoS enhancement scheme was designed to allow an end-to-
end QoS to be set up. This would then provide an effective delivery of a variety of
applications in the context of wired-cum-wireless networks, The next subsections
explain the mechanism of the proposed adaptive traffic allocation algorithm and the
manner in which WRR queue scheduling was adapted to enhance QoS.

7.3.1 Adaptive Traffic Allocation Algorithm

The adaptive traffic allocation algorithm was located in the access point in the wireless
side of the network. The algorithm is outlined in the flow chart shown in Figure 7-1.

The algorithm monitors the Packet Arrival Rate (PAR) of the uplink and downlink
traffic for each AC, in order to allocate traffic to an appropriate AC. The PAR
parameter was chosen because it signify how well real-time and non-real-time
applications are delivered. Also, it can be easily computed in real-time. This in turn
facilitates quick allocation of the arrived traffic to the most appropriate AC.

PAR is defined as the number of successfully received packets in a given time interval
(Wang et al., 2000). Equation (7.1) was used to calculate PAR:

PARi (t) = (7.1)

where, PARi (t ) is packet arrival rate during ith time interval. £ Pi (t) is the total
number of packet received during the ith interval, and tj is the time duration of the ith
interval.
138
Chapter 7 improvements in quality 01 service in com puter i^eiworics

Incoming packets

Calculate Packet Arrival Rate for each access

category (n) at the access point (i.e. PAR(nj),
where 0 < n < 3

Yes No

Shift all the lower Set traffic to the default

priority AC traffic to the AC settings of IEEE
next higher AC 802.l i e EDCA

Wait for preset time Transmit

interval packets

Figure 7-1. Adaptive traffic allocation algorithm flow chart.

During transmission, the algorithm calculates at each time interval, the PAR for each
AC (i.e., ACo to AC3). AC0 receives the highest priority traffic, while AC3 is for the
lowest. When the PAR at a higher AC is zero, the algorithm allocates the incoming
traffic to a lower priority AC to the next higher priority AC. This operation increases
the lower priority traffic’s ability to access the channel and thus it transmits its packets
at a faster rate. This also ensures the network resources do not remain idle.

The algorithm waits for a pre-set time interval and re-calculates the PAR associated
with traffic for each AC. If there is an application for transmission at a higher priority
AC, the lower priority traffic previously diverted to the higher AC, would be moved
139
L.napter / improvements in viiiamy 01 service 111 ^umpuier neiwui ks

back to its original default AC. However, if higher priority traffic started its
transmission during the pre-set time interval, it would be directed to its original AC and
the lower priority traffic would be moved back to its original AC after the current
interval is terminated. Therefore, the pre-set time interval must be chosen carefully to
accommodate the process. This is because in case of a large time interval, the higher
priority traffic would be negatively affected, whereas in case of short time interval, the
high computational load would be experienced (that is an issue for real-time
operations).

7.3.1 Integration of Weighted Round Robin Queuing

Mechanism

In the wired side of the network, WRR scheduling mechanism was integrated between
the router and the AP in order for the transmitted traffic to share the network resources,
based on QoS requirements. The WRR takes the incoming packets from a number of
traffic types and schedules them according to their allocated respective weights. The
allocation of weights considers the traffic priority, application's packet size, and its
transmission rate. In this study, four types of traffic were considered: VoIP and video as
time-sensitive applications and best effort traffic and background traffic as time-
insensitive applications. Time-sensitive applications were high priority and were
assigned 60% of the bandwidth, whereas best effort and background traffic were treated
as low priority and were allocated 40% of the bandwidth.

The respective weights for the aforementioned applications are shown in Table 7-1.
However, when one type of application was not transmitted at a particular time interval,
WRR divided the bandwidth allocated to that application to other applications being
transmitted according to their respective weights (Semeria, 2001).

Table 7-1. WRR queue weights.

^ ^ ^ A p p lic a tio n background

WRR VoIP Video best effort traffic
traffic

Weights 3 3 2 2

140
unapter / improvements in quality 01 service in computer i^eiworics

7.1 Network M odelling and Sim ulation

To validate the performance of the QoS enhancement scheme, wireless-cum-wired

network topologies with different network sizes (small, medium, and large) were
simulated using the Network Simulator- 2 (NS2). A small network was simulated using
8 unidirectional connections. Networks with a medium size were simulated using 16
and 24 unidirectional connections. A large network was simulated with 32
unidirectional connections. In all networks, half of the connections, transmitted traffic
from wireless to wired network, whereas the other half transmitted traffic from wired to
wireless. A network topology that illustrates wireless-cum-wired network is shown in
Figure 4-3 (See chapter 4).

The traffic transmitted over the simulated networks was: VoIP, video streaming, best
effort traffic, and background traffic. Constant Bit Rate (CBR) traffic was adapted to
model VoIP. VoIP was modelled with the G.711 voice encoding scheme. The video
streaming source was YUV QCIF (176 x 144) Foreman with 400 frames (YUV QCIF,
2012). Prior to transmission, each video frame was fragmented into packets that in turn
had a maximum length of 1024 bytes. The best-effort traffic was modelled using CBR
with different packet sizes and generation rates that corresponded to non- VoIP usage.
File Transfer Protocol (FTP) application was used for the background traffic. FTP was
transmitted over TCP (to ensure reliability), whereas other traffics were transmitted
using UDP (to ensure high transmission rate) transport protocol. Table 7-1 shows the
traffic characteristics of the aforementioned applications.

Table 7-2. Traffic characteristics.

* ^ ^ T ra ffic Type Back-ground

VoIP Video Best effort
Param eters^"' traffic

Packet size (Byte) 160 -692 1000 1000

Traffic type CBR VBR CBR FTP

Transmission rate
64 - 125 -
(kbps)

Two different scenarios were considered for each simulated network. The first scenario
excluded the presence of QoS enhancement scheme. The WLAN was based on IEEE
802.l i e EDCA, The main parameters that modelled the wireless channel were the

141
Chapter 7 improvements in quality ot service in computer i>eiworKs

default settings for IEEE 802.lie . These parameters are shown in Table 4-1. The
transmitted traffic (i.e. VoIP, video streaming, best effort traffic, and background
traffic) were mapped to the ACs to represent different levels of priority as shown in
Table 4-2. VoIP had highest priority, while the priority of the background traffic was
lowest. At the wired side of the network, First-in-First out (FIFO) queue scheduling
mechanism (queue size=50) was implemented between the router and the access point.
FIFO was chosen due to its simplicity that facilitates real-time operations.

However, in the second scenario where the QoS enhancement scheme was included, the
adaptive traffic allocation algorithm was operated at the wireless side of the AP. The
main parameter of adaptive allocation algorithm was the pre-set time interval which was
set up to be 2.5s. This value was chosen experimentally to provide best results. In the
wired side of the network, the queue scheduling mechanism implemented between the
router and the access point was Weighted Round Robin (WRR). WRR classified traffic
based on their QoS requirements. Time-sensitive applications had a higher priority than
time-insensitive applications. The main parameters of WRR are shown in Table 7-3.

Table 7-3. WRR Parameters.

WRR queue No.

1 2 3 4
Parameters
Background
Application type VoIP Video Best effort
traffic
WRR weights 3 3 2 2
Queue length 25 25 25 25
Allocated bandwidth 0.6 Mbps 0.6 Mbps 0.4 Mbps 0.4 Mbps

The network topology covered an area of 500m x 500m and the stations were
positioned randomly within this area. This setup remained unchanged during the
simulations. Simulations were repeated 10 times. Each time a different initial seed value
was used to randomly position the stations and to control which node transmitted first.
The randomness introduced using the different seeds avoided the bias of random
number generation. The results of the 10 simulations were then averaged. Simulation
duration was 300 seconds. These periods were considered sufficient to examine the
behaviour of IEEE 802.1 le protocol.

142
^napier / xinpruveiiieiiui 111 yuam j ui service 111 v^uinpuier nciwuiivs

7.2 Results and Discussion

This section is to demonstrate the effectiveness of the proposed QoS enhancement
scheme. The results obtained using the proposed QoS enhancement scheme (i.e.
adaptive allocation algorithm and WRR queuing scheduling mechanisms) were
compared with the results obtained using the legacy scheme (i.e. standard IEEE 802.1 le
EDCA and FIFO queuing mechanisms). The following subsections respectively explain
the comparison of delay, jitter, packet loss ratio, and the overall assessed QoS for the
transmitted traffic with and without the proposed QoS enhancement scheme.

7.2.1 Delay

In this study, the measurement of delay includes various types of delays such as queuing
delay, transmission delay, processing delay, and propagation delay. Figure 7-2 (a) - (d)
show respectively the average values of delay for VoIP, video streaming, best effort
traffic, and background traffic under different network loads. It is indicated from the
Figures that the average values of delay using legacy and QoS enhancement scheme
were increased as the number of connections increased. This is because increased
probability of packets collisions, which in turn caused the nodes to retransmit the
collided packets. In this case, the delay between the consecutive packets would be
further increased.

In addition, Figures 7-2 (a) - (d) indicate that both schemes treated the transmitted
traffic based on their QoS requirements. VoIP gained the lowest delay because it was
assigned by default to the highest priority access category (i.e. ACo), whereas the
background traffic experienced the highest delay among other traffics as it was assigned
by default to the lowest priority access category (i.e. AC 3).

The VoIP that was assigned to ACo had a small value for AIFS (i.e. equivalent to 2ps)
allowing ACo to start its backoff procedure after detecting the channel was idle for an
AIFS period. Also, small values for CWmin, and CWmax which were respectively 7 slots,
and 15 slots for ACo, made VoIP traffic to have a small waiting duration before
accessing the channel. In addition, large value of TXOP (i.e. equivalent to 3.008 ms)
allowed ACo to transmit multiple data frames continuously during that particular time
interval defined by TXOP. The values of aforementioned EDCA parameters gave VoIP
traffic the highest priority to access the channel. Therefore, the expected packet delay
would be lower than other transmitted traffic.
143
unapier / improvements 111 v£uamy ui oti vite 111 v^uiupuicr iicmuina

In contrast, the background traffic that was assigned to AC 3 experienced highest delay
among other transmitted traffics. For example, the delay values of the background
traffic at medium network loads (i.e. 16 connections) using legacy and QoS
enhancement scheme were receptively 186.6 ms, and 130.5 ms. These values were
much higher as compared with delay values of other traffic under the same network
conditions. The delay values of video streaming for instance when using legacy and
QoS enhancement scheme were respectively 17.1 ms, and 48.5 ms. This is because of
the values of EDCA parameters assigned to AC 3 . Large values of AIFS, CWmin, and
CWmax which were respectively 7 ps, 31 slots, and 1023 slots made the traffic assigned
to AC 3 to wait for long period of time before accessing the channel and then to start its
backoff procedure after detecting the channel was idle for an AIFS period.

However, Figures 7-2 (a) - (d) also indicate that the traffic transmitted using the
proposed QoS enhancement scheme experienced a lower delay for all ACs as compared
with the legacy network scheme. The average delay for video streaming, best effort
traffic, and background traffic were decreased by 57.6%, 63.6%, and 31.5%
respectively when the QoS enhancement scheme was introduced. The reductions in
delay values were due to the allocation of lower priority traffic to the next higher AC in
cases of Packet Arrival Rate (PAR) for the higher priority traffic was zero. For instance,
the packets of video streaming were transmitted by default using ACi which its EDCA
parameters value were 2 ps, 15 slots, and 31 slots for AIFS, CWmin, and CWmax
respectively. However, when QoS enhancement scheme was applied, the packets of
video streaming were transmitted for particular time interval using ACo when VoIP
traffic did not transmit at that particular time interval. The allocation of video streaming
traffic's to ACo (which its EDCA parameters value were 2 ps, 7 slots, and 15 slots for
AIFS, CWmin, and CWmax respectively) increased its ability to access the channel and
thus reduced the average value of delay.

Also, when QoS enhancement scheme was applied, the fairness in treating traffic with
different QoS requirements introduced by WRR at the wired side of the network had
been a factor in improving network performance. For example, the reduction o f 66.2%
in delay for VoIP was due to the implementation of WRR, as the adaptive allocation
algorithm did not shift the traffic transmitted by the highest priority AC.

144
Chapter 7 im provem en ts in v^uaiuy ui a er v itc m v.uiupuici n c m u in a

i Ip o S enhancem ent scheme

Legacy network

(a)

No. of connections

I IQoS enhancem ent scheme

Legacy network

(b)

No. o f connections

100
□ Q o S enhancem ent scheme
i
i

1
i
L
i
i

i
i

i
i

i
1 i

' m 1
i i
70
i t
I i
60
I i
i
i

(c) ^ 50
.2
i
i
« 40

30

20
i 1 B
10

1 No. of connections
32

□ QoS enhancem ent scheme

Legacy network

(d)

No. of connections

Figure 7-2. Bar chart representation of average delay with and without QoS
enhancement scheme for: (a) VoIP, (b) Video, (c) Best effort traffic, (d) Background
traffic.
145
C h a p te r / improvements in yuitiuj' ui oei vice in v^uni|mici ncmuina

7.2.2 Jitter

The average values of jitter for VoIP, video streaming, best effort traffic, and
background traffic under different network loads are shown respectively in Figures 7-3
(a) - (d). Each Figure provides a comparison between the jitter's values obtained using
the proposed QoS enhancement scheme and the legacy scheme under a light network
load (i.e. 8 connections), a medium network load (i.e. 16-24 connections), and a heavy
network load (i.e. 32 connections) respectively.

It is shown from the Figures that the values of jitter for all transmitted traffic gradually
increased as the network became more congested for both legacy and QoS enhancement
schemes. For example, the values of jitter for VoIP traffic under a light network load
were respectively 1.7 5ms, and 3.3 ms using QoS enhancement scheme and the legacy
scheme. Whereas, the jitter's values for VoIP traffic using QoS enhancement scheme
and the legacy scheme under a heavy network load were respectively 4.9 ms, and 5.7
ms.

