. .. . V , . . J. i..."
Eur: . . 4. . . . ‘ , . V , .1: WaEhawwmusw
. . , V . . , . 3. (”5.4 y. .3
.3; .wwu

 

.u
3.2: .1.
.

 

 

 

$33.”. 1.3.?

 

 

 

This is to certify that the
thesis entitled

METHODS FOR MAPPING SCALABLE VIDEO OVER
DIFFERENTIATED SERVICES NETWORKS

presented by

QAZI MUHAMMAD RASHID UL HAQ

has been accepted towards fulfillment
of the requirements for

M. S . degree in W
Engineering

 

 

 

Date .261- 2/ 257/2,

0-7639 MS U is an Affirmative Action/Equal Opportunity Institution

 

 

LIBRARY

Michigan State

University

 

PLACE IN RETURN BOX to remove this checkout from your record.
To AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE

DATE DUE

DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6/01 cJCIRC/DateDue.p65-p.15

METHODS FOR MAPPING SCALABLE VIDEO OVER

DIFFERENTIATED SERVICES NETWORKS

By

Qazi Muhammad Rashid Ul Haq

A THESIS

Submitted to
Michigan State University
in partial fulfillment of the requirements

for the degree of
MASTER OF SCIENCE
Department of Electrical and Computer Engineering

2002

ABSTRACT

METHODS FOR MAPPING SCALABLE VIDEO OVER
DIFFERENTIATED SERVICES NETWORKS

By
Qazi Muhammad Rashid Ul Haq

Differentiated Services (DiffServ), which are Quality-of-Service (QoS) based
networks for the Internet, provide diverse classes of reliable delivery of Internet packets
at a paid price. In this thesis, we propose methods for mapping scalable video streams
over the Assured Forwarding (AF) class of DiffServ. A key objective of the proposed
methods is to improve the overall video quality as measured by the packet loss ratio
(PLR) of the video layers in a scalable video stream. Furthermore, a ratio of the weighted
throughput of video packets to the associated cost is employed as a metric to evaluate the
performance of the proposed mapping methods. Two high-level mapping strategies are
described: a one-to-one strategy for mapping each video layer to a distinct AF priority
level, and a multiple-Io-one strategy for mapping multiple video layers to one AF priority
level. In order to achieve the desired weighted throughput of video packets in the later
strategy, we propose a linear pricing model (LPM) and an exponential pricing model
(EPM) for network bandwidth commitments. The aforementioned mapping strategies
were simulated using the Network Simulator (NS). These strategies were also analyzed
for various network overload conditions, which result as a consequence of bandwidth
over-commitment. Some of the key conclusions of this work include: (1) the EPM
provides the most desirable (weighted) PLR performance under zero over-commitment
conditions; (2) the LPM provides a desirable overall performance under a variety of

conditions.

To my Abbu Qazi Masroor Ul Haq and Ammi Sanjidah Masroor
~ and ~
their immeasurable prayers for every success in my life
~ and ~

their perpetual belief in my abilities

iii

ACKNOWLEDGEMENTS

I was very fortunate to have an able researcher and excellent teacher, Dr. Hayder
Radha as my advisor. I am most grateful to him for his scientific insights and guiding
strategies. His advice, criticism, encouragement, and support were invaluable throughout
the course of this research.

In addition, I would like to thank Dr. Percy Pierre and Dr. Michael A. Shanblatt for
their time and effort in being part of my committee. I would also like to extend special
thanks to Dr. Dimitri Loguinov for his interest, guidance as well as the insight he was
able to provide for this work during his short stay at Michigan State.

My gratitude is also extended to all my talented colleagues at the WAVES lab,
including, but not limited to, Irtiza, Shirish, Ali, Shahradha, and YongYing. Aparna R.
Gurijala deserves special thanks for guiding me during the final preparation of this thesis.
Moreover, I will always owe a special thanks to Khurram Bhai for providing me
invaluable advice throughout my graduate studies.

Words cannot express my thankfulness to Hashsham Bhai and Nihala Baji, for their
support, guidance and encouragement. They not only gave me a place in their home, but
also gave me a special place in their lives. I will not be able to forget and return their love
and care for the rest of my life. Their cute little sons Ahmad, Abdullah and Anser gave us
millions of moments to enjoy together.

I would also like to express my deepest appreciation to my family members back
home for their love, emotional support and sacrifice. Besides my parents, I owe special

thanks to Khalid Bhai, Lubna Baji, Tahir Bhai, Husna Baji, Anwar and my nephews

Haseeb, Habib, Faraz and specially Zoheb (arrived during this work, whom I never met).
Their support and encouragement helped make my graduate studies possible. I will never
be able to return what I have taken from all of them, family time!

Finally, my fiancee Bushra who came in my life at the beginning of this work, and
always remained their with an encouraging smile and love for me, all the time. I deeply
appreciate her love, care and patience for every decision I took during this whole time. I

will remain indebted to her, forever!

TABLE OF CONTENTS

LIST OF TABLES .............................................................................................. ix
LIST OF FIGURES .............................................................................................. x
1 Introduction ................................................................................................ 1

1.1 Motivation for This Work ........................................................................... 1

1.2 Problem Description ................................................................................... 4

1.3 Organization of the Thesis .......................................................................... 7
2 Background and Literature Review ........................................................... 9

2.1 Streaming Video over the Internet .............................................................. 9

2.1.1 Scalable Video over IP .................................................................. 11

2.1.2 Weight Distribution for Video Layers .......................................... 14

2.2 Differentiated Services for the Internet ..................................................... 16

2.2.1 DiffServ Architectural Overview .................................................. 17

2.2.2 Traffic Classification .................................................................... 18

2.2.2.1 DSCP and PHBs ............................................................ 19

2.2.3 Traffic Metering and Marking ...................................................... 22

2.2.4 Traffic Shaping ............................................................................. 24

2.3 Assured Forwarding PHB ......................................................................... 26

vi

2.4 Pricing and Cost Calculation for Assured Forwarding Class ................... 28

2.4.1 Flat Rate Pricing Technique .......................................................... 31
Mapping Strategies and Method of Analysis .......................................... 33
3.1 Packet Mapping ........................................................................................ 33
3.2 One-to-One Mapping ................................................................................ 36
3.3 Multiple-to—One Mapping ......................................................................... 37

3.3.1 Exponential Pricing Model (EPM) ............................................... 39

3.3.2 Linear Pricing Model (LPM) ........................................................ 40
3.4 Performance Evaluation ............................................................................ 41
Simulation Setup ...................................................................................... 47
4.1 Simulation Scenarios ................................................................................ 48
4.2 Network Topology and Statistics .............................................................. 48
4.3 Bitrate and Network Overload Configurations ......................................... 51
4.4 Traffic Conditioning Setup ....................................................................... 54

4.4.1 Traffic Metering and Marking ...................................................... 54

4.4.2 Traffic Shaping ............................................................................. 56
4.5 Weight Distribution for Video Layers ...................................................... 58
4.6 Network Fairness Criteria ......................................................................... 59
4.7 Summary of Simulation Steps .................................................................. 60
Simulation Results and Analysis ............................................................. 62

vii

5.1 Simulation Scenario 1: One-to-One Mapping ........................................... 62

5.2 Simulation Scenario H: Multiple-to-One Mapping ................................... 65
5.2.1 Simulation Scenario II (a): Multiple-to-One Mapping using Flat
Rate Pricing Model ....................................................................... 65
5.2.2 Simulation Scenario H (b): Multiple-to-One Mapping using
Exponential Pricing Model ........................................................... 69

5.2.3 Simulation Scenario H (c): Multiple-to-One Mapping using Linear

Pricing Model ................................................................................ 71

5.3 Comparison of the Pricing Models ........................................................... 75

5.4 A Comparative Performance Analysis ...................................................... 77
6 Conclusions .............................................................................................. 80
References............. ......................................................................................... 83
Appendix............... ........................................................................................ 87

viii

Table 2.1

Table 4.1

Table 4.2

Table 4.3

Table 4.4

Table 4.5

Table 5.1

Table 5.2

LIST OF TABLES

Assured Forwarding PHB code points ............................................................ 27
Statistics of Simulation Scenarios ................................................................... 48
Simulation Statistics ....................................................................................... 50
Distribution of bitrates (kbps) for video layers ............................................... 53
Selected values for trTCM parameters ........................................................... 55
Wei ghted-RED parameters for the simulations .............................................. 57
Comparison of PF loss in Simulation Scenarios ............................................ 78
Cost comparison of Simulation Scenarios ...................................................... 78

Figure 1-1
Figure 2-1
Figure 2—2
Figure 2-3
Figure 2-4
Figure 2-5
Figure 2-6
Figure 2—7
Figure 2-8
Figure 3-1
Figure 3-2
Figure 4-1
Figure 4-2
Figure 4-3
Figure 4-4
Figure 5-1
Figure 5-2
Figure 5-3
Figure 5-4

Figure 5-5

LIST OF FIGURES

Possible mappings of streaming video over DiffServ AF Class. .................... 7
Schematic diagram of streaming video encoder ........................................... 12

Layering of video stream (a) one BL and one EL (b) one BL and n-1 ELs.. 13

Weight distribution curve for video layers with a = 0.4, b = 0.7 .................. 15
A typical DiffServ Framework ...................................................................... l7
DiffServ Edge Router Components .............................................................. 18
Differentiated Services Code Point (DSCP) Field ........................................ 20
DiffServ Service Classes (PHB Tree) ........................................................... 21
A Typical RED Algorithm ............................................................................ 25
One—to-one Mapping ..................................................................................... 37
Multiple to One Mapping .............................................................................. 38
Network topology for the simulations ........................................................... 50
WRED parameters for priority queues .......................................................... 58
Employed weight distribution of scalable video layers ................................ 59

PLR of video sources at (a) scenario II(b) at 10% overload, and (b) scenario

II(c) at 50% overload ..................................................................................... 60
PLR for Simulation Scenario 1 ...................................................................... 63
Evaluated PF for Simulation Scenario 1 ........................................................ 64
PLR for Simulation Scenario II (a) ............................................................... 66
Evaluated PF for Simulation Scenario II (a) ................................................. 68
PLR for Simulation Scenario II (b) ............................................................... 70

Figure 5-6
Figure 5-7
Figure 5-8

Figure 5-9

Figure 5-10
Figure 5-11
Figure A-l

Figure A-2

Evaluated PF for Simulation Scenario II (b) ................................................. 71
PLR for Simulation Scenario II (c) ............................................................... 72
Evaluated PF for Simulation Scenario II(c) .................................................. 73

PLR and evaluated PF using two different fractional values of ,8 for Scenario

II(c) ................................................................................................................ 74
PLR comparison of pricing models at 0% overload ................................... 76
PLR comparison of pricing models at 100% overload ............................... 77
srTCM Algorithm flow diagram ................................................ 88
trTCM Algorithm flow diagram ................................................. 89

xi

1 Introduction

1.1 Motivation for This Work

Since its introduction as a packet network in the early seventies, the Internet [1] has
undergone a dramatic increase in its usage and enormous changes in its character. These
changes included the support of real-time and high-speed applications such as
audio/video streaming in addition to the traditional services like web browsing, email,
FTP and telnet.

The Intemet Protocol (IP) [2] promises a best eflort approach of data delivery [3][4].
Best effort means that IP finds the best possible path from source to destination for data
packet delivery without providing 100% guarantees that the data packets will actually be
delivered. In other words, and in general, IP routing protocol selects the shortest number
of hops to develop a path from source to destination without any delivery assurance.

The primary task of an IP router is to receive an incoming packet from one of its
interfaces and to forward the received packet to the next router. This routing is based on a
simple forwarding table lookup operation in conjunction with a set of routing tables that
are stored and updated periodically at each router. Therefore, depending on the packet’s
destination and port address, an IP router forwards the packet to the next hop router

according to the forwarding (routing) table. IP routers perform this task as fast as possible

on every packet they receive. When a large number of packets converge on the same
router within a very short time interval, and when the outbound links cannot service these
packets at the rate of their arrival, the packets get buffered at the router. This condition,
which results in excessive delays in packet delivery, is known as congestion. Congestion
could lead to packet loss events since routers start to discard incoming packets due to
limited buffer size.

Under ideal conditions, a packet network, such as the Internet, should guarantee the
delivery of every packet (i.e., high reliability) while supporting the desired sending rate
of any application. To achieve the first goal (i.e., packet delivery guarantees and high
reliability), transport-layer protocols, such as the well-known Transmission Control
Protocol (TCP) [7], was designed. However, TCP increases the reliability factor while
decreasing the sending data rate. Consequently, when TCP establishes an end-to-end
connection between the sender and the receiver, it requires the transmission of
acknowledgement for every packet from the receiver to deliver another set of packet(s).
This causes an application to reduce its sending rate. Delay-insensitive data like email,
ftp and WW can utilize TCP easily and with high reliability factor rather than
streaming real time applications.

Another transport layer protocol known as User Datagram Protocol (UDP) [5] gives
a provision to send data packets at the desired speed that is required by most interactive
real time streaming applications without any end-to-end connection establishment.
Consequently, relative to TCP, UDP reduces the transmission reliability while
maintaining a desired sending rate. Real time Protocol (RTP) [6], which was designed

for real time media applications, also recommends the use of UDP instead of TCP in

order to meet the transmission rate constraints required by real-time streaming
applications.

The employment of UDP over the Internet implies the occurrence of network
impairments such as packet loss events. These events are normally characterized through
the parameter packet loss ratio (PLR). PLR is ratio between the number of packets lost
(dropped) during transmission and the total number of packets sent by the sender. Other
network impairments, such as delay jitter and high roundtrip delays (also known as RTT)
occur over the Internet due to variation in the level of congestion. Since end-to-end
unreliable data delivery with jitter, high latency, and packet losses negatively affect the
user perceived media quality [10], Quality of Service (QoS) tools are being proposed.
QoS tools are expected to bring improvement in the network performance by providing
different PLR, latency and jitter characteristics to different types of packets [8][9].

While new QoS tools have been proposed and investigated, scalable video coding
techniques have been developed for transmission over networks that do not provide QoS
guarantees (such as the Internet). Both MPEG-2 [11][25] and MPEG-4 [26] [55]support a
variation of scalable compression methods. In such schemes, a video stream is composed
of a single base layer (BL) and one or more enhancement layers (ELs). The base layer
contains data required to produce a minimum acceptable quality video, where as every
enhancement layer, like the name states, enhances the video quality produced by the base
layer. Thus, packets containing the base layer always need a higher priority for
transmission than enhancement layer packets. For a layered video packet stream with n
layers, containing a base layer (i = 1) and n-1 enhancement layers, the probability P,- of

losing a packet in layer i should ideally be:

ping-+1 i=1,2,3....n (1.1)

or in terms of Packet Loss Ratio (PLR), the minimum requirement should be:

PLRi < Pure,-+1 (1.2)

Best effort Internet does not promise to satisfy the conditions mentioned in equations
(1.1) and (1.2) for streaming video applications. In order to ensure delivery of quality
video, every layer in a scalable video stream needs a distinct service behavior depending

upon its importance and priority.

1.2 Problem Description

Differentiated Services (DiffServ), which is one of the QoS routing protocols
suggested by the Internet Engineering Task Force (IETF) provides quantitative
classification of packets. Four major types of packet forwarding services (also known as
Per Hop Behaviors (PHB) Expedited Forwarding (EF) [12], Assured Forwarding (AF)
[13] and Best Effort (BE) are defined in the DiffServ charter. Depending upon the user-
service provider agreement known as Service Level Agreement (SLA), a packet can be
treated with one of these available packet-forwarding services.

