Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans

Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

Introduction to IP and ATM

Design and Performance

Introduction to IP and ATM

Design and Performance
With Applications Analysis Software
Second Edition

J M Pitts
J A Schormans
Queen Mary
University of London
UK

JOHN WILEY & SONS, LTD

Chichester

New York

Weinheim

Brisbane

Toronto

Singapore

First Edition published in 1996 as Introduction to ATM Design and Performance by John Wiley & Sons, Ltd.
Copyright 2000 by John Wiley & Sons, Ltd
Baffins Lane, Chichester,
West Sussex, PO19 1UD, England
National
01243 779777
International (C44) 1243 779777
e-mail (for orders and customer service enquiries): [email protected]
Visit our Home Page on http://www.wiley.co.uk or http://www.wiley.com
Reprinted March 2001
All Rights Reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted,
in any form or by any means, electronic, mechanical, photocopying, recording, scanning or otherwise, except
under the terms of the Copyright Designs and Patents Act 1988 or under the terms of a licence issued by
the Copyright Licensing Agency, 90 Tottenham Court Road, London, W1P 9HE, UK, without the permission
in writing of the Publisher, with the exception of any material supplied specifically for the purpose of being
entered and executed on a computer system, for exclusive use by the purchaser of the publication.
Neither the authors nor John Wiley & Sons Ltd accept any responsibility or liability for loss or damage
occasioned to any person or property through using the material, instructions, methods or ideas contained
herein, or acting or refraining from acting as a result of such use. The author(s) and Publisher expressly disclaim
all implied warranties, including merchantability or fitness for any particular purpose. There will be no duty
on the authors or Publisher to correct any errors or defects in the software.
Designations used by companies to distinguish their products are often claimed as trademarks. In all instances
where John Wiley & Sons is aware of a claim, the product names appear in initial capital or all capital
letters. Readers, however, should contact the appropriate companies for more complete information regarding
trademarks and registration.
Other Wiley Editorial Offices
John Wiley & Sons, Inc., 605 Third Avenue,
New York, NY 10158-0012, USA
Wiley-VCH Verlag GmbH
Pappelallee 3, D-69469 Weinheim, Germany
Jacaranda Wiley Ltd, 33 Park Road, Milton,
Queensland 4064, Australia
John Wiley & Sons (Canada) Ltd, 22 Worcester Road
Rexdale, Ontario, M9W 1L1, Canada
John Wiley & Sons (Asia) Pte Ltd, 2 Clementi Loop #02-01,
Jin Xing Distripark, Singapore 129809

British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library
ISBN 0471 49187 X
Typeset in 10 12 /12 12 pt Palatino by Laser Words, Chennai, India
Printed and bound in Great Britain by Bookcraft (Bath) Ltd
This book is printed on acid-free paper responsibly manufactured from sustainable forestry,
in which at least two trees are planted for each one used for paper production.

To
Suzanne, Rebekah, Verity and Barnabas
Jacqueline, Matthew and Daniel

Contents
Preface

PART I
1

xi

INTRODUCTORY TOPICS

An Introduction to the Technologies of IP and ATM

Circuit Switching
Packet Switching
Cell Switching and ATM
Connection-orientated Service
Connectionless Service and IP
Buffering in ATM switches and IP routers
Buffer Management
Traffic Control

3
5
7
8
9
11
11
13

Traffic Issues and Solutions

15

Delay and Loss Performance

15

Source models
Queueing behaviour

Coping with Multi-service Requirements: Differentiated Performance

16
18

30

Buffer sharing and partitioning

Cell and packet discard mechanisms
Queue scheduling mechanisms

30
32
35

Flows, Connections and Aggregates

37

Admission control mechanisms

Policing mechanisms
Dimensioning and configuration

37
40
41

Teletraffic Engineering

45

Sharing Resources
Mesh and Star Networks
Traffic Intensity
Performance
TCP: Traffic, Capacity and Performance
Variation of Traffic Intensity
Erlangs Lost Call Formula
Traffic Tables

45
45
47
49
49
50
52
53

viii

CONTENTS

Performance Evaluation
Methods of Performance Evaluation
Measurement
Predictive evaluation: analysis/simulation

Queueing Theory
Notation
Elementary relationships
The M/M/1 queue
The M/D/1/K queue
Delay in the M/M/1 and M/D/1 queueing systems

69
71
73
74
75
77

77
77

Traffic Models

81

Levels of Traffic Behaviour

Timing Information in Source Models
Time between Arrivals
Counting Arrivals
Rates of Flow

81
82
83
86
89

PART II

ATM QUEUEING AND TRAFFIC CONTROL

95

Basic Cell Switching

97

The Queueing Behaviour of ATM Cells in Output Buffers

Balance Equations for Buffering
Calculating the State Probability Distribution
Exact Analysis for FINITE Output Buffers
Delays

97
98
100
104
108

End-to-end delay

Cell-Scale Queueing
Cell-scale Queueing
Multiplexing Constant-bit-rate Traffic
Analysis of an Infinite Queue with Multiplexed CBR Input: The
ND/D/1
Heavy-traffic Approximation for the M/D/1 Queue
Heavy-traffic Approximation for the ND/D/1 Queue
Cell-scale Queueing in Switches

60
60
61
64
65

69

Cell-rate simulation

58

Fundamentals of Simulation

Accelerated Simulation

57
57
57

Discrete Time Simulation

Generating random numbers
M/D/1 queue simulator in Mathcad
Reaching steady state
Batch means and confidence intervals
Validation

57

Burst-Scale Queueing
ATM Queueing Behaviour
Burst-scale Queueing Behaviour

110

113
113
114
115
117
119
121

125
125
127

ix

CONTENTS

Fluid-flow Analysis of a Single Source Per-VC Queueing

Continuous Fluid-flow Approach
Discrete Fluid-flow Approach
Comparing the Discrete and Continuous Fluid-flow Approaches
Multiple ON/OFF Sources of the Same Type
The Bufferless Approach
The Burst-scale Delay Model

10

Connection Admission Control

The Traffic Contract
Admissible Load: The Cell-scale Constraint
A CAC algorithm based on M/D/1 analysis
A CAC algorithm based on ND/D/1 analysis
The cell-scale constraint in statistical-bit-rate transfer capability, based on
M/D/1 analysis

Admissible Load: The Burst Scale

A practical CAC scheme
Equivalent cell rate and linear CAC
Two-level CAC
Accounting for the burst-scale delay factor

CAC in The Standards

11

150
151
152
153
155

157
159
160
160
161

165

Usage Parameter Control

167
167
168
172

The leaky bucket

The problem of tolerances
Resources required for a worst-case ON/OFF cell stream from peak cell
rate UPC
Traffic shaping

Dual Leaky Buckets: The Leaky Cup and Saucer

Resources required for a worst-case ON/OFF cell stream from sustainable
cell rate UPC

Dimensioning
Combining The Burst and Cell Scales
Dimensioning The Buffer
Small buffers for cell-scale queueing
Large buffers for burst-scale queueing

Combining The Connection, Burst and Cell Scales

13

149

Protecting the Network

Controlling the Mean Cell Rate
Algorithms for UPC
Peak Cell Rate Control using the Leaky Bucket

12

129
129
131
136
139
141
145

Priority Control
Priorities
Space Priority and The Cell Loss Priority Bit
Partial Buffer Sharing
Increasing the admissible load
Dimensioning buffers for partial buffer sharing

172

173
176
178
182

182
184

187
187
190
193
198

200

205
205
205
207
214
215

Time Priority in ATM

218

Mean value analysis

219

CONTENTS

PART III
14

IP PERFORMANCE
AND TRAFFIC MANAGEMENT

Basic Packet Queueing

The Queueing Behaviour of Packets in an IP Router Buffer
Balance Equations for Packet Buffering: The Geo/Geo/1
Calculating the state probability distribution

Decay Rate Analysis

Using the decay rate to approximate the buffer overflow probability

Balance Equations for Packet Buffering: Excess-rate Queueing

Analysis
The excess-rate M/D/1, for application to voice-over-IP
The excess-rate solution for best-effort traffic

15

Resource Reservation

229
230
231

234
236

238
239
245

253
253
254
255

Parameterizing the two-state aggregate process

Analysing the queueing behaviour

257
259

IP Buffer Management
First-in First-out Buffering
Random Early Detection Probabilistic Packet Discard
Virtual Buffers and Scheduling Algorithms

17

229

Quality of Service and Traffic Aggregation

Characterizing an Aggregate of Packet Flows
Performance Analysis of Aggregate Packet Flows
Voice-over-IP, Revisited
Traffic Conditioning of Aggregate Flows

16

227

261
265

267
267
267
273

Precedence queueing
Weighted fair queueing

273
274

Buffer Space Partitioning

Shared Buffer Analysis

275
279

Self-similar Traffic

287

Self-similarity and Long-range-dependent Traffic

The Pareto Model of Activity
Impact of LRD Traffic on Queueing Behaviour
The Geo/Pareto/1 Queue

287
289
292
293

References

299

Index

301

Preface
In recent years, we have taught design and performance evaluation
techniques to undergraduates and postgraduates in the Department
of Electronic Engineering at Queen Mary, University of London
(http://www.elec.qmw.ac.uk/) and to graduates on various University
of London M.Sc. courses for industry. We have found that many
engineers and students of engineering experience difficulty in making
sense of teletraffic issues. This is partly because of the subject itself:
the technologies and standards are flexible, complicated, and always
evolving. However, some of the difficulties arise because of the advanced
mathematical models that have been applied to IP and ATM analysis.
The research literature, and many books reporting on it, is full of
differing analytical approaches applied to a bewildering array of traffic
mixes, buffer management mechanisms, switch designs, and traffic and
congestion control algorithms.
To counter this trend, our book, which is intended for use by students
both at final-year undergraduate, and at postgraduate level, and by practising engineers in the telecommunications and Internet world, provides
an introduction to the design and performance issues surrounding IP
and ATM. We cover performance evaluation by analysis and simulation,
presenting key formulas describing traffic and queueing behaviour, and
practical examples, with graphs and tables for the design of IP and ATM
networks.
In line with our general approach, derivations are included where they
demonstrate an intuitively simple technique; alternatively we give the
formula (and a reference) and then show how to apply it. As a bonus,
the formulas are available as Mathcad files (see below for details) so
there is no need to program them for yourself. In fact, many of the
graphs have the Mathcad code right beside them on the page. We have
ensured that the need for prior knowledge (in particular, probability
theory) has been kept to a minimum. We feel strongly that this enhances
the work, both as a textbook and as a design guide; it is far easier to

xii

PREFACE

make progress when you are not trying to deal with another subject in
the background.
For the second edition, we have added a substantial amount of new
material on IP traffic issues. Since the first edition, much work has
been done in the IP community to make the technology QoS-aware. In
essence, the techniques and mechanisms to do this are generic however,
they are often disguised by the use of confusing jargon in the different
communities. Of course, there are real differences in the technologies, but
the underlying approaches for providing guaranteed performance to a
wide range of service types are very similar.
We have introduced new ideas from our own research more accurate,
usable results and understandable derivations. These new ideas make
use of the excess-rate technique for queueing analysis, which we have
found applicable to a wide variety of queueing systems. Whilst we still
do not claim that the book is comprehensive, we do believe it presents
the essentials of design and performance analysis for both IP and ATM
technologies in an intuitive and understandable way.

Applications analysis software

Wheres the disk or CD? Unlike the first edition, we decided to put
all the Mathcad files on a web-site for the book. But in case you cant
immediately reach out and click on the Internet, most of the figures
in the book have the Mathcad code used to generate them alongside,
so take a look. Note that where Mathcad functions have been defined
for previous figures, they are not repeated, for clarity. So, check out
http://www.elec.qmw.ac.uk/ipatm/ Youll also find some homework
problems there.

Organization
In Chapter 1, we describe both IP and ATM technologies. On the surface
the technologies appear to be rather different, but both depend on similar
approaches to buffer management and traffic control in order to provide
performance guarantees to a wide variety of services. We highlight
the fundamental operations of both IP and ATM as they relate to the
underlying queueing and performance issues, rather than describe the
technologies and standards in detail. Chapter 2 is the executive summary
for the book: it gathers together the range of analytical solutions covered,
lists the parameters, and groups them according to their use in addressing
IP and ATM traffic issues. You may wish to skip over it on a first reading,
but use it afterwards as a ready reference.

PREFACE

xiii

Chapter 3 introduces the concept of resource sharing, which underpins the design and performance of any telecommunications technology,
in the context of circuit-switched networks. Here, we see the tradeoff between the economics of providing telecommunications capability
and satisfying the service requirements of the customer. To evaluate
the performance of shared resources, we need an understanding of
queueing theory. In Chapter 4, we introduce the fundamental concept
of a queue (or waiting line), its notation, and some elementary relationships, and apply these to the basic process of buffering, using ATM as
an example. This familiarizes the reader with the important measures of
delay and loss (whether of packets or cells), the typical orders of magnitude for these measures, and the use of approximations, without having
to struggle through analytical derivations at the same time. Simulation
is widely used to study performance and design issues, and Chapter 5
provides an introduction to the basic principles, including accelerated
techniques.
Chapter 6 describes a variety of simple traffic models, both for single
sources and for aggregate traffic, with sample parameter values typical
of IP and ATM. The distinction between levels of traffic behaviour,
particularly the cell/packet and burst levels is introduced, as well as
the different ways in which timing information is presented in source
models. Both these aspects are important in helping to simplify and
clarify the analysis of queueing behaviour.
In Part II, we turn to queueing and traffic control issues, with the specific
focus on ATM. Even if your main interest is in IP, we recommend you
read these chapters. The reason is not just that the queueing behaviour
is very similar (ATM cells and fixed-size packets look the same to a
queueing system), but because the development of an appreciation for
both the underlying queueing issues and the influence of key traffic
parameters builds in a more intuitive way.
In Chapter 7, we treat the queueing behaviour of ATM cells in output
buffers, taking the reader very carefully through the analytical derivation
of the queue state probability distribution, the cell loss probability, and
the cell delay distribution. The analytical approach used is a direct
probabilistic technique which is simple and intuitive, and key stages
in the derivation are illustrated graphically. This basic technique is
the underlying analytical approach applied in Chapter 13 to the more
complex issues of priority mechanisms, in Chapter 14 to basic packet
switching with variable-length packets, and in Chapter 17 to the problem
of queueing under self-similar traffic input.
Chapters 8 and 9 take the traffic models of Chapter 6 and the concept
of different levels of traffic behaviour, and apply them to the analysis of
ATM queueing. The distinction between cell-scale queueing (Chapter 8)
and burst-scale queueing (Chapter 9) is of fundamental importance

xiv

PREFACE

because it provides the basis for understanding and designing a traffic

control framework (based on the international standards for ATM) that
can handle integrated, multi-service traffic mixes. This framework is
described in Chapters 10, 11, 12 and 13. A key part of the treatment of
cell- and burst-scale queueing is the use of explicit formulas, based on
heavy-traffic approximate analysis; these formulas can be rearranged very
simply to illustrate the design of algorithms for connection admission
control (Chapter 10), usage parameter control (Chapter 11), and buffer
dimensioning (Chapter 12). In addition, Chapter 12 combines the celland burst-scale analysis with the connection level for link dimensioning,
by incorporating Erlangs loss analysis introduced in Chapter 3. In
Chapter 13, we build on the analytical approach, introduced in Chapter 7,
to cover space and time priority issues.
Part III deals with IP and its performance and traffic management.
Chapter 14 applies the simple queueing analysis from Chapter 7 to the
buffering of variable-size packets. A new approximate technique, based
on the notion of excess-rate, is developed for application to queueing
systems with fixed, bi-modal, or general packet size distributions. The
technique gives accurate results across the full range of load values, and
has wide applicability in both IP and ATM. The concept of decay rate is
introduced; decay rate is a very flexible tool for summarising queueing
behaviour, and is used to advantage in the following chapters. Chapter 15
addresses the issue of resource reservation for aggregate flows in IP. A full
burst-scale analysis, applicable to both delay-sensitive, and loss-sensitive
traffic, is developed by judicious parameterization of a simple two-state
model for aggregate packet flows. The analysis is used to produce design
curves for configuring token buckets: for traffic conditioning of behaviour
aggregates in Differentiated Services, or for queue scheduling of traffic
flows in the Integrated Services Architectures.
Chapter 16 addresses the topic of buffer management from an IP
perspective, relying heavily on the use of decay rate analysis from
previous chapters. Decay rates are used to illustrate the configuration
of thresholds in the probabilistic packet discard mechanism known as
RED (random early detection). The partitioning of buffer space and
service capacity into virtual buffers is introduced, and simple techniques for configuring buffer partitions, based on decay rate analysis, are
developed.
Finally, in Chapter 17, we give a simple introduction to the important,
and mathematically challenging, subjects of self-similarity and longrange dependence. We illustrate these issues with the Pareto distribution
as a traffic model, and show its impact on queueing behaviour using the
simple analysis developed in Chapter 7.

xv

PREFACE

Acknowledgements
This new edition has benefited from the comments and questions raised
by readers of the first edition, posted, e-mailed and telephoned from
around the world. We would like to thank our colleagues in the
Department of Electronic Engineering for a friendly, encouraging and
stimulating academic environment in which to work. But most important
of all are our families thank you for your patience, understanding and
support through thick and thin!

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

PART I
Introductory Topics

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

An Introduction
to the Technologies of IP
and ATM
the bare necessities

This chapter is intended as a brief introduction to the technologies of

the Asynchronous Transfer Mode (ATM) and the Internet Protocol (IP)
on the assumption that you will need some background information
before proceeding to the chapters on traffic engineering and design. If
you already have a good working knowledge you may wish to skip this
chapter, because we highlight the fundamental operation as it relates to
performance issues rather than describe the technologies and standards
in detail. For anyone wanting a deeper insight we refer to [1.1] for
a comprehensive introduction to the narrowband Integrated Services
Digital Network (ISDN), to [1.2] for a general introduction to ATM
(including its implications for interworking and evolution) and to [1.3]
for next-generation IP.

CIRCUIT SWITCHING
In traditional analogue circuit switching, a call is set-up on the basis that
it receives a path (from source to destination) that is its property for
the duration of the call, i.e. the whole of the bandwidth of the circuit
is available to the calling parties for the whole of the call. In a digital
circuit-switched system, the whole bit-rate of the line is assigned to a
call for only a single time slot per frame. This is called time division
multiplexing.
During the time period of a frame, the transmitting party will generate
a fixed number of bits of digital data (for example, 8 bits to represent

AN INTRODUCTION TO THE TECHNOLOGIES OF IP AND ATM

the level of an analogue telephony signal) and these bits will be grouped
together in the time slot allocated to that call. On a transmission link, the
same time slot in every frame is assigned to a call for the duration of that
call (Figure 1.1). So the time slot is identified by its position in the frame,
hence use of the name position multiplexing, although this term is not
used as much as time division multiplexing.
When a connection is set up, a route is found through the network and
that route remains fixed for the duration of the connection. The route will
probably traverse a number of switching nodes and require the use of
many transmission links to provide a circuit from source to destination.
The time slot position used by a call is likely to be different on each link.
The switches which interconnect the transmission links perform the time
slot interchange (as well as the space switching) necessary to provide
the through-connection (e.g. link M, time slot 2 switches to link N, time
slot 7 in Figure 1.2).
Direction of transmission
Another 8 bits of data

8 bits of data gathered

during previous frame
.

Time

Duration of frame
One frame contains 8 time slots,
each time slot contains 8 bits
Figure 1.1.

Link M
. . .

An Example of Time Division, or Position, Multiplexing

Time

Time slot interchange

Link N
. . .

Figure 1.2.

Time Slot Interchange

PACKET SWITCHING

In digital circuit-switched telephony networks, frames have a repetition

rate of 8000 frames per second (and so a duration of 125 s), and as there
are always 8 bits (one byte) per time slot, each channel has a bit-rate
of 64 kbit/s. With N time slots in each frame, the bit-rate of the line is
N 64 kbit/s. In practice, extra time slots or bits are added for control and
synchronization functions. So, for example, the widely used 30-channel
system has two extra time slots, giving a total of 32 time slots, and
thus a bit-rate of
30 C 2 64 D 2048 kbit/s. Some readers may be more
familiar with the 1544 kbit/s 24-channel system which has 1 extra bit per
frame.
The time division multiplexing concept can be applied recursively
by considering a 24- or 30-channel system as a single channel, each
frame of which occupies one time slot per frame of a higher-order
multiplexing system. This is the underlying principle in the synchronous
digital hierarchy (SDH), and an introduction to SDH can be found in [1.1].
The main performance issue for the user of a circuit-switched network
is whether, when a call is requested, there is a circuit available to the
required destination. Once a circuit is established, the user has available
a constant bit-rate with a fixed end-to-end delay. There is no error
detection or correction provided by the network on the circuit thats the
responsibility of the terminals at either end, if it is required. Nor is there
any per circuit overhead the whole bit-rate of the circuit is available for
user information.