In other words, as the number of active stations increased; the probability of collisions
increased accordingly due to increased competition between the stations. This forced the
MAC protocol to retransmit the collided packets. When the collided packets were
successfully received; the time between any two consecutive packets that were
successfully received at the destinations were increased, and subsequently increasing
the jitter.

However, Figures 7-3 (a) - (d), also indicates that the implementation of QoS
enhancement scheme reduced the jitter values of all transmitted traffic as compared
when the legacy scheme was applied. The average jitter for VoIP, video streaming, best
effort traffic, and background traffic were reduced by 36.3%, 36%, 20.5%, and 35.1%
respectively.

The reduction of jitter for traffic transmitting by lower priority access category (i.e. ACi
- AC3) was due to the adaptive allocation algorithm at the wireless network, and WRR
implemented at wired side of the network.

The jitter's reduction of best effort traffic for instance was due to the dynamic allocation
for CBR packets from the lower priority AC (i.e. AC 2) to the next higher AC (i.e. ACi)
when the ACi was not transmitting video traffic at that time. In other words, best effort
traffic were transmitted by default using AC 2 which its EDCA parameters were 3 ps, 31
146
 liu p iu v u iiu iid in y u r n iij u i a c t v i t c 111 V / U i u p u i c i iic m u ix v a

slots, and 1023 slots for AIFS, CWmin, and CWmax respectively. However, when QoS
enhancement scheme was applied, the CBR packets were transmitted for a particular
time interval using ACi when video traffic did not transmit at that time interval. The
allocation of CBR traffic to ACi which its EDCA parameters were 2 ps, 15 slots, and 31
slots for AIFS, CWmin, and CWmax respectively increased its ability to access the
channel and thus reduced the average jitter.

Also, the fairness of WRR to share network resources among transmitted traffic based
on their QoS requirements decreased the values of jitter. WRR allocated 20% o f the
bandwidth to CBR traffic making the packet collision to decrease accordingly. This is
unlike the FIFO mechanism in legacy scheme where high and low priority traffic were
treated in the same manner. The arrived packets dropped regardless of their priorities
when the queue of FIFO became full. As a result, the implementation of QoS
enhancement scheme reduced the possibility of CBR packets to collide and
subsequently reduced the variations in delays between any two consecutive CBR
packets.

The jitter for VoIP traffic which was reduced by 36.3%, when QoS enhancement
scheme was implemented was due to the implementation of WRR only. This is because
the adaptive allocation algorithm does not shift the traffic transmitted by default using
the highest priority AC (i.e. ACo).

As the highest priority traffic, 30% of the bandwidth was allocated to VoIP when WRR
was implemented. This is unlike FIFO mechanism in legacy scheme which made VoIP
as high priority traffic to be treated in the same manner as low priority traffic.
Subsequently, the arrived packets were dropped regardless of their priorities when the
buffer at the router became full.

The implementation of WRR in QoS enhancement scheme reduced the possibility of

competition between VoIP and other traffic to be transmitted throughout the router. This
reduced packets drop rate and subsequently reduced the variations in delays between
any two consecutive VoIP packets.

147
Chapter 7 improvements in yuaiiiy 01 service in L.ompuier iveiworKs

□ QoS enhancem ent scheme

I Legacy network

(a)

No. of connections

□ QoS enhancem ent scheme

Legacy network

(b)

No. of connections

□ QoS enhancem ent scheme

Legacy network

(c)

16 32
No. of connections

I IQoS enhancem ent schem e

Legacy network

(d) rao

No. of connections

Figure 7-3. Bar chart representation of average jitter with and without QoS
enhancement scheme for: (a) VoIP, (b) Video, (c) Best effort traffic, (d) Background
traffic.

148
unapier / i i u p i U Y C i u c i i U ) 111 v ^ u a i u j u i o e i v i c e i n ^ u m p u ic i n c iw u in s

7.2.3 Packet loss ratio

Figures 7-4 (a) - (d) depict comparison of packet loss ratio for VoIP, video streaming,
best effort traffic, and background traffic obtained using QoS enhancement scheme and
the legacy scheme.

The results indicate, the increase in packet loss ratio was based on the number of
connections, and the priority of the traffic for both schemes. For example, packet loss
ratio for video streaming increased as the number of connections increased as shown in
Figure 7-4 (b). Packet loss ratios for video streaming under a light load (i.e. 8

connections) were respectively 0.001%, and 0.1% for QoS enhancement scheme and the
legacy scheme. Whereas, packet loss ratios for video traffic using QoS enhancement
scheme and the legacy scheme under a heavy network load (i.e. 32 connections) were
respectively 1.9%, and 3.5%. This was caused by a high degree of competition between
the transmitting stations, which in turn increased the probability o f packets colliding.

The priority of the traffic also influenced packet loss. For example, Figure 7-4 (a) shows
the packet loss ratios for VoIP under a medium network load (i.e. 32 connections)
which were respectively 0.1%, and 0.25% using the QoS enhancement scheme and the
legacy scheme. Whereas, the packet loss ratios for the video application under the same
load were respectively 0.51%, and 1.71% using the QoS enhancement scheme and the
legacy scheme as shown in Figure 7-4 (b). This is because; VoIP traffic was assigned to
the highest priority AC (i.e. ACo) with small values of AIFS, CWmin, and CWmax.
Whereas video traffic was assigned to the lower priority AC (i.e. ACi) with large values
of AIFS, CWmin, and CWmax as compared with ACo. Small values of AIFS, CWmin,
and CWmax increased the traffic’s ability to access the channel and thus the AC
transmitted its packets faster.

Figure 7-4 (d) shows that the packet drop rate for FTP which was transmitted using the
lowest priority AC (i.e. AC 3) was lower than the packet loss ratio for CBR shown in
Figure 7-4 (c). For example, the packet loss ratios for FTP under a medium load (i.e. 16
connections) were respectively 0.21%, and 1.68% using QoS enhancement scheme and
the legacy scheme. Whereas, the packet loss ratios for CBR traffic under the same
network load were respectively 0.7%, and 2.05% using QoS enhancement scheme and
the legacy scheme.

149
Chapter 7 improvements in viuauty 01 service m t^umpuier ncivruiM

Although the CBR traffic was located at a higher priority AC (i.e. AC 2), FTP was
transmitted using Transmission Control Protocol (TCP) whereas CBR traffic
transmitted using User Datagram Protocol (UDP). TCP as a connection-oriented
protocol provides a reliable, ordered, and error-checked delivery of packets.

Figures 7-4 (a) - (d) also indicate the proposed QoS enhancement scheme reduced
packet loss ratio for all ACs as compared with legacy scheme. The overall packet loss
ratios for the VoIP, Video streaming, CBR traffic, and FTP application decreased by
72.2%, 6 8 %, 26.1%, and 81.9% respectively.

The reductions in packet loss ratios were due to two factors: the adaptive traffic
allocation algorithm at the wireless side of the AP, and the WRR queuing scheduling
mechanism which was implemented between the AP and the router at the wired side of
the network.

The former factor affected the video, CBR, and FTP traffic as it shifted traffic from a
lower AC to the next higher AC, whereas the latter factor reduced the packet loss for all
transmitted traffic due to its fairness distribution o f network resources. As an example,
the packet loss ratio for video streaming under a heavy network load (i.e. 32
connections) was 3.5% using the legacy scheme, whereas its packet loss ratio under the
same condition using QoS enhancement scheme was 1.9%. This reduction was due to
the dynamic allocation for the video packets from a lower priority AC (i.e. ACi) to the
highest AC (i.e. ACo) when the ACo was not transmitting VoIP traffic at that time. The
values of EDCA parameters for ACi were 2 ps, 15 slots, and 31 slots for AIFS, CWmin,
and CWmax respectively. Whereas, the values of AIFS, CWmin, and CWmax for ACo were
2ps, 7 slots, and 15 slots respectively. This operation increased video traffic’s ability to
access the wireless channel and transmit its packets faster.

The 20% of the bandwidth allocated by WRR to video traffic made the probability of
packet collision to be decreased. WRR implemented in QoS enhancement scheme
treated traffic based on their QoS requirements. This is unlike FIFO mechanism in the
legacy scheme where high and low priority traffic were treated in the same manner.

150
Chapter 7 Improvements in quality 01 service in computer iveiwurKs

0.7
]Q oS enhancem ent scheme
M Legacy network_________
0.6

0.5

(a)
-0 .3

0.2

0.1

_ J
No. o f connections

I iQoS enhancem ent scheme

Legacy netw ork

(b)

16 32
No. of connections

I IQoS enhancem ent schem e

Legacy netw ork

(c)

IQoS enhancem ent scheme

I Legacy network

(d) iu

K No. of connections

Figure 7-4. Bar chart representation of average packet loss with and without QoS
enhancement scheme for: (a) VoIP, (b) Video, (c) Best effort traffic, (d) Background
traffic.
151
Chapter 7 improvements in i^uamy 01 service 111 computer neiwui

7.2.4 Overall QoS

Figures 7-5 (a) - (d) show the overall QoS for VoIP, video streaming, best effort traffic,
and background traffic respectively when legacy and QoS enhancement schemes were
implemented.

The overall QoS was assessed using a supervised learning Multi-Layer Perceptron
(MLP) neural network (Dogman et al, 2012). In this technique, the QoS parameters (i.e.
delay, jitter, and packet loss ratio) for the transmitted applications were used as inputs to
the trained MLP. The MLP then quantified the overall QoS based on the values of
delay, jitter, and packet loss ratio taking into the account the QoS requirements for each
application. The overall QoS spanned between (0%-100%). A poor QoS did not exceed
33%, whereas a Good QoS could not be below 67%. More details about QoS
assessment using MLP can be found in section 6.3.4 (see chapter 6 ).

From Figures 7-5 (a) - (d), the overall QoS for all transmitted traffic were decreased as
long as the number of connections in the network was increased for both the legacy and
QoS enhancement schemes. For example, the values of overall QoS for the video
streaming under a light network load (i.e. 8 connections) were respectively 82.18%, and
42.46% using the QoS enhancement scheme and the legacy scheme. Whereas, the
overall QoS values for the video application using QoS enhancement scheme and the
legacy scheme under a heavy network load (i.e. 64 connections) were respectively
22.1%, and 15.01%. This was due to the gradual increase of delay, jitter, and packet loss
ratio for all transmitted traffic when the load on the network was increased. In both
schemes, as the number of active stations increased; the probability o f collisions
increased due to a high degree of competition between stations. This is in turn forced
the MAC protocol to retransmit the collided packets, and subsequently increased the
values of delay, jitter, and packet loss.

Also indicated from Figures 7-5 (a) - (d), is that the priority o f transmitted applications
could affect their overall QoS. For instance, the values of overall QoS for the VoIP
application under a medium network load (i.e. 32 connections) were respectively
75.46%, and 46.62% using the QoS enhancement scheme and the legacy scheme.
Whereas, the overall QoS values for CBR traffic using the QoS enhancement scheme
and the legacy scheme under the same network load were respectively 35.86%, and
21.55%. This is because VoIP was transmitted using the highest priority AC (i.e. ACo)
while the best effort traffic was transmitted by default using the lower priority AC (i.e.
152
i,napier / i m p i u v c m c u i s 111 y u m u j ' u i o c i y iw c 111 V / u m j i u i c i n c in u iiu

AC2). The values of EDCA parameters for ACo facilitated its traffic to access the
channel and transmitted at faster rate. The small values of AIFS, CWmin, and CWmax
which were respectively 2 ps, 7 slots, and 15 slots enabled VoIP transmitted by ACo to
have a small waiting period before accessing the medium. Also, the large value of
TXOP, which was equivalent to 3.01 ms, allowed ACo to transmit multiple data frames
continuously during that particular time interval defined by TXOP.

However, the transmission mechanism, and the QoS requirements for a transmitted
application could have an impact on its overall QoS rather than the traffic priority.
Although the background traffic was transmitted by default using the lowest priority AC
(i.e. AC3), its overall QoS was outperformed on the overall QoS of the video traffic
which was transmitted using A Q . For instance, the values of the overall QoS for
background traffic under heavy network load (i.e. 64 connections) were respectively
49.18%, and 33.48% using the QoS enhancement scheme and the legacy scheme.
Whereas, the overall QoS values for the video traffic using the QoS enhancement
scheme and the legacy scheme under the same network load were respectively 2 2 . 1 %,

and 15.01%. This is because the background traffic was transmitted using TCP, whereas
the video traffic was transmitted using UDP. TCP is a connection-oriented protocol that
provides a reliable, ordered, and error-checked delivery of packets transmitted between
stations, while UDP is a connectionless Internet protocol with no acknowledgment of
packet delivery. Also, the background traffic is tolerant of the QoS parameters such as
delay, and jitter as compared with the video traffic. The acceptable QoS requirements
for video application in terms of delay, jitter, and packet loss ratio should not exceed
150 ms, 10 ms, and 3% respectively.

Figures 7-5 (a) - (d) indicate that the overall QoS of traffic transmitted for all ACs using
the proposed QoS enhancement scheme was higher as compared with the legacy
network. For example, the average QoS for the VoIP, video, best effort traffic, and the
background traffic, when the network load was light (i.e. 8 connections), were improved
by 80.9%, 93.5%, 7.8%, and 29.4% respectively. The improvements in QoS were due to
allocating the lower priority traffic to the next higher AC, in cases of zero PAR for the
higher priority AC and the fairness introduced by WRR in treating the traffic based on
their QoS requirements.

153
Chapter 7 im provem en ts in v£uaiuy ui se rv ic e m \^oui|jui.ci n c m m iv s

I QoS enhancem ent scheme

Legacy network

No. of connections

I iQnS enhancem ent schem e

Legacy netw ork

■
No. of connections

IQ oS enhancem ent schem e

Legacy netw ork

No. of connections

I IQoS enhancem ent schem e

L egacy netw ork

No. of connections

Figure 7-5. Bar chart representation of overall assessed QoS with and without QoS
enhancement scheme for: (a) VoIP, (b) Video, (c) Best effort traffic, (d) Background
traffic.
154
^napier / im p iu v c iiie u ia 111 y u a m y u i o c i v ic e m c u n ijju iti i^ » n u iiu

Figure 7-6 shows a visual comparison of sample images from the Foreman video
transmitted over a medium network load (i.e. 24 connections) using the proposed QoS
enhancement scheme. It was observed that the image quality with the QoS enhancement
scheme was higher than the image quality without it. The number of received I, P, and
B frames for the transmitted Foreman video was greater when using the QoS
enhancement scheme as shown in Table 7-4. This improvement was due to the shift of
the video packets from ACi to ACo in the absence of VoIP and the fairness of
distributing network resources by WRR.