This work aims to analyze different strategies for the transmission of layered scalable
video streams, which can be used for popular streaming applications, over a DiffServ
network node. We consider strategies that meet the PLR constraint expressed in equation
(1.2). We also analyze the performance of these strategies in terms of the weighted
throughput over some “cost” parameter. This parameter provides a measure of the “cost”
paid for utilizing differential treatment of layered packet video streams like the ones
supported by the MPEG-4 Fine-Granularity-Scalable (FGS) video standard [31].

Although the EF is the premium service that provides a low jitter and low latency service

and is considered ideal for real-time interactive applications (e.g., voice or video
telephony), the potential high cost associated with this type of service does not justify its
use for popular streaming applications since these applications do not require very
stringent low-delay requirements. Consequently, in this thesis, we focused on utilizing
the AF service as a suitable framework for streaming applications. We believe that AF
strikes a good balance between requiring reasonable “cost” (when compared to EF) and
providing some level of PLR guarantees (when compared with best effort Internet).

The AF service class offers three levels of packet drop precedences (i.e., three priority
levels). Thus scalable video layers (say n) in a stream can utilize an AF class either by
allocating packets of every distinct layer to a distinct priority level as shown in Figure
l-l(a), or by providing service priority within one or two priority levels as shown in
Figure 1-1(b). The first approach (i.e., associating each video layer with a corresponding
AF priority level) is not suitable in all cases since DiffServ does not allow more than
three-drop priorities within a single class. However, this approach is applicable when a
video stream contains three or less number of layers. For the video streams containing
more than three layers (which is common in today’s popular video compression
techniques) we preferred the second approach. In this approach we utilize the two priority
levels (Level 1 and Level 2) within a single AF class to provide differential treatment for
a scalable video stream in order to achieve a desired packet loss behavior as defined by
(1.2). In this thesis we propose two mapping strategies for scalable video over AF (as
described later).

Moreover, DiffServ is a commitment based service network; hence it provides

differential forwarding treatments to the packets according to a “paid price”. The

differential treatment of a packet is based on a commitment between the service provider
and the user. This commitment allows a user (or sender) to utilize a committed amount of
network bandwidth for a reliable delivery of its packets. Under normal conditions, a
service provider does not allow a user (or users) to exceed the available network
bandwidth. If a service provider allows a user (or users) to exceed the available network
bandwidth, such phenomenon is called as over-commitment (or overbooking). This work
also analyzes the effect of such over-commitments on the PLR performance of scalable
video transmission. The performance of video transmission is analyzed under several
network overload levels, which are applied as over-commitments of network bandwidth,
ranging from 0% to 100%.

The paid price for the user-service provider commitments can be represented by a
corresponding set of “pricing parameters”. We used such pricing parameters to define
approaches to improve the user video throughput by controlling PLR in a DiffServ
domain, at several network overload levels, according to equation (1.2). To achieve this
goal, in addition to the proposed mapping strategies we also proposed two pricing models
besides the traditional flat rate pricing technique.

We present results obtained by performing a number of simulations on the Network
Simulator (NS) [14] (version ns-2.lb8a) for each proposed mapping strategy and pricing
model. After evaluating the video performance, we compared the associated overall
“cost” of each strategy. Finally, we proposed a low-cost framework to attain a desired
video quality transmission and a sustained performance for several network overload

levels.

Video D3 Video DS
Source Edge Source Edge
Router Router

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

ELn . Level n EL“ \ ~—> Level 1
}| > Level 2
Level 3
(Best Effort)
b Level 3
EL
2 ELz AF Class
EL] . Level 2 EL1 J
BL p Level 1 BL
(a) (b)

Figure 1-1 Possible mappings of streaming video over DiffServ AF Class.

1.3 Organization of the Thesis

The next chapter provides a brief overview of layered video over IP and their coding
techniques, and work done in recent years to bring improvement in video quality and
performance. Chapter 2 also includes a high-level overview of the architecture, elements
and working details of Differentiated Services for the Internet. Chapter 3 covers the
proposed mapping strategies and method of performance evaluation. The setup and
description of simulation work is covered in Chapter 4. Chapter 5 includes the simulation
results for network PLR behavior and performance analysis of the proposed mapping

strategies and pricing models for scalable video streams over DiffServ. Finally, Chapter 6

concludes this thesis with an outline of the key findings of this work and suggestions for

future work.

2 Background and Literature Review

This chapter provides some background material for three basic elements of this
thesis. It starts with a brief overview of streaming video over the Internet and some
description of scalable video transmissions over IP in the first section. The second section
provides a detailed description of the architecture, standards and working of
Differentiated Services (DiffServ) networks. It also includes a detailed overview of the
assured forwarding (AF) class of DiffServ. The Weighted Random Early Detection
(WRED) mechanism, an extension of Random Early Detection (RED) mechanism, is
employed as the packet traffic shaping (queuing) mechanism for the simulations of
DiffServ network. Hence, a brief overview of the basic WRED mechanism is also
included in this section. As this thesis also focuses on the performance analysis, in which
cost and pricing play a significant role, the traditional flat rate pricing technique and cost
evaluation methods for DiffServ and its assured forwarding (AF) class are discussed in

last section.

2.1 Streaming Video over the Internet
Streaming is a technique of transmitting real-time multimedia data, commonly audio

and video, from a host to a client where the client "plays" or decodes the data as it is

received. Streaming differs from downloading in that streaming lets the viewer play-back
the content in real-time after a short buffering delay, which could be up to few seconds.
This buffering allows the application to maintain a continuous playback even during
minor network congestions. UDP is the transport protocol used for interactive streaming
video applications over the Internet. The Real Time Protocol (RTP) [6] and Internet
Multicast Backbone (Mbone) [21] also recommends the use of UDP for video
transmissions. While giving a provision of sending data at any desired bitrate, UDP does
not specify any error recovery mechanism; therefore it is an unreliable transport protocol.
During congestion, the unreliable approach of UDP results in an increased packet loss
ratio (PLR), which consequently reduces the user-perceived video quality.

A number of approaches for UDP transmission improvement using adaptive rate
control methods were presented in [27]. In these approaches, the sending bitrate of
transmission can be reduced in response to a “packet loss” notification reported by the
receiver. For this purpose, use of real time control protocol (RTCP) is recommended.
Such adaptive rate control methods are helpful in improving the overall losses, but do not
provide any assurance to maintain the video quality and transmission performance during
multiple levels, ranging from 0% to 100% network congestion. Error concealment
methods, for example Forward Error Correction (FEC), and retransmission techniques for
multicast streaming [28] were also suggested to bring improvement in UDP
transmissions. The error concealment methods help to improve the user perceived video
quality, but are only applicable after loss notifications.

Focusing over the available approaches of video packet transmissions, motion JPEG

and certain wavelet-based schemes use the Intra-frame coding techniques. Whereas

10

popular video standards like H.261 [22], H.263 [23], MPEG1 [24], MPEG2 [25] and
MPEG4 [26] use inter-frame coding technique. Inter-frame coded video packet streams
are more sensitive to error and packet losses during their transmission. Every packet loss
during transmission affects the corresponding frame and error(s) in one frame propagates
to the rest. Mechanisms like High Priority Protection method (PEP?) [29] and Fine-
Grained Loss Protection (FGLP) [30] are suggested to bring improvement in the packet

loss resilience of MPEG video streams.

2.1.1 Scalable Video over IP

Typical scalable video applications (using standard like MPEG-4) send video streams
composed of one base layer (BL) and one or multiple enhancement layers (ELs). For this
purpose, a scalable video encoder generates two bit streams; one for the base layer and

the other(s) for the enhancement layer(s). These streams are encoded for transmission at

some bitrate rb for the base layer and re for the enhancement layer. Thus total bitrate

(Rmm) of a video packet stream is the sum of individual bitrates of base layer and

enhancement layer:

Rmax = rb + re (2.1)

Under normal conditions, the total bitrate remains less then the available network link
capacity (C):

Rmax < C (2.2)

Subsequently, the video server packetizes the video bit streams in packets of specified

size, and transmits packets over network link according to a corresponding bitrates. This

sequence of processes at the video server is illustrated by Figure 2-1.

11

At the receiver end, the data packets are received at their respective bitrates. The
decoder then reconstructs the original bit streams from such packets. Base layer packets
contain the data required to produce a minimum acceptable quality video. Whereas
enhancement layer provides additional information, which enhances the quality of video

produced by the base layer.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

BL bit
stream BL bit stream rb BL stream with
enCOder ...0100101011101 rate ’b
Packetizer
5:23;: EL stream with
encoder EL bit stream r, rate r,.
..0101000011101

 

Figure 2-1 Schematic diagram of streaming video encoder.

A typical scheme of encoding base layer and enhancement layer of a video frame is

shown in Figure 2-1. The video layers are encoded at their corresponding bitrates rb and
re. A single enhancement layer with a high sending bitrate re, can be divided into multiple

enhancement layers with lower bitrates r,, such that ri<re. A typical method is

partitioning of the high sending bitrate re of one enhancement layer into equally divided

(say n-l) small enhancement layers with lower sending bitrate r,-:

 

’i‘ = r" wherei=1--o(n—1)
n—1
n—1
:> re = Z ’2 (2.3)

12

Such partitioning of a high sending bitrate re of one enhancement layer into (n-l)

enhancement layers is shown in Figure 2-2.

The base layer and the enhancement layer(s) are inter-related with each other for a
particular frame as shown by the dotted box in Figure 2-2 (a), and Figure 2-2(b). The
occurrence of any packet loss during the transmission of base layer makes the whole
sequence of enhancement layers (for that particular frame) useless. Such dependence
scheme of video layers clearly requires the following simple priority scheme in terms of
PLR, during transmission as mentioned by equation (1.2):

PLR,- < PLR,“ (2.4)

where i is the number of corresponding video layer with base (i =1) as the first layer in

the sequence.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

[r ----------- : I.n-l
re ‘r'i E El?"
5 EL E EL EL EL r4 Eh
i f 3 1a.,
5 i r2 EL,
Yb '1’ : I1
5 BL i BL BL BL EL:
1 ............ I], if
BL
(21) (b)

Figure 2-2 Layering of video stream (a) one BL and one EL (b) one BL and n—1 ELs.

l3

2.1.2 Weight Distribution for Video Layers

As described later in this thesis, identifying the overall “cost” of a scalable-video
mapping-strategy over DiffServ networks requires a quantitative measure for the level-of-
importance associated with the different video layers. It is clear from the above
description of scalable video that the base-layer, for example, should encounter the
minimum (or a virtually null) PLR value. The first enhancement layer should encounter a
PLR smaller than the second enhancement layer PLR, and so on. This, however, does not
provide a direct quantitative measure for the relative importance of the different layers.

In this thesis, weight is a term used for indicating the comparative importance of one
video layer to another. The higher the weight measure of a particular video layer is, the
higher the “performance cost” (or performance penalty) that is associated with loosing
data from that layer.

Assigning a quantitative measure for the different layers of a scalable video stream is
a very challenging task. This is due to the variety of scalable video coding schemes, and
more importantly, is due to the wide range of possible video sequences, each of which
could have a different weight measure model [55]. Here, we resort to a generic weight
measure that captures a variety of prioritization models for scalable video layers.

In a scalable video stream containing one base layer and (n-1) enhancement layers,
one possible approach for assigning weight (w) is based on the following exponential

equation:

wj = ae_(j_l)b wherej = 1,2,3..(n — 1), 0 S a <1, b 2 0 (2.5)

where j is the corresponding enhancement layer.

14

In the above equation, the parameter a denotes the weight assigned to the first
enhancement layer, while variable b represents the exponential decay in the weight of
higher (less important) video layers. Since the base layer usually carries the highest
priority (or maximum weight), then the weight wo used for the base-layer should be
higher than the parameter a. In this thesis, we normalize this measure by using wo = 1 >
a.

As an example, a weight distribution curve for a scalable video stream containing one
base layer and 9 enhancement layers is shown in Figure 2-3. This curve is obtained by
using equation (2.5) with parameters a = 0.4 and b = 0.7. The same weight distribution

curve is used for the analysis later in this thesis.

Weight Distribution in Scalable Video Layers

 

 

 

 

I T I T T r I I
0.9 - ..
0.8 - l
0.7 - .1
0.6 t- _
E0
5 0.5 t' -l
3
0.4 r ‘
0.3 *- '4
0.2 " ‘
0.1 '- "
1 l l l I Y ‘ A
BL EL1 EL2 EL3 EL4 ELS EL6 EL7 ELB EL9

Video Layers

Figure 2-3 Weight distribution curve for video layers with a = 0.4, b = 0.7.

15

2.2 Differentiated Services for the Internet

Proposals that have been put forward by the IETF for improving the Internet QoS are
based on two types of approaches. The first approach is Fine Grained, which provides
QoS for individual applications and packet flows, and the second approach is Coarse
Grained, which provides QoS to the larger classes of data or aggregated packet traffic.
The proposals that belong to these two approaches are considered as deployable QoS
models in the form of Integrated Services (IntServ) [32] and DiffServ [33] respectively.
Due to the complexity associated with providing QoS guarantees to individual flows (i.e.,
IntServ), DiffServ has been receiving a great deal of attention as a realistic and practical
alternative for providing some form of QoS services over the global Internet.
Consequently, in this thesis, we limit our discussion to DiffServ-based scalable video
services.

Implementation of DiffServ routing over the Internet needs: (i) a proper definition of
the packet classes, (ii) method(s) of mapping each packet to a corresponding class, and
(iii) a suitable allocation of network resources to each class. These steps are referred to as
the standardization processes. From the anival of a packet at the DiffServ domain
boundary router, its classification, sorting, queuing and forwarding towards its
destination are performed under a number of “standardized” mechanisms [34]. IETF
DiffServ working group defined the directions to be taken for the implementation of such
mechanisms in the form of Internet drafts and requests for comments (RFCs), which are

easily accessible on the Internet [51].

16

2.2.1 DiffServ Architectural Overview

DiffServ network is a simple, well-defined set of building blocks [35]. Figure 2-4
shows a typical DiffServ domain. In addition to the links and the switches, a DiffServ
domain contains two types of network nodes that are referred to as, Edge Routers

(boundary nodes) and Core Routers.

   
   

DiffServ Domain

 

 

 

 

 

 

 

 

    

Edge Router/
Ingress Node Egress Node

    

Figure 2-4 A typical DiffServ Framework.

The edge routers are the DiffServ domain ingress/ egress nodes. This implies that
every packet entering or leaving the DiffServ domain needs to pass through an edge
router. An edge router is responsible for Trafi‘ic Conditioning, which includes the
classification and appropriate forwarding of packets according to their corresponding
Service Level Agreement (SLA). The SLA is a commitment that is established between
the service provider and the user (normally a sender). The primary task of an edge router
is to classify and shape the incoming packet stream from a user according to its
predefined service profile in the SLA. The SLA service profile contains the committed

rate and packet burst size values. The edge routers compare the incoming packet

17

parameters (for example bitrate) with a corresponding specified value in the SLA profile.
The packet(s) exceeding these profile values are classified as out of profile and are
eligible to get downgraded or even discarded at the DiffServ domain boundary. The in
profile packets are eligible to be forwarded towards their destinations according to their
given priorities. If required, core routers can also perform an additional traffic
classification.