PACKET SWITCHING
Lets now consider a generic packet-switching network, i.e. one intended
to represent the main characteristics of packet switching, rather than any
particular packet-switching system (later on in the chapter well look
more closely at the specifics of IP).
Instead of being organized into single eight-bit time slots which repeat
at regular intervals, data in a packet-switched network is organised into
packets comprising many bytes of user data (bytes may also be known as
octets). Packets can vary in size depending on how much data there is to
send, usually up to some predetermined limit (for example, 4096 bytes).
Each packet is then sent from node to node as a group of contiguous bits
fully occupying the link bit-rate for the duration of the packet. If there
is no packet to send, then nothing is sent on the link. When a packet
is ready, and the link is idle, then the packet can be sent immediately.
If the link is busy (another packet is currently being transmitted), then
the packet must wait in a buffer until the previous one has completed
transmission (Figure 1.3).
Each packet has a label to identify it as belonging to a particular
communication. Thus packets from different sources and to different

AN INTRODUCTION TO THE TECHNOLOGIES OF IP AND ATM

Packet
waiting
in buffer

Direction of transmission

Transmitted packet
...

Label Information
Link overhead

Packet being
transmitted
Link idle
...
Link overhead added to
beginning and end of
packet that is being
transmitted
Time

Figure 1.3.

An Example of Label Multiplexing

destinations can be multiplexed over the same link by being transmitted

one after the other. This is called label multiplexing. The label is used
at each node to select an outgoing link, routeing the packet across the
network. The outgoing link selected may be predetermined at the set-up
of the connection, or it may be varied according to traffic conditions (e.g.
take the least busy route). The former method ensures that packets arrive
in the order in which they were sent, whereas the latter method requires
the destination to be able to resequence out-of-order packets (in the event
that the delays on alternative routes are different).
Whichever routeing method is used, the packets destined for a particular link must be queued in the node prior to transmission. It is this
queueing which introduces variable delay to the packets. A system
of acknowledgements ensures that errored packets are not lost but are
retransmitted. This is done on a link-by-link basis, rather than end-to-end,
and contributes further to the variation in delay if a packet is corrupted
and needs retransmission. There is quite a significant per-packet overhead required for the error control and acknowledgement mechanisms,
in addition to the label. This overhead reduces the effective bit-rate available for the transfer of user information. The packet-plus-link overhead
is often (confusingly) called a frame. Note that it is not the same as a
frame in circuit switching.
A simple packet-switched network may continue to accept packets
without assessing whether it can cope with the extra traffic or not. Thus
it appears to be non-blocking, in contrast to a circuit-switched network

CELL SWITCHING AND ATM

which rejects (blocks) a connection request if there is no circuit available. The effect of this non-blocking operation is that packets experience
greater and greater delays across the network, as the load on the network
increases. As the load approaches the network capacity, the node buffers
become full, and further incoming packets cannot be stored. This triggers retransmission of those packets which only worsens the situation
by increasing the load; the successful throughput of packets decreases
significantly.
In order to maintain throughput, congestion control techniques, particularly flow control, are used. Their aim is to limit the rate at which sources
offer packets to the network. The flow control can be exercised on a linkby-link, or end-to-end basis. Thus a connection cannot be guaranteed any
particular bit-rate: it is allowed to send packets to the network as and
when it needs to, but if the network is congested then the network exerts
control by restricting this rate of flow.
The main performance issues for a user of a packet-switched network
are the delay experienced on any connection and the throughput. The
network operator aims to maximize throughput and limit the delay, even
in the presence of congestion. The user is able to send information on
demand, and the network provides error control through re-transmission
of packets on a link-by-link basis. Capacity is not dedicated to the
connection, but shared on a dynamic basis with other connections. The
capacity available to the user is reduced by the per-packet overheads
required for label multiplexing, flow and error control.

CELL SWITCHING AND ATM

Cell switching combines aspects of both circuit and packet switching. In
very simple terms, the ATM concept maintains the time-slotted nature
of transmission in circuit switching (but without the position in a frame
having any meaning) but increases the size of the data unit from one octet
(byte) to 53 octets. Alternatively, you could say that ATM maintains the
concept of a packet but restricts it to a fixed size of 53 octets, and requires
packet-synchronized transmission.
This group of 53 octets is called a cell. It contains 48 octets for user
data the information field and 5 octets of overhead the header. The
header contains a label to identify it as belonging to a particular connection. So ATM uses label multiplexing and not position multiplexing. But
what about the time slots? Well, these are called cell slots. An ATM link
operates a sort of conveyor belt of cell slots (Figure 1.4). If there is a cell
to send, then it must wait for the start of the next cell slot boundary the
next slot on the conveyor belt. The cell is not allowed to straddle two
slots. If there is no cell to send, then the cell slot is unused, i.e. it is empty.

AN INTRODUCTION TO THE TECHNOLOGIES OF IP AND ATM

Direction of transmission

Time

A cell, containing a
header field and an
information field

A full cell slot

An empty cell slot

The conveyor belt of cell slots

Figure 1.4.

The Conveyor Belt of Cells

There is no need for the concept of a repeating frame, as in circuit

switching, because the label in the header identifies the cell.

CONNECTION-ORIENTATED SERVICE
Lets take a more detailed look at the cell header in ATM. The label
consists of two components: the virtual channel identifier (VCI) and the
virtual path identifier (VPI). These identifiers do not have end-to-end
(user-to-user) significance; they identify a particular virtual channel (VC)
or virtual path (VP) on the link over which the cell is being transmitted.
When the cell arrives at the next node, the VCI and the VPI are used to
look up in the routeing table to what outgoing port the cell should be
switched and what new VCI and VPI values the cell should have. The
routeing table values are established at the set-up of a connection, and
remain constant for the duration of the connection, so the cells always take
the same route through the network, and the cell sequence integrity of
the connection is maintained. Hence ATM provides connection-orientated
service.
But surely only one label is needed to achieve this cell routeing
mechanism, and that would also make the routeing tables simpler: so
why have two types of identifier? The reason is for the flexibility gained
in handling connections. The basic equivalent to a circuit-switched or
packet-switched connection in ATM is the virtual channel connection
(VCC). This is established over a series of concatenated virtual channel
links. A virtual path is a bundle of virtual channel links, i.e. it groups a
number of VC links in parallel. This idea enables direct logical routes to
be established between two switching nodes that are not connected by a
direct physical link.
The best way to appreciate why this concept is so flexible is to consider
an example. Figure 1.5 shows three switching nodes connected in a
physical star structure to a cross-connect node. Over this physical
network, a logical network of three virtual paths has been established.

CONNECTIONLESS SERVICE AND IP

SWITCH

CROSS-CONNECT
physical link

physical link

VCI: 42

SWITCH

VPI: 12

VPI: 25

VCI: 42

Cross-connect converts
VPI values (e.g. 12 25)
but does not change VCI
values (e.g. 42)
SWITCH
Figure 1.5. Virtual Paths and Virtual Channels

These VPs provide a logical mesh structure of virtual channel links

between the switching nodes. The routeing table in the cross-connect
only deals with port numbers and VPIs the VCI values are neither read
nor are they altered. However, the routeing table in the switching nodes
deal with all three: port numbers, VPIs and VCIs.
In setting up a VCC, the cross-connect is effectively invisible; it does
not need to know about VCIs and is therefore not involved in the process.
If there was only one type of identifier in the ATM cell header, then either
direct physical links would be needed between each pair of switching
nodes to create a mesh network, or another switching node would be
required at the hub of the star network. This hub switching node would
then have to be involved in every connection set-up on the network.
Thus the VP concept brings significant benefits by enabling flexible
logical network structures to be created to suit the needs of the expected
traffic demands. It is also much simpler to change the logical network
structure than the physical structure. This can be done to reflect, for
example, time-of-day changes in demand to different destinations.
In some respects the VP/VC concept is rather similar to having a twolevel time division multiplexing hierarchy in a circuit-switched network.
It has extra advantages in that it is not bound by any particular framing
structure, and so the capacity used by the VPs and VCs can be allocated
in a very flexible manner.

CONNECTIONLESS SERVICE AND IP

So, ATM is a connection-orientated system: no user information can be
sent across an ATM network unless a VCC or VPC has already been
established. This has the advantage that real-time services, such as voice,
do not have to suffer further delays associated with re-ordering the

10

AN INTRODUCTION TO THE TECHNOLOGIES OF IP AND ATM

data on reception (in addition to the queueing and transmission delays

experienced en route). However, for native connectionless-type services,
the overhead of connection establishment, using signalling protocols, is
burdensome.
Studies of Internet traffic have shown that the majority of flows are
very short, comprising only a few packets, although the majority of
packets belong to a minority of (longer-term) flows. Thus for the majority
of flows, the overhead of signalling can exceed the amount of user
information to be sent. IP deals with this in a very flexible way: it
provides a connectionless service between end users in which successive
data units can follow different paths. At a router, each packet is treated
independently concerning the routeing decision (based on the destination
address in the IP packet header) for the next hop towards the destination.
This is ideal for transferring data on flows with small numbers of packets,
and also works well for large numbers of packets. Thus packets being
sent between the same source and destination points may follow quite
different paths from source to destination.
Routes in IP are able to adapt quickly to congestion or equipment
failure. Although from the point of view of each packet the service
is in essence unreliable, for communication between end users IP is
very robust. It is the transport-layer protocols, such as the transmission
control protocol (TCP), that make up for the inherent unreliability of
packet transfer in IP. TCP re-orders mis-sequenced packets, and detects
and recovers from packet loss (or excessive delay) through a system
of timers, acknowledgements and sequence numbers. It also provides a
credit-based flow control mechanism which reacts to network congestion
by reducing the rate at which packets are sent.
This is fine for elastic traffic, i.e. traffic such as email or file transfer
that can adjust to large changes in delay and throughput (so long as
the data eventually gets there), but not for stream, i.e. inelastic, traffic.
This latter requires at least a minimum bit-rate across the network to
be of any value. Voice telephony, at the normal rate of 64 kbit/s is an
example wherein (unless some extra sophistication is present) this is the
rate that must be supported otherwise the signal will suffer so much loss
as to render it unintelligible and therefore meaningless. Requirements for
inelastic traffic are difficult to meet, and impossible to guarantee, in an
environment of highly variable delay, throughput and congestion. This
is why they have traditionally been carried by technologies which are
connection-orientated.
So, how can an IP network cope with both types of traffic, elastic and
inelastic? The first requirement is to partition the traffic into groups that
can be given different treatment appropriate to their performance needs.
The second requirement is to provide the means to state their needs, and
mechanisms to reserve resources specifically for those different groups of

BUFFER MANAGEMENT

11

traffic. The Integrated Services Architecture (ISA), Resource Reservation

Protocol (RSVP), Differentiated Services (DiffServ), and Multiprotocol
Label Switching (MPLS) are a variety of means aimed at achieving
just that.

BUFFERING IN ATM SWITCHES AND IP ROUTERS

Both IP and ATM networks move data about in discrete units. Network
nodes, whether handling ATM cells or IP packets, have to merge traffic
streams from different sources and forward them to different destinations
via transmission links which the traffic shares for part of the journey. This
process involves the temporary storage of data in finite-sized buffers, the
actual pattern of arrivals causing queues to grow and diminish in size.
Thus, in either technology, the data units contend for output transmission
capacity, and in so doing form queues in buffers. In practice these buffers
can be located in different places within the devices (e.g. at the inputs,
outputs or crosspoints) but this is not of prime importance. The point is
that queues form when the number of arrivals over a period exceeds the
number of departures, and it is therefore the actual pattern of arrivals
that is of most significance.
Buffering, then, is common to both technologies. However, simply
providing buffers is not a good enough solution; it is necessary to
provide the quality of service (QoS) that users request (and have to pay
for). To ensure guaranteed QoS these buffers must be used intelligently,
and this means providing buffer management.

BUFFER MANAGEMENT
Both ATM and IP feature buffer management mechanisms that are
designed to enhance the capability of the networks. In essence, these
mechanisms deal with how cells or packets gain access to the finite
waiting area of the buffer and, once in that waiting area, how they gain
access to the server for onward transmission. The former deals with how
the buffer space is partitioned, and the discard policies in operation. The
latter deals with how the packets or cells are ordered and scheduled for
service, and how the service capacity is partitioned.
The key requirement is to provide partitions, i.e. virtual buffers, through
which different groups of traffic can be forwarded. In the extreme, a
virtual buffer is provided for each IP flow, or ATM connection, and it has
its own buffer space and service capacity allocation. This is called per-flow
or per-VC queueing. Typically, considerations of scale mean that traffic,
whether flows or connections, must be handled in aggregate through
virtual buffers, particularly in the core of the network. Terminology varies

12

AN INTRODUCTION TO THE TECHNOLOGIES OF IP AND ATM

(e.g. transfer capability in ATM, behaviour aggregate in DiffServ, traffic

trunk in MPLS), but the grouping tends to be according to traffic type, i.e.
those with similar performance requirements and traffic characteristics.
Discard policies provide the means to differentiate between the relative
magnitudes of data loss, and the extent and impact of loss on flows or
connections within an aggregate of traffic.
In ATM, the cell loss priority (CLP) bit in the cell header is used
to distinguish between two levels of what is called space priority. A
(virtual) buffer which is deemed full for low-priority cells can still allow
high-priority cells to enter. The effect is to increase the likelihood of loss
for low-priority cells compared with that for high-priority cells. Hence
the name cell loss priority bit. What use is this facility? Well, if a source
is able to distinguish between information that is absolutely vital to its
correct operation (e.g. video synchronization) and information which is
not quite so important (e.g. part of the video picture) then the network
can take advantage of the differences in cell loss requirement by accepting
more traffic on the network. Otherwise the network loading is restricted
by the more stringent cell loss requirement.
IP does have something similar: the type of service (ToS) field in IPv4
or the priority field in IPv6. Both fields have codes that specify different
levels of loss treatment. In addition, in IP there is a discard mechanism
called random early detection (RED) that anticipates congestion by
discarding packets probabilistically before the buffer becomes full. Packets
are discarded with increasing probability when the average queue size is
above a configurable threshold.
The rationale behind the RED mechanism derives from the particular
challenge of forwarding best-effort packet traffic: TCP, in particular,
can introduce unwelcome behaviour when the network (or part of it) is
congested. If a buffer is full and has to discard arriving packets from many
TCP connections, they will all enter their slow start phase. This reduces
the load through the buffer, but because it affects so many connections
it leads to a period of under-utilization. When all those TCP connections
come out of slow start at about the same time, there is a substantial
increase in traffic, causing congestion again in the buffer and more packet
discard. The principle behind RED is that it applies the brakes gradually:
in the early stages of congestion, only a few TCP connections are affected,
and this may be sufficient to reduce the load and avoid any further
increase in congestion. If the average queue size continues to increase,
then packets are discarded with increasing probability, and so more TCP
connections are affected. Once the average queue size exceeds an upper
threshold all arriving packets are discarded.
In addition to controlling admission to the buffers, ATM and IP feature
the ability to control the process that outputs data from the buffers the
buffer scheduling mechanism. This provides a means to differentiate

TRAFFIC CONTROL

13

between delay, as well as bit-rate, requirements: just as some traffic is

more time-sensitive, so some needs greater transmission capacity. In
both ATM and IP, the switches and routers can implement time priority
ordering (precedence queueing) and mechanisms such as weighted fair
queueing or round robin scheduling, to partition the service capacity
among the virtual buffers.

TRAFFIC CONTROL
We have seen that both IP and ATM provide temporary storage for
packets and cells in buffers across the network, introducing variable
delays, and on occasion, loss too. These buffers incorporate various mechanisms to enable the networks to cater for different types of traffic both
elastic and inelastic. As we have noted, part of the solution to this problem
is the use of buffer management strategies: partitioning and reserving
appropriate resources both buffer space and service capacity. However,
there is another part to the overall solution: traffic control. This allows
users to state their communications needs, and enables the network to
coordinate and monitor its corresponding provision.
Upon receiving a reservation request (for a connection in ATM, or a
flow in IP), a network assesses whether or not it can handle the traffic, in
addition to what has already been accepted on the network. This process
is rather more complicated than for circuit switching, because some of
the reservation requests will be from variable bit-rate (VBR) services, for
which the instantaneous bit-rate required will be varying in a random
manner over time, as indeed will be the capacity available because many
of the existing connections will also be VBR! So if a request arrives for
a time-varying amount of capacity, and the capacity available is also
varying with time, it is no longer a trivial problem to determine whether
the connection or flow should be accepted.
In practice such a system works in the following way: the user declares
values for some parameters which describe the traffic behaviour of the
requested connection or flow, as well as the loss and delay performance
required; the network then uses these traffic and performance values
to come to an accept/reject decision, and informs the user. If accepted,
the network has to ensure that the sequence of cells or packets corresponds to the declared traffic values. This whole process is aimed at
preventing congestion in the network and ensuring that the performance
requirements are met for all carried traffic.
The traffic and performance values agreed by the user and the network
form a traffic contract. The mechanism which makes the accept/reject
decision is the admission control function, and this resides in the ATM
switching nodes or IP routers in the network. A mechanism is also
necessary to ensure compliance with the traffic contract, i.e. the user

14

AN INTRODUCTION TO THE TECHNOLOGIES OF IP AND ATM

should not exceed the peak (or mean, or whatever) rate that was agreed
for the connection, flow, or aggregate. This mechanism is called usage
parameter control (UPC) in ATM and is situated on entry to the network.
If the user does exceed the traffic contract, then the UPC mechanism
takes action to protect the network from the effects of this excess, e.g.
discarding some of the cells from the non-compliant connection. A similar
mechanism in DiffServ for IP networks is called traffic conditioning, and
this involves packet metering, marking, shaping and dropping.
In order to design algorithms for these mechanisms, it is important that
we understand the characteristics of the traffic sources, and the effects
these sources have when they are multiplexed through buffers in the
network, in terms of the delay and loss incurred. How we design the
algorithms is very closely related to how large we make the buffers, and
what buffer management mechanisms are proposed. Buffer dimensioning
and management mechanisms depend on how we intend to handle the
different services and their performance requirements.

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

Traffic Issues and Solutions

short circuits, short packets

This chapter is the executive summary for the book: it provides a quick
way to find a range of analytical solutions for a variety of design and
performance issues relating to IP and ATM traffic problems. If you are
already familiar with performance evaluation and want a quick overview
of what the book has to offer, then read on. Otherwise, youll probably
find that its best to skip this chapter, and come back to it after you have
read the rest of the book youll then be able to use this chapter as a
ready reference.

DELAY AND LOSS PERFORMANCE

In cell- or packet-based networks, the fundamental behaviour affecting
performance is the queueing experienced by cells/packets traversing the
buffers within those switches or routers on the path(s) from source to
destination through the network. This queueing behaviour means that
cells/packets experience variations in the delay through a buffer and
also, if that delay becomes too large, loss.
At its simplest, a buffer has a fixed service rate, a finite capacity
for the temporary storage of cells or packets awaiting service, and
a first-infirst-out (FIFO) service discipline. Even in this simple case,
the queueing behaviour depends on the type and mix of traffic being
multiplexed through the buffer. So lets first look at the range of source
models covered in the book, and then well summarize the queueing
analysis results.

16

TRAFFIC ISSUES AND SOLUTIONS

Source models
Model:
Use:
Formula:
Parameters:
Location:

Model:
Use:
Formulas:
Parameters:
Location:

Model:
Use:
Formulas:
Parameters:

Location:

Model:
Use:
Formula:
Parameters:

Location:

negative exponential distribution

inter-arrival times, service times, for cells, packets, bursts, flows, calls
Prfinter-arrival time
tg D F
t D 1  et
t time
 rate of arrivals, or rate of service
Chapter 6, page 83

geometric distribution
inter-arrival times, service times, for cells, packets, bursts, flows, calls
Prfk time slots between arrivalsg D
1  pk1 p
Prf
k time slots between arrivalsg D 1 
1  pk
k time slots
p probability of an arrival, or end of service, in a time slot
Chapter 6, page 85

Poisson distribution
number of arrivals or amount of work, for octets, cells, packets, bursts,
flows, calls

 Tk T
Prfk arrivals in time Tg D
e
k!
T time
k number of arrivals, or amount of work
 rate of arrivals
Chapter 6, page 86

binomial distribution
number of arrivals (in time, or from a number of inputs) or amount of
work, for octets, cells, packets, bursts, flows, calls
Prfk arrivals in N time slotsg D

N!

1  pNk pk

N  k! k!

k number of arrivals, or amount of work

p probability of an arrival, in a time slot or from an input
N number of time slots, or number of inputs
Chapter 6, page 86

DELAY AND LOSS PERFORMANCE

Model:
Use:
Formulas:

Parameters:

Location:
Model:
Use:
Formulas:

Parameters:

Location:
Model:
Use:
Formulas:

17

Batch distribution
number of arrivals, or amount of work, for octets, cells, packets,
bursts, flows, calls
a
0 D 1  p
a
1 D p b
1
a
2 D p b
2
..
.
a
k D p b
k
..
.
a
M D p b
M
k number of arrivals
p probability there is a batch of arrivals in a time slot
b
k probability there are k arrivals in a batch (given that there is a
batch in a time slot)
M maximum number of arrivals in batch
Chapter 6, page 88
ONOFF two-state
rate of arrivals, for octets, cells, packets
1
Ton D E[on]
R
1
Toff D E[off]
C
R rate of arrivals
E[on] mean number of arrivals in ON state
C service rate, or rate of time-base
E[off] mean number of time units in OFF state
Chapter 6, page 91
Pareto distribution
number of arrivals, or amount of work, for octets, cells, packets, etc.