(a) (b)

Figure 7-6. Visual comparison of reconstructed Foreman video using: (a) QoS
enhancement scheme, (b) legacy network scheme.

Table 7-4. The Amount of sent, received, and lost Foreman video frames.

" “— ------------------ Frame Type

I P B Total
No. of frames —

Sent 45 89 266 400

Received with QoS enhancement scheme 45 89 266 400
Received without QoS enhancement scheme 42 73 228 343
Lost with QoS enhancement scheme 0 0 0 0
Lost without QoS enhancement scheme 3 16 ; 38 57

155
unapter / l iiip r u v e iiie u is m y u a i u j ' u i o c i v ic c m V /U iu jju iti n c i n u n u

7.3 Summary
A new QoS enhancement scheme that consisted of a combination of an adaptive access
category (AC) traffic allocation algorithm implemented on the wireless side o f a
simulated network, and the Weighted Round Robin (WRR) scheduling implemented on
its wired side was devised and its performance was evaluated. The traffic allocation
algorithm assigned traffic of a lower priority AC to the next higher priority AC in the
absence of any higher priority traffic to further improve the performance of IEEE
802.l i e EDCA standard, whereas WRR managed to fairly allocate network resources
among transmitted traffic types based on their QoS requirements. The performance of
the proposed scheme was compared with the standard IEEE 802.1 le EDCA and FIFO
queuing mechanisms (i.e. WLAN-wired network legacy scheme). The proposed scheme
significantly improved the QoS for transmitted applications. The average QoS for VoIP,
video, best effort traffic, and background traffic increased from their original values by
72.5%, 70.3%, 44.5%, and 45.2% respectively. The QoS proposed scheme allowed an
end-to-end QoS to be set up which in turn provided an improved delivery for a variety
of applications in the context of wired-cum-wireless networks.

156
 Chapter 8 Microcontroller Board Implementation

of Quality of Service Assessment System

8.1 Introduction
Network QoS assessment plays an important role in managing network resources and
ensuring that various applications receive an appropriate priority or sufficient resources.
Therefore, developing a hand-held system that accurately assesses QoS for different
applications is very valuable.

In this chapter, a network QoS monitoring system is designed and evaluated. It

incorporated the QoS assessment approach developed by (Dogman et al, 2012a) that
was based on regression model. More details about regression modelling based QoS
assessment is provided in Section 6.3.3 of Chapter 6 . The microcontroller board
MCB2300 KEIL ARM was used for the purpose of this study.

Following the measurement of the QoS parameters: delay, jitter and packet loss ratio of
multimedia applications, they are fed into the microcontroller board. The board then
analysed the parameters based on their transmission requirements and produced the
corresponding overall QoS. The performance of the system device is compared with
other QoS assessment methods (e.g. QoS assessment using Fuzzy Inference System
introduced by (Al-Sbou et al, 2006), and Neural Network QoS monitoring approach
proposed by (Dogman et al, 2012b). The results indicated that the developed system is
capable of accurately assessing QoS.

This chapter is organised as follows: the relevant studies about QoS monitoring tools
are discussed in section 8.2. Section 8.3 explains the MCB2300 KEIL ARM
microcontroller board. Section 8.4 outlines how the QoS assessment technique using
regression modelling was devised, and implemented on the MCB2300 KEIL ARM
microcontroller board. The experimental procedures are described in section 8.5. The
results are discussed in section 8 .6 . The conclusion is presented in Section 8.7.

157
V /iiit p ie r o 1T1 IL1 uvuiili u n ti n u p itiii^ iiia iiu ii ui y vl7 njjvajiiivu^ ujo^vau

8.2 Related Work

The growth in multimedia networks has led to the need for finding efficient ways to
monitor QoS accurately. QoS can be achieved by either (i) prioritising time sensitive
applications (such as video conferencing and voice over internet protocol) over time
insensitive applications (such as file transfer), or (ii) reserving network resources (such
as bandwidth) for the time sensitive applications prior to their transmission.

There are a number of QoS monitoring tools that have been proposed to monitor
network performance as in (Graham et al, 1998), (Zseby and Scheiner 2004), and
(Carvalho et al, 2009). A critical analysis of these tools can be found in Section 3.5,
Chapter 3. The existing network QoS monitoring tools have a number of shortcomings.
Some methods cannot determine directly the overall network QoS as in (Graham et al,
1998). Network managers have to do a variety of operations to assess the overall
network QoS. Other QoS monitoring tools are not stand-alone devices as in (Zseby and
Scheiner, 2004) and (Carvalho et al, 2009). From these limitations, the process of
monitoring QoS can be complicated, expensive, and time consuming.

Therefore, in this study, a hand-held system that accurately determines the overall
network QoS for multimedia applications was designed. The proposed device assessed
network QoS for multimedia applications taking into the account the QoS requirements
of these applications. A novel aspect of this study is that a microcontroller board with
integrated QoS monitoring tool is used.

The use of embedded systems in recent years has increased in such a way that it is used
in various mobile and multimedia applications. An efficient environment for
microcontroller applications is KEIL Microcontroller Development Kit (MDK-ARM)
(MDK-ARM, 2013). The features of KEIL MDK-ARM are its ease o f use, and the
manner it can be redesigned based on application's requirements without incurring
major Non-Recurring Engineering (NRE) costs.

MDK-ARM Microcontroller board has been used in a number of studies to implement

different applications. For instance, the digital implementation of a prototype DC motor
control system was performed based on MCB 2300 KEIL microcontroller (Pal et al,
2009). The study showed that MCB 2300 KEIL microcontroller had facilities for
concurrent programming and real-time control for fast handling of events in micro

158
V s iia p ic i o i v i i u u t u i i u u i i c i u u m u i m p ic iu c u u iu u ii v i y u o /la d c d a u ic iu j j o ic u i

manufacturing applications.

The algorithm which was proposed by (Thangaraj et al, 2006) enhanced MCB 2300
KEIL microcontroller performance. Most applications that run on embedded platform
were constrained by limited built in memory available for storing the application
programs. Therefore, the proposed algorithm reduced the KEIL ARM memory
requirement needed by an application program.

The contribution of this study is the implementation of the regression modelling used to
assess QoS which was reported by (Dogman et al, 2012a) on the KEIL ARM MCB2300
microcontroller board. The proposed system could work independently to assess the
QoS of multimedia applications based on their transmission requirements (Dogman et
al, 2012d), and (Dogman et al, 2013).

8.3 KEIL ARM MCB2300 Evaluation Board

The Keil MCB2300 Evaluation Board is designed to be a very flexible evaluation board
for the NXP LPC2300 family of microprocessors. Due to its facilities to create, test, and
run application programs, the MCB2300 evaluation board can be expanded to build
hardware prototypes. MCB2300 evaluation board operation can be described in terms of
its hardware and Development Kit.

8.3.1 MCB2300 Hardware Components

The hardware components and interfaces of the Keil MCB2300 evaluation board are
shown in Figure 8-1 (MDK-ARM, 2013). The interfaces on the Keil MCB2300
evaluation board provide an easy access to the on-chip peripherals.

8.3.2 MDK-ARM Microcontroller Development Kit

The MDK-ARM is a complete software development environment for ARM processor-

based devices. MDK-ARM is designed for microcontroller applications. Its simplicity
to learn and use makes it powerful software for most demanding embedded
applications. MDK-ARM Microcontroller Development Kit includes many components
as shown in Figure 8-2 (MDK-ARM, 2013). These components are:

• ARM C/C++ Compiler: this component allows the user to write ARM
applications in C or C++ and then compiles C/C++ source files to have the
159
Chapter 8 Microcontroller uoara implementation 01 y o a assessm eiu oysiem

efficiency and speed of assembly language. The ARM Compiler translates

C/C++ source files into re-locatable object modules which contain full symbolic
information for debugging with the p Vision Debugger.

• pVision Project Manager: pVision Integrated Development Environment

(DDE) is a complete simulation in one powerful environment. It provides
facilities such as source code editing, and program debugging. The pVision IDE
software is used to create, compile, download, debug, and run a program on the
MCB2300 board. When the program is downloaded and ran successfully, the
MCB2300 board could work independently as a standalone device.

• RTX Real-Time Operating System: the Keil RTX Real-Time Operating

System is designed for ARM devices. It allows the user to create programs that
simultaneously perform multiple functions. RTX Real-Time Operating System
aids to create good structured and easy maintained applications.

• CAN Driver: MDK tool kit includes a CAN interface layer which provides an
easy and quick approach to implement a CAN network. It also provides a
standard programming API for all supported microcontrollers.

• Flash File System: the Flash File System allows the embedded applications to
create, save, read, and modify files in standard storage devices such as ROM,
RAM, Flash ROM, and SD/MMC/SDHC Memory Cards.

• USB Host and Device Interface: MDK-Professional provides USB Host and
USB Device support for embedded systems. The USB Host library is an
embedded USB stack supporting USB MSC (Mass Storage Class) and HID
(Human Interface Device) classes. The USB Device Interface uses standard
device driver classes that are available with all Windows PCs.

• TCP/IP Networking Suite: the full TCP/IP Networking Suite is designed for
ARM processor-based microcontrollers. It is highly optimized, has a small code
footprint, and gives an excellent performance.

• Graphic User Interface Library (GUI): the GUI Library is a fully featured
graphics suite that makes it possible to add graphical user interfaces to
embedded applications. It supports a large number of displays such as
monochrome, grayscale and colour LCDs.
160
cnapier a Microcontroller r>oaru implementation ui yuo nsscMmcm opicm

USB-OTG* t6x2 LCD Panel SD Connector

USB Device
(Power)

Ethernet

CAN 2

CA N1

Speaker Port LEDs LPC2388 Potentiometer INTO &

or variant Rose!
Buttons

Figure 8-1. KEIL MCB2300 Evaluation Board.

f - \ ( \

ARM C/C++ Compiler pVision Project Manager

\ K /

r \

RTX Real-Time Operating System

\

/ \ f \

CAN Driver Flash File System

L J \ )
f > / \

USB Host USB Device

\ j )
( -\ f \

TCP/IP Networking Suite (GUI) Library

V J V y

Figure 8-2. MDK-ARM Microcontroller Development Kit.

8.4 QoS Assessment Implementation Using KEIL

MCB2300 ARM

A schematic diagram of computer network QoS monitoring system is shown in Figure

161
i^iiapici o IVlltl UtUUil UUCI D u aiu lllipiClllCUUlllUll U1 y u o /\93C99111C11I LJjaiCIIl

8-3. This used the regression model (explained in Section 6.3.3, Chapter 6) on a KEIL
MCB2300 microcontroller board.

The QoS parameters of transmitted multimedia applications (i.e. delay, jitter, and packet
loss ratio) were obtained from the generated trace files of the simulated network. These
were used as inputs to KEIL MCB2300 microcontroller in order to quantify the overall
QoS. The following subsections explain the QoS assessment technique using regression
model and its implementation on the MCB2300 KEIL ARM board microcontroller.

Delay Jitter Packet loss ratio

Proposed regression model to Implementation of the proposed regression

assess the overall QoS model on KEIL MCB2300 microcontroller

Assessed QoS

Figure 8-3. QoS monitoring System.

8.4.1 Proposed Regression Model to Assess QoS

In Section 6.3.3, Chapter 3, the QoS assessment technique using regression model was
proposed and evaluated.

In brief, in the devised regression model, the values of independent variables (x l t x 2, x 3)

were represented by delay, jitter, and packet loss ratio respectively, while the values of
dependent variable (y) were represented by the overall QoS.

The regression expression was developed based on QoS requirements listed in Table 2-
1 in order to provide the outputs that reflected the overall QoS. The QoS parameters and
162
unapter » M icrocontroller u u a ru iiiipieiiieiiuuiuu ui yuo amcsmuciu ojaicui

the overall QoS were then arranged in matrices in order to feed them to the regression
model as follows:

Q o S i 1 D ± A PLR{ \b o] ei
Q o S 2 1 ^2 J2 PL R 2 h e2
= +
b2
Q o S n . .1 D n Jn P LRn_ -b 3- .e n .

where Dp /*, PLRi, QoSit i = 1,2, ...,n are delay, jitter, packet loss ratio, and overall
QoS respectively. The regression coefficients b0, blt b2, b3 were determined from the
recorded data using equation (2.12). The vector of residual (i.e. error terms) was then
calculated using equation (2.13). In this study, the means of vector o f residual produced
from regression formula for VoIP and video traffics were zeros (i.e. mean o f error terms
,i = 1,2, ...,n was zero). This implied that the estimated regression model
determined was as expressed in equation ( 8 . 1 ):

Q o S i = b0 + hi * Dj + b2 * Ji + b3 * P L R t (8.1)

where Q o S i , D it J if P L R i i = 1 ,2 ,...,n are the overall QoS delay, jitter, packet loss
ratio for ith packet respectively.

8.4.2 Implementation of QoS Assessment Technique using KEIL

ARM Microcontroller

In this section, the QoS assessment technique using regression model is implemented on
Keil MCB2300 microcontroller. The code for developed QoS assessment technique
using regression model was written in C language in order to be implemented on
MCB2300 board.

The KEIL pVision was used to create, compile, download, debug, and run the C
program on the MCB2300 board. Once the C program operation was successfully
verified, the KEIL pVision downloaded the C code on the MCB2300 microcontroller.
The KEIL ARM MCB2300 microcontroller then was able to work as a standalone QoS
monitoring device.