Figure 2-5 shows the elements of an edge router, which are required to perform traffic
conditioning. These elements are: (i) Classifier, (ii) Meter, (iii) Marker, and (iv) Shaper/

Dropper. The function of each element is described in the subsequent sections.

In Profile Packets

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

, M 1‘
—" c ” Out of Profile
Packets
Shaper /
:> Classifier —‘/,___l\ Marker ——l/__l\ Dropper :
Incoming Outgoing

 

 

 

Packets without

. . . . Packet stream
initial marking

Packet stream

Figure 2-5 DiffServ Edge Router Components.

2.2.2 Traffic Classification

According to the DiffServ standard specifications [33][34], the classifications of
every incoming packet in a DiffServ domain can be performed either by using a
Behavioral Aggregate (BA) Classification [33] method, or by a Multi Field (MF)

Classification [49] method. A set of packets sharing the same level of network resources

18

across a DiffServ domain in one direction is referred to as behavioral aggregate (BA).
BA classification requires a proper definition of a DiflServ Code Point (DSCP) value in
the packet header [33]. Such code point helps an edge router classifier to identify and
allocate the corresponding service level to the packet.

The other method is multi field (MF) classification, which uses the existing fields in
the packet header (for example source address and destination address) rather than using
any specific codes for packet classification.

In this thesis, we focus on the BA classification method. This method gives a simple
one to one correspondence between DSCPs and the service level allocation. The DiffServ
module in the simulator used in this thesis (i.e., NS) also supports the same method of
packet classification. The details about the DSCPs and their corresponding service levels,

also known as Per Hop Behaviors (PHBs) [48], are discussed in the following section.

2.2.2.1 DSCP and PHBs
A DSCP is a 6-bit binary code in the packet header. The DSCP value [48] of a given

packet carries the information about the class and the level of service that should be
provided to that packet. Figure 2-6 shows an 8-bit DiffServ field used for the DiffServ
code point where six bits are defined for DSCP values and two bits are reserved for future
use, and consequently are labeled as currently unused (CU). In order to provide a
compatibility with the current IPv4, DiffServ uses the Type of service (ToS) byte in the IP
packet header for DSCP markings, whereas Class of Service (CoS) field in [P header of
the future IPv6 is proposed for DSCP.

Every defined DSCP value corresponds to a service level aggregate or a PHB. A PHB

is a service behavior of a DiffServ node, which refers to the appropriate packet

19

scheduling and forwarding of each packet sharing the same BA [49]. A service provider
is the one who decides the network resource allocation to a PHB for each class.
Depending upon the level of packet forwarding service provide by a DiffServ node, the
PHBs are divided into four categories. Each category is defined below, whereas a

complete PHB tree is shown in Figure 2-7.

o/r/z/‘3/4/5 6/7

 

 

DS§CP CU
Class Selector Currently
/ Unused
W
DSCP

Figure 2-6 Differentiated Services Code Point (DSCP) Field.

Default PHB: A code point ‘000000’ is assigned to default PHB, which gets the
traditional best effort service[34]. Also this default PHB is assigned to every packet
arriving at a DiffServ domain edge router without any code point in its header.

Class Selector PHB: A class selector PHB provides backward compatibility of
DiffServ with previous IP precedence schemes [33]. DSCP values in the form of
‘xxx000’ are termed as class selector PHBs. Thus, according to such format of DSCP

values for class selector PHB, the default PHB can also be classified as a class selector

PHB [49].

20

Expedited Forwarding (EF) PHB: The highest prioritized service within the
DiffServ PHB suite is expedited forwarding (EF) [12]. It can be termed as the “premium”
service class of DiffServ[34] because it provides a low latency-low jitter service, which is
considered ideal for voice and video transmissions over IP. Such premium service level
that is provided by the EF class also makes it the highest priced class in the DiffServ
PHB suite. Since EF has no further sub-classes defined, thus all packets carrying a DSCP
value for EF i.e. 101110 are treated equally[34][37]. Hence, no service differentiation is
expected between the packets utilizing the EF service.

In this thesis we focused on utilizing the assured forwarding (AF) class of services. A

detailed discussion regarding AF classes and services is presented in section 2.3.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

PHB
AF EF Class Selector Default / BE
AFl AFZ AF3 AF4
l—y AFll + AFZI + AF31 + AF4]
» AF12 _> AF72 AF32 + AF42
A1313 PM AF23 -> AF33 A1343

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 2-7 DiffServ Service Classes (PHB Tree).

21

2.2.3 Traffic Metering and Marking

When a packet enters in a DS domain with a corresponding DSCP in its header, an
edge router meters such packet by using a metering-marking algorithm. Metering is the
method of comparing (or testing) the packet parameters (such as bitrate, burst size etc.)
with the predefined SLA profile values. A SLA profile depicts the threshold values for
such parameters. As a result of metering, a packet may be categorized as an iii-profile or
an out-of-profile packet. If the packet’s parameters are found exceeding the given
threshold limits, the corresponding packet is categorized as out of profile, otherwise it
becomes an in-profrle packet. As mentioned earlier, an out-of-profile packet is eligible to
be downgraded to a lower priority level within an AF class, where it might get discarded.

Metering and marking are two sequel packet processes at an edge router. In other
words, marking of a packet “within” a DiffServ domain depends upon the result of packet
metering. Moreover, marking of a packet within a DiffServ domain is considered as ‘re-
marking’ unless a packet anives at an ingress node without any DSCP in its header.

There are three popular metering-marking algorithms designed for DiffServ:

1. Single Rate Three Color marker (srTCM) [38],

2. Two Rate Three Color marker (trTCM)[39], and

3. Time Sliding Window Three Color Marker (TSWTCM) [40].
TSWTCM [40][41] was specifically developed to support TCP traffic. The other two
algorithms srTCM and trTCM can be used for metering of UDP traffic. In this thesis we
used srTCM and trTCM as our marking algorithm. These algorithms are described in

Appendix at the end of this thesis.

22

There are two bitrate thresholds that are used in metering-marking algorithms:
Committed Information Rate (CIR) and Peak Information Rate (PIR)]. CIR is the
minimum network bandwidth, which a service provider is committed to provide a user
for a certain flow of packets, for a guaranteed delivery of packets. Basically, CIR is the
throughput rate that a user negotiates with a service provider, and the service provider
will guarantee that rate. One way the service provider guarantees CIR is by dropping
non-CIR traffic.

PIR is the peak rate of a packet flow that is allowed within a profile. Typically PIR is
not less than CIR, and not greater than the available link capacity C. CIR does not limit
the user to send data at the committed bitrate, but the user should not violate the
predefined threshold limits of CIR in order to keep the packets in-profrle.

Threshold parameters for limiting the burstiness of incoming data streams are:
Committed Burst Size (CBS) and Peak Burst Size (PBS). Burstiness is the maximum
amount of bits that a network agrees to support during a certain time interval. These burst

sizes are related to their corresponding rates by:

 

CIR ___ CBS
C (2 6)
PIR : P_BS
TP

where TC and Tp are the corresponding time intervals to evaluate CIR and PIR.

The two bitrates (CIR and PIR) along with their corresponding burst sizes (CBS and

PBS) work as two thresholds limits for incoming packet traffic in a DiffServ domain.

 

 

isrTCM algorithm does not employ PIR as a parameter, however the other two algorithms use CIR and PIR
both.

23

2.2.4 Traffic Shaping

Traflic Shapers are the packet queuing mechanisms at the edge router. After metering
and marking, packets are allowed in the buffer queues according to their respective
marked or re-marked DSCP. During this temporary storage period, service differentiation
is provided by means of a packet queuing algorithm.

Random Early Detection (RED) [15] is one of the congestion avoidance mechanisms
that has been successfully deployed in the Internet. It was initially proposed to provide
scalability in TCP transmissions. RED gives a provision to control the packet losses
during congestion by means of its parameters. While focusing over the AF class of
DiffServ, which defines three levels of packet drop precedences, RED can be utilized to
establish multiple queues. According to the predefined priorities of each queue, RED
parameters can be selected. Such scheme of deploying multiple RED queues with
different set of parameters for each queue is called multiple RED (MRED) or weighted
RED (WRED)[50][52]. WRED is a successfully deployed queuing mechanism in Cisco
routing software IOSTM release 12.2 [53] to support DiffServ. According to the WRED
mechanism, one physical RED queue is divided into multiple virtual RED queues; each
virtual RED queue use different set of RED parameter values and follows the basic RED
algorithm. The outline of RED algorithm is as follows:

RED Algorithm: The RED algorithm is based on continuous calculations performed
for every incoming packet to measure the average length (angn) of a packet queue in the
router. Typically, RED requires two queue length thresholds minm and max,;, and a drop

probability maxp. The RED algorithm for packet queuing and discard can be written as:

24

i. If the average length of the queue drops below the minimum threshold value
minm, RED allows the packet in the buffer, else

ii. If the average length of the queue becomes in between min,;. and max,;,, RED
discards incoming packets randomly with a drop probability of maxp , as
shown in Figure 2-8, else

iii. If the average queue length exceeds the maximum threshold max,;,, every

incoming packet gets discarded, until the average queue length drops below

the max”,

 

Pd

 

 

maxp

 

 

or I ’
mlnth maxlh

Average queue length (packets)

Figure 2-8 A typical RED Algorithm.

We used WRED to establish three packet queues, each dedicated to an AF priority
level, by applying a distinct set of RED parameters for each packet queue. Jacobson et al.

in [15] provided the guidelines as thumb rules for selection of the RED queue length

thresholds (i.e. min”, and max,;,). In a multiple queue environment, a given queue can be
assigned with a higher priority by using packet drop probability parameter maxp. A

packet queue assigned with the highest value of maxp should be the lowest priority queue.

25

In our case where we need to establish three priority queues, the parameter maxp can be

adjusted by using the following relation.

1
max,,(q_1)=Zmaxp(q_2)

1 (2.8)
maxp(q_1)=gmaxp(q_3); A>1, 6>A

In the above given relation, the values for parameters A, 5 and an initial value of
parameter marp of any queue can be selected in such a way that a highest priority will be

assigned to q_l, medium for q_2 and lowest for q_3. The setup and configuration of
WRED parameters employed for simulations is described in section 4.4.2.

Lyles et al. in [16] suggested that the deployment of RED is not a good choice over
the best effort Internet. However, the scalability features of RED (specially in case of
WRED) make it a good choice for its use in multiple queue networks, such as an AF
service class of DiffServ. Several other extensions in RED algorithm were proposed such
as RED with IN/OUT (R10) [17], RED+ [18], Adaptive RED (ARED) [19] and Flow
RED (FRED) [20]. Most of the work is done to bring improvements in the fair network
resource utilization during transmission, by using fair queuing methods like weighted fair
queuing (WFQ) [42][43][44]. Whereas IETF RFC 2963 [45] proposed rate-adaptive

shapers by using First In First Out (FIFO) queuing methods.

2.3 Assured Forwarding PHB

Assured Forwarding (AF) is a medium price service as compared to the EF and BE
services offered by DiffServ. There are four independent AF classes that are defined in

the DiffServ standard, where each AF class is in each DiffServ node allocated a certain

26

amount of forwarding resources (buffer space and bandwidth). Each AF class provides

three levels of packet drop precedences (or packet forwarding priorities) [13]. An AF
class can be represented as Any, where x denotes the number of AF class ranging from 1

to 4, and y denotes the packet priority (or drop precedence) level ranging from 1 to 3.
Priority level 1 in each AF class carries the highest priority (or lowest packet drop
precedence). Since each AF class is independent of the other classes, it may receive an
equal share of network resources, which is further distributed among three priority levels.
Therefore, for the remainder of the thesis, we focus on mapping scalable video onto a
single AF class and its corresponding three priority queues. The priority levels and their

corresponding standard DSCPs are given in Table 2.1.

Table 2.1 Assured Forwarding PHB code points.

 

 

 

 

 

 

 

 

 

Dmp AF 1 AF 2 AF 3 AF 4
Precedence
1) LOW AFll AFzr A1331 AF41
(Green) 001010 010010 011010 100010
2) Medium AF” AF22 AF32 AF42
(Yellow) 001100 010100 011100 100100
3) High AFB AF23 AF33 AF43
(Red) 001110 010110 011110 100110

 

 

A packet is termed as a colored packet if it is marked with an AF service class DSCP

in its header. A packet is called as a green packet if it is marked with the DSCP of the
highest priority level AFxl, a yellow packet if marked for the medium priority level Asz
and a red packet if marked for the lowest priority AFX3. If a packet initially marked as a

green packet becomes out-of—profile as a result of metering, it can be downgraded as a

yellow packet by re-marking a new DSCP at the ingress router. No priority level

27

upgrades are defined within a single DiffServ domain. Throughout this thesis we will use
these priority levels in terms of colors, as used by the standard DiffServ marking
algorithms. The effect of the number of drop precedences within an AF service class is
discussed in detail by Goyal et al in [36] and [42].

As mentioned earlier, the highest priority level within a given class is level 1 (green),
the medium priority level is 2 (yellow), and the lowest priority level is 3 (red). Thus the
probability Pd of loosing a colored packet during transmission can be expressed as:

Pd (green) < Pd (yellow) < Pd (red) (2.9)

In this thesis we analyzed performance as a function of the “cost” associated with the
services provided by each priority level of an AF class during transmission. The cost is a
measure of the price paid by the user for utilizing an AF class. The associated cost of
each priority level can be defined as a cost function c(x), where x is the corresponding
priority level utilized by a packet flow within a single AF class, then:

c( green) > c( yellow) > c(red) (2.10)

The pricing and cost calculation for an AF class is described in the next section.

2.4 Pricing and Cost Calculation for Assured Forwarding
Class

A DiffServ network provides service differentiation at its ingress nodes. This makes
sender-to-pay type pricing as the most appropriate pricing approach for the network
services [47]. This type of service pricing is a departure from the traditional Internet
receiver-paid flat rate model. The SLA between the sender and service provider adjust the

profile values for the network resource allocations.

28

As introduced earlier in Chapter 1, an important issue related with the pricing of these
services is the overbooking of channel bandwidth by the service providers. Service
providers usually overbook the capacity of their network channels, hoping that users
(customers) will not make excessive demands on the network at the same time. But if the
service provider nriscalculates, traffic will be dropped, even if it is guaranteed, although
the traffic with no guaranteed CIR will be dropped first. This phenomenon implies that,
for n packet flows over a channel of maximum capacity C (bps), the committed
bandwidth i.e. CIR should follows:

ll
ZCIR, < C (2.11)

i=1

If a DiffServ node starts operating over or very close to the available channel (or link)
capacity, the drop precedence in an AF class becomes redundant. Consequently, three-
drop precedence levels starts giving the same performance as two. We used this
phenomenon of overbooking by sending overload traffic (as committed traffic) to analyze
its effect on the transmission performance.