PrfX > xg D
x



F
x D 1 
x

C1

f
x D

E[x] D
1
(continued overleaf )

18

TRAFFIC ISSUES AND SOLUTIONS

Model:

Pareto distribution

Parameters:

Location:

minimum amount of work

x number of arrivals, or amount of work
power law decay
Chapter 17, page 289

Queueing behaviour
There are a number of basic queueing relationships which are true,
regardless of the pattern of arrivals or of service, assuming that the buffer
capacity is infinite (or that the loss is very low). For the basic FIFO queue,
there is a wide range of queueing analyses that can be applied to both
IP and ATM, according to the multiplexing scenario. These queueing
relationships and analyses are summarized below.
Model:
Use:
Formulas:

Parameters:

Location:
Model:
Use:

Formulas:

elementary relationships
queues with infinite buffer capacity
 Ds
w D  tw (known as Littles formula)
q D  tq (ditto)
tq D tw C s
qDwC
 mean number of arrivals per unit time
s mean service time for each customer
 utilization; fraction of time the server is busy
w mean number of customers waiting to be served
tw mean time a customer spends waiting for service
q mean number of customers in the system (waiting or being
served)
tq mean time a customer spends in the system
Chapter 4, page 61
M/M/1
classical continuous-time queueing model; NB: assumes variablesize customers, so more appropriate for IP, but has been used for
ATM

qD
1
(continued)

19

DELAY AND LOSS PERFORMANCE

Model:

Parameters:

Location:

Model:
Use:
Formulas:

M/M/1
s
tw D
1
Prfsystem size D xg D
1  x
Prfsystem size > xg D xC1
 utilization; load (as fraction of service rate) offered to system
q mean number in the system (waiting or being served)
tw mean time spent waiting for service
x buffer capacity in packets or cells
Chapter 4, page 62

batch arrivals, deterministic service, infinite buffer capacity

exact M/D/1, binomial/D/1, and arbitrary batch distributions
these can be applied to ATM, and to IP (with fixed packet sizes)
E[a] D 
s
0 D 1  E[a]
s
k  1  s
0 a
k  1 

k1


s
i a
k  i

iD1

s
k D

a
0
PrfUd D 1g D Ud
1 D s
0 C s
1
PrfUd D kg D Ud
k D s
k
1

a
i

iD0

Bd
k D
Td
k D

k


E[a]
k


Ud
j Bd
k  j

jD1

Td,n
k D

k


Td,n1
j Td,1
k  j

jD1

s
2
1  
a
k probability there are k arrivals in a time slot

for M/D/1: tw D
Parameters:

 utilization; load (as fraction of service rate) offered to system

E[a] mean number of arrivals per time slot
(continued overleaf )

20

TRAFFIC ISSUES AND SOLUTIONS

Model:

Location:

Model:
Use:
Formulas:

batch arrivals, deterministic service, infinite buffer capacity

s
k probability there are k in the system at the end of any slot
Ud
k probability there are k units of unfinished work in the buffer
Bd
k probability there are k arrivals ahead in arriving batch
Td
k probability that an arrival experiences total delay of k
Td,n
k probability that total delay through n buffers is k
s mean service time for each customer
tw mean time spent waiting for service
Chapter 7, pages 100, 109, 110; and Chapter 4, page 66 (M/D/1
waiting time)

batch arrivals, deterministic service, finite buffer capacity

exact M/D/1, binomial/D/1, and arbitrary batch distributions
these can be applied to ATM, and to IP (with fixed packet sizes)
A
k D 1  a
0  a
1   a
k  1
u
0 D 1
u
k  1  a
k  1 

k1


u
i a
k  i

iD1

u
k D

a
0

u
X D A
X
s
0 D

1
X


u
i

iD0

s
k D s
0 u
k

CLP D
Parameters:

E[a] 
1  s
0
E[a]

a
k probability there are k arrivals in a time slot

A
k probability there are at least k arrivals in a time slot
E[a] mean number of arrivals per time slot
s
k probability there are k cells in the system at the end of any slot
 utilization; load (as fraction of service rate) offered to system
CLP probability of loss (whether cells or packets)

Location:

Chapter 7, page 105

21

DELAY AND LOSS PERFORMANCE

Model:
Use:

Formulas:
Parameters:

ND/D/1
multiple constant-bit-rate (CBR) sources into deterministic server
this can be applied to ATM, and to IP (with fixed packet sizes)




 



N

N!
nx n
n  x Nn D  N C x
Q
x D
1

n!
N  n!
D
D
DnCx
nDxC1
x buffer capacity (in cells or packets)
N number of CBR sources
D period of CBR source (in service time slots)
Q
x probability that queue exceeds x (estimate for loss probability)

Location:

Model:
Use:

Formulas:

Parameters:

Chapter 8, page 116

M/D/1 heavy-traffic approximation

cell-scale queueing in ATM, basic packet queueing in IP (with fixed
packet sizes); NB: below 80% load, underestimates loss


Q
x D e

2x

1



1

x D  ln
Q
x
2
1
2x
D
2 x  ln
Q
x

x buffer capacity (in cells or packets)

 utilization; load (as fraction of service rate) offered to system
Q
x probability that queue exceeds x (estimate for loss probability)

Location:

Model:
Use:

Formulas:
Parameters:

Chapter 8, page 117

ND/D/1 heavy-traffic approximation

multiple constant-bit-rate (CBR) sources into deterministic server this
can be applied to ATM, and to IP (with fixed packet sizes); NB: below
80% load, underestimates performance


Q
x D e

x 1
2x N C 

x buffer capacity (in cells or packets)

(continued overleaf )

22

TRAFFIC ISSUES AND SOLUTIONS

Model:

Location:

Model:
Use:
Formulas:

Parameters:

Location:

ND/D/1 heavy-traffic approximation

 utilization; load (as fraction of service rate) offered to system
N number of CBR sources
Q
x probability that queue exceeds x (estimate for loss probability)
Chapter 8, page 120

Geo/Geo/1
basic discrete-time queueing model for IP (variable-size packets)
p
s
0 D 1 
q






p
p
1q k

s
k D 1 

q
1q
1p


k
p
1q
Q
k D
q
1p



p
1  q x/q
Q
x D
q
1p
q probability a packet completes service at the end of an octet slot
p probability a packet arrives in an octet slot
s
k probability there are k octets in system
Q
k probability that queue exceeds k octets
Q
x probability that queue exceeds x packets
Chapter 14, page 232

Model: excess-rate, Geometrically Approximated Poisson Process (GAPP),

M/D/1
Use: accurate approximation to M/D/1 can be applied to ATM, and to
IP (with fixed packet sizes)

 
k
 e  e  2 C  C e
 e  e  2 C  C e
Formulas: p
k D 1 

  1 C e
  1 C e

kC1
 e  e  2 C  C e
Q
k D
  1 C e
Parameters:  arrival rate of Poisson process
p
k probability an arriving excess-rate cell/packet finds k in the
system
(continued)

23

DELAY AND LOSS PERFORMANCE

Model: excess-rate, Geometrically Approximated Poisson Process (GAPP),

M/D/1
Q
k probability an arriving excess-rate cell/packet finds more than k
in the system
Location: Chapter 14, page 245

Model:
Use:
Formulas:

excess-rate GAPP analysis for bi-modal service distributions

accurate approximation to M/bi-modal/1 suitable for IP, with
bi-modal distribution to model short and long packets
E[a] D 
ps C
1  ps  n
a
0 D ps e C
1  ps  en
a
1 D ps  e C
1  ps  n  en


E[a]
1  a
1  1 C a
1 C
a
02
p
k D 1 
a
0
E[a]  1 C a
0

k
E[a]
1  a
1  1 C a
1 C
a
02

a
0
E[a]  1 C a
0


kC1
E[a]
1  a
1  1 C a
1 C
a
02
Q
k D
a
0
E[a]  1 C a
0

Parameters:

a
k probability there are k arrivals in a packet service time

E[a] mean number of arrivals per packet service time
 packet arrival rate of Poisson process (i.e. per time unit D short
packet)
ps proportion of short packets
n length of long packets (multiple of short packet)
p
k probability an arriving excess-rate packet finds k in the system
Q
k probability an arriving excess-rate packet finds more than k
in the system

Location:

Model:
Use:

Chapter 14, page 249

excess-rate GAPP analysis for M/G/1

accurate approximation to M/G/1 suitable for IP, with general
service time distribution to model variable-length packets
(continued overleaf )

24

TRAFFIC ISSUES AND SOLUTIONS

Model:
Formulas:

excess-rate GAPP analysis for M/G/1

E[a] D 

1


g
i D 

iD1

a
0 D

1


g
i ei

iD1

a
1 D

1


g
i i  ei

iD1


E[a]
1  a
1  1 C a
1 C
a
02
p
k D 1 
a
0
E[a]  1 C a
0

k
E[a]
1  a
1  1 C a
1 C
a
02

a
0
E[a]  1 C a
0


kC1
E[a]
1  a
1  1 C a
1 C
a
02
Q
k D
a
0
E[a]  1 C a
0

Parameters:

A
k probability there are k arrivals in a packet service time

E[a] mean number of arrivals per packet service time
 packet arrival rate of Poisson process (i.e. per unit time)
g
k probability a packet requires k units of time to be served
p
k probability an arriving excess-rate packet finds k in the system
Q
k probability an arriving excess-rate packet finds more than k in
the system

Location:
Model:
Use:
Formulas:

Chapter 14, page 249

ONOFF/D/1/K
basic continuous-time queueing model for IP or ATM, suitable for
per-flow or per-VC scenarios
Ton
D
Ton C Toff


CLPexcess-rate D

Parameters:

C  R e

X
CR
Ton
1
RCC

1   C 
R  C e
RC
CLPexcess-rate
CLP D
R
R ON rate
C service rate of buffer

X
CR
Ton
1
RCC

(continued)

25

DELAY AND LOSS PERFORMANCE

Model:

Location:
Model:
Use:
Formulas:

Parameters:

Location:
Model:
Use:

Formulas:

ONOFF/D/1/K
X buffer capacity in cells/packets
Ton mean duration in ON state
Toff mean duration in OFF state
activity factor of source (probability of being ON)
CLP loss probability
Chapter 9, page 130
ONOFF/D/1/K
basic discrete-time queueing model for IP or ATM, suitable for per-flow
or per-VC scenarios
1
aD1
Ton
R  C
1
sD1
Toff C
1




p
X D

X
s
1a
1C
1
a
sa
RC
RC
CLP D
CLPexcess-rate D
p
X
R
R
R ON rate
C service rate of buffer
X buffer capacity in cells/packets
Ton mean duration in ON state
Toff mean duration in OFF state
p
k D probability an excess-rate arrival finds k in the buffer
CLP loss probability
Chapter 9, page 136
multiple ONOFF sources bufferless analysis
burst-scale loss model for IP or ATM for delay-sensitive traffic,
or, combined with burst-scale delay analysis, for delay-insensitive
traffic
m
Ton
D
D
h
Ton C Toff
C
N0 D
h
(continued overleaf )

26

TRAFFIC ISSUES AND SOLUTIONS

Model:

multiple ONOFF sources bufferless analysis

pn D

N!
n
1  Nn
n!
N  n!
N

pn
n  N0 

Prfcell needs bufferg D

Parameters:

nDdN0 e

m mean rate of single source

h ON rate of single source
Ton mean duration in ON state for single source
Toff mean duration in OFF state for single source
activity factor of single source (probability of being ON)
C service rate of buffer
N0 minimum number of active sources for burst-scale queueing
N total number of ONOFF sources being multiplexed
pn D probability that n sources are active
Prfcell needs bufferg estimate of loss probability

Location:
Model:
Use:

Formulas:

Parameters:

Chapter 9, page 141

multiple ONOFF sources approximate bufferless analysis
burst-scale loss model for IP or ATM for delay-sensitive traffic,
or, combined with burst-scale delay analysis, for delay-insensitive
traffic
Nm
D
C
C
N0 D
h

 N0 bN0 c N
1
0

e
Prfcell needs bufferg

1  2 N0
bN0 c!
m mean rate of single source
h ON rate of single source
C service rate of buffer
N total number of ONOFF sources being multiplexed
 offered load as fraction of service rate
N0 minimum number of active sources for burst-scale queueing
Prfcell needs bufferg estimate of loss probability

Location:

Chapter 9, page 142

27

DELAY AND LOSS PERFORMANCE

Model:
Use:
Formulas:

multiple ONOFF sources burst-scale delay analysis

burst-scale queueing model for IP or ATM combined with burstscale loss (bufferless) analysis, for delay-insensitive traffic
N
D
Ton C Toff
b D Ton h
b
D
C
C
N0 D
h


X
13
 N0 b 4C1

Parameters:

Location:
Model:
Use:

Formulas:

CLPexcess-rate D e
N total number of ONOFF sources being multiplexed
Ton mean duration in ON state for single source
Toff mean duration in OFF state for single source
h ON rate of single source
C service rate of buffer
 number of bursts arriving per unit time
b mean number of cells/packets per burst
 offered load as fraction of service rate
N0 minimum number of active sources for burst-scale queueing
CLPexcess-rate excess-rate loss probability, i.e. conditioned on the
probability that the cell/packet needs a buffer
Chapter 9, page 146
multiple ONOFF sources excess-rate analysis
combined burst-scale loss and delay analysis suitable for IP and
ATM scenarios with multiple flows (e.g. RSVP), or variable-bitrate (VBR) traffic (e.g. SBR/VBR transfer capability)
C
N0 D
h
Ap
AD
h


  N

0 1
N0

r
N0

A
A
N0
A
N0
DD

N0 !
N0  A
r!
N0 !
N0  A
rD0

T
on D

h Ton
C  Ap
(continued overleaf )

28

TRAFFIC ISSUES AND SOLUTIONS

Model:

multiple ONOFF sources excess-rate analysis

Ap
C  Ap
1D
T
off  D T
on
D
Ap  D Ron
Roff D
1D

Ron D C C h

xC1
1
1

hD
T
on
Ron  C

Q
x D

C  Ap
1
T
off 
C  Roff 

Parameters:

Location:
Model:
Use:

Formulas:

h ON rate of flow in packet/s or cell/s

C service rate of buffer
N0 minimum number of active sources for burst-scale queueing
Ap overall mean load in packet/s or cell/s
A offered traffic in packet flows (equivalent to erlang occupancy
of circuits, each circuit of rate h)
D probability of a packet flow waiting, i.e. of being in excess rate
state
Ton mean duration of flow
T
on mean duration in excess-rate ON state
Ron mean input rate to buffer when in excess-rate ON state
T
off  mean duration in underload OFF state
Roff mean input rate to buffer when in underload OFF state
Q
x queue overflow probability for buffer size of x packets (estimate for loss probability)
Chapter 15, page 261
Geo/Pareto/1
discrete-time queueing model for LRD (long-range dependence)
traffic in IP or ATM can be viewed as batch arrival process with
Pareto-distributed number of packets, or geometric arrivals with
Pareto-distributed service times



1
b
1 D F
1.5  F
1 D 1 
1.5




1
1
b
x D F
x C 0.5  F
x  0.5 D

x  0.5
x C 0.5

BD
1
(continued)

29

DELAY AND LOSS PERFORMANCE

Model:

Geo/Pareto/1

qD
B
a
0 D 1  q
a
1 D q b
1
a
2 D q b
2
..
.
a
k D q b
k
E[a] D 
s
0 D 1  E[a]
s
k  1  s
0 a
k  1 
s
k D

Parameters:

Location:

k1


s
i a
k  i

iD1

a
0
x number of arrivals, or amount of work
a power-law decay
b
x probability that Pareto batch is of size x packets
B mean batch size in packets
 mean number of packets arriving per time unit
q probability that a batch arrives in a time unit
a
k probability there are k arrivals in a time unit
E[a] mean number of arrivals per time unit
s
k probability there are k in the system at the end of any time
unit
Chapter 17, page 293

Model: Geo/truncated Pareto/1

Use: discrete-time queueing model for LRD traffic in IP or ATM NB:
truncated Pareto distribution limits range of time scales of bursty
behaviour, giving more realistic LRD traffic model




 



 

1
1

1
1
xD1

0.5
X C 0.5



 



 
1
1
1
Formulas: b
x D


1

Xx>1

x  0.5
x C 0.5
X C 0.5

0
x>X

BD
x b
x
x

(continued overleaf )

30

TRAFFIC ISSUES AND SOLUTIONS

Model: Geo/truncated Pareto/1


qD
B
a
0 D 1  q
a
1 D q b
1
a
2 D q b
2
..
.
a
k D q b
k
E[a] D 
s
0 D 1  E[a]
s
k  1  s
0 a
k  1 
s
k D

k1


s
i a
k  i

iD1

a
0
Parameters: x number of arrivals, or amount of work
power-law decay
b
x probability that Pareto batch is of size x packets
B mean batch size in packets
 mean number of packets arriving per time unit
q probability that a batch arrives in a time unit
a
k probability there are k arrivals in a time unit
E[a] mean number of arrivals per time unit
s
k probability there are k in the system at the end of any time unit
Location: Chapter 17, page 298

COPING WITH MULTI-SERVICE REQUIREMENTS:

DIFFERENTIATED PERFORMANCE
A FIFO discipline does not allow different performance requirements to
be guaranteed by the network in best-effort IP all traffic suffers similar
delay and loss, and in ATM the most stringent requirement limits the
admissible load. The solution is to manage the buffer, both on entry
and at the exit this involves policies for partitioning and sharing the
buffer space and server capacity (e.g. per-flow/per-VC queueing), packet
and cell discard mechanisms, and queue scheduling (such as precedence
queueing and weighted fair queueing).

Buffer sharing and partitioning

With per-flow/per-VC queueing and weighted fair queueing, each virtual
buffer can be modelled as having its own server capacity and buffer

COPING WITH MULTI-SERVICE REQUIREMENTS: DIFFERENTIATED PERFORMANCE

31

space thus any of the analysis methods for FIFO queues (see previous
section) can be applied, as appropriate to the multiplexing scenario, and
traffic source(s). Typically these will give a decay rate for each virtual
buffer, which can then be used, along with the performance requirement,
to assess the partitioning of buffer space.
There is clearly benefit in partitioning to maintain different performance guarantees for a variety of service types sharing an output port.
However, the cost of partitioning is that it is not optimal when considering the overall loss situation at an output port: the loss of a cell or
packet from a full virtual buffer may not be necessary if buffer space is
shared. Indeed, buffer space can be shared across multiple output ports.
The results for both partitioning and sharing are summarized below.
Model:
Use:
Formulas:

buffer-space partitioning analysis

to allocate space to virtual buffers according to performance requirements, and decay rate (obtained from queueing analysis)
X

Xj

S1 dr1 1 D S2 dr2 2 D D Sj drj D D SV drVV

XD

V


Xj

jD1


log
Sj 
XC
log
drj 
log
Si 
jD1
Xi D


V
log
dri 

1
log
dri 
log
drj 
jD1
V


Parameters:

Location:

Model:
Use:

Formulas:

X total buffer space available

Xi buffer space allocation for virtual buffer i
Si overflow probability scaling factor for virtual buffer i
dri decay rate obtained from queueing analysis of virtual buffer i
Chapter 16, page 278

shared buffer analysis

to assess the performance improvement when sharing buffer space
across multiple output buffers, using decay rate (obtained from
queueing analysis)
p
k D
1  dr 
dr k
(continued overleaf )

32

TRAFFIC ISSUES AND SOLUTIONS

Model:

shared buffer analysis

PN
k D

k


PN1
j P1
k  j

jD0

QN
k D 1 

k


PN
j

jD0

negative binomial approximation:

QN
k  1 kCN1 CN1
dr k
1  dr N1 or
Parameters:

Location:

QN
k  1 ef
kCN1ln
kCN1kln
k
N1ln
N1Ckln
dr C
N1ln
1dr g
dr decay rate in individual buffer
p
k queue state probability for single (virtual) buffer, i.e. probability that individual buffer has k cells/packets
PN
k autoconvolution for N buffers sharing space, i.e. probability
that shared space has k cells/packets
QN
k overflow probability from shared buffer
Chapter 16, page 280

Cell and packet discard mechanisms

In ATM, the space priority mechanism enables a FIFO buffer to provide
two different levels of loss performance, based on a threshold level for
the queue size. Combined with virtual buffers, it is possible to provide
this type of loss differentiation to traffic classes, or indeed within a VC,
to the different CLP 0/CLP 1 flows.
In IP, random early detection (RED) provides a probabilistic discard
mechanism with the aim of alleviating congestion, and avoiding it if
possible addressing the global synchronization problems associated
with multiple TCP connections. RED also uses a threshold mechanism,
but unlike in ATM, this works in conjunction with the average queue size.
Analysis for both of these discard mechanisms is summarized below.
Model:
Use:
Formulas:

M/D/1/K with partial buffer sharing (PBS)

basic discrete-time queueing model for analysing PBS in ATM
a D ah C al
ak a
e
a
k D
k!
ah k a
ah
k D
e h
k!
(continued)