Figure 8-4 shows the pseudo code of the C program implemented on the MCB2300
microcontroller. First, the QoS parameters of multimedia applications (i.e. delay, jitter,
and packet loss ratio) were fed into the QoS monitoring device (i.e. programmed

163
Chapter 8 Microcontroller isoara implementation 01 assessment. opicm

MCB2300 board) through the SD connector. The Liquid Crystal Display (LCD) of the
QoS monitoring device indicated when the SD card was inserted and then the data
processing was initiated. The device then processed the QoS parameters file, read the
values of delay, jitter, and packet loss ratio, and assessed the overall QoS using equation
( 8 . 1).

Afterward, the device created an output file and recorded the values of delay, jitter, and
packet loss ratio, and their corresponding QoS value.

As soon as the QoS monitoring device completed the data processing, the overall QoS
was displayed on the LCD of the device.

//P rocess o f inserting SD card

If (SD card is not inserted) then
{ LCD output ("No SD-Card used")}
Else
( LCD output ("SD-Card used")
// Process o f reading QoS parameters
LCD output ( "Processing Data ")
Total QoS=0
Counter=0
Overall QoS=0
Open QoS parameters file
Create overall QoS file
While (Not an end-of-file)
{ Read delay, jitter, and packet loss ratio from QoS parameters file
//P rocess o f assessing QoS
Calculate QoS using equation (8.1)
/ / The output process
Write delay, jitter, packet loss ratio, and assessed QoS on overall QoS file
Total QoS= Total QoS + QoS
Counter is incremented by 1
}
LCD output ("Processing End")
Overall QoS= Total QoS / Counter
LCD output ("Overall QoS-", Overall QoS)
J________________________________________________

Figure 8-4. QoS Assessment Code.

164
 lv iiu u tu m iu u u D U d iu x i i i|i i c i u e i i u t u u i i u i ^ u o /iM C d d iu c iu

8.5 Experimental procedure

The aim of this section is to validate the performance of QoS monitoring system. This
section includes network modelling and simulation, and hardware and software for
experimental setup.

8.5.1 Modelling and Simulation

A wireless-cum-wired network topology was simulated using the Network Simulator- 2

(NS-2) in order to validate the performance of QoS monitoring system. The network
topology consisted of 8 wireless nodes, 2 wired nodes, and 2 base stations as illustrated
in Figure 8-5. The bandwidth of wired connections was 5 Mbps and 2 ms propagation
delay. The Wireless Local Area Network (WLAN) was based on IEEE 802.l i e standard
and implemented Enhanced Distributed Channel Access (EDCA) technique.

The queue management mechanism was Drop-Tail and the queue size was 50 packets.
A number of traffic types were transmitted over the simulated network. These were;
VoIP, video-conferencing, and best effort traffic. Constant Bit Rate (CBR) traffic was
adapted to model VoIP, videoconferencing, and data. VoIP modelled as G.711 voice
encoding scheme with 160 packet size and 64 kbps generation rate. The packet size of
the video traffic was 512 bytes and the inter-packet interval was 10 ms. This generated a
packet transmission rate of 384 kbps. The best effort traffic was modelled using
different packet sizes and the generation rates that corresponded to non­
videoconferencing or VoIP usage. All traffic were transmitted using UDP.

The simulation time was 500 seconds and was repeated 10 times for each experiment.
Each time a different initial seed was used in order to randomly manage which node
transmitted first as all the nodes were requested to transmit at a given time. The
randomness introduced using the different seeds avoided the bias of random number
generation.

The transmitted traffic (VoIP, video-conferencing, and best effort traffic) were mapped
into three access categories (ACs) to represent the three priority levels as shown in
Table 8-1. VoIP had the highest priority, whereas best effort traffic had the lowest
priority (IEEE Computer Society, 2005).

165
Chapter 8 Microcontroller isoara implementation 01 vjos Assessm ent system

802.1 le Network Wired Network

^ < 5^ <S*
Figure 8-5. The Simulated Network.

Table 8-1. IEEE 802.1 le Access Categories.

Parameters ACo ACi AC 2 best effort

VoIP V ideo-conferencing traffic
Arbitration Inter-Frame Space
2 2 3
(AIFS)
Minimum Contension
7 15 31
Window value (CWmin)
Maximum Contension
15 31 1023 !
Window value (CWmax)
Transmit Opportunity
3.01 6.02 0
TXOP (ms)

8.5.2 Hardware and Software Setup

This section explains the setup for the MCB2300 microcontroller board. This includes
the details about connecting and configuring procedures for the MCB2300 evaluation
board.

8.5.2.1 Hardware Setup

The following components were needed in this experiment in order to use the
MCB2300 Evaluation Kit:

• The MCB2300 microcontroller board.

• Compatible PC with at least one unused USB port in order to supply power to
the board and for downloading and debugging purpose.
• The Keil ULINK-ME USB-JTAG Adapter to run the Keil debugger using JTAG
emulation.

166
Chapter 8 Microcontroller board implementation 01 y o s Assessment system

• One USB cable.

The Keil ULINK-ME USB-JTAG Adapter which is shown in Figure 8-6 was first
connected to the MCB2300 board via its JTAG plug. Then the ULINK-ME USB-JTAG
Adapter was connected to the PC s USB port using a standard USB cable. This allows
the user to power to board, and download programs on the M CB2300 evaluation board.

Figure 8-6. The ULINK-ME Adapter.

Figure 8-7 shows the connection between the PC and the MCB2300 evaluation broad
using the Keil ULINK-ME USB-JTAG Adapter.

Figure 8-7. The connection between the PC and the MCB2300 using ULINK-ME
JTAG Adapter.

167
Chapter 8 Microcontroller isoara implementation 01 Assessment system

8.5.2.2 Software Setup

The MCB2300 Evaluation Kit uses the pVision IDE software tool. In this experiment,
the MCB2300 Evaluation Board was connected directly to the KEIL pVision IDE
software which was installed on the PC via the KEIL ULINK-ME USB-JTAG Adapter.

The Keil pVision IDE software was the front-end used with ULINK-ME adapter to
create, download, and test the embedded application on the MCB2300 microcontroller
board. Therefore, no additional software was required to run the board. Figure 8-8
shows the snapshot of pVision IDE software tool (MDK-ARM, 2013).

( 3 Bltnky - pVHJon

!** I* **» ’'J'* Ctbur) enit> *rrti lo o h JvCS VY«de« ueSp

am
_i J £
t? •• n tt o v t!:d „ -I - u • m ■ • * • , •
Preset . • x Disiiwn'.fcV 13
V.'CrtSpKt 9i: X2_v:u*
1I5C
- ADC_cwr*rr*tfV*iu*;
IM 1 Vr» Te r* G*d 2co#n
039 iiii iw x » 3 .iro.#9*cs. E xpcr I C T46S0$ 2070154s OICCCCC. [ ~ > T ] |o u
•»i A3 / 52:
, SUrtup Co* *93339':%hm nr/a
^ ) S T « 2 » iO » J «3*wc’ c< r* n r? r© sd;v r .r s .r J
393: It |A £ _ * c a l* d • • (
!j O p ttx s
O m O ’ CA 42B’ ~ :Mf TZ. r i
JS T M JIfJO lO e **. l
I itix - A£_*c*L«3;
■JJ RKK^tC <>!C«9C'CE H!t ~ H
OV 1 J
-A n -
- J It ‘try 1031
• JICD.4M.C
• JS c rtiL c
>:ntaC»cn CK. .u < Jj KEIL
Tools by ARM
- rjBWcy.C
t Sourct for <i + wSc.tn* (vo«J)
• J»*Y C
1OccumcnlKOn w h ile (
Q |jVision®4 0 m an (void)
♦ SysT<k_M*rvle- (vod)
_ )* 6 « rac tO rt AS v a D©v*!o ;> 'r* r 't £f»v»r©r>m*nt - <JGf*>c
•:, Gpo AD~* 0 m * n (v o d )
: J S rnutac. if'(W value d i z f ♦ S2Pre»cd (vod)
>SCJtuS CcOc AS C«py*Y«rtO19*7. 2005Krffetoorm.2005 2009AAMU4Alr**ifi ♦ S JP resttd (vod)
j3 S T M J » 1 0 * t le d Tks product ii procvctrd bjr US • li« r y r a p .* : sccoz ♦ J IC D .4 M .C
• j irvC'rtuKcn - j ftrtargct.c
■ •; s - ' -; : I , ♦ JWftjB*(mr«um_coc
■ i Source ♦ _ttvwch(#xch)
• JJG po.c if (CiOCiClSl { 0 ftno (fat* 0
Ciockls « 0; ♦ fo«c (me* f)

C*H St*A o x Mtmory 1

1 ________
VaWA0dr«s i MOmi [O
rtO
OO
OO
OO
♦ » □
LCD^t i t r 1 * VI *
♦» <VMtfett> 5*15000090: KCB3TX32
♦ AD„v«iv* ' *rr-v&e * 3*199595221 ............
A s * l <11 o»:ooooo4«i ............
♦ A0.wa*#-3 C*OOOOOC»
3*10000544i ............
LA A te C c = r« r'.* 3 V « ♦ A D .scaled « >
.......
OxlOOOOOAA: ...........
:* i9 9 0 0 3 c c» .....
C * :5 0 0 C 3 « I .....
U
S13
M BlUkSllttll SlUUMbll SlMttUU #
.wC
alS
?x* ij ......
n: io7ount m

Figure 8-8. pVision IDE software tool.

The project development cycle of the Keil pVision IDE is similar to any other software
development project. The development cycle started with creating a project for an
embedded application until downloading the application on target hardware. In this
experiment, the following stages were implemented to create an embedded application
with pVision IDE software:

1. Create a project, select the target device from the Device Data base, and configure
the tool settings: in this study, a new project file called "SD-File" was created from
168
Chapter 8 M ic ro c o n tro lle r Jtsoara im p ie m e n ia u o n u i y o a A b sessm cu i a j s ic u i

"New Project" option in "Project" Menu. pVision then created a new, empty project
file with the specified name. After a file name for the project was selected, a target
microcontroller was chosen. pVision had an option to select a target microcontroller.
The "Select Device for Target option" from "Project" Menu lists all the devices from
the pVision Device Database. The target device was MCB2300 microcontroller. This
step is very important, since pVision customizes the tool settings, peripherals, and
dialogs for a chosen device.

2. Create source files in C/C++ or Assembly: the "New" option from "File" Menu was
used to create a source code written in C language. The source code was saved in
"main.c" file. This file contained the pseudo code of QoS assessment shown in Figure
8-4. After a source file was created and saved, it was added to the project. The "SD-File"
project had a single source file called "main.c". This file contained a C function which
was used to read the QoS parameters of multimedia applications and assess the overall
QoS using regression model for QoS assessment.

3. Building the "SD-File" project with the Project Manager: There are several
commands were used from the "Project" Menu to compile and link the files in the
project. These were: "Translate File" command to compile the selected file in the
Project Workspace, "Build Target" command to compile files that have changed since
the last build and link them, and "Rebuild All Target Files" command to compile and
link all files in the project.

4. Downloading the "SD-File" program to MCB2300 board and test the linked
application: the "Configure Flash Tools" command from "Flash" Menu was used to
configure the target driver for flash programming. After the pVision IDE was
configured, the "Download" command from "Flash" Menu used the specified adapter
for Flash programming in order to flash the "SD-File" application program to the target
hardware (i.e. MCB2300 board). Blinking LED indicated successful download of the
program into the target hardware.

5. Debug the "SD-File" program application: the aim of debugging was to verify and
optimize the "SD-File" program application. The Options for Target - Debug dialog
was used to verify the configuration settings for the Debugger. The "Start/Stop Debug
Session" command from "Debug" Menu was used to debug "SD-File" program
application.

169
cnapier a lviicrucumruiier ouaiu xiiipiemciiuiuuii ui v uo /xascasiiiciu ajBicm

8.6 Results and Discussion

In this section, the feasibility and the functionality of the proposed QoS monitoring
device is demonstrated. The measured QoS results obtained using the proposed device
were compared with the results obtained using other QoS assessment techniques. These
techniques were QoS assessment using Fuzzy Inference System (FIS) introduced by
(Al-Sbou et al, 2006), and neural network QoS monitoring approach proposed by
(Dogman et al, 2012b).

Due to the high sensitivity of multimedia applications to the QoS parameters as shown
in Table 2-1 (See Chapter 2), in this chapter, the QoS for VoIP, and video traffic were
measured and evaluated. The following subsections respectively evaluate the assessed
QoS for the transmitted traffic (i.e. VoIP, and video traffic).

8.6.1 VoIP Traffic

After the traffic was configured and the network topology was simulated, the QoS
parameters (i.e. delay, jitter, and packet loss ratio) of VoIP were extracted from the
generated trace files of the simulated network. The QoS parameters were then used as
inputs to the KEIL MCB2300 microcontroller in order to quantify the QoS for VoIP.

Table 8-2 shows the QoS parameters of VoIP, and the evaluated QoS obtained from the
QoS Monitoring Device (QoS_MD), Fuzzy Inference System technique (FIS), and
Multi-Layer Perceptron neural network (MLP). The QoS assessment techniques as well
as the QoS monitoring device quantified the overall QoS based on the values of delay,
jitter, and Packet Loss Ratio (PLR) taking into the account the QoS requirements for
VoIP application. The overall QoS spanned between (0%-100%). Poor QoS did not
exceed 33%, whereas Good QoS could not be below 67%. More details about the
aforementioned QoS assessment techniques can be found in (Al-Sbou et al, 2006),
(Dogman et al, 2012b), (Dogman et al, 2012d), and (Dogman et al, 2013).