DiffServ follows a simple direct proportion of pricing and their related services.
While focusing on an AF class, it has three price classes; one belongs to each priority
level. These pricing classes can be implemented in two ways, either using flat rate
pricing scheme, which is the extreme capacity usage pricing where the available capacity
always equals to a fixed fraction of the link capacity, or by usage based pricing scheme.
A pricing scheme between these two extreme types was describe by Semret et al [46] as
explicit pricing of resources. The cost affecting parameters in a DiffServ network are
related to the bandwidth and buffer management. CIR and PIR are considered as the

major pricing and cost-affecting constraints in an AF class of DiffServ, when a certain

29

cost is associated with every increase in CIR value (whereas PIR is used as an upper

threshold which equals to the available bandwidth).
Let Kg, K“ and K, are the prices (as some 53) associated with every unit-committed

rate (bps) of the corresponding G number of green, Y—yellow and, R-red packet flows

utilizing an AF class. The corresponding sending bitrates of each green, yellow and red

packet flow are rg, r), and r,, along with their corresponding committed information rates

CIRg, CIRy and CIR, respectively. Then the individual cost cg, cy and c, of each colored

packet flow over an AF class can be calculated as follows:

G
Green packets c8 = Kg 2 min(rg ,CIRg)

g=1
1’
Yellow packets c), = K, 2 min(ry,CIRy) (2.12)
y=1
R
Red packets c, = K, 2 min(r,,CIR,)

r=l

Thus the total cost (CM) for the AF service utilized by the user can be given by the
following equation:

CA]: =cg +cy+cr (2.13)

By inserting the corresponding values of cg, c_,. and c, from equation (2.12) in (2.13) we

get,

G Y R
CA}: = Kg 2 CIRg +Ky Z CIR), +k, Z CIR, (2.14)
g=l y=l r=l

Every incoming colored packet needs to satisfy the upper and lower bound conditions
during the metering process, in order to maintain its status as in-profile. As mentioned by

[38][39] and [40] the marking (or re-marking) of an incoming packet is done according to

30

its predefined SLA profile values. The metering-marking algorithms at the edge router
use these values. According to the suggested DiffServ metering-marking algorithms (like

srTCM and trTCM) the upper and lower bound conditions for every incoming packet

with rate r,- can be given as:

Green packet fiows { (2.15)
r,- < PIRI‘
Yellow packet flows CIR,- < r,- S PIRl- (2.16)
I;- > CIRI'
Red packet flows (2.17)
rl- > PIR,‘

2.4.1 Flat Rate Pricing Technique
The most common pricing method for commitment based networks like DiffServ is

the flat rate pricing. Flat rate pricing technique estimates CIR of any packet flow i as a
fixed fraction of its bitrate r,-. The general equation of this model for a video stream

containing n video layers (or packet flows) can be expressed as:

CIR,- = pr,- i= l,2,3...n and p 20 (2.18)

Any selection of p in the above equation can be adjusted according to the desired
priority level of a packet flow. For instance, a simple approach for selecting p for
different colored packet flows utilizing an AF class according to the thresholds
mentioned in equations (2.15) to (2.17) should be:

Green packet flows p 21 (2.19)

Yellow packet flows 0 S p <1 (2.20)

31

Red packet flows 0 _<_ p <1 (2.21)

32

3 Mapping Strategies and Method of
Analysis

In this chapter we present different strategies developed to improve video quality
during the transmission of scalable video streams over an AP class of DiffServ networks.
The method for mapping video layers over any desired AF class is described in the first
section. Sections 3.2 and 3.3 describe the proposed strategies aimed to achieve service
differentiation for scalable video layers over an AF class. The same section also includes
a description of the proposed pricing models. These pricing models are developed to
control the packet loss ratio (PLR) of multiple video layers, utilizing a single priority
level of an AF class. Section 3.4 describes the method of performance evaluation for

scalable video streams over an AF class.

3.1 Packet Mapping

Mapping is a term, which we use to illustrate the selection of a DiffServ class that is
employed for transmitting a packet. A packet is mapped to a DiffServ class by marking a
DiffServ code point (DSCP) value in its packet header “before” its transmission. By

marking all the packets belonging to a given video layer with a similar DSCP, the

33

complete video layer can be mapped to a corresponding DiffServ class. This work
focuses on mapping of the video layers to the priority levels of an AF class.

The mapping of various (say n) video layers in a scalable video stream over an AF
class can be divided into two categories. The first category is One-to-One mapping,
which is applicable to video streams containing up to three video layers i.e. n S 3. The
second category is Multiple-to-One mapping, which is applicable to video streams
containing more than three layers i.e. n>3. Under each of the above two mapping
categories, one can define a variety of approaches for mapping scalable video layers onto
AF priority levels. We refer to these approaches as mapping strategies (as explained
below). Moreover, it is important to note that some scalable video coding methods have
the flexibility of generating any desired number of enhancement layers. Consequently,
these methods can employ strategies from the one-to-one or multiple-to-one mapping
category. An example of such scalable video streams is the ones generated by the MPEG-
4 FGS coding method.

As described earlier, in order to improve the quality of a scalable video stream
containing upto n video layers that are transmitted over a lossy packet network, the PLR
and throughput (T) associated with a video layer i, should meet the following constraints:

PLR,- < PLR,+1 i=1,2,3...n

(3.1)
Ti >Ti+l

All mapping strategies in this thesis are developed with an objective of achieving the
network PLR constraint that is expressed in equation (3.1). For each mapping strategy,
the Committed Information Rate (CIR) is used as a “commitment parameter” as well as a
“cost parameter” for the service agreement. In other words, we consider CIR as a

measure of cost since the higher the value of CIR that is stipulated by the service level

34

agreement (SLA), the higher the cost that is associated with this agreement. Meanwhile,
the higher the CIR, the lower the PLR that can be achieved. Consequently, the mapping
strategies, which are defined as functions of the CIR parameter, represent a form of
pricing models. Hence, below, we use the two expressions: mapping strategy and pricing
model interchangeably. More importantly, CIR plays a key role in expressing and
defining the proposed strategies for mapping scalable video over AF DiffServ classes. As
described later in this chapter, by defining different functions of CIR for the different
video layers, one can achieve different levels of treatment (by a DiffServ node) for these
video layers. This will provide the differentiation in PLR performance that is expressed in
equation (3.1).

In this thesis, we consider three strategies for mapping scalable video layers to the AF
priority levels. We refer to these mapping strategies as flat rate, linear, and exponential.
(These strategies are described in detail below). For the one-to-one mapping category, we
only employ the flat rate based technique. This is due to two reasons: (1) The small
number of layers associated with each AF priority level (only one) in the one-to-one
mapping category significantly limits the number of mapping strategies that one can use;
and (2) The flat-rate based one-to-one mapping represents a benchmark for the simplest
approach for mapping scalable video layers over DiffServ. For multiple-to-one mapping,
linear and exponential pricing models (mapping strategies) are also used in addition to
the flat rate pricing technique. After the above high-level description of the different
mapping strategies and categories, for the remainder of the thesis, we will use the two

expressions mapping strategy and mapping category interchangeably.

35

Before proceeding further, we highlight one more item regarding the identification of
the different packet flows of scalable video. In order to provide identification for the
packets belonging to different video layers during transmission, the use of a unique flow-
id for each video layer is proposed. A flow-id is a set of 16-bit address field (containing
source and destination IP addresses) in an IP packet header and a 32-bit address field
(containing the corresponding source and destination port addresses) in the UDP packet
header. Thus all packets belonging to a single video layer are assigned the same flow-id.
Use of a distinct flow-id helps in identifying each video layer as a distinct packet flow.
Hence, a unique set of profile values (containing rate thresholds like CIR and PIR) can be
defined for each layer. As described earlier in section 2.2.3, these profile values are used

by the metering-marking algorithm at the DiffServ ingress router for traffic conditioning.

3.2 One-to-One Mapping
This mapping category is applicable to video streams that are composed of one base

layer (BL) and a maximum of two enhancement layers (ELI and EL2). The base layer is
transmitted with a bitrate rb, whereas the corresponding bitrates of the two enhancement
layers are r, and r2, respectively. Each video layer is mapped to one of the three priority
levels of an AF class as shown in Figure 3-1. The packets belonging to the base layer are
mapped to the highest priority level AF,” as green packets, and the packets belonging to
ELl and EL2 are mapped to Asz and AFX3 as yellow and red packets respectively. By

applying this strategy, we intend to obtain a distinct network packet forwarding behavior

for each layer governed by equation (3.1).

36

For this mapping category, the pricing model of choice is flat rate based pricing
model. As we described earlier, a pricing model deals with the network bandwidth
commitment between the user and the service provider. Thus, for maximum utilization of
the available bandwidth allocated to a single layer mapped onto a single priority level, no
other bandwidth commitment model could perform better than flat rate based pricing
model. Also, the limited number of one video layer per priority level does not allow using
linear or exponential model, as they will reduce the possible throughput while leaving
some of the available bandwidth unused. Furthermore, we will use same approach for the
calculating CIR of base layer in every strategy. This is because the base layer is the only

layer utilizing the highest priority level in both mapping categories in the remaining

 

 

 

 

 

 

 

 

 

 

 

 

thesis.
r2 5 AFx
EL;
P--1
r I _’ AFxl
l
I
EL
' pun nnnnnnnnn : nnnnnnnn Ian-y AFX2
I
It, I
—-—l> Br - - - —> Am

 

 

 

 

 

 

 

 

 

 

 

Figure 3-1 One-to-one Mapping.

3.3 Multiple-to-One Mapping
Under this category, the number of video layers (i.e. n) in a stream may be greater
than the available AF priority levels (i.e. n > 3). This mapping category can be

considered as a general case, which is applicable for any number of video layers in a

37

stream. Under this mapping category, a video stream containing one base layer and n-1

enhancement layers is mapped onto two priority levels of an AF class. As shown in

Figure 3-2, the packets belonging to the base layer are mapped to the AF,“ priority level

and packets belonging to all enhancement layers are mapped to the Asz priority level.

This mapping scheme enables all base layer packets to be recognized as green packets

and packets belonging to all enhancement layer as yellow packets.

With this mapping strategy, the lowest possible PLR is expected in the base layer.

Therefore, during congestion the green (BL) packets would retain the lowest probability

of being discarded from the buffer queues. Hence, this strategy minimizes the loss of

green packets at the cost of losing more yellow packets.

rn-

r4
f3
r2
F1

rb

 

 

AFx

 

AF,”

 

 

 

 

 

 

BL

AFX 3

 

 

 

 

 

 

 

 

 

Figure 3-2 Multiple to One Mapping.

All enhancement layers in this mapping category are using a single AF priority level.

This may allow the network to treat every enhancement layer equally regardless of their

corresponding weight. Thus, an equal PLR for each enhancement layer may ensue, which

is undesirable. In order to avoid this situation and achieve the PLR constraints for each

38

enhancement layer according to the inequality in equation (3.1), we propose two pricing
models as strategies for calculating the CIR for each uniformly mapped enhancement
layer. As mentioned earlier, these pricing models are developed as a function of CIR,
while keeping in mind the inversely proportional relationship between the applied CIR
for network service agreement and the PLR of the corresponding packet flow. Thus by
using a higher CIR value for a given enhancement layer will result in a lower PLR. In
order to obtain a higher PLR for every increasing enhancement layer, the corresponding
CIR should be decreased monotonically. To achieve such a CIR scheme, we propose two
strategies: linear and exponential. These approaches are described as two possible pricing

models in the subsequent sections.

3.3.1 Exponential Pricing Model (EPM)

Using this model, we aim to evaluate CIR values for the given (n-l) enhancement
layers in an exponentially decreasing manner, starting from the highest value for the first
enhancement layer to the lowest value for the last enhancement layer in a scalable video
stream. Therefore, using this approach an exponentially increasing PLR response is
expected for an increasing number of enhancement layers, which satisfy equation 3.1.

As mentioned in equation (2.16), the CIR of a yellow packet flow, which constitutes
an enhancement layer in this case, should be less than its corresponding bitrate in order to
remain in—profile. We used the following equation to estimate a corresponding CIR value,
for a given enhancement layer i with sending bitrate r,-. This CIR estimate will always

remain less than the given threshold r,- for (n-l) enhancement layers:

CIR, = r,- —(r,-)“i; i: 1,2,3..(n—1) (3.2)

39

In the above equation, a is used as an exponential pricing factor, which we define as a
ratio of the i” enhancement layer to the total number of enhancement layers in the video
stream, i.e.,

i
03':

 

; i=1,2,3...(n—1) (3.3)
n-

As i will approach (n-l), the CIR will decrease exponentially, which consequently
will increase the corresponding PLR.

It is important to note that the above exponential mapping strategy will lead to a

differentiation among the enhancement layers’ CIR values even if the rate r,~ = r is

constant for all of these layers.

3.3.2 Linear Pricing Model (LPM)
The second strategy we propose is a linear decay for estimation of the corresponding

smaller CIR value for each increasing enhancement layer. We assume that all

enhancement layers have the same sending bitrate i.e. r,- =n+1= r, according to the

typical method of rate distribution as discussed in section 2.1.1. In this approach, a CIR
value for a given enhancement layer is evaluated by subtracting the product of a constant
parameter B and the enhancement layer number i from its bitrate r. Similar to the EPM,
the general threshold condition for the yellow packet flows (as given by equation 2.13) is
also followed in this approach. The CIR value for a given enhancement layer i with

bitrate r can be expressed as follows:

CIR,- = r-ifl, i: 1,2,3..(n—1) (3.4)

The constant ,8 in equation (3.4) is the linear pricing factor, which can be defined as a

ratio of the common bitrate r and the total number of the enhancement layers. According

40

to this definition, the estimated range of linear gradient ,8 for a video stream containing
(n-l) enhancement layers is:

r

O<BS

 

3.5
n _1 ( )
Any selection of ,8 in the above-defined range leads to a linearly decreasing CIR with
increasing number of an enhancement layer in the video stream. The LPM provides an
equal difference between the CIR values for any two consecutive enhancement layers in a

scalable video stream, i.e.,

C1R,,+1 = CIR" - ,6 (3.6)

3.4 Performance Evaluation

The mapping categories described in the previous two sections (3.2 and 3.3) are
developed with an objective to improve the delivered quality of scalable video streams
over the Internet. This improvement in quality is achieved by attaining a desirably lower
PLR for a given layer in a video stream, according to its predefined weight. Since
DiffServ is a commitment-based network, this improvement in quality is also associated
with some price. On the other hand, improvement in the quality of received video over a
lossy packet network typically depends on successful reception of packets belonging to
the video layers with higher associated weights as compared to others with relatively
lower weights. This has compelled us to consider the performance of the proposed
strategies as, (1) not only a function of the user perceived throughput but also as a
function of the associated weight of every received layer, and as (2) a function of the
price paid for Diffserv. Thus, the performance of a certain mapping strategy can be

assessed collectively as a measure of the user perceived quality and the price associated

41

with it. In this thesis, we measure the performance of a given mapping strategy in terms

of the Performance Factor (PF), which can be defined as the ratio of the weighted
throughput (T,,.) of a given scalable video stream and the total cost (CM) associated with
its transmission over an AF class of DiffServ. Mathematically, PF (Pf) can be expressed

as follows:

 

T .
pf = ‘t (3.7)

Hence, we evaluate the PF for each of the above mapping strategies to validate their
performance and effectiveness. The above equation shows that for a given mapping
technique the PF is maximized by increasing the weighted throughput of the video stream
while minimizing the cost associated with it. We will evaluate the PF for each mapping
strategy as a comparison parameter in Chapter 5. Furthermore, this measure is also used
in analyzing the affect of the channel over-commitment on the performance of each
mapping strategy.