COPING WITH MULTI-SERVICE REQUIREMENTS: DIFFERENTIATED PERFORMANCE

Model:

33

M/D/1/K with partial buffer sharing (PBS)

A
k D 1 

k1


a
j

jD0

Ah
k D 1 

k1


jD0
1


ah
j

n
imn

i  m!
ah
al

a
i

a
m, n D
n!

i

m

n!
a
a
iDmCn

mCn1
1


0
a
m C n C i
A
m, n D 1 
a
i 
0

n1


jD0

iD0

iD0

n C i!

j!
n C i  j!

ah
a

j
nCij 

al
a

u
0 D 1

1  a
0
u
1 D
a
0
below the threshold:
k1

A
k C
u
i A
k  i C 1
iD1

u
k D

a
0
A
M C

M1


u
i A0
M  i, 1

iD1

u
M D

ah
0

above the threshold:

A0
M, k  M C

M1


fu
i A0
M  i, k  M C 1g

iD1
k1

C

fu
i Ah
k  i C 1g

iDM

u
k D

A0
M, X  M
u
X D
ah
0
1
s
0 D X

u
i

ah
0

iD0

s
k D s
0 u
k

(continued overleaf )

34

TRAFFIC ISSUES AND SOLUTIONS

Model:

M/D/1/K with partial buffer sharing (PBS)

al C ah 
1  s
0
al C ah
M1
X


lh
j D
s
i a0
M  i, X  M C j C
s
i ah
X  i C j
CLP D

iD0


j

CLPh D
ll
j D

iDM

j lh
j

M1


ah

iD0

j

CLPl D
Parameters:

Location:

1


s
i

al
a


rDMiCj
X


a
r

r 
M  i!

r 
M  i  j! j!

ah
a

r
Mij

s
i al
j

iDM

j ll
j

al
a mean arrival rate of both high- and low-priority arrivals
al mean arrival rate of low-priority arrivals
ah mean arrival rate of high-priority arrivals
a
k probability there are k arrivals in one time slot
al
k probability there are k low-priority arrivals in one time slot
ah
k probability there are k high-priority arrivals in one time slot
A
k probability at least k cells arrive in one time slot
Ah
k probability at least k high-priority cells arrive in one time
slot
a0
m, n probability that m cells of either low or high priority are
admitted, up to the threshold, and a further n high-priority cells
are admitted above the threshold
A0
m, n probability that m cells of either low or high priority are
admitted, up to the threshold, and at least a further n high-priority
cells are admitted above the threshold
s
k probability there are k cells in the system
lh
j probability that j high-priority cells are lost in a time slot
ll
j probability that j low-priority cells are lost in a time slot
M PBS threshold (cells)
X buffer capacity in cells
CLP overall cell loss probability
CLPh cell loss probability for high-priority cells
CLPl cell loss probability for low-priority cells
Chapter 13, page 207

COPING WITH MULTI-SERVICE REQUIREMENTS: DIFFERENTIATED PERFORMANCE

35

Model:

M/D/1/K with PBS margin between high- and low-priority CLP

Use:

rule of thumb for configuring PBS threshold, given high-priority

load
CLPmargin D 10
XM for ah D 0.25

Formulas:

CLPmargin D 102
XM for ah D 0.04

Parameters:

ah mean arrival rate of high-priority arrivals

M PBS threshold (cells)
X buffer capacity in cells
CLPmargin difference between high- and low-priority cell loss
probability

Location:

Chapter 16, page 218

Model:

threshold indicators for random early detection (RED) mechanism

Use:

mean queue size based on decay rate analysis, to aid in configuring

EWMA thresholds: min and max
p
k D
1  dr 
dr k

Formulas:

dr
1  dr
dr decay rate based on appropriate queueing analysis (see earlier,
e.g. use both basic packet-scale and burst-scale decay rates)

qD
Parameters:

p
k probability that queue contains k packets

q mean queue size
Location:

Chapter 16, page 270

Queue scheduling mechanisms

Whether in IP or in ATM, queue scheduling mechanisms involve partitioning the service capacity of the output buffer. Weighted fair queueing
allocates a specified proportion of the server to each virtual buffer to
analyse this, each virtual buffer can be treated as an independent FIFO
queue with a fixed service capacity equal to its allocation. Thus any of the
analysis methods for FIFO queues (see previous section) can be applied.
Precedence queueing, i.e. time priority, requires specific analysis, because
the partitions between different priority levels are not fixed. The amount
of service capacity seen by lower-priority traffic depends on how much
has been used by higher-priority traffic. The queueing analysis for time
priorities is summarized below.

36

TRAFFIC ISSUES AND SOLUTIONS

Model:
Use:
Formulas:

Parameters:
Location:
Model:
Use:
Formulas:

M/D/1 with time priorities mean value analysis

basic discrete-time queueing model for analysing time priorities in
ATM also applies to IP with fixed packet sizes
a1 C a2
w1 D
2
1  a1 
a1 C a2
w1 a1 C
2
w2 D
1  a1  a2
ai mean arrival rate of priority-i arrivals (highest priority is 1)
wi mean waiting time of priority-i arrivals
Chapter 13, page 219
M/D/1 with time priorities waiting-time distribution analysis
basic discrete-time queueing model for analysing time priorities in
ATM also applies to IP with fixed packet sizes
E[a] D 
s
0 D 1  E[a]
k1

s
i a
k  i
s
k  1  s
0 a
k  1 
iD1

s
k D

a
0

a1 k a
e 1
k!
a2 k a
a2
k D
e 2
k!
k

1
a2
i

a1
k D

iD0

b
k D

a2
u
0 D s
0 C s
1
u
k D s
k C 1 for k > 0

k
i


u
k  i
b
j a1
i  j
v
k D
iD0

jD0

a1 xk ka
a1
k, x D
e 1
x!
w
0 D v
0
k

v
i a1
k  i, k i
w
k D

iD1

for k > 0
(continued)

37

FLOWS, CONNECTIONS AND AGGREGATES

Model:

M/D/1 with time priorities waiting-time distribution analysis

Parameters:

E[a] mean arrival rate of both high- and low-priority arrivals

ai mean arrival rate of priority-i arrivals (highest priority is 1)
a
k probability there are k arrivals in one time slot
ai
k probability there are k arrivals of priority i in one time slot
s
k probability of k cells/packets in queue
u
k probability of k units of unfinished work in queue
b
k probability there are k arrivals ahead in arriving batch
v
k virtual waiting time distribution
a1
k, x probability that k high-priority cells arrive in x time slots
w
k D Prfa priority-2 cell must wait k time slots before it enters
serviceg

Location:

Chapter 13, page 222

FLOWS, CONNECTIONS AND AGGREGATES

To provide end-to-end performance guarantees, traffic contracts must be
established, resources reserved, and traffic flows monitored to ensure
compliance with the contract. At the heart of these functions is the
assessment of performance resulting from traffic flowing over network
resources, i.e. various forms of queueing analysis.

Admission control mechanisms

Whether in ATM or in (QoS-aware) IP, an admission control function
has to assess whether or not a new connection or flow can be admitted.
This function must take into account the commitments that the network
is currently supporting and the resources available for any new flows.
To do this, it needs to be aware of the structure and configuration of the
buffers (i.e. the scheduling, discard and partitioning mechanisms), and
make an appropriate assessment of whether the requested performance
can be guaranteed. This requires queueing analysis, as summarized in
previous sections in this chapter, in a form suitable for fast, accurate
responses. Examples of these forms are summarized below.
Model:
Use:

admissible load based on ND/D/1 analysis

admission control in ATM (for DBR, i.e. cell-scale component) or IP
(e.g. voice-over-IP with fixed packet sizes) for constant-bit-rate traffic
(continued overleaf )

38

TRAFFIC ISSUES AND SOLUTIONS

Model:
Formulas:

admissible load based on ND/D/1 analysis

If
2 x2

nC1

ln


min
CLPi 

iD1!nC1

then accept if
n
hi
hnC1 
C

1
C
C
iD1
If

nC1>
ln

then accept if

2 x2

min
CLPi 

iD1!nC1

n
hnC1 
hi
C
C
C
iD1

2 x
n C 1

2 x
n C 1  2 x2 C
n C 1 ln

Parameters:

Location:
Model:
Use:
Formulas:


min
CLPi 

iD1!nC1

The above is based on heavy-traffic approximation; for alternative

based on more accurate ND/D/1 analysis, see Tables 10.2 and 10.3
hi fixed arrival rate of ith flow or connection
CLPi loss requirement of ith flow or connection
n number of existing flows or connections
C service rate of (virtual) buffer
x buffer capacity in packets or cells
Chapter 10, page 153
admissible load based on M/D/1 analysis
admission control in ATM (for DBR, and cell-scale component in SBR)
or IP (e.g. voice-over-IP with fixed packet sizes)
n
hnC1  hi
2x



C


C
C
iD1
2 x  ln
min
CLPi 
iD1!nC1

Parameters:

right-hand side of inequality test is based on heavy traffic approximation; for alternative based on exact M/D/1 analysis, see Table 10.1
hi fixed arrival rate of ith flow or connection
CLPi loss requirement of ith flow or connection
n number of existing flows or connections
(continued)

39

FLOWS, CONNECTIONS AND AGGREGATES

Model:

admissible load based on M/D/1 analysis

C service rate of (virtual) buffer
x buffer capacity in packets or cells

Location:
Model:
Use:

Formulas:

Chapter 10, page 152

admissible load based on burst-scale loss analysis
admission control in ATM (burst-scale component in SBR) or IP (e.g.
voice-over-IP with silence suppression) for delay-sensitive VBR
traffic
C
N0 D
h
n
mnC1 
mi
C


CLP, N0 
C
C
iD1
where  (CLP, N0 ) is based on e.g.
1

 N0 bN0 c N
0

e

1  2 N0
bN0 c!
and can be found in Table 10.4

CLP

Parameters:

mi mean arrival rate of ith flow or connection

CLP most stringent loss requirement of all flows or connections
n number of existing flows or connections
C service rate of (virtual) buffer
h maximum individual peak arrival rate of all flows or connections
N0 minimum number of active sources at maximum peak rate for
burst-scale queueing
 admissible load

Location:
Model:
Use:
Formulas:

Chapter 10, page 159

admissible load based on burst-scale delay analysis
admission control in ATM (burst-scale delay component in SBR) or IP,
for delay-insensitive VBR traffic
Table 10.5 is based on burst-scale delay analysis:

3
CLPexcess-rate D e

X
1
 N0 b 4C1

and should be used in conjunction with burst-scale loss, Table 10.4

(continued overleaf )

40

TRAFFIC ISSUES AND SOLUTIONS

Model:

admissible load based on burst-scale delay analysis

Parameters:

b mean burst duration rate (maximum over all flows or connections)

X buffer capacity in cells or packets
N0 minimum number of active sources at maximum peak rate for
burst-scale queueing
CLPexcess-rate loss component attributed to burst-scale delay analysis
 admissible load

Location:

Chapter 10, page 162

Policing mechanisms
Once admitted, the flow of packets or cells is monitored (to ensure
compliance with the traffic contract) by a policing function, usually a
token, or leaky, bucket. In IP, the token bucket is typically integrated into
the queue scheduling mechanism, whereas in ATM, the leaky bucket is
normally a separate function on entry to the network. Both can be assessed
using a variety of forms of queueing analysis (see earlier), in order to
know how to configure the buckets appropriately. Some examples of
analysis are summarized below.
Model:
Use:
Formulas:

Parameters:

worst-case analysis of leaky bucket for CDV tolerances

allowance for impact on load of CDV tolerance in contract for CBR
traffic using DBR transfer capability in ATM
"
#
"
#

/
D1C
MBS D 1 C
T
D1



X
2
CD
MBS
D



MBS ln
CLP
D 2
X
T inter-arrival time at peak cell rate
cell slot time
D inter-arrival time at peak cell rate, in units of the cell slot time
 cell delay variation tolerance
MBS maximum burst size allowed through by leaky bucket
X buffer capacity of output port downstream in network
 admissible load assuming worst-case arrival pattern

Location:

Chapter 11, page 179

FLOWS, CONNECTIONS AND AGGREGATES

Model:

worst-case analysis of dual leaky bucket (leaky cup and saucer)

Use:
Formulas:

Parameters:

Location:
Model:

allowance for impact on load of dual leaky bucket for VBR traffic using
SBR transfer capability in ATM
"
#
IBT
MBS D 1 C
T
 TPCR

SCR

X
2
CD
MBS
D



MBS ln
CLP
D 2
X
TPCR inter-arrival time at peak cell rate
TSCR inter-arrival time at sustainable cell rate (SCR)
IBT intrinsic burst tolerance
D inter-arrival time at sustainable cell rate, in units of the cell slot
time
MBS maximum burst size allowed through by leaky bucket for SCR
X buffer capacity of output port downstream in network
 admissible load assuming worst-case arrival pattern of maximumsized bursts arriving at cell slot rate
Chapter 11, page 183
multiple ONOFF sources excess-rate analysis

Use:
Formulas:

Parameters:

Location:

41

token bucket configuration for traffic conditioning of aggregate IP

flows
see multiple ONOFF sources excess-rate analysis summarized in
Chapter 2, page 27
example relationships between R and B, given different loss probability
estimates, are shown in Figure 15.7
R token bucket rate allocation is the same as C, service rate of buffer
B token bucket capacity is the same as x, the buffer size
both R and B would normally be scaled from packets to octets
Chapter 15, page 266

Dimensioning and configuration

On the longer time scale, the provisioning, dimensioning and configuration of network resources (transmission links, buffer partitions, switching
capacities, etc.) are needed to match the expected user demand. This too
makes use of queueing analysis to determine probability of blocking or

42

TRAFFIC ISSUES AND SOLUTIONS

congestion at the connection, or aggregate level. Results for buffer and

link dimensioning are summarized below.
Model:
Use:
Formulas:

Parameters:

Location:
Model:
Use:

Formulas:
Parameters:

M/D/1 delay and loss analysis

dimensioning small (virtual) buffers for delay-sensitive traffic in IP
or ATM
s
tq D s C
2
1  
see batch arrivals, deterministic service, infinite buffer capacity,
summarized in Chapter 2, page 19
Table 12.2 shows exact buffer dimensions, mean and maximum
delay, given offered load and loss probability requirement
s service time for a cell or packet
 maximum offered load expected
tq mean delay
Chapter 12, page 193
burst-scale delay and loss analysis
dimensioning large (virtual) buffers for delay-insensitive traffic in IP
or ATM



CLPtarget
ln
X
4C1
CLPbsl

D
3
b

1  
N0
X buffer capacity in cells or packets
b mean number of cells/packets per burst
 maximum offered load expected
N0 minimum number of active sources for burst-scale queueing
(calculated from service rate of buffer, C, and maximum arrival rate,
h, for an individual flow or connection)
CLPtarget overall loss probability as design target
CLPbsl loss probability contribution from burst-scale loss analysis

Location:
Model:
Use:

Chapter 12, page 198

Erlangs loss analysis
dimensioning maximum number of simultaneous flows or connections
in IP or ATM, for blocking probability requirement
(continued)

FLOWS, CONNECTIONS AND AGGREGATES

Model:
Formulas:

43

Erlangs loss analysis

ADah



AN
A1
A2
AN
BD
1C
C
C C
N!
1!
2!
N!
see Table 12.4 for full erlang traffic table
link dimension is then
C D ECR N

Parameters:

h mean duration of flow or connection

a expected arrival rate of flows or connections
A offered traffic in erlangs
B grade of service (blocking probability) required
N number of equivalent circuits
ECR equivalent cell rate, i.e. the cell or packet rate required per
equivalent circuit (from queueing analysis suitable for traffic type)
C total service rate required for (virtual) buffer

Location:

Chapter 3, page 52

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

Teletraffic Engineering
the economic and service arguments

SHARING RESOURCES
A simple answer to the question Why have a network? is To communicate information between people. A slightly more detailed answer would
be: To communicate information between all people who would want to
exchange information, when they want to. Teletraffic engineering addresses
the problems caused by sharing of network resources among the population of users; it is used to answer questions like: How much traffic
needs to be handled? What level of performance should be maintained?
What type of, and how many, resources are required? How should the
resources be organized to handle traffic?

MESH AND STAR NETWORKS

Consider a very simple example: a telephone network in which a separate
path (with a handset on each end) is provided between every pair of
users. For N users, this means having N
N
1/2 paths and N
N
1
telephone handsets. A simple cost-saving measure would be to replace
the N
1 handsets per user with just one handset and a 1 to N
1 switch
(Figure 3.1). A total of N handsets and N switches is required, along with
the N
N
1/2 paths. If all N users are communicating over the network
at the same time, i.e. there are N/2 simultaneous calls (or
N
1/2
if N is odd), then 1/
N
1 of the paths and all of the handsets and
switches would be in use. So in a network with 120 users, for example,
the maximum path utilization is just under 1%, and handset and switch
utilization are both 100%.

46

TELETRAFFIC ENGINEERING

Contrast this with a star network, where each user has a single handset
connected to two N to 1 switches, and the poles of the switches are
connected by a single path (Figure 3.2). In this example, there are N
handsets, N C 1 paths, and 2 switches. However, only 2 users may
communicate at any one time, i.e. 3/
N C 1 of the paths, 2/N of the
handsets and both of the switches would be in use. So for a network with
120 users, the maximum values are: path utilization is just under 3%,
handset utilization is just under 2% and switch utilization is 100%.
In the course of one day, suppose that each one of the 120 users
initiates on average two 3-minute calls. Thus the total traffic volume is
120 2 3 D 720 call minutes, i.e. 12 hours of calls. Both star and mesh
networks can handle this amount of traffic; the mesh network can carry
up to 60 calls simultaneously; the star network carries only 1 call at a
time. The mesh network provides the maximum capability for immediate
communication, but at the expense of many paths and switches. The star
network provides the minimum capability for communication between
any two users at minimum cost, but at the inconvenience of having to
wait to use the network.
The capacity of the star network could be increased by installing M
switching units, where each unit comprises two N to 1 switches linked
by a single path (Figure 3.3). Thus, with N/2 switching units, the star
network would have the same communication capability as the mesh
network, with the same number of switches and handsets, but requiring
only 3N/2 paths. Even in this case, though, the size becomes impractical as

N
N.(N-1)/2
N
Figure 3.1.

The Mesh Network

N
N+1
2
Figure 3.2.

The Star Network

47

TRAFFIC INTENSITY

N
N +M
2M
Figure 3.3.

The Star Network with M Switching Units

N increases, such that reorganization and further sharing of the switching

capacity becomes necessary.

TRAFFIC INTENSITY
Traffic volume is defined as the total call holding time for all calls, i.e. the
number of calls multiplied by the mean holding time per call. This is not
very helpful in determining the total number of paths or switching units
required. We need a measure that gives some indication of the average
workload we are applying to the network.
Traffic intensity is defined in two ways, depending on whether we
are concerned with the workload applied to the network (offered traffic),
or the work done by the network (carried traffic). The offered traffic
intensity is defined as:
ch
AD
T
where c is number of call attempts in time period T, and h is the mean
call holding time (the average call duration). Note that if we let T equal h
then the offered traffic intensity is just the number of call attempts during
the mean call holding time. The rate of call attempts, also called the call
arrival rate, is given by
c
aD
T
So the offered traffic intensity can also be expressed as
ADah
For any specific pattern of call attempts, there may be insufficient paths
to satisfy all of the call attempts; this is particularly obvious in the case
of the star network in Figure 3.2 which has just one path available. A call
attempt made when the network is full is blocked (lost) and cannot be
carried. If, during time period T, cc calls are carried and cl calls are lost,
then the total number of call attempts is
c D cc C cl

48

TELETRAFFIC ENGINEERING

We then have
AD

cc C cl  h
DCCL
T

where C, the carried traffic, is given by

CD

cc h
T

and L, the lost traffic, is given by

LD

cl h
T

The blocked calls contributing to the lost traffic intensity obviously do

not last for any length of time. The lost traffic intensity, as defined, is thus
a theoretical intensity which would exist if there were infinite resources
available. Hence the lost traffic cannot be measured, although the number
of lost calls can. The carried traffic intensity can be measured, and is the
average number of paths in use simultaneously (this is intuitive, as we
have already stated that it should be a measure of the work being done by
the network). As theoretical concepts, however, we shall see that offered,
lost and carried traffic prove to be very useful indeed.
Traffic intensity is a dimensionless quantity. It is given the honorary
dimension of erlangs in memory of Anders K. Erlang, the founder of
traffic theory: one erlang of traffic is written as 1 E. Lets put some
numbers in the formulas. In our previous example we had 240 calls over
the period of a day, and an average call duration of 3 minutes. Suppose
24 calls are unsuccessful, then c D 240, cc D 216, and cl D 24. Thus
240 3
D 0.5 E
24 60
24 3
D 0.05 E
LD
24 60
216 3
D 0.45 E
CD
24 60

AD

Later in this chapter we will introduce a formula which relates A and L

according to the number of available paths, N.
It is important to keep in mind that one erlang (1 E) implicitly represents
a quantity of bandwidth, e.g. a 64 kbit/s circuit, being used continuously.
For circuit-switched telephone networks, it is unnecessary to make this
explicit: one telephone call occupies one circuit for the duration of one
call. However, if we need to handle traffic with many different bandwidth
demands, traffic intensity is rather more difficult to define.