From Table 8-2, it can be observed that the QoS values reflected the corresponding QoS
parameters. In other words, as the values of QoS parameters increased, the values of
overall QoS decreased accordingly. For example, when the values of delay, jitter, and
packet loss ratio were 16.9ms, 0.8ms, and 0% respectively, the values of QoS for VoIP
were in Good region, i.e. 92.5% for QoS monitoring system, 89.6% for FIS based
technique and 96.9% for MLP neural network based technique.
170
unapter a M icrocontroller n o a r a iiiip ieiiieiiia iio ii 01 A ssessm ent o p ic in

In contrast, when the values of QoS parameters were high (i.e. 422.3ms for delay, 5.0
ms for jitter, and 6% for packet loss ratio), the corresponding QoS were at the Poor
level, i.e. 1.8% for QoS monitoring system, 9.3% for FIS based technique, and 4.2%
using MLP neural network. This was due to the gradual increase o f delay, jitter, and
packet loss ratio for VoIP traffic when the load on the network was increased. As the
number of active stations increased; the probability of collisions increased due to the
high degree of competition between stations. This is in turn forced the MAC protocol to
retransmit the collided packets, and subsequently increased the values of delay, jitter,
and packet loss.

Table 8-2. QoS Parameters, and the evaluated QoS of VoIP application using QoS
Monitoring Device (QoS_MD), Fuzzy Inference System (FIS), and Multi-Layer
Perceptron (MLP).

QoS Parameters Evaluated QoS

Delay (ms) Jitter (ms) PLR (%) QoS_MT FIS MLP

11.5 5.0 0.0 28.1 9.3 26.0

16.9 0.8 0.0 92.5 89.6 96.9

16.1 3.1 2.0 50.7 47.7 46.8

17.1 2.9 2.0 53.7 51.1 52.9

117.4 5.0 5.7 7.7 9.3 11.4

286.9 5.0 6.0 4.0 9.3 7.3

24.3 2.5 2.0 59.7 54.1 61.5

105.8 5.0 6.0 6.9 9.3 8.7

422.3 5.0 6.0 1.8 9.3 4.2

600.0 5.0 6.0 0.01 9.3 L!

As shown in Table 8-2, the results obtained from QoS monitoring device were
compared with other QoS assessment methods (i.e. QoS assessment using FIS technique
introduced by (Al-Sbou et al, 2006), and MLP neural network QoS monitoring
approach proposed by (Dogman et al, 2012b)). The correlation coefficient (R ) was used
to evaluate the accuracy of the proposed device compared with the other QoS
assessment techniques. When using the correlation coefficient, the magnitude of R is
between 0 and 1. The magnitude closest to 1 indicates a perfect correlation, whereas a
171
Chapter 8 Microcontroller noara impiemeniauon 0 1 yoa Assessm ent aysiem

correlation less than 0.5 would be described as weak correlation. More details about
measuring accuracy using correlation coefficient can be found in Section 6.3.5 (see
Chapter 6)

In this study, the values of the correlation coefficients between QoS determined using
the monitoring tool, and the other QoS assessment methods were: 0.97 for FIS
technique, and 0.99 for MLP neural network.

From the values of evaluated QoS provided in Table 8-2, it can be concluded that all
approaches provided results which were closely comparable. Although some outputs
were slightly different, they were in the same QoS region (i.e. Poor, Average, or Good).
The discrepancies were due to the fact that each method followed a different scheme to
determine QoS. However, the values of QoS obtained from the proposed QoS
monitoring device spanned between 1-100%, whereas the range o f QoS values produced
by FIS was between 10-90%. These indicate that the QoS monitoring system was
accurately in its operation.

8.6.2 Video Traffic

The QoS parameters of video traffic were also fed to the QoS monitoring device (i.e.
programed KEIL MCB2300 microcontroller board) to assess the overall QoS. The
extracted delay, jitter, and packet loss ratio o f video traffic were combined by the QoS
monitoring device which in turn produced its overall QoS.

As shown in Table 8-3, the QoS parameters of video application were processed using
QoS monitoring system, Fuzzy Inference System technique (FIS), and Multi-Layer
Perceptron neural network (MLP). The QoS assessment processes for these techniques
including the QoS monitoring system were based on the QoS requirements for video
application. A good overall QoS which ranged between 67-100%) corresponded to low
value of QoS parameters (i.e. delay < 1 5 0 ms, jitter < 1 0 ms, and packet loss ratio <
1%), An average QoS (i.e.33% < QoS < 67%) corresponded to medium QoS
parameters (i.e. 150 < delay < 400 ms, 10 < jitter < 20 ms, and 1% < packet loss ratio
< 2%), and high QoS parameters (i.e. delay > 400 ms, jitter > 20 ms, and packet loss
ratio > 2%) corresponded to a poor QoS (i.e. QoS < 33%).

Therefore, it can be observed from Table 8-3 that the values of evaluated QoS are
reflecting the corresponding QoS parameters. Low values of QoS parameters produce
172
cnapter 8 M icrocontroller n o a ra im plem en tation 01 v jo s A ssessm ent sy stem

high values of overall QoS and vice versa. For example, when the values of delay, jitter,
and packet loss ratio were 55.2ms, 11.32ms, and 0% respectively, the values of
expected QoS for video application were 77.14% when QoS monitoring device was
used, 80.69% when FIS technique was used, and 81.73% when using MLP neural
network. The expected QoS values were all in Good QoS region. In contrast, when the
values of QoS parameters were high (i.e. 533.4ms for delay, 24.5ms for jitter, and 3%
for packet loss ratio), the corresponding QoS were in poor region. The values of QoS
were 7.81% using QoS monitoring device, 9.81% using FIS technique, and 5.27% using
MLP neural network.
Table 8-3. QoS Parameters, and the evaluated QoS of video application using QoS
Monitoring Device (QoS_MD), Fuzzy Inference System (FIS), and Multi-Layer
Perceptron (MLP).

QoS Parameters Evaluated QoS

Delay (ms) Jitter (ms) PLR (%) QoS_MT FIS MLP

55.20 11.32 0.00 77.14 80.69 81.73

! 215.34 1.41 0.00 82.42 84.41 79.37

41.46 19.58 0.00 57.35 54.81 58.21

235.44 6.68 0.00 66.08 78.52 73.36

81.65 29.60 0.00 16.03 9.29 12.30

420.45 5.49 0.00 45.39 30.23 42.15

452.07 12.20 0.00 23.81 21.02 32.81

344.27 30.00 0.00 8.79 9.28 8.67

600.00 9.74 0.00 11.21 9.39 11.65

533.35 24.54 3.00 7.81 9.81 5.27

From Table 8-3, the results obtained from the QoS monitoring device, QoS assessment
using FIS technique, and MLP neural network QoS monitoring approach were
comparable. The correlation coefficient (R ) was used to evaluate the accuracy of the
proposed QoS device as compared with the other QoS assessment techniques.

In this study, the values of the correlation coefficients between QoS determined using
the monitoring system, and the other QoS assessment methods were: 0.98 for FIS
technique, and 0.99 for MLP neural network.
173
<^napier o iviicrucumruiicr dusiu liiipiemeiiutuuii ui y u j rtsstssmciu oyaicm

From the values of evaluated QoS obtained from the QoS monitoring system, QoS
assessment using FIS technique, and MLP neural network QoS monitoring approach as
provided in Table 8-3, it can be concluded that all approaches provided results which
were closely comparable. Nevertheless, some outputs were slightly different. These
divergences were because each method followed a different scheme to determine QoS.
However, the values of QoS obtained from the proposed QoS monitoring system
spanned between 1-100%, whereas the range of QoS values produced by FIS was
between 10-90%. This indicates that the QoS monitoring device could be more accurate.

8.7 Summary
In this study, a portable hand-held system to assess the QoS for multimedia applications
was designed and evaluated. Our developed QoS assessment technique which was
based on the regression model was implemented on the MCB2300 KEIL ARM
microcontroller board. The proposed system analysed the QoS parameters for
multimedia applications to measure the overall QoS. The QoS parameters (i.e. delay,
jitter, and packet loss ratio) were fed into the proposed device which in turn produced a
single value that represented the overall QoS.

The QoS assessment results were highly correlated with results obtined from a number
of previously developed QoS assessment methods. This indicated the correctness of the
developed system in monitoring QoS.

Further discussion about the results obtained from the developed approaches in this
thesis will be presented in the next chapter. Chapter 9 will conclude the thesis and
provide recommendations for future work.

174
 Chapter 9 Conclusions, and Future Work

9.1 Conclusions
In this research, network QoS management referred to evaluation and improvement of
QoS provided by wired and wireless computer networks. Therefore, the main focus was
on development of techniques to evaluate QoS in multimedia networks and the use of
this information as part of network management to improve its performance.

In chapter 5, statistical adaptive sampling techniques to adjust sampling rate based on

traffic's statistics were developed and evaluated. Three adaptive statistical sampling
techniques were proposed to sample multimedia traffic. The sampling rates of the three
devised sampling techniques were controlled using three different mechanisms: simple
linear adjustment mechanism, quarter adjustment mechanism, and fuzzy inference
system. The proposed adaptive statistical sampling techniques decreased the sampling
rate when the statistics of the traffic did not significantly change over time (i.e. steady
traffic) and increased the sampling rate when the statistics of the traffic significantly
changed with time (i.e. time varying traffic). A comparison of adaptive statistical
sampling techniques versus conventional sampling techniques (i.e. systematic sampling,
stratified sampling, and random sampling) was also carried out.

The main QoS parameters of VoIP traffic (i.e. throughput, delay, jitter, and packet loss
ratio) with their sampled versions using adaptive and non-adaptive sampling techniques
were discussed in chapter 5. It was concluded from the findings in chapter 5 that the
sampled versions of throughput, delay, jitter, and packet loss ratio obtained using the
adaptive statistical sampling approaches were closer to the actual population than the
non-adaptive sampling approaches (i.e. systematic, stratified, and random sampling).
For instance, the bias values of sampled jitter versions obtained from adaptive sampling
based on fuzzy approach, linear adjustment approach, and quarter adjustment approach
were closer to zero as compared with non-adaptive sampling approaches as shown in
Figure 5-12 (See chapter 5).

This concludes that the developed adaptive statistical sampling methods were more
effective than conventional sampling methods in representing the traffic. This is because
the sample interval was adjusted during the sampling process in case of adaptive

175
Chapter 9 conclusions, ana ruiure vvo i k

statistical sampling approaches based on linear adjustment mechanism, quarter

adjustment mechanism, and FIS, whenever the calculated overall traffic statistic
changed significantly over time. Conversely, the sampling rates of the conventional
sampling techniques were either constant as in systematic sampling or changed
randomly as in stratified and random samplings. The fixed and random sampling rates
resulted in a significant discrepancy between the actual data and its sampled version.
The advantages of the proposed adaptive statistical sampling techniques were the ease
of implementation and their ability to be implemented in real time.

Chapter 6 of this research is about development of techniques to analyse and assess QoS
parameters of multimedia networks accurately in real time. The multimedia QoS
information collected by adaptive statistical sampling techniques was considered.

Two innovative QoS evaluation approaches that combine analysis and measurement
techniques were developed. The first approach combined FCM and regression model to
analyse and assess QoS of multimedia applications in a simulated network. The second
approach analysed and assessed QoS in multimedia applications using a combination of
supervised and unsupervised neural networks.

The contribution of chapter 6 was to analyse and classify network QoS parameters (i.e.
delay, jitter, and packet loss ratio) using either Fuzzy C-means (FCM) or Self
Organizing Maps (SOM). Another contribution was to propose QoS assessment
techniques. The proposed techniques assess QoS in a manner similar to human subjects
and quantified the QoS without the necessity for complex mathematical models. Also,
the proposed QoS assessment techniques did not add significant extra load to the
network. The proposed assessment techniques were based on the traffic generated from
the proposed analysis techniques. A regression model was developed and a multi-layer
perceptron (MLP) was trained to combine the QoS parameters (i.e. delay, jitter, and
packet loss ratio) for each QoS class identified by SOM or FCM to estimate the overall
QoS. Both regression model and MLP were capable of combining QoS parameters (i.e.
delay, jitter, and packet loss ratio) to provide overall QoS.

The accuracy of QoS evaluation approaches was examined using typical values of QoS
parameters for VoIP, and a video. The findings of chapter 6 showed that FCM and
Kohonen network classified the values of QoS parameters of transmitted VoIP and
video into clusters representing Low, Medium, and High values of QoS as illustrated in
Figures 6-8, 6-9, 6-12, and 6-13. The regression model and MLP in turn combined the
176
uiajJier y VsUHClllMUIla, illlU CU1UIC VVU11V

QoS parameters (i.e. delay, jitter, and packet loss ratio) for each centre of generated
clusters and produced a single value that represented the overall QoS. The overall QoS
was an accurate indication of network performance as indicated in Tables 6-3, and 6-4.

Chapter 7 of this thesis considers the use of QoS information as part of multimedia
network management to improve its performance. Therefore, another object of this
study was about deployment network QoS enhancement which can be an effective
solution for multimedia applications to be shared under finite network resources.

A new QoS enhancement scheme for WLAN-wired networks was proposed. The
devised enhancement scheme consisted of an adaptive Access Category (AC) traffic
allocation algorithm which was incorporated into the network's wireless side to improve
the performance of IEEE 802.1 le Enhanced Distributed Channel Access (EDCA)
protocol, and a Weighted Round Robin (WRR) queuing scheduling mechanism that was
incorporated into the wired side of the network.

The algorithm considered the Packet Arrival Rate (PAR) of the uplink and downlink
traffic for each access category (AC), in order to allocate traffic to an appropriate AC.
PAR was chosen because it is an important QoS parameter that affects most real-time
and non-real-time applications. In addition, PAR can be easily computed while
applications are being transmitted. This in turn facilitated quick allocation o f the arrived
traffic to the most appropriate AC. Once PAR values for the up-link and down-link
traffic for each AC were determined, the algorithm dynamically allocated the traffic of a
lower priority AC to the next higher AC, when the higher AC was not receiving traffic
at the time. On the wired side of the network, a weighted round robin (WRR) shared the
network resources, based on the traffic’s quality o f service requirements.