The calculation of the weighted throughput and the total cost measures for the

evaluation of PF are described below:
Weighted Throughput Measure: The mapping strategies discussed in the previous
sections were developed with an objective of achieving a desirable PLR for different
video layers according to their predefined weights in the video stream. The PLR of a
packet flow is related to the channel throughput (T) according to the following
relationship:

T = (l—PLR) (3.8).

The weighted throughput of a given video layer can be expressed as a product of its

throughput, bitrate, and its predefined weight. Thus for a scalable video stream composed

42

of one base layer and (n-1) enhancement layers, the weighted throughput of the video

stream can be determined by using the following relation.

ll
Tw = 20mm), (3.9)
i=1
where, T,-, r,- and w) are the corresponding throughput, bitrate and pre-defined weight of
the i” video layer respectively.

Total Cost Measure: In order to measure the total associated cost, we measured the
cost for each video layer utilizing an AF class of DiffServ in terms of its CIR. Thus, the
cost associated with green (cg), yellow (CV) and red (Cr) packet flows mapped onto an AF
class can be estimated as:

G
cg = Kg ZCIR,
i=1
Y
c}, = KyZCIR, (3.10)
i=1

R
c, = 192 CIR,-

i=1

where K,C is the bandwidth price (as some S) associated with a unit CIR (bps) for a given

AF priority level x. G, Y and R being the numbers of (mapped) green, yellow and red
layers respectively.
Hence, the sum of individual costs for all packet flows (i.e., green, yellow and red

video layers) gives the total cost C,” for the entire packet stream (in our case, the scalable

video stream) utilizing an AF class.

.+c, (3.11)

CAF 3C8 +C)

43

Thus if a packet stream contains G packet flows mapped as green, Y packet flows
mapped as yellow and R packet flows mapped as red, then the total cost (CAF) can be

calculated as:
G g Y ) R
1 _ ' 1‘
CM _ KgZCIRi + KyZCIRi +K,ZCIRi (3.12)
(:1 [=1 [:1
where CIth is the CIR for the i"' layer mapped to an AF priority level x.
I
In equation (3.12), the unknown parameters are the prices of the service K, for the x"'
AF class. Since the price associated with an AF class may vary with the service provider,
no specific values can be assigned. One can decide a price range for providing such a
priority packet-forwarding service. Typically only a relationship between the prices for
each AF priority level can be determined based on their assigned priorities:
Kg>Ky>Kr (3.13)
However, the above relationship is not helpful in deriving any meaningful values, which

are needed for subsequent analysis. Due to the absence of any specific relationship
between the prices of AF priority levels, we express both Kg and K, in terms of Ky as

described below.
Similar to the technique used in Section 2.2.4 to implement the three WRED priority

queues for AF priority levels, we can inter-relate the WRED packet drop probability

parameter (maxp) for each AF priority queue (i.e. green, yellow, and red) as follows:

1
max p (green) = Z maxp (yellow)

1 (3.14)
maxp(green) =Emaxp(red); A >1, 6 > A

Equation (3.14) shows that during transmission, a green packet receives A times more

importance than a yellow packet and 5 times more importance than a red packet. Thus,

every unit Kg for transmission of green packets is equivalent to A unit Ky and 5 unit Kr.

So the prices K8 and K, can be expressed in terms of Ky as:

Kg =AKy
A (3.15)
K,- 33K),

Thus, the total cost (CH) of a packet stream utilizing AF services can be obtained by

substituting (3.15) in (3.12):

C? Y R
A
_ 8 y r
em. —AKyZCIRi +KyZCIRi +gKyZCIRi (3.16)
[:1 [=1 i=1
or,
(3 g Y ‘ [A R
_ r
CA]: .Ky AZCIRI +ZiCIRi +E-ZICIRi (3.17)
i=1 i: i=

For convenience, we use Kyzl for all calculations. As a result equation (3.17) simplifies

[OZ

(3 Y R
. A
_ 8 )
CAF _Ai:EICIRi +£1CIRI +5i=EICIRir (3.18)

Hence, by inserting the corresponding values of weighted throughput and total cost in
equation (3.7), for a packet stream containing G green, Y yellow and R red packet flows,
the PF can be evaluated as follows:

I1
zmhwz')

pf —_- G i=Y1 (3.19)

A R
g y r
A2 CIRi + 2 UR +_§_§:1CIRI

i=1 i=1

 

45

where n is the total number of packet flows (i.e. n=G+Y+R) and CIR?‘ is the CIR for the
l

i” layer mapped in AF priority level x.

For given video-layers weighting values w,- and bitrates r,~, one can evaluate the
performance factor Pf of the desired mapping strategy if the PLR,- values (for the different
video layers) are known for that strategy. In this work, we employed the well-known
network simulator (NS), which is the most popular software simulation tool that are used
for Intemet-based performance evaluation, to estimate the PLR,- values of the different

video layers and for the different mapping strategies described above.

46

4 Simulation Setup

In this chapter we present the setup that is used for the simulation of the proposed
mapping categories and their associated pricing models. A high-level description of the
simulation scenarios are provided in section 4.1. The second section describes the
DiffServ network topology and the configuration of video sources used in the
simulations. Bitrate calculations for each packet stream transmitted over the DiffServ
network is described in section 4.3. An objective of this thesis was the analysis of the
proposed strategies at various levels of network over-commitment, where the available
network bandwidth is intentionally over-booked by the service provider. This situation is
implemented in the simulations by overloading the network links. The details of the
network link overloading are also described in the same section. Section 4.4 describes the
parameters used for traffic conditioning of video streams at the DiffServ network’s edge
router. Before conducting the extensive DiffServ simulations for the different strategies
described above, we validated the employed network topology by running two randomly
chosen simulation scenarios. These validation results are presented in section 4.6. This

chapter concludes with a short summary of the simulation steps that are used in this work.

47

4.1 Simulation Scenarios
The proposed mapping categories in Chapter 3 are used as the two basic simulation
scenarios namely: Scenario I for one-to-one mapping category, and Scenario II for

multiple-to-one mapping category. A description of these simulation scenarios is listed in

 

 

 

 

 

 

 

Table 4.1.
Table 4.1 Statistics of Simulation Scenarios.
Simulation n Mapping Strategy
Scenario (number of video
layers) Video Layer AF Priority Level

BL AF“

I 3 EU AF12

EL2 A]:13

11 10 BL AF“

ELl-EL9 AF12

 

 

 

 

 

 

Simulation scenario II is further divided into three sub-scenarios based on the
calculation method of the enhancement layer CIRs. Sub-scenario [1(a) is used as a
reference model, in which the flat rate based technique is used. The other two sub-
scenarios are: scenario II(b) for the exponential pricing model (EPM), and scenario II(c)

for the linear pricing model (LPM).

4.2 Network Topology and Statistics

We used Network Simulator (NS) version ns-2.1b8a [14] to simulate the proposed
mapping strategies. NS is a popular network simulation tool that is extensively employed
for network related research, in general, and Internet-based simulation in particular. Only

NS ns-2.1b8a and later versions are equipped with the DiffServ module [54] and with

48

standard traffic conditioning algorithms (e.g. metering-marking algorithms: srTCM,
trTCM and TSWTCM) and traffic shaping (or queuing) mechanisms (e.g. FIFO, RED
and WRED etc.). Except for a few modifications made in the back-end code of NS, that
were specific to our requirements, we mostly utilized the standard built-in sub-routines
and algorithms for our simulations. Thus, a configuration of the built—in sub-routines by
calculated parameter values are used for most of the simulations.

The network topology used for the simulations is shown in Figure 4-1. All network
links in the figure are T1 links, which can support transmission of data up to 1.5 Mbps. A
DiffServ domain is established by using one ingress edge router, one core router, and one
egress edge router. Five video sources, attached with the ingress router, are configured to
send video packet streams toward their respective destinations (or clients), which are
attached to the egress edge router. Each video source is configured to send equal size (i.e.
512 bytes) UDP packets at some constant bitrate (CBR). A separate packet stream is
generated for each video layer by using a separate UDP agent in the source. In this thesis
we use CBR packet streams as a limiting case, due to the prior knowledge of available
bandwidth. However, the scalable video coding methods like MPEG-4 FGS are capable
to send packet streams (or video layers) at any desirable bitrate, depending upon the
availability of network bandwidth.

Each packet stream from a video source contains two types of packets: the base layer
(BL) packets and the enhancement layer (EL) packets. The method employed for the
calculation of bitrates for each video stream and their subsequent video layers is

described in the next section.

49

e l -I-. Irma”. " ~

 

Table 4.2 Simulation Statistics.

 

 

Simulation Parameter Value

Simulation Time 60 sec

Links between Network nodes 1.5 Mbps, 5ms delay
Packet Size 512 bytes

Maximum number of ELs 9

Number of Video sources 5

Buffer queue length 70 packets

 

 

 

Before transmission over the network link, every video source is configured to mark

each packet with: i) a flow-id (to identify the video layer), and ii) a corresponding DSCP
of an AF priority level (i.e. AF,” (green), Asz (yellow) or AFX3 (red)) depending on the

applied mapping strategy.

       
 
 

-_.'J‘

   
 
 

ES

Video Sorce

{E

Vida—Source 2

7.- r
t
-
_ i
r t ‘1
‘
\

Video Source 3 . I" i;

' deo Client 1

 
 

  
 
 

Difl'Serv Domain _-_J

ideolient 1

 

 

Video Client 3

  
 
 

 
  
 
  
 

  
 

 

f Edge uteri Core Router Edge Router! - ,
= [E] ingress Node Egress Node _ Jr;

        

I

 
 

.Lu,

Video Clie

 

Video Source 4

  

  

—
L
— -
—
‘

VideSource 5

..',w

9
- \i
J“
i
...« r‘l_
'0

.-FT, .

4
l
k 1

L-“
:ina‘.

Video Clint 5

Figure 4-1 Network topology for the simulations.

50

4.3 Bitrate and Network Overload Configurations

The network topology discussed in the previous section has been simulated under two
network conditions: (1) when the bandwidth of the network links is only committed to
their full capacity i.e. no overload traffic, and (2) when the network link bandwidth is
over-committed (ranging from 10% to 100% overload traffic). These two network
conditions are referred to as normal and overload conditions.

Each over-commitment scenario has been simulated by sending overload video traffic
towards the DiffServ domain. For this purpose, the bitrate of each video stream (and its
subsequent video layers) is calculated and applied to the network in order to create a
given level of network overload. The same bitrates are then used in the committed
information rate (CIR) computations using the corresponding pricing model (i.e., flat
rate, EPM or LPM) according to the mapping category in use.

In the first phase of experiments, the previously described DiffServ domain is
simulated under normal condition, without any overload, to obtain the network PLR
response for the proposed mapping strategies. In the second phase of experiments, the
network is simulated under three distinct overload levels of 10%, 50% and 100%

respectively. In order to estimate the individual bitrates of each video layer at every
network overload level, the maximum bitrate (Rmax) of the video stream sent by each

video source is calculated first as a function of the ingress edge router’s outbound link

capacity C as follows:

Rm... {mm (4.1)

51

where s denotes the number of video sources attached to the edge router, and v is the
network overload as a fraction that ranges between zero and one.
For instance, in order to overload the edge router by 10%, we used v = 0.1 in equation

(4.1). Hence, for a given link capacity of 1.5 Mbps and five video sources (i.e. s = 5), the
bitrate Rm,” is calculated to be 330 kbps for each video source. Similarly, at no overload
the available link capacity is equally committed (or allocated) to each video source with

Rmm. = 300 kbps.

As described in the previous section, each video source is configured to send two

types of packet streams, i.e. base layer and enhancement layer packet streams at bitrates
rb and re respectively. However, depending on the number of enhancement layers, the
number of packet streams from each video source may vary from 1 to n. Thus, the
maximum bitrate (Rmm) of a video source equals the sum of the individual packet

streams’ bitrates that are generated from the video source.

Under normal conditions, the available bandwidth of edge router’s outbound link is

distributed equally to each video source. Hence, the default value for Rm“, at no-overload

network condition is 300 kbps for C = 1.5 Mbps. The bitrate of the base layer (rb) is kept

constant at 64 kbps in each simulation scenario at every network overload level, while the
remaining allocated bandwidth is used for the enhancement layer(s). By using a constant

bitrate for the base layer, the overloading of network is achieved by a variation of the
enhancement layer(s) bitrate(s). The distribution of bitrates rb, and re in an is listed in

Table 4.3 for each network overload condition.

52

The bitrate re in Table 4.3 represents the total bitrate of the enhancement layer packet
stream(s). During simulations when more than one (say It) enhancement layers are used,

the bitrate r,- of any i"' enhancement layer is kept fixed at:

r,- =r—e, where l<iSn (4.2)

n
Therefore, the rate distribution of the enhancement layers is according to a simple

uniform allocation method (as discussed in section 2.1.1).

Table 4.3 Distribution of bitrates (kbps) for video layers.

 

 

 

 

33:31:21; 0% 10% 50% 100%
Rmax 300 330 450 600
rb 64 64 64 64
re 236 266 386 536

 

 

 

 

 

 

 

As mentioned earlier in section 3.2, we used the flat rate technique for calculating the
CIR when a single layer is mapped to a single priority level. Since the base layer is

always mapped to the highest priority level as green packet flow in each scenario, the
base layer CIR (CIRb) is always calculated using the flat rate pricing equation (2.18) with

p = 1.01 (this value of p conforms to the general threshold condition for green packet

flows as indicated in equation (2.15)) as follows:

CIRb =1.01q, (4.3)

Similarly in scenario I, the two enhancement layer(s) are also mapped to the two
distinct priority levels as yellow and red packet flows, and thus the corresponding Cle
are calculated using the flat rate pricing equation (2.18) with p = 0.99 as follows:

CIRe = 0.99 re (4.4)

where re is the rate of the corresponding enhancement layer.

53

The equations (4.3) and (4.4) satisfy the necessary conditions mentioned in equations
(2.15) to (2.17) for green, yellow and red packet flows, respectively. Equation (4.4) is

also used for calculating the CIR of each enhancement layer in simulation scenario II (a).

4.4 Traffic Conditioning Setup

As described in sections 2.2.2 to 2.2.4, the traffic conditioning is a set of sequel
processes of metering, marking, and shaping of the incoming packet streams at the
DiffServ boundary node (i.e. ingress edge router). The selected simulation parameters
and their statistics for these traffic-conditioning processes are described in the subsequent

sections.

4.4.1 Traffic Metering and Marking

A metering-marking algorithm at the edge router is used to “meter” every incoming
packet with a given set of service level agreement (SLA) profile values. DiffServ module
in NS supports both UDP/CBR compatible standard metering-marking algorithms:
srTCM and trTCM [54]. Due to the similarity of the parameters in both algorithms [see
Appendix], similar results were expected. This similar expected behavior is also due the
fact that the proposed mapping strategies are simulated by varying only the CIR
parameter values, while keeping the other parameters constant. Thus, two sets of
simulations were performed. Results showed that the two sets of PLR values that were
generated from using both markers were similar. Therefore, results generated using only

trTCM are included in this thesis.