49

TCP: TRAFFIC, CAPACITY AND PERFORMANCE

One way of taking the service bandwidth into account is to use

the MbitE/s (the megabit-erlang-per-second) as a measure of traffic
intensity. Thus 1 E of 64 kbit/s digital telephony is represented as
0.064 MbitE/s (in each direction of communication). We shall see later,
though, that finding a single value for the service bandwidth of variablerate traffic is not an easy matter. Suffice to say that we need to know the
call arrival rate and the average call duration to give the traffic flow in
erlangs, and also the fact that some bandwidth is implicitly associated
with the traffic flow for each different type of traffic.

PERFORMANCE
The two different network structures, mesh and star, illustrate how the
same volume of traffic can be handled very differently. With the star
network, users may have to wait significantly longer for service (which,
in a circuit-switched network, can mean repeated attempts by a user to
establish a call). A comparison of the waiting time and the delay that users
will tolerate (before they give up and become customers of a competing
network operator) enables us to assess the adequacy of the network. The
waiting time is a measure of performance, as is the loss of a customer.
This also shows a general principle about the flow of traffic:introducing
delay reduces the flow, and a reduced traffic flow requires fewer
resources. The challenge is to find an optimum value of the delay
introduced in order to balance the traffic demand, the performance
requirements, and the amount (and cost) of network resources. We
will see that much teletraffic engineering is concerned with assessing
the traffic flow of cells or packets being carried through the delaying
mechanism of the buffer.

TCP: TRAFFIC, CAPACITY AND PERFORMANCE

So we have identified three elements: the capacity of a network and its
constituent parts; the amount of traffic to be carried on the network; and
the requirements associated with that traffic, in terms of the performance
offered by the network to users (see Figure 3.4). One of these elements
Performance

Capacity

Traffic
Figure 3.4. Traffic, Capacity and Performance

50

TELETRAFFIC ENGINEERING

may be fixed in order to determine how the others vary with each
other, or two elements may be fixed in order to find a value for the
third. For example, the emphasis in dimensioning is on determining the
capacity required, given specific traffic demand and performance targets.
Performance engineering aims at assessing the feasibility of a particular
network design (or, more commonly, an aspect or part of a network)
under different traffic conditions; hence the emphasis is on varying the
traffic and measuring the performance for a given capacity (network
design). Admission control procedures for calls in an ATM network
have the capacity and performance requirements fixed, with the aim of
assessing how much, and what mix of, traffic can be accepted by the
network.
In summary, a network provides the ability to communicate information between users, with the aim of providing an effective service at
reasonable cost. It is uneconomic to provide separate paths between every
pair of users. There is thus a need to share paths, and provide users with
the means to access these paths when required. A network comprises
building blocks (switches, terminal equipment, transmission paths), each
of which has a finite capacity for transferring information. Whether or not
this capacity is adequate depends on the demand from users for transferring information, and the requirements that users place on that transfer.
Teletraffic engineering is concerned with the relationships between these
three elements of traffic, capacity and performance.

VARIATION OF TRAFFIC INTENSITY

It is important not to fall into the trap of thinking that a traffic intensity
of x erlangs can always be carried on x circuits. The occurrence of any
particular pattern of calls is a matter of chance, and the traffic intensity
measures the average, not the variation in, traffic during a particular
period. The general principle is that more circuits will be needed on a
route than the numerical value of the traffic intensity.
Figure 3.5 shows a typical distribution of the number of call attempts
per unit time (including the Mathcad code to generate the graph). If we
let this unit time be equal to the average call duration, then the average
number of call attempts per unit time is numerically equal to the offered
traffic intensity. In the case shown it is 2.5 E.
At this stage, dont worry about the specific formula for the Poissondistribution. The key point is that this distribution is an example which
describes the time-varying nature of traffic for a constant average intensity. We could define this average, as before, over the period of one day.
But is this sensible? What if 240 calls occur during a day, but 200 of
the 240 calls occur between 10 a.m. and 11 a.m.? Then the offered traffic

51

ERLANGS LOST CALL FORMULA

Number of call attempts per unit time

0

10

15

101

Probability

102
103
104
105
106
107

Poisson (k, ) :D

k

e
k!

i :D 0.. 15
xi :D i
yi :D Poisson (i, 2.5)

Figure 3.5. Graph of the Distribution of Demand for an Offered Traffic Intensity of
2.5 E, and the Mathcad Code to Generate (x, y) Values for Plotting the Graph

intensity for this hour is

AD

200 3
D 10 E
60

This is significantly larger than the daily average, which we calculated

earlier to be 0.5 E. The larger figure for offered traffic gives a better
indication of the number of circuits needed. This is because traffic
intensity, in practice, varies from a low level during the night to one
or more peaks during the day, and a network operator must provide
enough circuits to ensure that the performance requirements are met
when the traffic is at its peak during the busiest period of the day. The
busy hour is defined as a period when the intensity is at a maximum
over an uninterrupted period of 60 minutes. Note that the busy-hour
traffic is still an average: it is an average over the time scale of the
busy hour (recall that this is then the maximum over the time scale of
a day).

52

TELETRAFFIC ENGINEERING

ERLANGS LOST CALL FORMULA

In 1917, Erlang published a teletraffic dimensioning method for circuitswitched networks. He developed a formula which expressed the probability, B, of a call being blocked, as a function of the applied (offered)
traffic intensity, A, and the number of circuits available, N:
AN
BD
N!



A2
AN
A1
C
C C
1C
1!
2!
N!

B is also the proportion of offered traffic that is lost. Hence

BD

L
A

Number of circuits, N
10

15

Probability of call blocking, B

100
101
102
103
104
105
106
107

AN
B
A, N :D N!
N r

A
rD0 r!

i :D 0.. 15
xi :D i
yi :D B
2.5 , i

Figure 3.6. Graph of the Probability of Call Blocking for A D 2.5 E, and the Mathcad
Code to Generate
x, y Values for Plotting the Graph

53

TRAFFIC TABLES

where L is the lost traffic, as before. A derivation of Erlangs formula can

be found in [3.1, 3.2].
The most important assumption made concerns the pattern of
arrivals calls occur individually and collectively at random. This
means they are as likely to occur at one time as at any other time. This type
of arrival process is called Poisson traffic. The Poisson distribution gives
the probability that a certain number of calls arrive during a particular
time interval. We will look at this distribution in more detail in Chapter 6.
First, let us plot B against N when A D 2.5 E. This is shown in Figure 3.6,
with the Mathcad code used to generate the results.
We can read from the graph that the blocking probability is B D 0.01
when the number of circuits is N D 7. Thus we can use this graph for
dimensioning: choose the required probability of blocking and find the
number of circuits corresponding to this on the graph. But we dont
want to have to produce graphs for every possible value of offered
traffic.

TRAFFIC TABLES
The problem is that Erlangs lost call formula gives the call blocking (i.e.
loss) probability, B, given a certain number, N, of trunks being offered
a certain amount, A, of traffic. But the dimensioning question comes
the other way around: with a certain amount, A, of traffic offered, how
many trunks, N, are required to give a blocking probability of B? It is
not possible to express N in terms of B, so traffic tables, like the one in
Table 3.1, have been produced (using iteration), and are widely used, to
simplify this calculation.
Table 3.1. Table of Traffic which May Be Offered,
Based on Erlangs Lost Call Formula

Probability of blocking, B

Number of
trunks, N

0.02

1
2
3
4
5
6
7
8
9
10

0.02
0.22
0.60
1.1
1.7
2.3
2.9
3.6
4.3
5.1

0.01

0.005

0.001

0.005
0.105
0.35
0.7
1.1
1.6
2.2
2.7
3.3
4.0

0.001
0.046
0.19
0.44
0.8
1.1
1.6
2.1
2.6
3.1

offered traffic, A:
0.01
0.15
0.45
0.9
1.4
1.9
2.5
3.1
3.8
4.5

54

TELETRAFFIC ENGINEERING

The blocking probability specified is used to select the correct column,

and then we track down the column to a row whose value is equal
to or just exceeds the required offered traffic intensity. The value of N
for this row is the minimum number of circuits needed to satisfy the
required demand at the specified probability of call blocking. From the
columns of data in the traffic table, it can be seen that as the number
of circuits increases, the average loading of each circuit increases, for a
fixed call-blocking probability. This is plotted in Figure 3.7 (note that for
simplicity we approximate the circuit loading by the average offered traffic
per circuit, A/N). So, for example, if we have 10 circuits arranged into
two groups of 5, then for a blocking probability of 0.001 we can load each
group with 0.8 E, i.e. a total of 1.6 E. If all 10 circuits are put together into
one group, then 3.1 E can be offered for the same probability of blocking
of 0.001. In the first case the offered traffic per circuit is 0.16 E; in the
second it is 0.31 E. Thus the larger the group of circuits, the better the
circuit loading efficiency.

Average offered traffic

per circuit

0.6

B = 0.02

0.5

B = 0.01

0.4
B = 0.001

0.3

B = 0.005

0.2
0.1
0

Figure 3.7.

3
Offered traffic

Loading Efficiency of Circuits

Percentage overload
0

50

100

150

Probability of call
blocking

0.1

0.01

0.001

10 circuits

7 circuits

Figure 3.8. Overload Capability of Circuit Groups

4 circuits

TRAFFIC TABLES

55

However, if we consider how a group of circuits perform under overload, there are disadvantages in having large groups. Here, we use the
rows of data from Table 3.1 and plot, in Figure 3.8, the blocking probability against the percentage increase in offered traffic over the offered
traffic for B D 0.001. Small groups of circuits do better under overload
conditions than larger groups; this is because the inefficient small groups
have more waste capacity to deal with unexpected overload and the
deterioration in the blocking probability is small. For a large group of
circuits this deterioration can be substantial.

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

Performance Evaluation
hows it going?

METHODS OF PERFORMANCE EVALUATION

If we are to design a network, we need to know whether the equipment
is going to be used to best effect, and to achieve this we will need to
be able to evaluate its performance. Methods for performance evaluation
fall into two categories: measurement techniques, and predictive techniques; with the latter category comprising mathematical analysis and
simulation.

Measurement
Measurement methods require real networks to be available for experimentation. The advantage of direct measurement of network performance
is that no detail of network operation is excluded: the actual operation of
the real network is being monitored and measured. However, there are
some constraints. A revenue-earning network cannot be exercised to its
limits of performance because customers are likely to complain and take
their business elsewhere. An experimental network may be limited in the
number and type of traffic sources available, thus restricting the range of
realistic experimental conditions.

Predictive evaluation: analysis/simulation

In comparing analysis and simulation, the main factors to consider are
the accuracy of results, the time to produce results, and the overall cost
of using the method (this includes development as well as use).
One advantage of analytical solutions is that they can be used reasonably quickly. However, the need to be able to solve the model restricts

58

PERFORMANCE EVALUATION

the range of system or traffic characteristics that can be included. This

can result in right answers to the wrong problem, if the model has to
be changed so much from reality to make it tractable. Thus analysis is
often used to produce an approximation of a system, with results being
produced relatively quickly and cheaply.
Networks of almost arbitrary complexity can be investigated using
simulation: systems may be modelled to the required level of precision.
Very often, simulation is the only feasible method because of the nature
of the problem and because analytical techniques become too difficult to
handle. However, simulation can be costly to develop and run, and it is
time-consuming, particularly when very rare events (such as ATM cell
loss) are being measured (although accelerated simulation techniques
can reduce the time and cost involved).

QUEUEING THEORY
Analysis of the queueing process is a fundamental part of performance
evaluation, because queues (or waiting lines) form in telecommunications systems whenever customers contend for limited resources. In
technologies such as ATM or IP not only do connections contest, and
may be made to queue, but each accepted connection consists of a stream
of cells or packets and these also must queue at the switching nodes or
routers as they traverse the network.
We will use a queue then as a mathematical expression of the idea of
resource contention (Figure 4.1): customers arrive at a queueing system
needing a certain amount of service; they wait for service, if it is not
immediately available, in a storage area (called a buffer, queue, or
waiting line); and having waited a certain length of time, they are
served and leave the system. Note that the term customers is the general
expression you will encounter in queueing theory terminology and it is
q, number of customers in system

Customers
arriving
with rate

w, number of
customers waiting

r, utilization

Buffer

Server

tw, waiting time

s, service time

tq, system time

Figure 4.1.

The Queueing System

Customers
leaving

59

QUEUEING THEORY

used to mean anything that queues; in ATM or IP, the customers can be
cells, packets, bursts, flows, or connections. In the rest of this chapter, the
queueing systems refer to ATM buffers and the customers are cells.
Any queueing system is described by the arrival pattern of customers,
the service pattern of customers, the number of service channels, and
the system capacity. The arrival pattern of customers is the input to
a queueing system and can sometimes be specified just as the average
number of arrivals per unit of time (mean arrival rate,
) or by the average
time between arrivals (mean inter-arrival time). The simplest input any
queueing system can have is deterministic, in which the arrival pattern
is one customer every t time units, i.e. an arrival rate of 1/t. So, for a
64 kbit/s constant bit-rate (CBR) service, if all 48 octets of the information
field are filled then the cell rate is 167 cell/s, and the inter-arrival time
is 6 ms. If the arrival pattern is stochastic (i.e. it varies in some random
fashion over time), then further characterization is required, e.g. the
probability distribution of the time between arrivals. Arrivals may come
in batches instead of singly, and the size of these batches may vary. We
will look at a selection of arrival patterns in Chapter 6.
The service pattern of customers, as with arrival patterns, can be
described as either a rate, , of serving customers, or as the time, s,
required to service a customer. There is one important difference: service
time or service rate are conditioned on the system not being empty. If it is
empty, the service facility is said to be idle. However, when an ATM cell
buffer is empty, a continuous stream of empty cell slots is transmitted.
Thus the server is synchronized and deterministic; this is illustrated in
Figure 1.4
In the mathematical analysis of an ATM buffer, the synchronization is
often neglected thus a cell is assumed to enter service immediately upon
entry to an empty buffer, instead of waiting until the beginning of the
next free slot. For a 155.52 Mbit/s link, the cell slot rate is 366 792 cell/s
and the service time per cell is 2.726 s. However, 1 in every 27 cell slots is
used for operations and maintenance (OAM) cells for various monitoring
and measurement duties. Thus the cell slot rate available for traffic is
26
366 792 D 353 208 cell/s
27
which can be approximated as a service time per cell of 2.831 s.
The number of service channels refers to the number of servers that
can serve customers simultaneously. Multi-channel systems may differ
according to the organization of the queue(s): each server may have its
own queue, or there may be only one queue for all the servers. This is of
particular interest when analysing different ATM switch designs.
The system capacity consists of the waiting area and the number of
service channels, and may be finite or infinite. Obviously in a real system

60

PERFORMANCE EVALUATION

the capacity must be finite. However, assuming infinite capacity can

simplify the analysis and still be of value in describing ATM queueing
behaviour.

Notation
Kendalls notation, A/B/X/Y/Z, is widely used to describe queueing
systems:
A
B
X
Y
Z

specifies the inter-arrival time distribution

specifies the service time distribution
specifies the number of service channels
specifies the system capacity, and
specifies the queue discipline

An example is the M/D/1 queue. Here the M refers to a memoryless,

or Markov, process, i.e. negative exponential inter-arrival times. The D
means that the service time is always the same: fixed or deterministic
(hence the D), and 1 refers to a single server. The Y/Z part of the notation
is omitted when the system capacity is infinite and the queue discipline is
first-come first-served. We will introduce abbreviations for other arrival
and service processes as we need them.

Elementary relationships
Table 4.1 summarizes the notation commonly used for the various
elements of a queueing process. This notation is not standardized, so
beware. . . for example, q may be used, either to mean the average
number of customers in the system, or the average number waiting to
be served (unless otherwise stated, we will use it to mean the average
number in the system).
There are some basic queueing relationships which are true, assuming
that the system capacity is infinite, but regardless of the arrival or service
Table 4.1.

Notation


s

q
tq
w
tw

Commonly Used Notation for Queueing Systems

Description
mean number of arrivals per unit time
mean service time for each customer
utilization; fraction of time the server is busy
mean number of customers in the system (waiting or being served)
mean time a customer spends in the system
mean number of customers waiting to be served
mean time a customer spends waiting for service

61

QUEUEING THEORY

patterns and the number of channels or the queue discipline. The utilization, , is equal to the product of the mean arrival rate and the mean
service time, i.e.
 D
s
for a single-server queue. With one thousand 64 kbit/s CBR sources, the
arrival rate is 166 667 cell/s. We have calculated that the service time of a
cell is 2.831 s, so the utilization, , is 0.472.
The mean number of customers in the queue is related to the average
time spent waiting in the queue by a formula called Littles formula (often
written as L D
W). In our notation this is:
w D
tw
So, if the mean waiting time is 50 s, then the average queue length
is 8.333 cells. This relationship also applies to the average number of
customers in the system:
q D
tq
The mean time in the system is simply equal to the sum of the mean
service time and waiting time, i.e.
tq D tw C s
which, in our example, gives a value of 52.831 s. The mean number of
customers in a single-server system is given by
qDwC
which gives a value of 8.805 cells.

The M/M/1 queue

We can continue with the example of N CBR sources feeding an ATM
buffer by making two assumptions, but the example will at least give us
a context for choosing various parameter values. The first assumption is
that the cell arrival pattern from N CBR sources can be approximated by
negative exponential inter-arrival times. This is the same as saying that
the arrivals are described by a Poisson process. This process just looks at
the arrival pattern from a different perspective. Instead of specifying a
time duration, the Poisson distribution counts the number of arrivals in
a time interval.
The second assumption is that the service times of these cells are
described by a negative exponential distribution. In Chapter 8 we will
see that the first assumption can be justified for large N. Given the fact
that ATM uses fixed-length cells (and hence fixed service times), the

62

PERFORMANCE EVALUATION

second assumption is not very accurate! Nonetheless, we can use this

example to illustrate some important points about queueing systems.
So, how large should we make the ATM buffer? Remember that the
M/M/1 queueing system assumes infinite buffer space, but we can get
some idea by considering the average number of cells in the system,
which is given by

qD
1
In our example, the utilization resulting from 1000 CBR sources is
0.472, which gives an average system size of 0.894 cell. Subtracting the
utilization from this gives us the average waiting space that is used,
0.422 cell. This is not a very helpful result for dimensioning an ATM
buffer; we would expect to provide at least some waiting space in excess
of 1 cell. But if we look at a graph (Figure 4.2) of q against , as 
varies from 0 to 1, then we can draw a very useful conclusion. The key
characteristic is the knee in the curve around 80% to 90% utilization,

Average number in queueing system

50

40

30

20

10

0
0.0

0.2

0.4
Utilization

0.6

0.8

1.0


1
i :D 0.. 999
i
xi :D
1000
yi :D q xi 

q  :D

Figure 4.2. Graph of the Average Number of Cells in the M/M/1 Queueing System,
and the Mathcad Code to Generate (x, y) Values for Plotting the Graph

63

QUEUEING THEORY

which suggests that it is best to operate the system below 80% utilization
to avoid large queues building up.
But we still do not have any idea of how large to make the ATM buffer.
The next step is to look at the distribution of system size which is given by
Prfsystem size D xg D 1  x
Figure 4.3 shows this distribution for a range of different utilization
values, including the value of 0.472 which is our particular example. In
this case we can read from the graph that the probability associated with
a system size of 10 cells is 0.0003.
From this we might conclude that a buffer length of 10 cells would
not be adequate to meet the cell loss probability (CLP) requirements of
ATM which are often quoted as being 108 or less. For the system size
probability to be less than 108 , the system size needs to be 24 cells;
the actual probability is 7.89 109 . In making this deduction, we have
approximated the CLP by the probability that the buffer has reached a

Queue size
10

15

100

= 0.2
= 0.4
= 0.472
= 0.6
= 0.8

101
Probability

20

102
103
104
105

PrSystemSizeisX , x :D 1   x
i :D 0.. 20
xi :D i
y1i :D PrSystemSizeisX 0.2, xi 
y2i :D PrSystemSizeisX 0.4, xi 
y3i :D PrSystemSizeisX 0.472, xi 
y4i :D PrSystemSizeisX 0.6, xi 
y5i :D PrSystemSizeisX 0.8, xi 
Figure 4.3. Graph of the System State Distribution for the M/M/1 Queue, and the
Mathcad Code to Generate (x, y) Values for Plotting the Graph

64

PERFORMANCE EVALUATION

particular level in our infinite buffer model. This assumes that an infinite
buffer model is a good model of a finite buffer, and that Prfsystem size D
xg is a reasonable approximation to the loss from a finite queue of size x.
Before we leave the M/M/1, lets look at another approximation to the
CLP. This is the probability that the system size exceeds x. This is found by
summing the state probabilities up to and including that for x, and then
subtracting this sum from 1 (this is a simpler task than summing from
x C 1 up to infinity). The equation for this turns out to be very simple:
Prfsystem size > xg D xC1
When x D 24 and  D 0.472, this equation gives a value of 7.06 109
which is very close to the previous estimate.
Now Figure 4.4 compares the results for the two approximations,
Prfsystem size D xg and Prfsystem size
xg, with the actual loss probability from the M/M/1/K system, for a system size of 24 cells, with the
utilization varying from 0 to 1. What we find is that all three approaches
give very similar results over most utilization values, diverging only
when the utilization approaches 100%. For the example utilization value
of 0.472, there is in fact very little difference. The main point to note
here is that an infinite queue can provide a useful approximation for a
finite one.