The performance of the proposed scheme was compared with the standard IEEE
802.1 le EDCA and FIFO queuing mechanisms (i.e. WLAN-wired network legacy
scheme). The incorporation of the scheme enhanced the performance o f the WLAN-
wired network and significantly improved the QoS for transmitted applications. The
average QoS for VoIP, video, best effort traffic, and background traffic were increased
from their original values by 72.5%, 70.3%, 44.5%, and 45.2% respectively.

The QoS scheme proposed allowed an end-to-end QoS to be set up. This in turn
provided an improved delivery of a variety of applications in the context of wired-cum-
wireless networks.

177
Chapter 9 i^unciusiuus, aim r u iu re v t u i iv

In chapter 8, a network QoS monitoring device was designed and evaluated. The
existing QoS assessment devices do not determine directly the overall network QoS.
Network managers have to carry out a variety of tasks to assess the overall network
QoS. This makes the process complicated, expensive, and time consuming. Therefore,
developing a hand-held system that accurately assesses QoS for multimedia applications
is very valuable. The proposed QoS monitoring device used the QoS assessment
approach which was based on regression model which is described in section 6.3.3,
Chapter 6. The QoS assessment approach was implemented on the MCB2300 KEIL
ARM microcontroller board to design a hand-held device that assessed QoS of
multimedia applications.

The QoS parameters (delay, jitter and packet loss ratio) for multimedia applications
were fed into the proposed device to determine their overall QoS. The QoS monitoring
device analysed to QoS parameters of multimedia applications based on their QoS
requirements, and then produced an output that reflected their overall QoS. The
performance of the proposed device was compared with other QoS assessment
methods, it was observed from the findings that the overall QoS values obtained from
the proposed device were highly correlated with the QoS values obtained from QoS
assessment using Fuzzy Inference System introduced by (Al-Sbou et al, 2006), and
Neural Network QoS monitoring approach proposed by (Dogman et al, 2012b). The
obtained results indicated the effectiveness of the developed device in monitoring
multimedia QoS accurately.

9.2 Future Work

Although many solutions to improve multimedia network operation were developed in
this study, there remains several ongoing research and development follow ups. These
include:

• Implementation of the Proposed Approaches in Physical Networks: The

execution and validation of the proposed approaches were carried out by simulations
in this study. The implementations these approaches in physical networks can further
demonstrate their effectiveness.
• Incorporating QoS into Call Admission Control (CAC): The overall assessed
QoS should be used to manage the utilisation of available network resources. The
value of QoS can be integrated into CAC algorithm in future studies. The purpose of

178
u ia p ie r y cunciusiuiis, aim r u i u r e vvorK

CAC is to determine whether a new traffic should be admitted into the network. This
operation depends on many factors. The factors could include the overall QoS, QoS
requirements for new traffic, and the state of the network. Consequently, the QoS
requirements of admitted traffic will be satisfied.
• Proposed Adaptive Sampling Techniques and QoS Evaluation Methods over
other Packet Networks: Although the focus of this thesis was on wired-wireless
networks, the sampling approaches and the QoS evaluation methods can be applied
to other networks such as Mobile Ad hoc Networks (MANETs). This is because
adaptive sampling and QoS evaluation approaches are based on the traffic QoS
requirements, and measured QoS parameters.

• Implementation of Adaptive Statistical Sampling Algorithm into Hardware:

Investigating how an adaptive statistical sampling approach can be implemented into
hardware as a System-on-Chip (SoC) is another area of further development. The
chip could be integrated into network monitoring points such as routers in order to
sample transmitted traffic.
• Extend the Investigation of Queuing Mechanism: The impact of queuing
mechanisms on application's QoS is a source o f further studies. There are many
queuing mechanisms. The advantages of these mechanisms, and their effects on
traditional and real time applications should be studied further.

Finally, this thesis contributed significantly to the field of QoS management in

multimedia computer networks. The developed techniques drew a realistic scenario
about the process of managing QoS in multimedia networks. Also, they provided a firm
basis and useful insights on how to effectively design future QoS management
solutions.

179
 References

Abdel-Hady, M. and Ward, R. (2007), A Framework for Evaluating Video

Transmission over Wireless Ad Hoc Networks, IEEE Pacific Rim Conference on
Communications, Computers and Signal Processing (PACRIM), pp. 78 - 81.

Abdul Aziz, S. and Parthiban, J. (2006), Fuzzy Logic, [Online] Last access on 08 March
2013 at URL:
http://www.doc.ic.ac.uk/~nd/surprise_96/journal/vol4/sbaa/report.fuzzysets.html

Abeysekera, H., Matsuda, T., and Takine, T. (2009), Dynamic Contention Window
Control Scheme in IEEE 802.l i e Wireless LANs, In IEEE 69th Vehicular Technology
Conference, (VTC), pp. 1-5.

Abraham, A. (2005) Artificial neural network, Handbook o f Measuring System Design,

John W ily and Sons, pp. 901-908.

Aho, A.V., Kemighan, B.W., and Weinberger, P.J. (1988), The AWK Programming
Language, Addison-Wesley.

Akinaga, Y., Kaneda, S., Shinagawa, N., and Miura, A. (2005), A Proposal for a Mobile
Communication Traffic Forecasting Method using Time Series Analysis for Multi­
variant Data, IEEE Global Telecommunications Conference (GLOBECOM), pp. 1119 -
1124.

Alahmadi, A. A., and Madkour, M. A. (2008), Performance Evaluation o f the IEEE

802.1 le EDCA Access Method, IEEE International Conference on Innovations in
Information Technology (IIT), pp. 490-494.

Alkahtani, A., Woodward, M., and Al-Begain, K. (2003), An Overview of Quality of

Service QoS and QoS Routing in Communication Networks, In Proceedings o f 4th
PGNet Conference, Liverpool, pp. 236-244.

Al-Kuwaiti, M., Kyriakopoulos, N., Hussein, S. (2007), QoS Mapping: A Framework

Model for Mapping Network Loss to Application Loss, IEEE International Conference
on Signal Processing and Communications (ICSPC), pp. 1243-1246.

Al-Sbou, Y (2006), Quality of Service Assessment and Analysis o f Wireless

Multimedia Networks, PhD Thesis, Sheffield Hallam University, UK.

Al-Sbou, Y (2010), Fuzzy Logic Estimation System of Quality o f Service for

Multimedia Transmission, International Journal o f QoS Issues in Networking
(IJQoSIN), Vol. 1, No. 1, pp. 1 - 9.

Al-Sbou, Y., Saatchi, R., Al-khayatt, S., and Strachan, R. (2006), Quality of Service
Assessment of Multimedia Traffic over Wireless Ad hoc Networks, International
Symposium on Communication Systems, Networks and Digital Signal Processing
(CSNDSP), pp. 129-133.

180
Al-Sbou, Y.A., Saatchi, R., Al-Khayatt, S., Strachan, R., Ayyash, M., and Saraireh, M.,
(2008), A Novel Quality of Service Assessment of Multimedia Traffic over Wireless Ad
hoc Networks, The Second International Conference on Next Generation Mobile
Applications, Services and Technologies, (NGMAST), pp. 479-484.

Arshdeep, K., Amrit, K. (2012), Comparison of Mamdani-Type and Sugeno-Type

Fuzzy Inference Systems for Air Conditioning System, International Journal o f Soft
Computing and Engineering (IJSCE), Vol. 2, Issue 2, pp. 323 -325.

Bajaj, S., Breslau, L., Estrin, D., Fall, K., Floyd, S., Haidar, P., and Zappala, D. (1999),
Improving Simulation for Network Research. Defence Advanced Research Projects
Agency (DARPA), Virtual Internetwork Testbed project (VINT), pp. 1-11.

Baldwin, R. (1999), Improving the Real-Time Performance of a Wireless Local Area

Network, PhD Thesis, Virginia Polytechnic Institute and State University, Blacksburg,
Virginia.

Balogh, T., and Medvecky M. (2011), Performance Evaluation of WFQ, WF2Q+ and
WRR Queue Scheduling Algorithms, The 34th International Conference on
Telecommunications and Signal Processing (TSP), pp. 136-140.

Black, D., Blake, S., Carlson, M., Davies, E., Wang, Z., and Weiss, W. (1998), An
Architecture for Differentiated Services. Informational RFC 2475, pp. 1-35

Box, G., and Jenkins, G, (1976) Time Series Analysis: Forecasting and Control,
Holden-Day.

Braden, R., Clark, D., and Shenker, S. (1994), Integrated Services in the Internet
Architecture: an Overview. Informational RFC 1633, pp. 1-32

Brase, C. H., and Brase, C. P. (2010), Understanding Basic Statistics, The 5th edition
Brooks/Cole.

Brauer, F., Ehsan, M. S., and Kubin, G. (2008), Subjective Evaluation of

Conversational Multimedia Quality in IP Networks, IEEE 10th Workshop on Multimedia
Signal Processing, pp. 872-876.

Brekne, T., Clemetsen, M., Heegaard, P., Ingvaldsen, T., and Viken, B. (2002), State of
the Art in Performance Monitoring and Measurements, pp. 1-96.

Caky, P., Klimo, M., Paluch, P., and Skvarek, O. (2006), End-to-End VoIP Quality
Measurement, Acta Electrotechnica et Informatica, Vol. 6, No. 1, pp. 1-5.

Caro, G. (2003), Analysis of Simulation Environments for Mobile Ad hoc Networks,

Microsoft Academic Search, Technical Report. No. IDSIA-24-03.

Carvalho, P., Lima S. R., Ferreira, A., Freitas, E., and Leifao, F. (2009), Providing
Cost-Effective QoS Monitoring in Multiservice Networks, Next Generation Internet
Networks (NGI), pp. 1 - 8 .

Chaabane, S., Sayadi, M., Fnaiech, F., and Brassart, E. (2008), Colour Image
Segmentation using Automatic Threshold and the Fuzzy C-Means Techniques, The 14th
IEEE Mediterranean Electrotechnical Conference (MELECON), pp. 857 - 861.

181
Chan, A., Zeng, K., Mohapatra, P., Lee, S., and Banerjee, S. (2010), Metrics for
Evaluating Video Streaming Quality in Lossy IEEE 802.11 Wireless Networks, IEEE
INFOCOM, pp. 1 - 9

Chatterjee, S., and Hadi, A. (2006), Regression Analysis by Example, 4th edition. John
Wiley & Sons, Inc, pp. 53-84.

Chen, B., Hu, J., Duan, L., and Gu, Y. (2009), Network Administrator Assistance
System Based on Fuzzy C-Means Analysis, Journal of Advanced Computational
Intelligence and Intelligent Informatics, pp. 91-96.

Cheong, F., and Lai, R. (1999), QoS Specfication and Mapping for Distributed
Multimedia Systems: A Survey o f Issues. The Journal of Systems and Software, Vol.
45, Issue. 2, pp. 127-139.

Chung and Claypool (2004) NS By Example, [Online] Last access at 15 February 2010
at URL: http://nile.wpi.edu/NS/

Cirstea, M., Dinu, A., Khor, J., and McCormick, M. (2002), Neural and Fuzzy Logic
Control of Drives and Power Systems. Oxford.

Claffy, K.C., Polyzos, G.C., and Braun, H.W. (1993), Application of Sampling
Methodologies to Network Traffic Characterization, Conference Proceedings on
Communications Architectures, Protocols and Applications, pp. 194-203.

Cranley, N., and Davis, M. (2005), Performance Evaluation of Video Streaming with
Background Traffic over IEEE 802.11 WLAN Networks, In Proceedings o f the 1st
ACM Workshop on Wireless Multimedia Networking and Performance Modeling, New
York, NY, USA, ACM, pp. 131-139.

Crawley, E., Nair, R., Rajagopalan, B., and Sandick, H. (1998), A Framework for QoS-
Based Routing in the Internet, RFC2386, pp. 1-37.

D ’Antonio, S., Esposito, M., Gargiulo, M., Romano, S., and Ventre, G. (2003), A
Component-Based Approach to SLA Monitoring in Premium IP Networks, In First
International Workshop on Inter-Domain Performance and Simulation IPS, Salzburg,
Austria, pp. 1-8.

Dogman, A., Saatchi, R., and Al-Khayatt, S. (2010a), An Adaptive Statistical Sampling
Technique for Computer Network Traffic, In Proceedings o f the 7th IEEE, IET
International Symposium on Communication Systems Networks and Digital Signal
Processing (CSNDSP10), 21st -23rd July, University o f Northumbria, Newcastle, UK,
pp. 479 - 483.

Dogman, A., Saatchi, R., and Al-Khayatt, S. ( 2 0 1 0 b ) , Evaluation of Adaptive Statistical

Sampling versus Random Sampling for Video Traffic, In Proceedings o f the I I th
International Arab Conference on Information Technology (ACIT10), 14th-16th
December, University of Benghazi, Benghazi, Libya.

Dogman, A., Saatchi, R., Nwaizu, H., and Al-Khayatt, S. (2011), Adaptive Statistical
Sampling of VoIP Traffic in WLAN and Wired Networks using Fuzzy Inference
System, In Proceedings of the 7th IEEE International Wireless Communications and
Mobile Computing Conference (IWCMC11), 4th - 8th July, Bahcesehir University,
Istanbul, Turkey, pp. 1731 - 1736.

182
Dogman, A., Saatchi, R., and Al-Khayatt, S. (2012a), Quality of Service Evaluation
using a Combination of Fuzzy C-Means and Regression Model, International Journal
o f Electronics and Electrical Engineering, World Academy o f Science Engineering and
Technology (WASET12), Vol. 6, pp. 58 - 65.

Dogman, A., Saatchi, R., and Al-Khayatt, S. (2012b), Evaluation of Computer Network
Quality of Service Using Neural Networks, In Proceedings o f IEEE Symposium on
Business, Engineering and Industrial Applications (ISBEIA12), 23rd - 26th September,
Bandung, Indonesia, pp. 217 - 222.

Dogman, A., Saatchi, R., and Al-Khayatt, S. (2012c). Improving Quality of Service in
IEEE 802.1 le Enhanced Distributed Channel Access Protocol, In Proceedings o f the 8th
IEEE International Symposium on Communication Systems Networks and Digital
Signal Processing (CSNDSP12), 18th - 20th July, Poznan University, Poznan, Poland,
pp. 1 - 6.