54

As described in Chapter 3, the bitrate r,- of each incoming colored packet i is metered

using a given set of bitrate threshold values, CIR and PIR:

Green ackets 4.5

Yellow packets CIR,- < r,- S PIRl- (4.6)

Red packets (4.7)
r;- > PIR

Recall from chapter 2 that as a result of metering, an incoming colored packet (i.e.
marked initially with a DSCP of an AF priority level) is categorized either as an in-
profile or an out-of-profile packet. If the rate of an incoming packet is found conformal to
the given threshold limits of CIR and PIR, it is categorized as an in—profile packet. Such
packets are forwarded directly to their respective buffer queues. Whereas, if an incoming
packet is found exceeding the given rate threshold, it will be categorized as an out~of-
profile packet. Each out-of—profile packet is eligible to be re-marked with the DSCP of a
lower AF priority level before being forwarded towards the buffer queues. In other
words, the drop probability of an out-of—profile packet increases due to re—marking. A
summary of metering-marking algorithm parameters selected for the simulations are

listed in Table 4.4.

Table 4.4 Selected values for trTCM parameters.

 

 

 

 

 

trTCM Comments/ Value

parameter

CIR Evaluated by using pricing models (flat,
exponential or linear)

CBS 2 packets (1024 bytes)

PIR Green and Yellow packets: 1.5 Mbps
Red packets: 0.99 x (bitrate of red packet flow)

PBS 2 packets (1024 bytes)

 

 

 

 

55

The CBS and PBS parameters used in trTCM and srTCM algorithms are defined to be
the size of token bucket at the ingress router. The use of these token buckets is described
in Appendix. For the T1 links used in our simulations, we choose a suitable token bucket
size of 2 packets i.e. 1024 bytes. These parameter values are also kept constant

throughout the simulations.

4.4.2 Traffic Shaping

As introduced in Chapter 2, we preferred WRED as traffic shaping mechanism. The
reason for selecting WRED is the provision of establishing multiple RED queues and
control over the packet drop probability in each queue, which allows an assignment of
priorities to each buffer queue with respect to the given AF priority level. The employed
version of NS for our simulations also supports the WRED queuing mechanism.

According to this mechanism, one physical RED queue is divided into three virtual
RED queues, each dedicated to the packets belonging to a given AF priority level; green,

yellow, or red. Length of the physical RED queue, for the given T1 links, is kept constant
at 70 packets throughout the simulations, while the queue length thresholds (i.e. min”, and
max,;,) for each virtual RED queue is selected according to the “rule of thumb” suggested

by Jacobson et al. in [15]. In order to provide service differentiation and priority, a
maximum length of 40 packets is selected for green packet queue, a maximum length of

20 packets for yellow packet queue and a maximum length of 10 packets for red packet

56

queue’. Similarly a different packet drop probability (maxp) is allocated to each virtual

queue with respect to its priority level. Thus, by using the relationship (2.8) and its

modification in equation (3.14) for AF priority queues:

max p (green) 2 i max p (yellow)
1 (4.8)
maxp(green) == gmaxpved); A Z 1, (3 Z A

For our simulations, we selected A = 5, and 5 = 10. By selecting the highest value of

maxp = 0.2 for red packets, we obtained a medium maxp (2 0.1) for yellow packets, and

the lowest maxp (= 0.02) for the green packets. The selected WRED parameter values for

all simulations are listed in Table 4.5, and are shown in Figure 4-2. However, one can
select other parameter values within the specified range, which may ensue different
values for PLR. However, the behavioral trend (in terms of PLR variations among the
different video layers) will remain the same if the queue priorities will assign in the same

proportion.

Table 4.5 Wei ghted-RED parameters for the simulations.

 

 

 

 

Queue min”. max”. max,

Green 20 40 0.02

Yellow 10 20 0. 1
Red 5 10 0.2

 

 

 

 

 

 

 

I This technique is adopted from Cisco IOSTM release 12.2 specifications. which implement service
differentiation in RED queues by varying the lower threshold (minm) while keeping the upper threshold

(maxm) constant.

57

 

— Green

_ Yellow

— Red
0.5

 

0.2
0.1 /
0.02 ..............................................................
0 / 0/ / >

0 5 15 20 25 30 35 40
Average Queue length (packets)

 

 

 

 

 

 

Figure 4—2 WRED parameters for priority queues.

4.5 Weight Distribution for Video Layers

Due to the absence of any typical weight distribution curve definition for video layers
in scalable video coding standards [24][25][26], we used an exponential weight
distribution curve (see Section 2.1.2) for the performance factor (PF) evaluations. The
weight distribution curve (shown in Figure 4-3) is obtained by using a = 0.4, b = 0.7 in
equation (2.5). However, one can use other combinations for the parameters a and b in
equation (2.5) to obtain different weight distribution curves for the video layers in a
scalable video stream. The selection of a = 0.4 denotes that the first enhancement layer
(EU) is responsible for bringing 40% improvement in the video quality produced by the
base layer. The curve in Figure 4-3 represents the weight distribution for a video stream

containing one base layer and nine enhancement layers i.e. n = 10.

58

Weight Distribution in Scalable Video Layers

 

 

 

 

0.9:- 't
0.8» «
0.7i- _(
0.6” ..
E.
i3 05" '4
:5
O4" ‘
0,3» -‘
0.2 " ..
01" -1
r r 1 1 1 1 ‘ A
%L ELI EL2 EL3 EL4 ELS EL6 EL7 ELB EL9

Video Layers

Figure 4-3 Employed weight distribution of scalable video layers.

4.6 Network Fairness Criteria

As discussed in section 4.3, the available bandwidth (and also the over committed
bandwidth) is equally distributed to each attached video source. Thus, as a pre-simulation
test of the adopted network topology, an experiment was performed to ascertain the fair
distribution of network resources to all video sources. By a fair distribution we mean that
the allocated network resources (like bandwidth and buffer) are equally shared by each
video source. With a fair distribution of network resources, each video source is expected
to suffer similar PLR. For this purpose, we simulated the network for two randomly

chosen scenarios, at different network overload levels. The results in Figure 4-4 show

59

that the video source location does not significantly affect the obtained PLR response.

Hence we will use the data obtained from only one source for all subsequent analyses.

PLR (%)

Fairness of Network Resource Allocation

 

100

90-

80-

70-

50'

40"

30"

20-

10*

 

I

I' T I I T

 

I - - Video Source 1
——- Video Source 2
Video Source 3

— Video Source 4 .

 

_Video Source 5

 

Y T

 

 

%L

1 l l l 1 L 1 1
EL1 EL2 EL3 EL4 EL5 EL6 EL7 EL8 EL9

Video Layers

PLR (%)

 

 

 

 

 

 

100 I I I I T l I I
--‘ Video Source 1
Video Source 2
90 ' Video Source 3 7
. -— Video Sowce 4
| — Video Sourcei
80 '- '-
70 _ / _
60 r -
50 '- .
40 ’- "
30 >- a
20 *- -
10 r- -
l l I L 1 l L 1
BL EL1 EL2 EL3 EL4 ELS EL6 EL7 EL8 EL9

Video Layers

Figure 4-4 PLR of video sources at (a) scenario II(b) at 10% overload, and (b) scenario
II(c) at 50% overload.

4.7 Summary of Simulation Steps

Each scenario has been simulated using the following sequence of steps.

STEP 1:

STEP 2:

Calculation of Rmr for each video source and for every network overload

level.

Calculation of the base layer and the enhancement layer(s) bitrates rb and

re, for each network overload level.

60

STEP 3:

STEP 4:

STEP 5:

STEP 6:

STEP 7:

STEP 8:

CIR calculation for the corresponding base layer and enhancement layer
bitrates, for the simulated pricing model at various network overload
levels.

NS code configuration: Assignment of a flow-id for each video layer.

NS code configuration: Assignment of a corresponding DSCP value for
each video layer, in accordance with the simulation scenario.

Running NS simulation for exactly 60 seconds, by application of the
computed bitrates and CIRs for each scenario.

Calculation of PLR for each video layer.

PF evaluation using the network PLRs determined in step 7.

61

5 Simulation Results and Analysis

This chapter describes the results obtained from the simulations of the proposed
mapping categories. The results of each simulated category are presented in the first two
sections. In each section, the network packet loss ratio (PLR) response for the simulated
video stream is presented first, and the evaluated performance factor (PF) as a measure of
the performance of the mapping strategy is presented next.

Section 5.1 contains the simulation results for scenario I (i.e. one-to-one mapping).
The next section (i.e.,5.2) contains the simulation results of scenario H (i.e. multiple-to-
one mapping). Further three subsections (i.e. 5.2.1 to 5.2.3) illustrate the results obtain by
applying the proposed three pricing models (i.e., flat, exponential, and linear). A
comparison of the PLR results obtained from using the different pricing models of the
multiple—to-one strategy is presented in section 5.3. The last section concludes the
simulation results with an analysis of the evaluated performance and cost for each

simulated scenario.

5.1 Simulation Scenario 1: One-to-One Mapping
The PLR response, obtained by the simulations of scenario I at various network
overload condition ranging from 0% to 100% overload, is shown in Figure 5-1. Under

normal conditions, i.e. at 0% overload, the PLR response in Figure 5-1 shows similar

62

packet forwarding behaviors for three distinctively mapped packet flows i.e. one base
layer (BL) and two enhancement layers (EL1 and EL2). All BL packets received the
highest forwarding priority and suffer no losses during transmission as expected. Similar
PLR behavior is also observed for EL1. Being the lowest prioritized packet flow, the
maximum packet losses occurred in EL2. The results for this network condition show a

fulfillment of the desired PLR constraints as depicted in equation (3.1).

Network PLR Response for Scenario I

_.1_ I
—.— 6W
«r 10% Overload .
.-.... 50% Overload I," —
--+-- 100% Overload

 

100

  
 

90*-

 

I
l

80

I
l

7 O
6 O l- -

v, "
I I"
I, I.
5 O I .-'
’- II ._-‘ -l
0' l”
o “.-
,.

PLR (%)

I"-
a’. I."
40 .-' _
F 5’ _."
I’. J.
I .-‘
I. 2'

I
l

30

I

10

.-'
.....
.....
....
.....
.....
0"
.....
,.-
..
.c'
...r
or-
-.--
..
.o’
o..-
..
..
-.
.0-
..~'

.....
o ’9

 

 

 

 

L 511 EL2
Video Layers

Figure 5-1 PLR for Simulation Scenario 1.

As shown in Figure 5-1, an increase in overload does not affect the transmission of
the base layer. This behavior indicates that during severe network overload conditions,
BL packets will retain the highest priority of being transmitted without any losses. For
EL1, the PLR increased significantly (i.e. upto 12%) at 100% overload. Notice that no

packet losses appeared in EL1 up to 50% overload. As expected, the maximum packet

63

losses occurred during the transmission of EL2, which was mapped as the lowest priority
red packet flow. The PLR of EL2 increased from 2.3% to 99.5% with a corresponding
increase in the network overload from 0% to 100%. Hence, equation (3.1) is satisfied at

every network overload level.

Performance Factor: Scenario I

E0 ::| -------------- I -------------- I -------------- l
8%

10% 50% 100%
Overload Level

 

 

 

Figure 5-2 Evaluated PF for Simulation Scenario 1.

The PF calculated using the above PLR results for this scenario is shown in Figure
5-2. An expected decay in the PF value is observed during the transition of network
overload from 0% to 50%, and from 50% to 100% overload. However, the PF value
degraded more during the second transition due to an increased number of packet losses

in EL1, which is the most important enhancement layer due to its maximum weight.

64

5.2 Simulation Scenario II: Multiple-to-One Mapping

5.2.1 Simulation Scenario II (a): Multiple-to—One Mapping
using Flat Rate Pricing Model

The PLR response obtained by simulating the multiple-to-one mapping using the flat
rate pricing is shown in Figure 5-3. The base layer (BL), being the highest prioritized
packet flow, suffered no packet losses at any network overload state. Under normal
network conditions, all equally prioritized enhancement layers (EL1-EL9) suffered some
packet losses. Since, in this case, the network commitments were based on the flat rate
pricing model, the price for each enhancement layer is paid to transmit 99% of the
delivered packets, and thus, not more than 1% packet losses were expected during the
transmission of each enhancement layer. The higher—than expected PLR values shown in
the figure, indicates queue overflow at the ingress router. Since every packet belonging to
any of the enhancement layers (from every video source) is buffered in the same queue
(i.e., same yellow packet queue due to the flat pricing strategy), and because (virtually)
all of these packets remained “in-profile”, then these packets created random queue

congestion. Therefore, the queuing mechanism (WRED) dropped a large number of
packets randomly, according to the specified drop probability parameter maxp for the

queue. Consequently, every enhancement layer suffered more than 1% packet losses.
However, a closer to flat response is visible for the enhancement layers only at 0%
overload, which does not comply with the required transmission quality as given by

PLRr < PLR: + 1.

65

For an increasing overload from 10% and above, an undesirable and random PLR
behavior is obtained as shown in Figure 5-3. Neither the obtained PLR for all equally
mapped (and equally paid) enhancement layers in the video stream met the required
relationship i.e. PLRr < PLRm, nor a flat PLR response is obtained at these overload

levels during the simulation of this scenario.

Network PLR Response for Scenario II(a)

 

 

 

 

 

 

 

 

‘l 00 r r r ,
—9— 0% Overload
_ --I- 10% Overload 2
90 "O"- 50% Overload J
--+~ 100% Overloa_d__
80 - -
7O - a
h, Group A Group B
A 60 r i ................. 4 ............... .4)
50 i A--.. I- """"""""" 9- ............... r """"""
E 50 r- if! "--~.- ............. O ................ 0'.-.-- “1‘
E . .........................
4o - ................................ . q
:5, u ............ ‘ .........
--------- . .......... , ........................
3O __ 5:5: ............... ‘.. ....... l ............ .4
20 - ~
5". I 1"- -n
10 by; ‘‘‘‘‘‘ T" -------- o ........... 1""1 ______ ' ........... ' """""""" r -.
1"!" +f #:¥ 3 *7”
L EL1 EL2 EL3 EL4 EL5 EL6 EL7 EL8 EL9
Vrdeo Layers

Figure 5-3 PLR for Simulation Scenario II (a).

Under overload conditions, the PLR behavior of enhancement layers shows an
anomalous behavior, which divides the enhancement layers into two groups. First group,
say ‘group A’ contains lower enhancement layers starting from EL1, and the second

group of layers, say ‘group B’ contains the upper enhancement layers that starts from the

66

next layer above ‘group A’ up to the last enhancement layer EL9. The vertical lines in
Figure 5-3 bifurcate the curves into these groups. The PLR of enhancement layers in
‘group A’ are found to be monotonically decreasing with a maximum PLR in EL1. This
behavior shows that PLR of enhancement layers in this group did not fulfill the
requirement for quality transmission of video as given by equation (3.1), i.e.

PLR, < PLR instead it shows a contradictory behavior i.e. PLR, 2 PLR

(i+l) °

According

(1+1),
to Figure 5-3, ‘group A’ includes enhancement layers EL1 to EL5 at 10% and 50%
overloads, and EL1 to EL4 at 100% overload. On the other hand, the PLR values for the
enhancement layers in the second group, i.e. ‘group B’ satisfy the PLR constraints

PLR, < PLR

((+1) '

The evaluated PF values for this scenario at four network overload levels are shown
in Figure 5-4. In this scenario, the video streams contained nine enhancement layers,
instead of two enhancement layers as in scenario I. However, the total bitrate of the
enhancement layers is the same in both scenarios, but the ratio of individual enhancement
layer bitrates in scenario I and II is 4.5:1 (i.e., 9:2). A similar ratio (i.e. 4.521) is true for
the number of packets belonging to each enhancement layer. Thus, the reduced number
of packets sent for EL1 and EL2 (with same weights) in this scenario (i.e., scenario IIa),
affects the evaluated PF less significantly (i.e., reduced by almost half as compared to the
evaluated PF values in scenario I). This explanation is true for all simulation results of
scenario II since the same number of enhancement layers and same weight distribution is

used in subsequent two scenarios.