The M/D/1/K queue

So lets now modify our second assumption, about service times, and
instead of being described by a negative exponential distribution we will
model the cells as they are of fixed length. The only assumption we
will make now is that they enter service whenever the server is idle,
rather than waiting for the next cell slot. The first assumption, about

Estimate of cell loss

1E+00
1E02
1E04

0.2

Utilization
0.4
0.6

0.8

1
Pr {x>24}
M/M/1/24
Pr{x=24}

1E06
1E08
1E10
1E12

Figure 4.4. Comparison of CLP Estimates for Finite M/M/1 Queueing System

65

QUEUEING THEORY

arrival times, remains the same. We will deal with a finite queue directly,
rather than approximating it to an infinite queue. This, then, is called the
M/D/1/K queueing system.
The solution for this system is described in Chapter 7. Figure 4.5
compares the cell loss from the M/D/1/K with the M/M/1 CLP estimator, Prfsystem size D xg, when the system size is 10. As before, the
utilization ranges from 0 to 1. At the utilization of interest, 0.472, the
difference between the cell loss results is about two orders of magnitude.
So we need to remember that performance evaluation answers can
be rather sensitive to the choice of model, and that this means they will
always be, to some extent, open to debate. For the cell loss probability in
the M/D/1/K to be less than 108 , the system size needs to be a minimum
of 15 cells, and the actual CLP (if it is 15 cells) is 4.34 109 . So, by using
a more accurate model of the system (compared to the M/M/1), we
can save on designed buffer space, or alternatively, if we use a system
size of 24 cells, the utilization can be increased to 66.8%, rather than
47.2%. This increase corresponds to 415 extra 64 kbit/s simultaneous
CBR connections.
It is also worth noting from Figure 4.5 that the cell loss probabilities are
very close for high utilizations, i.e. the difference between the two models,
with their very different service time assumptions, becomes almost
negligible under heavy traffic conditions. In later chapters we present
some useful heavy traffic results which can be used for performance
evaluation of ATM, where applicable.

Delay in the M/M/1 and M/D/1 queueing systems

ATM features both cell loss and cell delay as key performance measures,
and so far we have only considered loss. However, delay is particularly

0.2

Utilization
0.4
0.6

0.8

Cell loss probability

1
0.01
0.0001
1E06

M/M/1
M/D/1/10

1E08
1E10
1E12

Figure 4.5. Comparison of M/D/1/K and M/M/1 Cell Loss Results

66

PERFORMANCE EVALUATION

important to real-time services, e.g. voice and video. Littles result allows
us to calculate the average waiting time from the average number waiting
in the queue and the arrival rate. If we apply this analysis to the example
of 1000 CBR connections multiplexed together, we obtain the following:
tw D

w
0.422
D
D 2.532 s


166 667

The average time in the system is then

tq D tw C s D 2.532 C 2.831 D 5.363 s
Another way of obtaining the same result is to use the waiting time
formula for the M/M/1 queue. This is
tw D

s
1

For the M/D/1 queue, there is a similar waiting time formula:

tw D

s
2 1  

In both cases we need to add the service time (cell transmission time) to
obtain the overall delay through the system. But the main point to note is
that the average waiting time in the M/D/1 queue (which works out as
1.265 s in our example) is half that for the M/M/1 queue.
Figure 4.6 shows the average waiting time against utilization for both
queue models. The straight line shows the cell service time. Notice how it
dominates the delay up to about 60% utilization. We can take as a useful
rule of thumb that the average delay arising from queueing across a
network will be approximately twice the sum of the service times. This
assumes, of course, that the utilization in any queue will be no more
than about 60%. For the total end-to-end delay, we must also add in the
propagation times on the transmission links.
So, are these significant values? Well, yes, but, taken alone, they
are not sufficient. We should remember that they are averages, and
cells will actually experience delays both larger and smaller. Delay is
particularly important when we consider the end-to-end characteristics
of connections; all the cells in a connection will have to pass through a
series of buffers, each of which will delay them by some random amount
depending on the number of cells already in the buffer on arrival. This
will result in certain cells being delayed more than others, so-called delay
jitter, or cell delay variation (CDV).

67

QUEUEING THEORY

Average waiting time (in s)

50
40
30
20
M/M/1
10
M/D/1
0
0.0

0.2

0.4
0.6
Utilization

0.8

1.0

s
1
s
MD1tw , s :D
2 1  
i :D 0.. 999
i
xi :D
1000
y1i :D MM1tw xi , 2.831
y2i :D MD1tw xi , 2.831
y3i :D 2.831
MM1tw , s :D

Figure 4.6. Graph of the Average Waiting Times for M/M/1 and M/D/1 Queues,
and the Mathcad Code to Generate (x, y) Values for Plotting the Graph
Other traffic
Monitored call
3

2
Longer
interval

Figure 4.7.

1
Shorter
interval

Variation in Delay for Cells Passing through a Buffer

A pictorial illustration of this is shown in Figure 4.7. Here, we show

only the cells of the connection we are monitoring; there is, of course,
other traffic to contribute to the queueing in the buffer. The second cell
experiences a shorter delay than the first and third cells. This produces
a smaller interval between cells 1 and 2, and a longer interval between
cells 2 and 3. Variation in delay can be a particular problem for usage
parameter control, and we will look at this issue again in Chapter 11.

68

PERFORMANCE EVALUATION

Waiting time (in cell slots)

4
6
8

10

1E+00

Probability

1E01

Number
of
buffers

1E02

10

1E03

3
2

1E04

1E05

Figure 4.8.

End-to-End Waiting Time Distributions

So much for illustrations, what of concrete examples? If again we use

our CBR example (1000 multiplexed CBR 64 kbit/s sources), we can use
more of the theory associated with the M/D/1 queue to predict the result
of passing this stream of cells through a succession of similar queues, and
plot the resulting waiting time distribution. The probabilities associated
with a cell in the stream being delayed by x time slots having passed
through 1, 2, 3 and 10 similar buffers are shown in Figure 4.8. To generate
these results, we have assumed that each buffer is independent of all
the others, and that they are all loaded at 0.472. The results clearly show
the trend for the delay distribution to flatten as the number of buffers
increases: as you might expect, the more buffers the cells pass through,
the more the probabilities associated with long waits and with short waits
tend to equal out.

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

Fundamentals of Simulation
those vital statistics

DISCRETE TIME SIMULATION

This chapter is intended as an introduction to simulation and, in particular,
its application to cell- and packet-based queueing. For anyone wanting
a more comprehensive treatment of the subject of simulation in general,
we refer to [5.1].
We will introduce the subject of simulation by concentrating on a
discrete version of the M/D/1 queue, applicable to the study of ATM cell
buffering. There are two basic ways to simulate such a queue:

discrete time advance

discrete event advance

In the former, the simulator moves from time instant i to time instant
i C 1 regardless of whether the system state has changed, e.g. if the
M/D/1 queue is empty at i it could still be empty at i C 1 and the
program will still only advance the clock to time i C 1. These instants
can correspond to cell slots in ATM. In discrete-event simulation, the
simulator clock is advanced to the next time for which there is a change
in the state of the simulation model, e.g. a cell arrival or departure at the
M/D/1 queue.
So we have a choice: discrete time advance or discrete event advance.
The latter can run more quickly because it will cut out the slot-toslot transitions when the queue is empty, but the former is easier
to understand in the context of ATM because it is simpler to implement and it models the cell buffer from the point of view of the
server process, i.e. the conveyor belt of cell slots (see Figure 1.4).
We will concentrate on the discrete time advance mechanism in this
introduction.

70

FUNDAMENTALS OF SIMULATION

In the case of the synchronized M/D/1 queue the obvious events

between which the simulator can jump are the end of time slot instants,
and so the simulator needs to model the following algorithm:
Ki D max
0, Ki1 C Ai  1
where
Ki = number of cells in modelled system at end of time slot i
Ai = number of cells arriving to the system during time slot i
This algorithm can be expressed as a simulation program in the following
pseudocode:
BEGIN
initialize variables
i, A, K, arrival rate, time slot limit, histogram[]
WHILE (i < time slot limit)
generate new arrivals
A := Poisson(arrival rate)
K := K + A
serve a waiting cell
IF K > 0 THEN
K := K 1
ELSE
K := 0
store results
histogram[K] := histogram[K] + 1
advance time to next time slot
i := i + 1
END WHILE
END

The main program loop implements the discrete time advance mechanism
in the form of a loop counter, i. The beginning of the loop corresponds
to the start of time slot i, and the first section generate new arrivals calls
function Poisson which returns a random non-negative integer for the
number of cell arrivals during this current time slot. We model the queue
with an arrivals-first buffer management strategy, so the service instants
occur at the end of the time slot after any arrivals. This is dealt with by
the second section, serve a waiting cell, which decrements the queue state
variable K, if it is greater than 0, i.e. if the queue is not empty. At this
point, in store results we record the state of the queue in a histogram. This
is simply a count of the number of times the queue is in state K, for each
possible value of K, (see Figure 5.1), and can be converted to an estimate
of the state probability distribution by dividing each value in the array
histogram[] by the total number of time slots in the simulation run.

71

Histogram(K)

DISCRETE TIME SIMULATION

102

101

100

5
queue state, K

10

Figure 5.1. An Example of a Histogram of the Queue State (for a Simulation Run of
1000 Time Slots)

Generating random numbers

The function Poisson generates a random non-negative integer number
of cell arrivals according to a Poisson distribution with a particular
arrival rate. This is achieved in two parts: generate random numbers that
are uniformly distributed over the range 0 to 1; convert these random
numbers to be Poisson-distributed. Lets assume that we have a function generate random number which implements the first part. The
following pseudocode converts the random numbers from having a
uniform distribution to having a Poisson distribution.
FUNCTION X = Poisson(arrival rate)
initialize variables
a := earrival rate
b := 1
j := 1
REPEAT
j := j + 1
U := generate random number
b := b U
UNTIL
b < a
return result
X := j
END FUNCTION

The REPEAT loop corresponds to the generation of cells, and the loop
records the number of cells in the batch in variable j, returning the
final total in variable X. Remember that with this particular simulation
program we are not interested in the arrival time of each cell within the
slot, but in the number of arrivals during a slot.

72

FUNDAMENTALS OF SIMULATION

But how do we generate the random numbers themselves? A good

random number generator (RNG) should produce a sequence of numbers
which are uniformly distributed on the range [0, 1] and which do
not exhibit any correlation between the generated numbers. It must
be fast and avoid the need for much storage. An important property of the random number sequence is that it must be reproducible;
this aids debugging, and can be used to increase the precision of the
results.
A typical RNG is of the form:
Ui D
a Ui1 C c mod
m
where Ui is the ith random number generated, and m (the modulus), a
(the multiplier) and c (the increment) are all non-negative integers, as is
U0 , the initial value which is called the seed. The values should satisfy
0 < m, a < m, c < m and U0 < m. In practice m is chosen to be very large,
say 109 .
Obviously, once the RNG produces a value for Ui which it has
produced before, the sequence of numbers being generated will repeat,
and unwanted correlations will begin to appear in the simulator results.
An important characteristic of an RNG is the length of the sequence
before it repeats; this is called the period. The values of m and c are
chosen, in part, to maximize this period. The WichmannHill algorithm
combines three of these basic generators to produce a random number
generator which exhibits exceptional performance. The pseudocode for
this algorithm is:
FUNCTION U = generate random number
x :=
171 x mod
30269
y :=
172 y mod
30307
z :=
170 z mod
30323
U :=
x/30269 +
y/30307 +
z/30323
temp := trunc
U
U := U temp
END FUNCTION

The period is of particular relevance for ATM traffic studies, where rare
events can occur with probabilities as low as 1010 (e.g. lost cells). Once
an RNG repeats its sequence, unwanted correlations will begin to appear
in the results, depending on how the random number sequence has been
applied. In our discrete time advance simulation, we are simulating time
slot by time slot, where each time slot can have 0 or more cell arrivals.
The RNG is called once per time slot, and then once for each cell arrival

DISCRETE TIME SIMULATION

73

during the time slot. With the discrete event advance approach, a cellby-cell simulator would call the RNG once per cell arrival to generate the
inter-arrival time to the next cell.
The WichmannHill algorithm has a period of about 7 1012 . Thus,
so long as the number of units simulated does not exceed the period
of 7 1012 , this RNG algorithm can be applied. The computing time
required to simulate this number of cells is impractical anyway, so we can
be confident that this RNG algorithm will not introduce correlation due
to repetition of the random number sequence. Note that the period of the
WichmannHill algorithm is significantly better than many of the random
number generators that are supplied in general-purpose programming
languages. So, check carefully before you use a built-in RNG.
Note that there are other ways in which correlations can appear in a
sequence of random numbers. For more details, see [5.1].

M/D/1 queue simulator in Mathcad

The following Mathcad code implements the discrete time advance
simulator pseudocode for the M/D/1 queue. Note that the WHILE loop
in the main pseudocode is vectorized (using range variable i), as is the
REPEAT loop in the Poisson function pseudocode (using range variable j).
An example of the histogram of queue state results is shown in Figure 5.1
(plotting histogramK against Kbins ).
initialize variables
time slotlimit := 1000
arrivalrate := 0 . 5
i := 1 .. time slotlimit
maxK :D 10
K0 :D 0

generate new arrivals

a :D earrivalrate
bi , 0 :D 1
j :D 1 .. 10
!
bi , j :D
rnd
1 bi , j1 
cellsi ,

j :D if
bi , j < a , 0 , 1
Ai :D j cellsi , j

serve a waiting cell

Ki :D max[[0
Ki1 C Ai   1]]

74

FUNDAMENTALS OF SIMULATION

store results

i Ai
time slotlimit
actualload D 0.495

actualload :D

q :D 0 , 1 ..

maxK

Kbinsq :D q
histogramK :D hist (Kbins , K)

end of simulation

Reaching steady state

When do we stop a simulation? This is not a trivial question, and if, for
example, we want to find the cell loss probability in an M/D/1/K model,
then the probability we are seeking is actually a steady-state probability:
the long-run proportion of cells lost during period T as T ! 1. Since
we cannot actually wait that long, we must have some prior idea about
how long it takes for the simulator to reach a good approximation to
steady-state.
A simulation is said to be in steady state, not when the performance measurements become constant, but when the distribution of the
measurements becomes (close to being) invariant with time. In particular,
the simulation needs to be sufficiently long that the effect of the initial
state of the system on the results is negligible. Lets take an example.
Recall from Chapter 4 that we can use the probability that the queue size
is greater than K, denoted Q
K, as an estimate of the cell loss from a finite
queue of size K. Suppose that the queue length is 2. We can calculate Q
2
from the histogram data recorded in our simulation program thus:
1


Q
2 D

histogram[K]

KD3

or, alternatively as
i
Q
2 D

2


histogram[K]

KD0

If we start our M/D/1 simulator, and plot Q
2 for it as this value evolves
over time, we will see something like that which is shown in Figure 5.2.

75

DISCRETE TIME SIMULATION

1E+00

1E01

Time slot number

1000 2000 3000 4000 5000 6000 7000 8000 9000 10000

Actual value

Q(2)
1E02

Simulation
measurements

1E03
Figure 5.2.

Evolution of Q
2 for the Simulated M/D/1

Here, the simulator calculates a measurement result for Q
2 every 1000
time slots; that is to say it provides an estimate of Q
2 every 1000 slots. But
from Figure 5.2 we can see that there are transient measurements, and
that these strongly reflect the initial system state. It is possible to cut out
these measurements in the calculation of steady state results; however,
it is not easy to identify when the transient phase is finished. We might
consider the first 7000 slots as the transient period in our example.

Batch means and confidence intervals

The output from one run of a simulation is a sequence of measurements
which depends on the particular sequence of random numbers used. In
the example we have been considering, we store results at the end of
each time slot, then, at intervals of 1000 time slots, we output a value
for Q
2. But we do not take the last value to be output as the final
result of the simulation run. The sequence of measurements of Q
2
needs to be evaluated statistically in order to provide reliable results for
the steady-state value of Q
2.
Suppose that we take j D 1 to N measurements of Q
2. First, we can
obtain an estimate of the mean value by calculating
N


O
Q
2
D

Q
2j

jD1

Then we need an estimate of how the measurements vary over the

set. We can construct an estimate of the confidence interval for Q
2 by

76

FUNDAMENTALS OF SIMULATION

calculating
N


O
Q
2
z/2

2
O

Q
2j  Q
2

jD1

N
N  1

where z/2 is obtained from standard normal tables and 1  is the

degree of confidence.
A confidence interval quantifies the confidence that can be ascribed
to the results from a simulation experiment, in a statistical sense. For
example, a 90% confidence interval (i.e. D 0.1) means that for 90%
of the simulation runs for which an interval is calculated, the actual
value for the measure of interest falls within the calculated interval (see
Figure 5.3). On the other 10% of occasions, the actual value falls outside
the calculated interval. The actual percentage of times that a confidence
interval does span the correct value is called the coverage.
There are a number of different methods for organizing simulation
experiments so that confidence intervals can be calculated from the
measurements. The method of independent replications uses N estimates
obtained from N independent simulation runs. In the method of batch
means, one single run is divided into N batches (each batch of a certain
fixed number, L, of observations) from which N estimates are calculated.
The value of L is crucial, because it determines the correlation between
batches: considering our M/D/1 example again, if L is too small then the
system state at the end of Nj will be heavily influenced by (correlated with)
the system state at the end of Nj1 . The regenerative method also uses a
single run, but depends on the definition of a regenerative state a state
after which the process repeats probabilistically. Determining an appropriate regenerative state can be difficult, and it can be time-consuming
to obtain a sufficient number of points at which the simulation passes
through such states, in order to calculate valid confidence intervals.
1

10

Experiments

Actual value

Actual value falls

within confidence
interval obtained
from experiment
Figure 5.3.

Actual value falls

outside confidence
interval obtained
from experiment
Confidence Intervals and Coverage

ACCELERATED SIMULATION

77

The main advantage of methods involving just a single run is that

only one transient phase needs to be discarded. Determining the best
length for the simulation run(s) is a problem for all three methods. This is
because, if the runs are too short, they can produce confidence intervals
with actual coverage considerably lower than desired. However, this has
to be balanced against the need to limit the sample size to minimize the
time (and hence, cost) of the simulation; so the emphasis is on finding
a sufficient sample size. In addressing this problem, an alternative to
the arbitrary fixed sample size approach is to increase the sample size
sequentially until an acceptable confidence interval can be constructed.

Validation
Any simulation model will need to be checked to ensure that it works. This
can be a problem: a very general program that is capable of analysing
a large number of scenarios will be impossible to test in all of them,
especially as it would probably have been developed to solve systems
that have no analytical solution to check against. However, even for the
most general of simulators it will be possible to test certain simple models
that do have analytical solutions, e.g. the M/D/1.

ACCELERATED SIMULATION
In the discussion on random number generation we mentioned that the
computing time required to simulate 1012 cells is impractical, although
cell loss probabilities of 1010 are typically specified for ATM buffers. In
fact, most published simulation results for ATM extend no further than
probabilities of 105 or so.
How can a simulation be accelerated in order to be able to measure
such rare events? There are three main ways to achieve this: use more
computing power, particularly in the form of parallel processing; use
statistical techniques to make better use of the simulation measurements;
and decompose the simulation model into connection, burst and cell
scales and use only those time scales that are relevant to the study.
We will focus on the last approach because it extends the analytical
understanding of the cell and burst scales that we develop in later
chapters and applies it to the process of simulation. In particular, burstscale queueing behaviour can be modelled by a technique called cell-rate
simulation.

Cell-rate simulation
The basic unit of traffic with cell-rate simulation is a burst of cells.
This is defined as a fixed cell-rate lasting for a particular time period

78

FUNDAMENTALS OF SIMULATION

Aell rate

A burst of
fixed cell rate

An event marks
a change from
one fixed cell
rate to another

Time

Figure 5.4. The Basic Unit of Traffic in Cell-Rate Simulation

during which it is assumed that the inter-arrival times do not vary (see
Figure 5.4). Thus instead of an event being the arrival or service of a cell,
an event marks the change from one fixed cell-rate to another. Hence
traffic sources in a cell-rate simulator must produce a sequence of bursts
of cells. Such traffic sources, based on a cell-rate description are covered
in Chapter 6.
The multiplexing of bursts from different sources through an ATM
buffer has to take into account the simultaneous nature of these bursts.
Bursts from different sources will overlap in time and a change in the rate
of just one source can affect the output rates of all the other VCs passing
through the buffer.
An ATM buffer is described by two parameters: the maximum number
of cells it can hold, i.e. its buffer capacity; and the constant rate at which
cells are served, i.e. its cell service-rate. The state of a queue, at any
moment in time, is determined by the combination of the input rates
of all the VCs, the current size of the queue, and the queue parameter
values.
The flow of traffic through a queue is described by input, output,
queueing and loss rates (see Figure 5.5). Over any time period, all cells
input to the buffer must be accounted for; they are or served queued or
lost. At any time, the rates for each VC, and for all VCs, must balance:
input rate D output rate C queueing rate C loss rate
When the queue is empty, the output rates of VCs are equal to their
input rates, the total input rate is less than the service rate, and so there
is no burst-scale queueing.