Dogman, A., Saatchi, R., and Al-Khayatt, S. (2012d), Computer Network Quality of
Service Monitoring Using KEIL ARM Microcontroller, In Proceedings o f Regional
Annual Fundamental Science Symposium (RAFSS12), 10th - 13th December, Persada
Johor International Convention Centre, Johor Bahru, Malaysia.

Dogman, A., Saatchi, R., and Al-Khayatt, S. (2013), Network Quality o f Service
Assessment Implementation Using a Microcontroller Board, Malaysian Journal of
Fundamental and Applied Sciences, Vol.9, No.2, pp. 57-61.

Eberhart, R. C., and Dobbins, R. W., (1990) Neural Network PC Tools: A Practical
Guide, Academic Press, ISBN 0-12-228640-5.

Elashheb, A. (2012) Performance Evaluation of AODV and DSDV Routing Protocol in

wireless sensor network Environment, International Conference on Computer Networks
and Communication Systems (CNCS), pp 55 -62

Epiphaniou, G., Maple, C., Sant, P., and Reeve, M. (2010), Affects of queuing
mechanisms on RTP traffic: comparative analysis of jitter, end-to-end delay and packet
loss. The 10th IEEE International Conference on Availability, Reliability, and Security
ARES'10, pp. 33-40.

Fall, K., and Varadhan, K. (2011) The NS Manual, Virtual Internetwork Testbed project
(VINT).

Farrel, A. (2008) Network quality of service, know it all, Pearson Education, London,
UK.

Floyd, S., Allman, M. (2008) Comments on the Usefulness of Simple Best-Effort

Traffic, Informational RFC 5290, pp 1 - 19

Fraleigh, C., Tobagi, F. and Diot, C. (2003) Provisioning IP backbone networks to

support latency sensitive traffic, In Twenty-Second Annual Joint Conference o f the
IEEE Computer and Communications Societies IEEE INFOCOM. pp. 375-385.

Fraley C, and Raftery, A. (1998). “How Many Clusters? Which Clustering Method? -
Answers via Model-based Cluster Analysis.” Computer Journal, 41, pp 578-588.

Frantti, T., and Jutila, M. (2009), Embedded fuzzy expert system for adaptive weighted
fair queuing. Expert Systems with Applications, 36(8), 11390-11397

183
Gan, Y., Zhang, Y. and Qian, D. (2009) Adaptive sampling measurement for high speed
network traffic flow, Proceedings o f the 5th International Conference on Wireless
Communications, Networking and Mobile Computing IEEE Press, pp. 4085-4088.

Georgoulas, S., Trimintzios, P. and Pavlou, G. (2004) Admission Control Placement in

Differentiated Services Networks, Ninth International Symposium on Computers and
Communications (ISCC), pp 816 - 821

Giertl, J., Baca, J., Jakab, F. and Andoga, R. (2008) Adaptive sampling in measuring
traffic parameters in a computer network using a fuzzy regulator and a neural network,
Cybernetics and Systems Analysis, vol. 44, no. 3, pp. 348-356.

Giertl, J., Jakab, F., Baca, J., Andoga, R., and Mirilovic, M. (2006) Contribution to
adaptive sampling of QoS parameters in computer networks, Acta Electrotechnica et
Informatica, vol. 6, no. 1, pp. 1-8.

GloMoSim. GloMoSim Manual (2012) [Online] last access at 25 January 2012 at URL:
http://pcl.cs.ucla.edu/projects/glomosim/GloMoSimManual.html

Gozdecki, J., Jajszczyk, A. and Stankiewicz, R. (2003) Quality o f service terminology

in IP networks, IEEE Communications Magazine, vol. 41, no. 3, pp. 153-159.

Graciva, R. S., Aparicio, A. C., and Pascual, J. D. (2008) Network performance

assessment using adaptive traffic Sampling, In proceedings o f the 7th International
IFIP-TC6 Networking Conference on Ad-Hoc and Sensor Networks, Wireless Networks,
and Next Generation Internet, pp 252-263

Graham, I., Donnelly, S., Martin, S., Martens, J., and Cleary, J (1998) Nonintrusive and
Accurate Measurement of Unidirectional Delay and Delay Variation on the Internet,
Proceeding of INET, Switzerland

Gupta, S. and Saket, R. (2011) Peformance Metric Comparison of AODV AND DSDV
Routing Protocols in MANETs using NS-2, International Journal o f Research and
Reviews in Applied Sciences, Vol. 7, Issue 3, pp 339 - 350

Haman, A, and Geogranas, N (2008) Comparison of Mamdani and Sugeno Fuzzy

Inference Systems for Evaluating the Quality of Experienceof Hapto-Audio-Visual
Applications, IEEE International Workshop on Haptic Audio Visual Environments and
their Applications (HAVE), pp 8 7 - 9 2

Han, J and Kamber, M (2006) Data Mining: Concepts and Techniques, 2nd edition.

Hasegawa, G., Matsuo, T., Murata, M., and Miyahara, H. (2002) Comparisons of packet
scheduling algorithms for fair service among connections on the Internet, Journal o f
High Speed Networks, 72(1), 1-27.

Haykin, S. (1999) Neural Networks: A Comprehensive Foundation, 2nd ed. Englewood

Cliffs, NJ: Prentice-Hall.

He, Y., Ma, X., (2011). Deterministic backoff: Towards efficient polling for 802.1 le
HCCA in wireless home networks, IEEE Transactions on Mobile Computing 10 (12) pp
1726 - 1740.

Heckmann, O., Pandit, K., Schmitt, J., Hoffmann, M., and Jobmann, M. (2002)
LETSQoS milestone 2, network QoS , LETSQoS project, Darmstadt University o f
Technology and Technical University Munich, pp. 1-106.
184
Hernandez, E.A., Chidester, M.C. and George, A.D. (2001) Adaptive sampling for
network management, Journal o f Network and Systems Management, vol. 9, no. 4, pp.
409-434.

Hoang, D., Kumar, R., and Panda, S. (2010) Fuzzy C-means clustering protocol for
wireless sensor networks, IEEE International Symposium on Industrial Electronics
ISIE, pp 3477-3482.

IEEE Computer Society LAN/MAN Standards Committee (1999), Wireless LAN

medium access control (MAC) and physical layer (PHY) specifications.

IEEE Computer Society LAN/MAN Standards Committee (2005), Wireless LAN

Medium Access Control (MAC) and Physical Layer (PHY) specifications Amendment
8: Medium Access Control (MAC) Quality o f Service Enhancements.

ITU-T E.800 (1994) Terms and definitions related to quality of service and network
performance including dependability, Telecommunication Standardization Sector o f
ITU, pp. 1-57.

ITU-T P.910 (2008) Subjective video quality assessment methods for multimedia
applications, Telecommunication Standardization Sector o f ITU. pp. 1-42.

ITU-T P.911 (1998) Subjective audiovisual quality assessment methods for multimedia
applications, Telecommunication Standardization Sector o f ITU. pp. 1-27.

ITU-T, G.1010 (2001) End-user multimedia QoS categories, Telecommunication

Standardization Sector of ITU. pp. 1-18.

Jain, R. (1991) The Art of Computer Systems Performance Analysis: Techniques for
Experimental Design, Measurement, Simulation, and Modelling. Wiley- New York,
ISBN: 0471503363.

Jain, R. (2008) A Survey of Network Simulation Tools: Current Status and Future
Developments, [Online] at URL: http://www.cse.wustl.edu/~jain/cse567-08/ftp/simtools

Jantzen, J., (1998) Design of Fuzzy Controllers. Technical report, University of

Denmark, Department of Automation, Vol: 236.

Karlik, B and Olgac A., (2010) Performance analysis of various activation functions in
generalized mlp architectures of neural networks. International Journal o f Artificial
Intelligence and Expert Systems, Vol. 1, Issue 4., pp 111-122.

Ke,C.H., Shieh, C.K., Hwang, W.S., and Ziviani, A (2008) An Evaluation Framework
for More Realistic Simulations of MPEG Video Transmission, Journal o f Information
Science and Engineering, Vol.24, No. 1, pp 425-440

Khoukhi, L., and Cherkaoui, S. (2008) Experimenting with Fuzzy Logic for QoS
Management in Mobile Ad Hoc Networks. HCSNS International Journal of Computer
Science and Network Security, VOL.8 No.8, P 372-386

Kiziloren, T., and Germen, E. (2007) Network traffic classification with self organizing
maps, International Symposium on Computer and Information Sciences ISCIS, pp. 1-5.

Klir, J. and Yuan, B. (1995) Fuzzy Sets and Fuzzy Logic: Theory and Applications.
Prentice Hall Inc. ISBN: 0131011715.

185
Kohonen, T. (1982) Self-organized formation of topologically correct feature maps.
Biological Cybernetics, pp: 59-69.

Koutsakis, P. (2009) Dynamic versus Static Traffic Policing: A New Approach for
Videoconference Traffic over Wireless Cellular Network, IEEE TRANSACTIONS ON
MOBILE COMPUTING, VOL. 8, NO. 9, pp 1153-1166

Kumar, D., Ryu, Y. and Jang, H. (2008) Quality of service (QoS) of voice over MAC
protocol 802.11 using NS-2.ACM, International Multimedia Conference, P. 39-44.
ISBN:978-l-60558-319-8.

Kurose, J. and Ross, K. (2005) Computer networking a top-down approach, Fifth

edition, Addison-Wesley, Reading.

Lavington, S., Hagras, H., and Dewhurst, N. (1999) Using a MLP to predict packet loss
during real-time video transmission, Tech Report, University o f Essex, pp 1-10.

Lee, K. Y., Cho, K. S., and Ryu, W. (2011), Efficient QoS Scheduling Algorithm for
Multimedia Services in IEEE 802.1 le WLAN, IEEE Vehicular Technology Conference
(VTC Fall), pp. 1-6).

Lei, X., Yifei, G., Sheng, L., Wei, H., and Di, X (2012) Research on the Improved FCM
Cluster Method in the Hotspots Analysis on Web, The 2nd IEEE International
Conference on Instrumentation, Measurement, Computer, Communication and Control
(IMCCC), pp 478 - 481.

Lekcharoen S. (2007) Performance and Evaluation of Adaptive Backoff Schemes in

Traffic Shaping over High Speed Network, Proceedings of Asia-Pacific Conference on
Communications, pp 241 -245

Liang, H. M., Ke, C. H., Shieh, C. K., Hwang, W. S., and Chilamkurti, N. K. (2006)
“Performance evaluation of 802.1 le EDCF in infrastructure mode with real audio/video
traffic”, IEEE International Conference on Networking and Services ICNS, pp. 92-97.

Lin, C. H., Shieh, C. K., Ke, C. H., Chilamkurti, N. K., and Zeadally, S. (2009), An
adaptive cross-layer mapping algorithm for MPEG-4 video transmission over IEEE
802.11 e WLAN. Telecommunication Systems, 42(3-4), 223-234.

Lindfield, G. R., and Penny, J. E. T. (2012) Numerical methods: using MATLAB,

Elsevier Science & Technology.

Liu, L., Jin, X., Min, G., and Xu, L. (2012), Real-Time Diagnosis of Network Anomaly
Based on Statistical Traffic Analysis. In IEEE 11th International Conference on Trust,
Security and Privacy in Computing and Communications (TrustCom), pp. 264-270.

Ljung, L. (1999) System Identification: Thoery for the User, 2nd edition, Prentice Hall.

Lucio et al (2003) OPNET Modeler and Ns-2: Comparing The Accuracy of Network
Simulators for Packet-Level Analysis Using a Network Testbed. International
Conference on Simulation, Modelling and Optimization (ICOSMO). P 1-8.

Ma, W., Huang, C. and Yan, J. (2004) Adaptive sampling for network performance
measurement under voice traffic, IEEE International Conference on Communications,
vol. 2, pp. 1129-1134.

186
Ma, W., Yan, J. and Huang, C. (2003) Adaptive sampling methods for network
performance metrics measurement and evaluation in MPLS-based IP networks, In
Proceedings of the Canadian Conference on Electrical and Computer Engineering
Cite seer, pp. 1005-1008.

Malan, G., and Jahanian, F. (1998) An extensible probe architecture for network
protocol performance measurement, ACM SIGCOMM Computer Communication
Review, vol. 28, no. 4, pp. 215-227, 1998.

Malhotra, M., Varma, D. P., Pramod, P. J., and Jain, D. K. (2011), StreamTrack: A tool
for QoS monitoring o f multimedia sessions. The 6th IEEE International Conference on
In Telecommunication Systems, Services, and Applications (TSSA), pp. 121-125.

Markopoulou, A., Tobagi, F., and Karam, M. (2003) Assessing the Quality o f Voice
Communications over Internet Backbones, IEEE/ACM TRANSACTIONS ON
NETWORKING, VOL. 11, NO. 5, pp 747 - 760

Mathworks (2012(a)) Fuzzy Logic Toolbox. [Online] last access on 18 February 2013 at
URL: http://www.mathworks.com/access/helpdesk/help/toolbox/fuzzy

Mathworks (2012(b)) Neural Netowrk Toolbox. [Online] last access on 08 March 2013
at URL: http://www.mathworks.co.uk/help/nnet/index.html

MATLAB, The Language of Technical Computing (2012) [Online] last access at 17

March 2012 at URL: http://www.mathworks.co.uk/products/matlab/

MDK-ARM Microcontroller Development Kit [Online] last access on 15 September

2013 at URL: http://www.keil.com/arm/mdk.asp.

Menth, M., Martin, R. and Charzinski, J. (2006) Capacity over provisioning for
networks with resilience requirements, In Proceedings o f the 2006 Conference on
Applications, Technologies, Architectures, and Protocols fo r Computer
Communications, New York, NY, USA, ACM, pp. 87-98.

Miaji, Y., and Hassan, S. (2010), Comparative Simulation of Scheduling Mechanism in

Packet Switching Network. The 2n IEEE International Conference on Network
Applications Protocols and Services (NETAPPS), pp. 141-147.