67

Performance Factor: Scenario II(a)

 

0.25 ........................................................... _

0.2 ------------ 7 ——————————— ~ ________________________________ L

 

 

 

 

 

 

8% 1 0% 100%
Overload Level

Figure 5-4 Evaluated PF for Simulation Scenario II (a).

Except at 0% overload, the maximum number of packet losses occurred in EL1
during transmission at every network overload state. With an increase in network
overload, the PLR for EL1 increased from 3.9% at 0% overload up to 63.3% at 100%
overload. Similarly, the PLR of EL2 also increased from 5.4% to 55.7% as network was
overloaded from 0% up to 100%. The effect of such high PLR in the two highest priority
enhancement layers is clearly visible in Figure 5-4, where the PF value degraded from

0.15 at 0% overload to 0.09 at 100% overload.

68

5.2.2 Simulation Scenario II (b): Multiple-to-One Mapping
using Exponential Pricing Model

The EPM was proposed to achieve an exponentially increasing PLR behavior for the

nine equally mapped enhancement layers, and consequently meet the PLR constraints:

PLR, < PLRW). The PLR response obtained by the simulations of this scenario is shown

in Figure 5-5. The exponentially increasing PLR curve for all video layers is obtained at
0% network overload, which shows an improved quality of delivered video as compared
to the previous scenario II (a). The base layer received highest forwarding priority and
suffered no losses during transmission at any network overload level.

Similar to the PLR behavior for the previous scenario H (a) at network overload
conditions, the enhancement layers can be categorized into two groups according to their
PLR response in this scenario. ‘Group A’ contains the enhancement layers which shows a
monotonically decreasing PLR behavior starting from a maximum value in EL1. This
group includes enhancement layers EL1 to EL3 at 10% overload, EL1 to EL5 at 50%
overload, and EL1 to EL6 at 100% overload. The desired relationship in PLR of the

video layers, i.e. PLR, < PLR

(M), is observed in ‘group B’ during every network overload
condition.

At 0% and 10% overload conditions, i.e. minimal overload levels, maximum packet
losses are seem to be pushed towards the upper most layer(s). However, this model is

found to be unsuccessful in maintaining the over all exponentially increasing PLR during

higher network overload states i.e. 10%, 50% and 100% overload.

69

Network PLR Response for Scenario II(b)

x I I I I I I
—o— 0% Overload

--I-- 10% Overload

.0.... 50% Overload

~+~ueagemmq

 

100

 

   
 

90

80r

60

I
‘C.’
‘I

PLR (%)
()1
‘3

ii A.
40 _ l

I

20

.
If
I- ’1
-‘.-‘
I:
'0
!:

 

 

 

 

L EL EL2 EL3 EL4 EL5 EL6 EL7 EL8 EL9
VideoLayers

Figure 5-5 PLR for Simulation Scenario II (b).

The PF evaluated by using the PLR results of this scenario is shown in Figure 5-6.
The PF value decreased nominally, with increasing network overload from 0% to 10%.
With further increase in network overload, the PF value decreased from 0.16 (at 0%
overload) to 0.09 (at 100% overload). By using the EPM, the maximum PLR is observed
in EL9, at every network overload level. However, with an increase in network overload,
the PLR in EL1 also increased from 1.04% at 0% overload up to 59.3% at 100%
overload. Also, the PLR in EL2 increased from 1.3% to 48.8% with the corresponding
increase in network overload from 0% to 100%. Consequently, the expected qualitative
performance behavior is observed only up to a nominal overload of 10%, but the

performance degraded significantly at higher network overloads.

70

Performance Factor: Scenario II(b)

 

0.25 .......................................................... ,

 

 

 

 

 

 

Overload Level

Figure 5-6 Evaluated PF for Simulation Scenario II (b).

5.2.3 Simulation Scenario II (c): Multiple-to-One Mapping
using Linear Pricing Model

The network PLR response obtained by using the LPM for multiple-to-one mapping
strategy is shown in Figure 5-7. With an increase in the network overload, the linearity
manifested itself in the PLR of the enhancement layers. The PLR response became linear,
as the network was overloaded up to 100%. As shown in Figure 5-7, the base layer does
not suffer any packet losses during its transmission at any network overload condition.
Whereas, the desired PLR constraints as given by equation (3.1), i.e. PLRr < PLRr + 1, are

satisfied by all video layers at every network overload level, ranging from 0% to 100%.

71

Network PLR Response for Scenario II(c)

 

 

 

 

1 oo - . , y 1 . 1 . ,
“- 0% Overload '
--o- 10% Overload / I.“
90 e .0.... 50% Overload If a
-4». 100% Overload """""
.- A"
80- .../r ..... q
of”. pi" ,u,’
70 I ."I 0
0. 60- -
Q ."l ....-
m 56- _
"'1 ,1 . '-
9‘ 40 I .../Ir”:.0...-”I..- -l
'.'.""'- “...-.... . V"
30- ..r 0
IIIII .A i n "x
20- .” fl
.1", “3"...“ I {.y ’’’’’’ ‘ ' — - "
1 O P- I’.‘ ,.-""‘ . e’f I .1
1'13“" ’ . . u
i I' '1,
.’.-'
,.

 

 

 

L EL1 EL2 EL3 EL4 EL5 EL6 EL7 EL8 EL9
Video Layers

Figure 5-7 PLR for Simulation Scenario II (c).

As shown in Figure 5-7, higher PLR values appeared in the enhancement layers EL2
to EL8 even during nominal network overloads of 0% and 10%. Such high PLR values
were not observed in these layers when applying the flat rate and the EPM. Meanwhile,
EL1 did not suffer any significant packet losses as compared to the corresponding packet
losses for the prior discussed pricing models during 10%, 50% and 100% overloads
conditions.

For the resulting PLR, the corresponding PF in Figure 5-8 shows a similar behavior to
the other scenarios i.e. a corresponding decrease in PF with increasing network overload.
However, the PF value did not significantly degrade during the transition of the network
overload from 0% to 100% overload. Best performance in this scenario is observed at 0%

overload, which degraded slightly from 0.19 down to 0.17 at 100% overload. This shows

72

that the performance and quality of transmitted video can be better sustained by applying
network bandwidth commitments based on the LPM, as compared to the other pricing

techniques used in this thesis.

Performance Factor: Scenario II(c)

 

0.25 —————————————————————————— ~ ~ ______________________________ __

 

 

 

 

 

 

8% 10% 100%
Overload Level

Figure 5-8 Evaluated PF for Simulation Scenario II(c).

The results in Figure 5-7 and Figure 5-8 were obtained by using the maximum

possible value of linear pricing factor i.e., B = —r_1 , in equation (3.4). The PLR results
it —

show a desired PLR behavior for all video layers in the stream as depicted by the
equation (3.1). Additionally a sustained performance is observed during all the simulated
network overload conditions. Moreover, we analyze the network behavior and
performance using two different fractional values of B, i.e. 0.1 B and 0.5 B, for the same
simulation setup, to explore any possibility for a further reduced cost when using the

LPM. The results obtained by the simulations are shown in Figure 5—9.

73

Network PLR Response using 0.15 Network PLR Response using 0.53

 

 

 

   

 

 

 

 

 

 

 

 

 

 

100 . . _. . 100 ,0
'—o—0%o.L l-'-0%0-L
W -+- 10% o.L 80" e. 10% CL ,
...... 50% O.L . 50% CL
-+- 100%O.L l-~- 100%0 L ”..-..-«l
a? 60- £3 60' ..- _____ c, ............. 41
v v ''''''''''' ‘I.
“ -,- m I"- """" v. """" O ...... . J ......... C... h
: 40> g 40’ '3. a ........ a
' ,- ..... --------
if ------- . ----- l
20 ' 20 . !'
3L EL1 EL2 EL3 EL4 EL5 euo EL7 EL8 EL9 L EL1 EL2 EL3 EL4 EL5 Ebb EL7 EL8 EL9
Video Layers Video Layers
Performance Factor using 0.15 Performance Factor using 0.56
0.25 —————————————————————————————
0.2
g: 0.15

 

 

 

 

 

 

 

 

 

Overload Level Overload Level

Figure 5-9 PLR and evaluated PF using two different fractional values of B for Scenario
II(c).

As shown in Figure 5-9, while using 0.1 B, relatively lower PLR is obtained for
enhancement layers at 0% and 10% overload as compared to the previous case. As we
increase the network overload to 50% and 100%, the undesirable PLR curve
characteristics set in, similar to the prior cases of the flat rate and EPM. Increase in the
number of packet losses in the lower enhancement layers (EL1 to EL3) also affects the
evaluated PF value, which degraded from 0.15 at 0% overload to 0.11 at 100% overload.

In the second case, using 0.5 B, we obtained approximately a linearly increasing PLR
for video layers at nominal network overloads of 0% and 10%. During higher network

overloads of 50% and 100%, again EL1 and EL2 suffered an undesirably large number of

74

packet losses. As a result, the evaluated PF degraded from 0.16 at 0% overload down to
0.13 at 100% overload. The PF values obtained under normal network condition, i.e. at
0% overload, in both cases (i.e. 0.15 by using 0.1 B and 0.16 by using 0.5 B) are also
significantly smaller than the PF value (0.19) for the same network overload condition

using maximum possible B.

5.3 Comparison of the Pricing Models

The criterion for the estimation of the delivered video quality was based on the
relation PLRr<PLRr+r. Under normal conditions, i.e. at no network overload, only
exponential and linear pricing models achieved this criterion. On the other hand, flat rate
model showed higher PLR values for lower (important) enhancement layers during
similar network overload condition. As the network was heavily overloaded to 50% or
100%, lower enhancement layers suffered from a high PLR. The PLR curve of three
pricing models (i.e. flat rate, EPM and LPM) at two extreme network overload conditions
of 0% and 100% are shown in Figure 5-10 and Figure 5-11.

As shown in Figure 5-10, the PLR response for EPM in the upper layers increased
exponentially with a very low change in PLR values for the lower enhancement layers.
This behavior depicts an ideal transmission of scalable video streams over lossy packet
networks such as the Internet, where all possible packet losses are pushed towards the
layer(s) with minimum weight. Whereas, a comparison of the PLR in the most important
video layers, in the video stream, is shown by a circle in Figure 5-10. On the other hand,
in case of LPM, the PLR started increasing from EL2, whereas a medium PLR value is

appeared in EL1 as compared to the other two pricing models.

75

PLR Comparison of Pricing Models at 0% Overload

 

100

I I

 

+ Flat Rate
-~&~ Exponential .'
90 -"_*'°_L.‘9€ai__ 5‘

 

 

I
l

80

I
l

.’ I
'I
.I I
I I
O l' l
i - I
x '
r
.I r
I I
‘ l

I, '
a, I
.’ I
' l
6 0 h .0. ‘ d
a l
I' '
‘ I
0’ I
I

n I
a. l
I
x’ .
l d
I
l
.
I

a
I"
a"
a‘
l"-
'.
.
,0

PLR (%)

I
l

40

30

I
.‘.
\
K.
‘Q
‘l
l

20

.0'
O
.o'

10 - ------ ‘ 4

_,_-+__--/- .... T ........... T fit 1 L
EL4 EL5 EL6 EL7 EL8 EL9
Video Layers

   

 

 

 

 

Figure 5-10 PLR comparison of pricing models at 0% overload.

At 100% network overload, as depicted in Figure 5-11, an apparently low PLR is
observed using LPM for enhancement layers EL1 to EL4 as compared to the other two
pricing models. A circle in Figure 5-11 shows this variation in PLR. Consequently the
upper enhancement layers EL5 to EL9 suffered a high PLR for this scenario.

Hence, the results obtained from each simulated scenario of multiple-to-one mapping
strategy illustrates that the SLA based on LPM shows a low PLR in the lowest
enhancement layer EL1, which is the next most important layer in the video stream after
the base layer. This behavior was observed even during heavy network overload
conditions. EPM demonstrates the desired PLR response with each increasing video layer

only during normal overload conditions. This condition is supposed to be ideal for

76

scalable video transmissions, where every expected packet loss is pushed towards the last
video layer of the stream.

We analyze the network behavior in terms of its performance factor, which in turn is
affected by the cost. Due to the high cost paid for lower enhancement layers, EPM does
not show any performance advantage over the LPM. The cost comparison of each model

is presented in the next section.

PLR Comparison of Pricing Models at 100% Overload

 

 

 

 

 

1 00 I I T I I r I I '. ‘
+ Flat Rate .-" I,
"Ar- Exponential fg
90 ~ “It-~- Linear "x ..... _‘
./’
80 - Ix ..... .1
...!" If;
70 I. _;'x. ‘, -1
A 60 I x "I I
50 'I ~‘\ .o' x
V .' 1"! +_ ”A;
E 50 h ‘ ..x’"
7“. _- ..v" _____-"
a" 40 f ,,,,,,,,, i ------------ .-“" q
0 .x'
30 ,4!” ‘
5",
20 '- it}. _
x"
1 0 r .-/ _
.-
n”‘
I

 

 

 

L EL1 EL2 EL3 EL4 EL5 EL6 EL7 EL8 EL9
Video Layers

Figure 5-11 PLR comparison of pricing models at 100% overload.

5.4 A Comparative Performance Analysis
The performance analysis presented here is based on two key factors: firstly, the PF
loss occurred in every strategy when the network is overloaded from 0% to 100%, and

secondly the total cost associated with each scenario. The PF loss provides a measure for

77

the change in overall performance as the network conditions change. Based on this
measure, it is desirable that the PF changes gracefully as the network conditions change.
Below, we also show the (absolute measured) cost in terms of the bitrate commitment of
network (i.e. CIR) according to Equation (3.16).

In order to evaluate the performance loss (P/L) during the transition of network

overload from 0% to 100%, we used the following equation:

Pf (100)
Pf (0)

where Pj(0) and Pj(100) are the corresponding PF values at 0% and 100% network

 

PjL = l— (%) (4.9)

overload, respectively. The estimated PF loss and total cost (c,) for each simulated

mapping strategy are listed in Table 5.1 and Table 5.2 respectively.

 

 

 

 

 

 

Table 5.1 Comparison of PF loss in Simulation Scenarios.
Simulation
Scenario Pf(0) Pf<100) Pfl (%)
I 0.2671 0.2181 18.34
II(a) 0.1495 0.0969 35.18
II(b) 0.1616 0.1110 31.31
II(c) 0.1935 0.1743 9.92

 

 

 

 

 

Table 5.2 Cost comparison of Simulation Scenarios.

 

 

 

 

 

 

Simulation Cu (bps) at clap (bps) at cAF (bPS) at CAP (bps) at
Scenario 0% overload 10% overload 50% overload 100% overload
I 501982 524480 614471 726959
II(a) 560971 590968 710956 860941
II(b) 522272 547621 649224 776553
II(c) 429885 443216 496543 563218

 

 

 

 

 

78

 

On the basis of the data in Table 5.1 and Table 5.2, we recommend LPM as a
framework to attain (1) better quality delivery of scalable video stream at minimum cost,
and (2) sustained performance, with a performance loss of only 9.92 % at 100% network
overload.