79

ACCELERATED SIMULATION

Input rate

ve

Output rate

Queueing rate
+ve
Loss rate
Figure 5.5.

The Balance of Cell Rates in the Queueing Model for Cell-Rate Simulation

output rate D input rate

queueing rate D loss rate D 0
In a real ATM system there will of course be cell-scale queueing, but this
behaviour is not modelled by cell-rate simulation. When the combined
input rate exceeds the service rate, the queue size begins to increase at a
rate determined by the difference between the input rate and service rate
of the queue.
queueing rate D input rate  service rate
For an individual VC, its share of the total queueing rate corresponds to
its share of the total input rate. Once the queue becomes full, the total
queueing rate is zero and the loss rate is equal to the difference in the
input rate and service rate.
loss rate D input rate  service rate
Although this appears to be a simple model for the combined cell rates,
it is more complicated when individual VCs are considered. An input
change to a full buffer, when the total input rate exceeds the service rate,
has an impact not only on the loss rate but also on all the individual VC
queueing rates. Also, the effect of a change to the input rate of a VC, i.e. an
event at the input to the queue, is not immediately apparent on the output,
if there are cells queued. At the time of the input event, only the queueing
and/or loss rates change. The change appears on the output only after
the cells which are currently in the queue have been served. Then, at the
time of this output event, the queueing and output rates change.
It is beyond the scope of this book to describe the cell-rate simulation technique in more detail (the interested reader is referred to [5.2]);
however, we present some results in Figure 5.6 which illustrate the accelerated nature of the technique. In comparison with cell-by-cell simulation,
cell-rate simulation shows significant speed increases, varying from 10
times to over 10 000 times faster. The speed improvement increases in

80

FUNDAMENTALS OF SIMULATION

100000
Low utilization (40%)

10000

High utilization (80%)

1000
100
10
1
10

100

1000

10000

100000

Burst length (cells)

Figure 5.6.
tion

Speed Increase of Cell-Rate Simulation Relative to Cell-by-Cell Simula-

proportion to the average number of cells in a fixed-rate burst, and also

increases the lower the utilization and hence also the lower the cell
loss. This is because it focuses processing effort on the traffic behaviour
which dominates the cell loss: the burst-scale queueing behaviour. So
cell-rate simulation enables the low cell-loss probabilities required of
ATM networks to be measured within reasonable computing times.

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

Traffic Models
youve got a source

LEVELS OF TRAFFIC BEHAVIOUR

So, what kind of traffic behaviour are we interested in for ATM, or IP? In
Chapter 3 we looked at the flow of calls in a circuit-switched telephony
network, and in Chapter 4 we extended this to consider the flow of
cells through an ATM buffer. In both cases, the time between arrivals
(whether calls or cells) was given by a negative exponential distribution:
that is to say, arrivals formed a Poisson process. But although the same
source model is used, different types of behaviour are being modelled.
In the first case the behaviour concerns the use made of the telephony
service by customers in terms of how often the service is used, and
for how long. In the second case, the focus is at the level below the call
time scale, i.e. the characteristic behaviour of the service as a flow of cells
or, indeed, packets. Figure 6.1 distinguishes these two different types of
behaviour by considering four different time scales of activity:

calendar: daily, weekly and seasonal variations

connection: set-up and clear events delimit the connection duration,

which is typically in the range 100 to 1000 seconds

burst: the behaviour of a transmitting user, characterized as a cell (or

packet) flow rate, over an interval during which that rate is assumed
constant. For telephony, the talk-spurt on/off characteristics have
durations ranging from a fraction of a second to a few seconds. In IP,
similar time scales apply to packet flows.

cell/packet: the behaviour of cell or packet generation at the lowest

level, concerned with the time interval between arrivals (e.g. multiples
of 2.831 s at 155.52 Mbit/s in ATM)

82

TRAFFIC MODELS

Time-scale of activity
Dimensioning
Use made
of the
service

Calendar
Connection

Burst

Characteristic
behaviour
of the
service

Cell
Performance Engineering
Figure 6.1. Levels of Traffic Behaviour

This analysis of traffic behaviour helps in distinguishing the primary

objectives of dimensioning and performance engineering. Dimensioning
focuses on the organization and provision of sufficient equipment in the
network to meet the needs of services used by subscribers (i.e. at the
calendar and connection levels); it does require knowledge of the service
characteristics, but this is in aggregate form and not necessarily to a great
level of detail. Performance engineering, however, focuses on the detail
of how the network resources are able to support services (i.e. assessing
the limits of performance); this requires consideration of the detail of
service characteristics (primarily at the cell and burst levels), as well as
information about typical service mixes how much voice, video and
data traffic is being transported on any link (which would be obtained
from a study of service use).

TIMING INFORMATION IN SOURCE MODELS

A source model describes how traffic, whether cells, bursts or connections,
emanates from a user. As we have already seen, the same source model
can be applied to different time scales of activity, but the Poisson process
is not the only one used for ATM or IP. Source models may be classified in
a variety of ways: continuous time or discrete time, inter-arrival time or
counting process, state-based or distribution-based, and we will consider
some of these in the rest of this chapter. It is worth noting that some
models are associated with a particular queue modelling method, an
example being fluid flow analysis.
A distinguishing feature of source models is the way the timing
information is presented. Figure 6.2 shows the three different ways in
the context of an example ATM cell stream: as the number of cell slots
between arrivals (the inter-arrival times are 5, 7, 3 and 5 slots in this

83

TIME BETWEEN ARRIVALS

Cell slots between arrivals

Cells in block of 25 cell slots

20% of cell slot rate
Time

Figure 6.2. Timing Information for an Example ATM Cell Stream

example); as a count of the number of arrivals within a specified period

(here, it is 5 cells in 25 cell slots); and as a cell rate, which in this case is
20% of the cell slot rate.

TIME BETWEEN ARRIVALS

Inter-arrival times can be specified by either a fixed value, or some arbitrary probability distribution of values, for the time between successive
arrivals (whether cells or connections). These values may be in continuous time, taking on any real value, or in discrete time, for example an
integer multiple of a discrete time period such as the transmission time
of a cell, e.g. 2.831 s.
A negative-exponential distribution of inter-arrival times is the prime
example of a continuous-time process because of the memoryless property. This name arises from the fact that, if the time is now t1 , the
probability of there being k arrivals in the interval t1 ! t2 is independent
of the interval, t, since the last arrival (Figure 6.3). It is this property that
allows the development of some of the simple formulas for queues.
The probability that the inter-arrival time is less than or equal to t is
given by the equation
Prfinter-arrival time
tg D Ft D 1  et
Arrival instant

Time
t
t1
Figure 6.3.

t2

The Memoryless Property of the Negative Exponential Distribution

84

TRAFFIC MODELS

5e006

1e005

Time
1.5e005

2e005

2.5e005

F(t)

0.1

0.01

F, t :D 1  et
i :D 1.. 250
x1i :D i 107
y1i :D F166667, x1i 
j :D 1 .. 8
x2j :D j 2 831 106
y2j :D 1
Figure 6.4. Graph of the Negative Exponential Distribution for a Load of 0.472, and
the Mathcad Code to Generate x, y Values for Plotting the Graph

where the arrival rate is . This distribution, Ft, is shown in Figure 6.4
for a load of 47.2% (i.e. the 1000 CBR source example from Chapter 4). The
arrival rate is 166 667 cell/s which corresponds to an average inter-arrival
time of 6 s. The cell slot intervals are also shown every 2.831 s on the
time axis.
The discrete time equivalent is to have a geometrically distributed
number of time slots between arrivals (Figure 6.5), where that number is
counted from the end of the first cell to the end of the next cell to arrive.
...

k Time slots between cell arrivals

Figure 6.5.
Arrivals

Time

Inter-Arrival Times Specified as the Number of Time Slots between

85

TIME BETWEEN ARRIVALS

Obviously a cell rate of 1 cell per time slot has an inter-arrival time of
1 cell slot, i.e. no empty cell slots between arrivals. The probability that a
cell time slot contains a cell is a constant, which we will call p. Hence a
time slot is empty with probability 1  p. The probability that there are k
time slots between arrivals is given by
Prfk time slots between arrivalsg D 1  pk1 p
i.e. k  1 empty time slots, followed by one full time slot. This is the
geometric distribution, the discrete time equivalent of the negative exponential distribution. The geometric distribution is often introduced in
text books in terms of the throwing of dice or coins, hence it is thought
Time
0

5e006

1e005

1.5e005

2e005

2.5e005

Probability

0.1

0.01

F, t :D 1  et
Geometric (p, k) :D 1  1  pk
i :D 1..250
x1i :D i 107
y1i :D F166667, x1i 
j :D 1..8
y2j :D 1
j :D 1..8
y3j :D Geometric 166667 2.831 106 , j
Figure 6.6. A Comparison of Negative Exponential and Geometric Distributions,
and the Mathcad Code to Generate x, y Values for Plotting the Graph

86

TRAFFIC MODELS

of as having k  1 failures (empty time slots, to us), followed by one

success (a cell arrival). The mean of the distribution is the inverse of
the probability of success, i.e. 1/p. Note that the geometric distribution
also has a memoryless property in that the value of p for time slot n
remains constant however many arrivals there have been in the previous
n  1 slots.
Figure 6.6 compares the geometric and negative exponential distributions for a load of 47.2% (i.e. for the geometric distribution, p D 0.472,
with a time base of 2.831 s; and for the negative exponential distribution,  D 166 667 cell/s, as before). These are cumulative distributions
(like Figure 6.4), and they show the probability that the inter-arrival
time is less than or equal to a certain value on the time axis. This
time axis is sub-divided into cell slots for ease of comparison. The
cumulative geometric distribution begins at time slot k D 1 and adds
Prfk time slots between arrivalsg for each subsequent value of k.
Prf
k time slots between arrivalsg D 1  1  pk

COUNTING ARRIVALS
An alternative way of presenting timing information about an arrival
process is by counting the number of arrivals in a defined time interval.
There is an equivalence here with the inter-arrival time approach in
continuous time: negative exponential distributed inter-arrival times
form a Poisson process:
Prfk arrivals in time Tg D

 Tk T
e
k!

where  is the arrival rate.

In discrete time, geometric inter-arrival times form a Bernoulli process,
where the probability of one arrival in a time slot is p and the probability
of no arrival in a time slot is 1  p. If we consider more than one time slot,
then the number of arrivals in N slots is binomially distributed:
Prfk arrivals in N time slotsg D

N!
1  pNk pk
N  k! k!

and p is the average number of arrivals per time slot.

How are these distributions used to model ATM or IP systems?
Consider the example of an ATM source that is generating cell arrivals
as a Poisson process; the cells are then buffered, and transmitted in
the usual way for ATM as a cell stream in synchronized slots (see
Figure 6.7). The Poisson process represents cells arriving from the source

87

COUNTING ARRIVALS

Buffer
Source

Negative exponential
distribution for time
between arrivals

Figure 6.7.
Stream

Geometrically distributed
number of time slots
between cells in
synchronized cell stream

The Bernoulli Output Process as an Approximation to a Poisson Arrival

to the buffer, at a cell arrival rate of  cells per time slot. At the buffer
output, a cell occupies time slot i with probability p as we previously
defined for the Bernoulli process. Now if  is the cell arrival rate and p
is the output cell rate (both in terms of number of cells per time slot),
and if we are not losing any cells in our (infinite) buffer, we must have
that  D p.
Note that the output process of an ATM buffer of infinite length, fed
by a Poisson source is not actually a Bernoulli process. The reason is that
the queue introduces dependence from slot to slot. If there are cells in the
buffer, then the probability that no cell is served at the next cell slot is 0,
whereas for the Bernoulli process it is 1  p. So, although the output cell
stream is not a memoryless process, the Bernoulli process is still a useful
approximate model, variations of which are frequently encountered in
teletraffic engineering for ATM and for IP.
The limitation of the negative exponential and geometric inter-arrival
processes is that they do not incorporate all of the important characteristics of typical traffic, as will become apparent later.
Certain forms of switch analysis assume batch-arrival processes: here,
instead of a single arrival with probability p, we get a group (the batch),
and the number in the group can have any distribution. This form of
arrival process can also be considered in this category of counting arrivals.
For example, at a buffer in an ATM switch, a batch of arrivals up to some
maximum, M, arrive from different parts of the switch during a time slot.
This can be thought of as counting the same number of arrivals as cells in
the batch during that time slot. The Bernoulli process with batch arrivals
is characterized by having an independent and identically distributed
number of arrivals per discrete time period. This is defined in two parts:
the presence of a batch
Prfthere is a batch of arrivals in a time slotg D p
or the absence of a batch

88

TRAFFIC MODELS

Prfthere is no batch of arrivals in a time slotg D 1  p

and the distribution of the number of cells in a batch:
bk D Prfthere are k cells in a batch given that there is a batch in the
time slotg
Note that k is greater than 0. This description of the arrival process can
be rearranged to give the overall distribution of the number of arrivals
per slot, ak, as follows:
a0 D 1  p
a1 D p b1
a2 D p b2
..
.
ak D p bk
..
.
aM D p bM
This form of input is used in the switching analysis described in Chapter 7
and the basic packet queueing analysis described in Chapter 14. It is a
general form which can be used for both Poisson and binomial input
distributions, as well as arbitrary distributions. Indeed, in Chapter 17 we
use a batch arrival process to model long-range dependent traffic, with
Pareto-distributed batch sizes.
In the case of a Poisson input distribution, the time duration T is one
time slot, and if  is the arrival rate in cells per time slot, then
ak D

k 
e
k!

For the binomial distribution, we now want the probability that there
are k arrivals from M inputs where each input has a probability, p, of
producing a cell arrival in any time slot. Thus
ak D

M!
1  pMk pk
M  k! k!

and the total arrival rate is M p cells per time slot. Figure 6.8 shows
what happens when the total arrival rate is fixed at 0.95 cells per time

89

RATES OF FLOW

k arrivals in one time slot

3
4
5
6
7

10

100
101

Poisson
M=100
M=20
M=10

Probability

102
103
104
105
106
10 7
108

Poisson k ,  :D

k 
e
k!

Binomial k , M , p :D

M!
1  pMk pk
M  k! k!

i :D 0 .. 10
xi :D i
y1i :D Poisson x

i , 0.95


0.95
100


0.95
y3i :D Binomial xi , 20,
20 


0.95
y4i :D Binomial xi , 10,
10
y2i :D Binomial

xi , 100,

Figure 6.8. A Comparison of Binomial and Poisson Distributions, and the Mathcad
Code to Generate x, y Values for Plotting the Graph

slot and the numbers of inputs are 10, 20 and 100 (and so p is 0.095,
0.0475 and 0.0095 respectively). The binomial distribution tends towards
the Poisson distribution, and in fact in the limit as N ! 1 and p ! 0 the
distributions are the same.

RATES OF FLOW
The simplest form of source using a rate description is the periodic
arrival stream. We have already met an example of this in 64 kbit/s CBR

90

TRAFFIC MODELS

telephony, which has a cell rate of 167 cell/s in ATM. The next step is
to consider an ONOFF source, where the process switches between a
silent state, producing no cells, and a state which produces a particular
fixed rate of cells. Sources with durations (in the ON and OFF states)
distributed as negative exponentials have been most frequently studied,
and have been applied to data traffic, to packet-speech traffic, and as a
general model for bursty traffic in an ATM multiplexor.
Figure 6.9 shows a typical teletraffic model for an ONOFF source.
During the time in which the source is on (called the sojourn time in
the active state), the source generates cells at a rate of R. After each cell,
another cell is generated with probability a, or the source changes to the
silent state with probability 1  a. Similarly, in the silent state, the source
generates another empty time slot with probability s, or moves to the
active state with probability 1  s. This type of source generates cells in
patterns like that shown in Figure 6.10; for this pattern, R is equal to half
of the cell slot rate. Note that there are empty slots during the active state;
these occur if the cell arrival rate, R, is less than the cell slot rate.
We can view the ONOFF source in a different way. Instead of showing
the cell generation process and empty time slot process explicitly as
Bernoulli processes, we can simply describe the active state as having a
geometrically distributed number of cell arrivals, and the silent state as
having a geometrically distributed number of cell slots. The mean number
of cells in an active state, E[on], is equal to the inverse of the probability
of exiting the active state, i.e. 1/1  a cells. The mean number of empty
Pr{no} = 1- a
SILENT STATE
Silent for
another time
slot?
Pr{yes} = s

Pr{yes} = a
ACTIVE STATE
Generate
another cell
arrival?

Pr{no} = 1-s
Figure 6.9.

An ONOFF Source Model

ACTIVE

SILENT

1/R

1/C

ACTIVE

Time

Figure 6.10. Cell Pattern for an ONOFF Source Model

91

RATES OF FLOW

1
SILENT STATE

ACTIVE STATE

Generate empty
slotsat a rate of C
E[off] = 1/(1-s)

Generate cells at
a rate of R
E[on] =1/(1-a)
1

Figure 6.11.

An Alternative Representation of the ONOFF Source Model

cell slots in a silent state, E[off], is equal to 1/1  s cell slots. At the
end of a sojourn period in a state, the process switches to the other state
with probability 1. Figure 6.11 shows this alternative representation of
the ONOFF source model.
It is important to note that the geometric distributions for the active
and silent states have different time bases. For the active state the unit of
time is 1/R, i.e. the cell inter-arrival time. Thus the mean duration in the
active state is
1
Ton D E[on]
R
For the silent state the unit of time is 1/C, where C is the cell slot rate;
thus the mean duration in the silent state is
Toff D

1
E[off]
C

The alternative representation of Figure 6.11 can then be generalized by

allowing arbitrary distributions for the number of cells generated in an
active period, and also for the number of empty slots generated in a silent
period.
Before leaving the ONOFF source, lets apply it to a practical example:
silence-suppressed telephony (no cells are transmitted during periods in
which the speaker is silent). Typical figures (found by measurement)
for the mean ON and OFF periods are 0.96 second and 1.69 seconds
respectively. Cells are generated from a 64 kbit/s telephony source at
a rate of R D 167 cell/s and the cell slot rate of a 155.52 Mbit/s link is
C D 353 208 cell/s. Thus the mean number of cells produced in an active
state is
E[on] D R 0.96 D 160 cells
and the mean number of empty slots in a silent state is
E[off] D C 1.69 D 596 921 cell slots
This gives the model shown in Figure 6.12.

92

TRAFFIC MODELS

1
SILENT STATE

ACTIVE STATE

C = 353208 slot/s

R = 167 cell/s

E[off] = 596921 slots

E[on] = 160 cells

1

Figure 6.12.

ONOFF Source Model for Silence-Suppressed Telephony

We can also calculate values of parameters a and s for the model in

Figure 6.9. We know that
E[on] D
so

1
D 160
1a

aD1

1
D 0.993 75
160

E[off] D

1
D 596 921
1s

and

so
sD1

1
D 0.999 998 324 7
596 921

The ONOFF source is just a particular example of a state-based model

in which the arrival rate in a state is fixed, there are just two states,
and the period of time spent in a state (the sojourn time) is negative
exponentially, geometrically, or arbitrarily distributed. We can generalize
this to incorporate N states, with fixed rates in each state. These multistate models (called modulated deterministic processes) are useful for
modelling a number of ONOFF sources multiplexed together, or a
single, more complex, traffic source such as video.
If we allow the sojourn times to have arbitrary distributions, the
resulting process is called a Generally Modulated Deterministic Process
(GMDP). If the state durations are exponentially distributed then the
process is called a Markov Modulated Deterministic Process (MMDP). In
this case, each state produces a geometrically distributed number of cells
during any sojourn period. This is because, having generated arrival i, it
generates arrival i C 1 with a probability given by the probability that the
sojourn time does not end before the time of the next arrival. This probability is a constant if sojourn periods are exponentially distributed because
of the memoryless property of the negative exponential distribution.
We do not need to restrict the model to having a constant arrival rate
in each state: if the arrival process per state is a Poisson process, and the

93

RATES OF FLOW

STATE 3
p(1,3)

p(2,3)

p(3,1)

p(3,2)
p(1,2)

STATE 1

STATE 2
p (2,1

Figure 6.13. The Three-State GMDP

mean of the Poisson distribution is determined by the state the model is

in, then we have an MMPP, which is useful for representing an aggregate
cell arrival process.
For all these state processes, at the end of a sojourn in state i, a transition
is made to another state j; this transition is governed by an N N matrix
of transition probabilities, pi, j i 6D j. Figure 6.13 illustrates a multi-state
model, with three states, and with the transition probabilities from state
i to state j shown as pi, j.
For a comprehensive review of traffic models, the reader is referred
to [6.1].