Mishra, R., and Sharma, V. (2003) QoS routing in MPLS networks using active
measurements, IEEE Conference on Convergent Technologies fo r Asia-Pacific Region
(TENCON), Citeseer, pp. 323- 327.

Mishra, R., and Sharma, V. (2003) QoS routing in MPLS networks using active
measurements, IEEE Conference on Convergent Technologies fo r Asia-Pacific Region
(TENCON), Citeseer, pp. 323- 327.

Mohamed, S. (2002) A study o f real-time packet video quality using random neural
networks, IEEE Transactions on Circuits and Systems for Video Technology, Vol. 12,
No. 12, pp 1071-1083.

Mohamed, S., Rubino, G., Cervantes, F. and Afifi, H. (2001) Real-time video quality
assessment in packet networks: A neural network model, IEEE Transactions on Circuits
and Systems fo r Video Technology, pp. 1-21

187
Molina-Jimenez, C., Shrivastava, S., Crowcroft, J. and Gevros, P. (2004) On the
monitoring of contractual service level agreements, In First IEEE International
Workshop on Electronic Contracting, pp. 1-8.

Muyeen, S. M., and Al-Durra, A (2013) Modeling and Control Strategies of Fuzzy
Logic Controlled Inverter System for Grid Interconnected Variable Speed Wind
Generator, IEEE System Journal, issue 99, pp 1-8.

Naoum-Sawaya, J., Ghaddar, B. (2005) A Fuzzy Logic Approach for Adjusting The
Contention Window Size in IEEE 80 2 .l i e Wireless Ad Hoc Networks. International
Symposium on Communication, Control, and Single Processing (ISCCSP). P 1-4

Nascimento, S., Mirkin, B., and. Moura-Pires, F (2000) A fuzzy clustering model of
data and fuzzy c-means, IEEE International Conference on Fuzzy Systems, vol. 1, pp.
302-307.

Nogueira, A., Salvador, P., and Valadas, R. (2006) Predicting the quality of service of
wireless LANs using neural networks, Proceedings of the 9th ACM International
Symposium on Modeling Analysis and Simulation of Wireless and Mobile Systems
ACM, New York, NY, USA, pp. 52-60.

Nortal Networks (2003) Introduction to quality of service, White Paper, pp. 1-14.

NS Network Simulator 2 (2012) [Online] last access at 17 January 2012 at URL:

http://www.isi.edu/nsnam/ns/

Olios, G., and Vida, R. (2009) Adaptive regression algorithm for distributed dynamic
clustering in wireless sensor networks, IEEE International Federation fo r Information
Processing IFIP, pp 1-5.

OMNeT++ (2012) [Online] last access at 03 February 2012 at URL:

http://www.omnetpp.org/

Pal,T., Shekhar, C., and Sharma, H. D. (2009) “Design and Implementation of

Embedded Speed Controller on ARM for Micro-manufacturing Applications,” IEEE
International Conference on Advances in Computing, Control, and Telecommunication
Technologies, pp. 406 - 410.

OPNET (2012) OPNET Technologies [Online] last access at 10 February 2012 at URL:
http://www.opnet.com/

Palomar, D., Skehill, R., Rice, I., Picovici, D., and Me Grath, S. (2008) Objective
assessment of audio quality, IET Signals and Systems Conference, ISSC. Irish, pp. 37-
42.

Park, S., Kim, K., Kim, D., Choi, S., and Hong, S. (2003) “Collaborative QoS
architecture between DiffServ and 802.1 le wireless LAN,” The 57th IEEE Semi annual
Vehicular Technology Conference VTC, Vol 2, pp 945-949.

Parker, J. and Hall, L. Bezdek, J. (2012) Comparison of Scalable Fuzzy Clustering

Methods, IEEE World Congress on Computational Intelligence (WCCI), Brisbane,
Australia, pp 1 - 9.

188
Patcha, A., and Park, J. M. (2006) An adaptive sampling algorithm with applications to
denial-of-service attack detection, 15th IEEE International Conference on Computer
Communications and Networks, ICCCN, pp. 11-16

Patel, T., Ogale, V., Baek, S., Cui, N. and Park, R. (2003) Capacity estimation of VoIP
channels on wireless networks, Dept o f Electrical and Computer Engineering, The
University o f Texas at Austin, pp. 1-37.

Peng, F. Peng, B. and Qian, D. (2010) “Performance analysis of IEEE 802.l i e

enhanced distributed channel access”, IET Communications, Vol. 4, Issue. 6, pp. 7 2 8 -
738.

Pias, M., and Wilbur, S. (2001) “EdgeMeter distributed network metering model”, IEEE
Global Telecommunications Conference, GLOBECOM, pp.2517-2521.

Politis, A. Mavridis, I. Manitsaris, A. and Hilas, C. (2011) “X-EDCA: A cross-layer

MAC-centric mechanism for efficient multimedia transmission in congested IEEE
802.l i e infrastructure networks” IEEE International Wireless Communications and
Mobile Computing Conference IWCMC, pp. 1724-1730.

Priestly, M., (1988) Non Linear and Non Stationary Time Series Analysis, Academic
Press.

Radhakrishnan, K., and Larijani, H. (2011) Evaluating perceived voice quality on

packet networks using different random neural network architectures, Performance
Evaluation, Vol. 68, pp 347-360.

Rashid, M. M., Hossain, E., and Bhargava, V. K. (2007), HCCA scheduler design for
guaranteed QoS in IEEE 802.l i e based WLANs, IEEE Wireless Communications and
Networking Conference, (WCNC), pp. 1538-1543.

Rauf, B., Amjad, M.F. and Ahmed, K. (2009) “Performance evaluation of IEEE 802.11
DCF in comparison with IEEE 802.1 le EDCA”, IEEE International Conference for
Internet Technology and Secured Transactions, ICITST, pp. 1-6.

Rokach, L. and Maimon, O (2005) Clustering Methods, Data Mining and knowledge
discovery handbook, pp 321 - 352.

Ross (1995) Fuzzy Logic with Engineering Applications. McGraw-Hill, USA. ISBN
007-113637-1

S. Park, S. Kim, K., Kim, D., Choi, S., and Hong, S. (2003) “Collaborative QoS
architecture between DiffServ and 802.l i e wireless LAN”, The 57th IEEE Semi annual
Vehicular Technology Conference VTC, Vol 2, pp. 945-949.

Said, A., and R. Saatchi, R. (2009) A Performance Analusis of AODV AND DSR
Routing Protocols in Mobile Wireless Networks, The Mediterranean Journal o f
Computers and Networks, Vol. 5, No. 1, pp 17 - 26

Saliba, A.J., Beresford, M.A., Ivanovich, M. and Fitzpatrick, P. (2005) User-perceived

quality of service in wireless data networks, Personal and Ubiquitous Computing, vol.
9, no. 6, pp. 413-422.

Saraireh, M. (2006) Medium access control mechanisms for quality of service in

wireless computer networks, PhD Thesis, Sheffield Hallam University, UK.

189
Saraireh, M. Saatchi, R. Al-Khayatt, S. and Strachan, R. (2008) “Assessment and
improvement of quality of service using fuzzy logic and hybrid genetic-fuzzy
approaches”, ACM., vol 27, issue 2-3, P.95 - 1 1 1 , ISSN:0269-2821.

Saraireh, M., Saatchi, R., Al-Khayatt, S., and Strachan, R. (2007) Assessment and
improvement of quality of service using fuzzy logic and hybrid genetic-fuzzy
approaches, ACM., vol. 27, issue 2-3, pp. 95 - 111.

Sarkar, N, and Halim, S (2011) A Review of Simulation of Telecommunication

Networks: Simulators, Classification, Comparison, Methodologies, and
Recommendations, Journal of Selected Areas in Telecommunications (JSAT), pp 10 -
17
Semeria, C. (2001) “Supporting differentiated service classes: queue scheduling
disciplines”, White Paper Juniper Networks, pp. 1-27, 2001.

Senkindu, S. and Chan, H.A. (2008) “Enabling end-to-end quality o f service in a

WLAN-wired network,” IEEE Military Communications Conference, MILCOM, pp.l-
7.

Shah, I. (2001) Brining comprehensive quality of service capabilities to next generation

networks, white paper, pp 1-20

Shah, S., Khandre, A., Shirole, M., and Bhole, G. (2008) Performance Evaluation of Ad
Hoc Routing Protocols Using NS2 Simulation”, Conference of Mobile and Pervasive
Computing (CMPC), pp 167 - 171

Siraj, S., Gupta, A., and Badgujar, R (2012) Network Simulation Tools Survey,
International Journal of Advanced Research in Computer and Communication
Engineering, Vol. 1, Issue 4, pp 201 - 210

Skyrianoglou, D., Passas, N., Salkintzis, A. and Zervas, E. (2002) “A generic adaption
layer for differentiated services and improved performance in wireless networks”, The
13th IEEE International Symposium on Personal, Indoor and Mobile Radio
Communications PIMRC, Vol 3, pp. 1141-1145.

Sun, L. and Ifeachor, E. (2002) Perceived speech quality prediction for voice over IP-
based networks, In IEEE International Conference on Communications, ICC, pp. 2573-
2577.

Sweet, S. and Grace-Martin, K. (2010) Data Analysis with SPSS: A First Course in
Applied Statistics, 4th edition.

Szigeti, T., and Hatting, C. (2005). End-to-end qos network design. Cisco Systems.

Teyeb, O., Sprensen, T., Mogensen, P., and Wigard, J. (2006) Subjective evaluation of
packet service performance in UMTS and heterogeneous networks, The 2nd ACM
International Workshop on Quality o f Service and Security fo r Wireless and Mobile
Networks ACM, New York, NY, USA, pp. 95-102.

Timo, L., Hannu, K., and Tapani, H. (2002) Profiling network applications with fuzzy
C-means clustering and SOM, International Conference on Fuzzy Systems and
Knowledge Discovery, pp 1-5.

190
Ting, B., Yong, W., and Xiaoling, T. (2010) Network traffic classification based on
kernel self-organizing maps, International Conference on Intelligent Computing and
Integrated Systems ICISS, pp. 310 - 314.

Thangaraj, S., Gummadi, S., and Radhakrishnan, S. (2006) “Enhancement in ARM

Code Optimization for Memory Constrained Embedded Systems,” IEEE Advanced
Computing and Communications, pp. 483 - 486.

Vandaele, W., (1983) Applied Time Series and Box-Jenkins Models, Academic Press.

Vesanto, J., Himberg, J., Alhoniemi, E., and, Parhankangas, J. (1999) Self-organizing
map in matlab: the SOM toolbox, In Proceedings ofM atlab DSP Conference, Espoo,
Finland, pp. 35-40.

Villalon, J., Cuenca, P., and Orozco-Barbosa, L. (2007), on the capabilities of IEEE
802.l i e for multimedia communications over heterogeneous 802.11/802.l i e WLANs,
Telecommunication Systems, 36(1-3), 27-38.

Wang, P.Y., Yemini, Y., Florissi, D., Zinky, J. and Florissi, P. (2000) Experimental
QoS performances of multimedia applications, IEEE INFOCOM, Citeseer, pp. 970-979.

Wang, T., Liang, Z., and Zhao, C. (2009) A detection method for routing attacks of
wireless sensor network based on fuzzy C-means clustering, IEEE Sixth International
Conference on Fuzzy Systems and Knowledge Discovery FSKD, pp445-449.

Weingartner, E., Lehn, H., and Wehrle, K. (2009) A performance comparison of recent
network simulators, IEEE International Conference on Communications, ICC, pp 1 - 5

Xia W., and Wang R. (2010) Prediction Model of Network Security Situation Based on
Regression Analysis, IEEE International Conference on Wireless Communications,
Networking and Information Security (WCNIS). pp 616 - 619.

Xie, X. and Beni, G. G. (1991) A validity measure for fuzzy clustering, IEEE Pattern
Analysis and Machine Intelligence, pp 841-847.

Xin, Q., Hong, L., and Fang, L. (2009). A modified FLC adaptive sampling method, In
Proceedings o f IEEE International Conference on Communications and Mobile
Computing (CMC09), 6th - 8th Januarey, China University of Petroleum, China, pp 515-
520.

Yu, Z., and Meng, F. (2009) An improved algorithm for packet fair queuing scheduling,
2nd IEEE International Conference on Computer Science and Information Technology,
ICCSIT, pp 153-157

YUV QCIF Reference Videos. (2012, January 7) [Online]. Available at:

http://www.tkn.tu-berlin.de/research/evalvid/qcif.html

Zadeh, L. (1965) Fuzzy sets, in Information and Control, vol. 8, Pages: 338-353.

Zaknich, A. (2003) Neural networks for intelligent signal processing, World Scientific
Pub Co Inc, New Jersey.

Zhai, H., Chen, X. and Fang, Y. (2005) How well can the IEEE 802.11 wireless LAN
support quality of service?, IEEE Transactions on Wireless Communications, vol. 4, no.
6, pp. 3084-3094.

191
 Zhang, L. and Zhang, T. (2007) Small packet threshold adaptive sampling for network
management, In International Conference on Mechatronics and Automation ICMA. pp.
811-815.

Zhou, R. and Sik-Jang, K. (2008) Adaptive MPEG-4 Video Streaming over IP

networks, the 23 International Technical Conference on Circuits, Systems, Computer,
and Communication, (ITCCSCC), pp 637 - 640.

Zseby, T. (2002) Deployment of sampling methods for SLA validation with non-
intrusive measurements, Pass and Active Measurement workshop (PAM), USA, pp.84-
94.

Zseby, T. (2004) Comparison of sampling methods for non-intrusive SLA validation,

Second Workshop on End-to-End Monitoring Techniques and Services E2EMON, pp.
1- 8 .

Zseby, T. and Scheiner, F. (2004) QoS Monitoring and Measurement Benchmarking,

Information Society Technology. [Online] Last accessed on 06 April 2013 at URL:
www.ist-mome.org/documents/M_and_M_benchmark.pdf

1i

| 192