As described in previous section, the EPM shows an ideal PLR behavior for scalable
video streaming at normal network condition. This pricing model, however, did not show
any consistency in performance with an increasing network overload. Also the cost
associated (from Table 5.2) with the performance and quality delivered by the EPM is
highest among all the strategies in consideration, even at normal network condition. As
compared to the LPM the mean ratio of the total costs for both strategies was also found
to be 1: 0.779. By a comparison of the PLR obtained at no overload, the EPM along with
the given weight distribution and a high cost provides the best quality of video to the
receiver(s). However, this pricing model is also not capable of preserving the desired
PLR behavior when the network bandwidth is over-committed.

One-to-one mapping was used in this thesis as a specific case where the generated
scalable video streams contained only three layers. In order to utilize the available
bandwidth of the network links we used high bitrate video layers in this mapping
category. Therefore an overall better performance at every network overload (by using a
high bitrate video stream) was obtained, but the performance was not consistent when the
network overload was increased from 0% to 100%. A loss of 18.34% in the performance

was observed.

79

6 Conclusions

The main objective of this thesis was to develop methods that provide high quality
video transmission over DiffServ. One-to-one and multiple-to-one mapping strategies and
related pricing models were proposed and simulated to realize this goal. DiffServ is a
commitment-based network, where cost is associated with every transmitted packet. For
this reason we analyzed the performance of each mapping strategy and the corresponding
pricing models in terms of the cost and the delivered quality i.e., weighted throughput.
Another main objective of this thesis was to analyze the effect of network bandwidth
over-commitments on the two mapping categories.

The mapping schemes for streaming video were developed for a medium priced class
of assured forwarding (AF) service. For this purpose “flow based” video transmission
was proposed to apply multiple service level agreements (SLAs) to a single video stream.
Since performance of every packet stream over DiffServ is also affected by the associated
cost of the network transmission services, we used the bandwidth commitment parameter
committed information rate (CIR), to control the packet losses (or throughput).

During this work we used four network overload conditions to illustrate the over-
commitments by the service provider(s) of DiffServ. The simulations performed at 0%

overload represent the normal condition where a service provider is committing the

80

channel bandwidth without exceeding its maximum capacity. The two mapping
categories were mainly developed to improve the video quality at normal network
conditions. Exponential pricing model (EPM) and Linear pricing model (LPM) were
designed for network bandwidth commitments, according to the given weight distribution
of video layers. These pricing models demonstrated desired throughput results and thus
implied an improvement in delivered video quality when compared to the current best-
effort Internet. The associated cost for the application of LPM and EPM was also found
to be less than that obtained using the traditional flat rate technique, which proves that a
better quality video is achievable without paying a higher cost.

One main reason for analyzing the effect of over-commitments on the delivered
quality and performance was to present a framework in which a network can guarantee a
stable quality even when the network bandwidth exceeds the available channel capacity.
The suggested EPM was able to provide the desired performance and quality during only
low over-commitment levels (i.e. up to 10% overload), whereas the LPM showed results
that provide assured quality and better performance upto 100% of over-commitment
levels.

The simulation results of the proposed low-cost framework, i.e. LPM for multiple-to-
one mapping strategy, is capable of preserving the video quality (with minimal
performance loss of 9.92%), even as the network link is over-committed upto 100%
(which is unusual in practice). Based on the parameter values used, our simulations show
that the LPM could provide graceful (and relatively minor) degradation in overall quality
while achieving higher utilization of the link capacity. The multiple-to-one mapping

scheme is applicable to video streams containing several video layers. On the other hand,

81

the one-to-one mapping strategy can be reference to as a high-cost framework to obtain a
consistent quality and performance at least up to 50% over-committed channels.

Future Work Suggestions for future work involve the utilization of improved traffic
shaping mechanisms and metering-marking algorithms, to control the PLR within a AF
class. A single tradeoff pricing structure to provide the benefits of EPM and LPM during
severe network congestion levels could also provide further extension in this work.

In this thesis constant bit rate (CBR) video stream are used in our simulations.
Scalable mapping strategies and bandwidth commitment models could be designed to
accommodate variable bit rate (VBR) video streams to attain a high quality video
transmission and performance for previously known and for emerging video coding

standards and methods.

82

[1]

[2]

[3]

[4]

[5]
[6]
[7]

[8]

[9]

[10]

[ll]

[12]
[13]

[14]

References

Holzmann, G.J., B. Pehrson. The Early history of Data Networks. IEEE
Computer Society Press, Los Alarnitos, CA, 1995.

IETF RFC 791, Internet Protocol, DARPA Internet Program Protocol
Specification, Sep 1981.

Stevens W. Richard TCP/1P Illustrated. Volume 1: The Protocols. Addison
Wesley, Reading, MA, September 1998.

Comer, D.E. Intemetworking with TCP/IP. Volume I : Principles, Protocols, and
Architecture. Prentice Hall, Eaglewood Cliffs, NJ, 3rd Edition, 1995.

IETF RFC 768, User Datagram Protocol, Aug 1980
IETF RFC 1889, RTP: A transport Protocol for Real time Applications, Jan 1996

Jacobson V. Congestion Avoidance and Control. Proceedings of the
SIGCOMM’88 pg. 314-329, Aug. 1988.

Ferguson, Paul, and Huston, Geoff. Quality of Service: Delivering QoS on the
Internet and in Corporate Networks. New York: John Wiley & Sons, 1998.

IETF RFC 2386, "A Framework for QoS-Based Routing in the Internet."

J. Boyce, R. Gaglianello, Packet Loss efiects on MPEG video sent over the public
Internet, Proceedings of ACM Multimedia Conference, UK, 1998.

Draft International Standards, Generic Coding of Moving Pictures and Associated
Audio Information: Video, ISO/[EC 13818-2, Nov 1994

IETF RFC 3246, An Expedited Forwarding PHB, March 2002.
IETF RFC 2597, Assured Forwarding PHB Group, June 1999.

The Network Simulator NS2, www.isi.edu/nsnam/dist

83

[15]

[16]

[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24]

[25]

[26]

[27]

[28]

Floyd, S. and V. Jacobson. Random Early Detection Gateways for Congestion
Avoidance, [BEE/ACM Transactions on Networking, vol. 1 No. 4, Aug 1993.

Martin May, Jean Bolot, Christopher Diot, and Bryan Lyles. Reasons not to
deploy RED. In Proceedings of the IEEE/IFIP International Workshop on Quality
of Service (IWQoS'99), June 1999

D. Clark and W. Fang. “Explicit allocation of best-eflort packet delivery service
IEEE/ACM Transactions on Networking, 6(4):362--373, August 1998.

T. Ziegler, S. Fdida, U. Hofmann, "RED+ Gateways for Identification and
Discrimination of unfriendly best-eflort Flows in the Internet", Proceedings of

IFIP Broadband Communications 99, November 1999

Feng, W., D. Kandlur, D. Saha, and K. Shin A Self—Configuring RED Gateway,
Proceedings from INFOCOMM 99, March 1999.

Lin D., R. Morris, Dynamics of Random Early Detection, Proceedings from ACM
SIGCOMM 97, Oct 1997.

H.Eriksson, MBONE: The Multicast Backbone, Comm ACM 37 (8) Aug 1994.
ITU-T Recommendation H.261,Video Codec for Audio-visual services at Px 64
Kbits, Mar 1993.

ITU-T Recommendation H.263, Video Coding for Low bitrate Communication,
March 1996.

International Organization for Standardization JT Cl/SC29/WG11 CODING OF
MOVING PICTURES AND AUDIO MPEG 96/June 1996

International Organization for Standardization JT C1/SC29/WGll CODING OF
MOVING PICTURES AND AUDIO MPEG 00/October 2000

International Organization for Standardization JTCl/SC29/WG11 CODING OF

MOVING PICTURES AND AUDIO MPEG March 2002

I. Busse, B. Deffner, H. Schulzrinne, Dynamic QoS control of multimedia
applications based on RTP, Computer Communications 19 pgs. 49-68, Jan 1996.

Y. Wang, Q.F. Zhu, Error Control and concealement for video communication- a
review, Proc. IEEE 86 (5) May 1998.

84

[29]

[30]

[31]

[32]

[33]

[34]

[35]

[36]

[37]

[38]
[39]

[40]

[41]

[421

[43]

[44]

J. M. Boyce, Packet Loss Resielient transmission of MPEG Video over the
Internet, Signal Processing: Image Communication 15, pgs 7-24, 1999.

M. Schaar, H. Radha, Unequal Packet Loss Resielience for Fine-Granular-
Scalibility Video, IEEE Transactions on Multimedia, 3 (4) ,Dec 2001.

H. Radha et. al, Scalable Internet Video using MPEG-4, Signal Processing: Image
Communication 15(1999) pg. 95-126

IETF RFC 1633, Integrated Services in the Internet Architecture: an Overview,
June 1994.

IETF RFC 2474,Definition of the Differentiated Services Field (DS Field) in the
Ipv4 and Ipv6 headers, Dec 1998.

IETF RFC 2475, An Architecture for Diflerentiated Services, December 1998.

IETF RFC 2430, A Provider Architecture for Diflerentiated Services and Traffic
Engineering (PASTE), October 1998.

M. Goyal, A. durresi, R. Jain, Performance Analysis of Assured Forwarding,
IETF Internet Draft draft-goyal—diffserv-afstudy-OO

Xipeng Xiao, Lionel M. Ni, Internet QoS: the Big Picture, IEEE Network,
l3(2):8-18, Mar-April 1999.

IETF RFC 2697, A Single Rate Three Color Marker, September 1999.
IETF RFC 2698, A Two Rate Three Color Marker, September 1999.

IETF RFC 2859, A Time Sliding Window Three Colour Marker (TS WTCM), June
2000.

Mohammed Atiquzzaman, Hongjun Su, ItszCM: A New Aggregate Marker to
Improve Fairness in DiffServ, presented in [PS 2001, (a part of Globecom 2001),
November 2001.

M.Goyal, A. Durresi, R. Jain, Chunlie Liu, Eflect of Number of Drop Precedences
in Assured Forwarding, IETF draft-goyal-dpstdy-diffserv-01.txt, June 1999.

Mohamed El-Gendy, Kang G. Shin Equation-Based Packet Marking for Assured
Forwarding Services, selected for INFOCOMM 2002.

I. Andrikopoulos, L. Wood and G. Pavlou, A Fair Traffic Conditioner for the
Assured Service in a Difi‘erentiated Services Internet, IEEE [CC

85

[45]

[46]

[47]

[48]

[49]

[50]

[51]

[52]

[53]
[54]

[55]

IETF RFC 2963, A Rate Adaptive Shaper for Diflerentiated Services, October
2000

N. Semret, R. Liao, A. Campbell, and A. Lazar, "Pricing, Provisioning and
Peering: Dynamic Markets for Differentiated Internet Services and Implications
for Network Interconnections", IEEE Journal on Selected Areas of
Communications, Vol. 18, No. 12, pp. 2499-2513, December 2001.

M. S. Borella, V. Upadhyay and I. Sidhu. Pricing Framework for a Differential
Services Internet. European Transactions on Telecommunications, Vol. 10(2),
March/April 1999.

IETF RFC 3140, Per Hop Behavior Identification Codes, June 2001

IETF RFC 3086, Definition of Diflerentiated Services Per Domain Behaviors
and Rules for their Specification, April 2001

U. Bodin, S. Schalen and S. Pink, “Load Tolerant Differentiation with Active
Queue management”, SIGCOMM Computer Communications vol. 30 n3 , July
2000

IETF website, http://www.ietf.org

S.Sahu, P. Nain, D. Towsley, C. Diot and V. Firoiu, “On Achievable Service
Difierentiation With Token Bucket Marking for TCP”, In proceedings of ACM
SIGMETRICS, June 2000.

Cisco official web site, http://www.cisco.com

NS Manual, http://www.isi.edu/nsnam/ns/ns-documentation.html

H. Radha, M. van der Schaar, Y. Chen, "The MPEG-4 Fine-Granted Scalable

Video Coding Methodfor Multimedia Streaming over IP," IEEE Transactions on
Multimedia, March 2001.

86

Appendix

DiffServ Metering-Marking Algorithms
The three types of markers suggested by IETF belongs to two different categories:
1. Token Bucket Markers, includes
i. Single Rate Three Color Marker (srTCM) [38]
ii. Two Rate Three Color Marker (trTCM)[39]
2. Time Slide Window Markers
iii. Time Slide Window Three Color Marker (TSWTCM) [40]
We used srTCM and trTCM marking algorithms for AF packet metering and marking

during the simulations. The standard specification of these algorithms is described below:

Single Rate Three Color Marker (srTCM)

Two token buckets C and P with their sizes CBS (bytes) and PBS (bytes) respectively
are used in this algorithm. An incoming packet with rate B (bytes per second) is marked
as a green or yellow packets depending on the number of available tokens. The srTCM

algorithm can be described by the following figure.

87

 

¢C IR

Tc

 

CBS

 

 

 

C

PBS

iC IR

 

 

Tp

 

 

P

 

 

Incoming packets with rate =r, size=B

 

 

 

>

 

Figure A-1

srTCM Algorithm logic diagram.

The token buckets are initially full at time 0, i.e Token Count Tc(0) =CBS and Token

count Tp(0)=PBS. Thereafter the token count Tp is incremented by one CIR times per

second upto PBS, and token count Tc is incremented by one CIR times per second upto

CBS. For every incoming packet of size B and rate r, the srTCM algorithm can be

summarized as:

STEP 1. If Tc(t) — B >=0, the packet is Green, and Tc is decremented by B,

else

STEP 2. If Tc(t) —B>=0, the packet is Yellow, and Te is decremented by B,

else

STEP 3. The packet is Red and neither Tc nor Te is decremented.

The srTCM works in both color blind as well as color aware modes. Marking of a

packet with a given color requires that there be enough tokens of that color available to

accommodate the entire packet. Committed Information Rate (CIR) is rate used to refill

the token bucket in bytes per second.

88

Two Rate Three Color Marker (trTCM)

The only difference between trTCM and srTCM is that this marker use two rates and
two token bucket sizes to meter the incoming packet. Two token bucket C and P are used
with sizes CBS (bytes) and PBS (bytes) respectively. Tc and Tp are their respective token

count belongs to each token bucket.

 

¢PIR lCIR

PBS TP (:33 TC

P C
Incoming packets with rate =r, size=B >

Figure A-2 trTCM Algorithm logic diagram.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

The token buckets P and C are initially full at time 0, i.e. the Token Count
Tc(0)=CBS and token count Tp(0)=PBS. Thereafter the token count Tp is incremented by
one PIR times per second upto PBS, and token count Tc is incremented by one CIR times
per second upto CBS. An incoming packet with rate r (bytes per second) is marked as a
yellow or red packet depending on the number of available tokens.

For every incoming packet of size B and rate r, the trTCM algorithms can be given as:
STEP 1. If Tp(t) —B < 0, the packet is Red, else
STEP 2. If Tc(t) —B < 0, the packet is Yellow and Tp is decremented by B,
else

STEP 3. The packet is Green and both Tp and Tc are decremented by B.

89

Two rates i.e. Committed Information Rate (CIR) and Peak Information Rate (PIR)
are used to re-fill their respective token buckets. The PBS and CBS are measured in bytes
and they always configured to be greater than 0. On the other hand PIR must not be less

than CIR.

9O

MICHIGAN S IATE UNIVERS

lull/ll)willlllllllll‘illlll

3 1 93 02427 1235