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

PART II
ATM Queueing and
Traffic Control

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

Basic Cell Switching

up against the buffers

THE QUEUEING BEHAVIOUR OF ATM CELLS IN OUTPUT

BUFFERS
In Chapter 3, we saw how teletraffic engineering results have been
used to dimension circuit-switched telecommunications networks. ATM
is a connection-orientated telecommunications network, and we can
(correctly) anticipate being able to use these methods to investigate the
connection-level behaviour of ATM traffic. However, the major difference
between circuit-switched networks and ATM is that ATM connections
consist of a cell stream, where the time between these cells will usually
be variable (at whichever point in the network that you measure them).
We now need to consider what may happen to such a cell stream as it
travels through an ATM switch (it will, in general, pass through many
such switches as it crosses the network).
The purpose of an ATM switch is to route arriving cells to the appropriate output. A variety of techniques have been proposed and developed
to do switching [7.1], but the most common uses output buffering. We
will therefore concentrate our analysis on the behaviour of the output
buffers in ATM switches. There are three different types of behaviour in
which we are interested: the state probabilities, by which we mean the
proportion of time that a queue is in a particular state (being in state k
means the queue contains k cells) over a very long period of time (i.e.
the steady-state probabilities); the cell loss probability, by which we mean
the proportion of cells lost over a very long period of time; and the cell
waiting-time probabilities, by which we mean the probabilities associated
with a cell being delayed k time slots.
To analyse these different types of behaviour, we need to be aware of
the timing of events in the output buffer. In ATM, the cell service is of
fixed duration, equal to a single time slot, and synchronized so that a cell

98

BASIC CELL SWITCHING

A batch of cells arriving

during time slot n
n1

n+1

Time (slotted)
Departure instant
for cell in service
during time slot n 1
Figure 7.1.
Strategy

Departure instant
for cell in service
during time slot n

Timing of Events in the Buffer: the Arrivals-First Buffer Management

enters service at the beginning of a time slot. The cell departs at the end
of a time slot, and this is synchronized with the start of service of the
next cell (or empty time slot, if there is nothing waiting in the buffer).
Cells arrive during time slots, as shown in Figure 7.1. The exact instants
of arrival are unimportant, but we will assume that any arrivals in a time
slot occur before the departure instant for the cell in service during the
time slot. This is called an arrivals-first buffer management strategy. We
will also assume that if a cell arrives during time slot n, the earliest it can
be transmitted (served) is during time slot n C 1.
For our analysis, we will use a Bernoulli process with batch arrivals,
characterized by an independent and identically distributed batch of k
arrivals (k D 0, 1, 2, . . .) in each cell slot:
ak D Prfk arrivals in a cell slotg
It is particularly important to note that the state probabilities refer to the
state of the queue at moments in time that are usually called the end of
time-slot instants. These instants are after the arrivals (if there are any)
and after the departure (if there is one); indeed they are usually defined
to be at a time t after the end of the slot, where t ! 0.

BALANCE EQUATIONS FOR BUFFERING

The effect of random arrivals on the queue is shown in Figure 7.2. For the
buffer to contain i cells at the end of any time slot it could have contained
any one of 0, 1, . . . , i C 1 at the end of the previous slot. State i can be reached

99

BALANCE EQUATIONS FOR BUFFERING

i+2
i+1

a(0)

i
a(1)
a(i-2)
a(i-1)
a(i)
a(i)

.
.
.
3
2
1
0

Figure 7.2. How to Reach State i at the End of a Time Slot from States at the End of
the Previous Slot

from any of the states 0 up to i by a precise number of arrivals, i down to

1 (with probability ai . . . a1) as expressed in the figure (note that not
all the transitions are shown). To move from i C 1 to i requires that there
are no arrivals, the probability of which is expressed as a0; this then
reflects the completion of service of a cell during the current time slot.
We define the state probability, i.e. the probability of being in state k, as
sk D Prfthere are k cells in the queueing system at the end of any
time slotg
and again (as in Chapter 4) we begin by making the simplifying assumption that the queue has infinite capacity. This means we can find the
system empty probability, s0 from simple traffic theory. We know
from Chapter 3 that
LDAC
where L is the lost traffic, A is the offered traffic and C is the carried
traffic. But if the queue is infinite, then there is no loss (L D 0), so
ADC
This time, though, we are dealing with a stream of cells, not calls. Thus
our offered traffic is numerically equal to , the mean arrival rate of
cells in cell/s (because the cell service time, s, is one time slot), and the
carried traffic is the mean number of cells served per second, i.e. it is the
utilization divided by the service time per cell, so
D


s

100

BASIC CELL SWITCHING

If we now consider the service time of a cell to be one time slot, for
simplicity, then the average number of arrivals per time slot is denoted
E[a] (which is the mean of the arrival distribution ak), and the average
number of cells carried per time slot is the utilization. Thus
E[a] D 
But the utilization is just the steady-state probability that the system is
not empty, so
E[a] D  D 1  s0
and therefore
s0 D 1  E[a]
So from just the arrival rate (without any knowledge of the arrival
distribution ak) we are able to determine the probability that the system
is empty at the end of any time slot. It is worth noting that, if the applied
cell arrival rate is greater than the cell service rate (one cell per time
slot), then
s0 < 0
which is a very silly answer! Obviously then we need to ensure that cells
are not arriving faster (on average) than the system is able to transmit
them. If E[a]
1 cell per time slot, then it is said that the queueing system
is unstable, and the number of cells in the buffer will simply grow in an
unbounded fashion.

CALCULATING THE STATE PROBABILITY DISTRIBUTION

We can build on this value, s0, by going back to the idea of adding all
the ways in which it is possible to end up in any particular state. Starting
with state 0 (the system is empty), this can be reached from a system state
of either 1 or 0, as shown in Figure 7.3. This is saying that the system can
be in state 0 at the end of slot n  1, with no arrivals in slot n, or it can be
in state 1 at the end of slot n  1, with no arrivals in slot n, and at the end
of slot n, the system will be in state 0.
We can write an equation to express this relationship:
s0 D s0 a0 C s1 a0
1

a(0)

a(0)

Figure 7.3. How to Reach State 0 at the End of a Time Slot

101

CALCULATING THE STATE PROBABILITY DISTRIBUTION

You may ask how it can be that sk applies as the state probabilities for
the end of time slot n  1 and time slot n. Well, the answer lies in the fact
that these are steady-state (sometimes called long-run) probabilities,
and, on the assumption that the buffer has been active for a very long
period, the probability distribution for the queue at the end of time slot
n  1 is the same as the probability distribution for the end of time slot n.
Our equation can be rearranged to give a formula for s1:
s1 D s0

1  a0
a0

In a similar way, we can find a formula for s2 by writing a balance

equation for s1:
s1 D s0 a1 C s1 a1 C s2 a0
Again, this is expressing the probability of having 1 in the queueing
system at the end of slot n, in terms of having 0, 1 or 2 in the system
at the end of slot n  1, along with the appropriate number of arrivals
(Figure 7.4). Remember, though, that any arrivals during the current time
slot cannot be served during this slot.
Rearranging the equation gives:
s2 D

s1  s0 a1  s1 a1

a0

We can continue with this process to find a similar expression for the
general state, k.
sk  1 D s0 ak  1 C s1 ak  1 C s2 ak  2 C C sk  1
a1 C sk a0
which, when rearranged, gives:
sk  1  s0 ak  1 

k1

si ak  i

iD1

sk D

a0

2
a(1)
a(1)

1
0

a(0)

Figure 7.4. How to Reach State 1 at the End of a Time Slot

102

BASIC CELL SWITCHING

Queue size
0

10

15

20

25

30

10

State probability

101

Poisson
Binomial

102
103
104
105
106

k 
e
k! 
Binomial k, M, P :D  0 if k > M

M!


1  pMK pk
 M  K! k!
k :D 0.. 30
Poisson k,  :D

if k  M

aPk :D Poisson k, 0.8

aBk :D Binomial k, 8, 0.1

1  Ea
infiniteQX, a, Ea :D  s0


1  a0 
 s1
s0
if X > 0

a0


 for k 2 2.. X
if X > 1




k1




si aki
sk1  s0 ak1 


iD1
 sk

a0

s
xk :D k
y1 :D infiniteQ30, aP, 0.8
y2 :D infiniteQ30, aB, 0.8
Figure 7.5. Graph of the State Probability Distributions for an Infinite Queue with
Binomial and Poisson Input, and the Mathcad Code to Generate (x, y) Values for
Plotting the Graph

103

CALCULATING THE STATE PROBABILITY DISTRIBUTION

Because we have used the simplifying assumption that the queue length
is infinite, we can, theoretically, make k as large as we like. In practice,
how large we can make it will depend upon the value of sk that results
from this calculation, and the program used to implement this algorithm
(depending on the relative precision of the real-number representation
being used).
Now what about results? What does this state distribution look like?
Well, in part this will depend on the actual input distribution, the values
of ak, so we can start by obtaining results for the two input distributions
discussed in Chapter 6: the binomial and the Poisson. Specifically, let us
Buffer capacity, X

10

15

20

25

30

100

Pr{queue size > X}

101

Poisson
Binomial

102
103
104
105
106

QX, s :D  qx0
1  s0

 for i 2 1.. X
if X > 0


 qx
qxi1  si

i

 qx
xk :D k
yP :D infiniteQ30, aP, 0.8
yB :D infiniteQ30, aB, 0.8
y1 :D Q30, yP
y2 :D Q30, yB

Figure 7.6. Graph of the Approximation to the Cell Loss by the Probability that the
Queue State Exceeds X, and the Mathcad Code to Generate (x, y) Values for Plotting
the Graph

104

BASIC CELL SWITCHING

assume an output-buffered switch, and plot the state probabilities for

an infinite queue at one of the output buffers; the arrival rate per input
is 0.1 (i.e. the probability that an input port contains a cell destined for
the output buffer in question is 0.1 for any time slot) and M D 8 input
and output ports. Thus we have a binomial distribution with parameters
M D 8, p D 0.1, compared to a Poisson distribution with mean arrival rate
of M p D 0.8 cells per time slot. Both are shown in Figure 7.5.
What then of cell loss? Well, with an infinite queue we will not actually
have any; in the next section we will deal exactly with the cell loss
probability (CLP) from a finite queue of capacity X. Before we do so, it
is worth considering approximations for the CLP found from the infinite
buffer case. As with Chapter 4, we can use the probability that there are
more than X cells in the infinite buffer as an approximation for the CLP.
In Figure 7.6 we plot this value, for both the binomial and Poisson cases
considered previously, over a range of buffer length values.

EXACT ANALYSIS FOR FINITE OUTPUT BUFFERS

Having considered infinite buffers, we now want to quantify exactly the
effect of a finite buffer, such as we would actually find acting as the output
buffer in a switch. We want to know how the CLP at this queue varies
with the buffer capacity, X, and to do this we need to use the balance
equation technique. However, this time we cannot find s0 directly, by
equating carried traffic and offered traffic, because there will be some lost
traffic, and it is this that we need to find!
So initially we use the same approach as for the infinite queue,
temporarily ignoring the fact that we do not know s0:
s1 D s0

1  a0
a0

sk  1  s0 ak  1 

sk D

k1

si ak  i

iD1

a0

For the system to become full with the arrivals-first buffer management
strategy, there is actually only one way in which this can happen at the end
of time-slot instants: to be full at the end of time slot i, the buffer must begin
slot i empty, and have X or more cells arrive in the slot. If the system is
non-empty at the start, then just before the end of the slot (given enough
arrivals) the system will be full, but when the cell departure occurs at
the slot end, there will be X  1 cells left, and not X. So for the full state,
we have:
sX D s0 AX

105

EXACT ANALYSIS FOR FINITE OUTPUT BUFFERS

where
Ak D 1  a0  a1   ak  1
So Ak is the probability that at least k cells arrive in a slot. Now we face
the problem that, without the value for s0, we cannot evaluate sk for
k > 0. What we do is to define a new variable, uk, as follows:
uk D

sk
s0

so
u0 D 1
Then
u1 D

1  a0
a0
uk  1  ak  1 

k1

ui ak  i

iD1

uk D

a0

uX D AX
and all the values of uk, 0  k  X, can be evaluated! Then using the
fact that all the state probabilities must sum to 1, i.e.
X

si D 1

iD0

we have

X


si
iD0

1
ui
D
D
s0
s0
iD0
X

so
s0 D

1
X

ui

iD0

The other values of sk, for k > 0, can then be found from the definition
of uk:
sk D s0 uk
Now we can apply the basic traffic theory again, using the relationship
between offered, carried and lost traffic at the cell level, i.e.
LDAC

106

BASIC CELL SWITCHING

Queue size
0

10

101

State probability

102
103

Poisson
Binomial

104
105
106
107
108
109
1010

k : D 0.. 10
aPk :D Poisson k, 0.8
aBk :D Binominal k, 8, 0.1

1
finiteQstate (X, a) :D  u0


1  a0 
 u1

a0

 for k 2 2.. X  1
if X > 2




k1




ui ak1
uk1  ak1 


iD1
 uk

a0

X1



u
1
ai if X > 1
 x

iD0


1

 s0
X




ui


iD0

 for k 2 1.. X


 sk
s0 uk


s
xk :D k
y1 :D finiteQstate (10, aP)
y2 :D finiteQstate (10, aB)
Figure 7.7. Graph of the State Probability Distribution for a Finite Queue of 10 Cells
and a Load of 80%, and the Mathcad Code to Generate (x, y) Values for Plotting the
Graph

107

EXACT ANALYSIS FOR FINITE OUTPUT BUFFERS

Buffer capacity, X
0

10

15

20

25

30

101

Cell loss probability

Poisson
Binomial
102
103
104
105
106

k : D 0.. 30
aPk :D Poisson k, 0.8
aBk :D Binominal k, 8, 0.1

finiteQloss (X, a, Ea) :D  u0

 u1


 for






 uk



 ux




 s0






 for

 sk


 CLP


i :D 2.. 30
xi :D i
y1i :D finiteQloss (xi , aP, 0.8)
y2i :D finiteQloss (xi , aB, 0.8)

1
1  a0 
a0
k 2 2.. X  1


if X > 2

k1


ui aki
uk1  ak1 
iD1

1

X1

a0
ai

if X > 1

iD0

1
X

ui

iD0

k 2 1.. X
s0 uk
Ea  1  s0 
Ea

Figure 7.8. Graph of the Exact Cell Loss Probability against System Capacity X for
a Load of 80%

108

BASIC CELL SWITCHING

As before, we consider the service time of a cell to be one time slot, for
simplicity; then the average number of arrivals per time slot is E[a] and
the average number of cells carried per time slot is the utilization. Thus
L D E[a]   D E[a]  1  s0
and the cell loss probability is just the ratio of lost traffic to offered traffic:
CLP D

E[a]  1  s0
E[a]

Figure 7.7 shows the state probability distribution for an output buffer
of capacity 10 cells (which includes the server) being fed from our 8
Bernoulli sources each having p D 0.1 as before. The total load is 80%.
Notice that the probability of the buffer being full is very low in the
Poisson case, and zero in the binomial case. This is because the arrivalsfirst strategy needs 10 cells to arrive at an empty queue in order for the
queue to fill up; the maximum batch size with 8 Bernoulli sources is
8 cells.
Now we can generate the exact cell loss probabilities for finite buffers.
Figure 7.8 plots the exact CLP value for binomial and Poisson input to a
finite queue of system capacity X, where X varies from 2 up to 30 cells.
Now compare this with Figure 7.6.

DELAYS
We looked at waiting times in M/M/1 and M/D/1 queueing systems in
Chapter 4. Waiting time plus service time gives the system time, which is
the overall delay through the queueing system. So, how do we work out
the probabilities associated with particular delays in the output buffers
of an ATM switch? Notice first that the delay experienced by a cell, which
we will call cell C, in a buffer has two components: the delay due to the
unfinished work (cells) in the buffer when cell C arrives, Ud ; and the
delay caused by the other cells in the batch in which C arrives, Bd .
T d D Ud C B d
where Td is the total delay from the arrival of C until the completion of
its transmission (the total system time).
In effect we have already determined Ud ; these values are given by the
state probabilities as follows:
PrfUd D 1g D Ud 1 D s0 C s1
Remember that we assumed that each cell will be delayed by at least 1
time slot, the slot in which it is transmitted. For all k > 1 we have the

109

DELAYS

relationship:
PrfUd D kg D Ud k D sk
The formula for Bd k D PrfBd D kg accounts for the position of C within
the batch as well:
k


1
ai
iD0

Bd k D

E[a]

Note that this equation is covered in more depth in Chapter 13.

Now the total delay, Td k, consists of all the following possibilities:
Td k D PrfUd D 1 and Bd D k  1g C PrfUd D 2 and Bd D k  2g C
and we account for them all by convolving the two components of delay,
using the following formula:
Td k D

k

Ud j Bd k  j

jD1

We plot the cell delay probabilities for the example we have been
considering (binomial and Poisson input processes, p D 0.1 and M D 8,
 D 0.8) in Figure 7.9.
Delay (time slots)

Probability of delay

10

0.1
Poisson
Binomial
0.01

0.001
Figure 7.9.

Cell Delay Probabilities for a Finite Buffer of Size 10 Cells with a Load of 80%

110

BASIC CELL SWITCHING

End-to-end delay
To find the cell delay variation through a number of switches, we convolve
the cell delay distribution for a single buffer with itself. Let
Td,n k D Prftotal delay through n buffers D kg
Then, for two switches the delay distribution is given by
Td,2 k D

k

Td,1 j Td,1 k  j

jD1

There is one very important assumption we are making: that the arrivals
to each buffer are independent of each other. This is definitely not the
case if all the traffic through the first buffer goes through the second
one. In practice, it is likely that only a small proportion will do so; the
bulk of the traffic will be routed elsewhere. This situation is shown in
Figure 7.10.
We can extend our calculation for 2 switches by applying it recursively
to find the delay through n buffers:
Td,n k D

k

Td,n1 j Td,1 k  j

jD1

Figure 7.11 shows the end-to-end delay distributions for 1, 2, 3, 5, 7 and

9 buffers, where the buffers have identical but independent binomial
arrival distributions, each buffer is finite with a size of 10 cells, and
the load offered to each buffer is 80%. Lines are shown as well as
markers on the graph to help identify each distribution; obviously,
the delay can only take integer values. As we found in Chapter 4, the
delay distribution flattens as the number of buffers increases. Note
that this is a delay distribution, which includes one time slot for the
server in each buffer; in Figure 4.8, it is the end-to-end waiting time
Other traffic, routed elsewhere
Buffer 1

Through traffic

Buffer n

Through traffic

Figure 7.10. Independence Assumption for End-to-End Delay Distribution:

Through Traffic is a Small Proportion of Total Traffic Arriving at Each Buffer

111

DELAYS

End to end delay (time slots)

0
1E+00

10

20

30

40

50

n=1

Probability of delay

1E01 n=2
n=3
1E02
n=5
1E03
n=7
1E04
n=9
1E05
Figure 7.11. End-to-End Delay Distributions for 1, 2, 3, 5, 7 and 9 Buffers, with a Load of 80%

distribution which is shown. So, for example, in the distribution for

end-to-end delay through 9 buffers, the smallest delay is 9 time slots
(and the largest delay is 90 time slots, although this is not shown in
Figure 7.11).

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

Cell-Scale Queueing
dealing with the jitters

CELL-SCALE QUEUEING
In Chapter 4 we considered a situation in which a large collection of CBR
voice sources all send their cells to a single buffer. We stated that it was
reasonably accurate under certain circumstances (when the number of
sources is large enough) to model the total cell-arrival process from all
the voice sources as a Poisson process.
Now a Poisson process is a single statistical model from which the
detailed information about the behaviour of the individual sources has
been lost, quite deliberately, in order to achieve simplicity. The process
features a random number (a batch) of arrivals per slot (see Figure 8.1)
where this batch can vary as 0, 1, 2, . . . , 1.
So we could say that in, for example, slot n C 4, the process has
overloaded the queueing system because two cells have arrived one
more than the buffer can transmit. Again, in slot n C 5 the buffer has
been overloaded by three cells in the slot. So the process provides short
periods during which its instantaneous arrival rate is greater than the cell
service rate; indeed, if this did not happen, there would be no need for a
buffer.
But what does this mean for our N CBR sources? Each source is at a
constant rate of 167 cell/s, so the cell rate will never individually exceed
the service rate of the buffer; and provided N 167 < 353 208 cell/s, the
total cell rate will not do so either. The maximum number of sources
is 353 208/167 D 2115 or, put another way, each source produces one
cell every 2115 time slots. However, the sources are not necessarily
arranged such that a cell from each one arrives in its own time slot;
indeed, although the probability is not high, all the sources could be
(accidentally) synchronized such that all the cells arrive in the same slot.
In fact, for our example of multiplexing 2115 CBR sources, it is possible

114

CELL-SCALE QUEUEING

Number of arrivals in a slot

5
4
3
2
1
0
n

n+1

Figure 8.1.

n+2

n+3

n+4 n+5 n+6

Time slot number

n+7

n+8 n+9

n+10

A Random Number of Arrivals per Time Slot

for any number of cells varying from 0 up to 2115 to arrive in the same
slot. The queueing behaviour which arises from this is called cell-scale
queueing.

MULTIPLEXING CONSTANT-BIT-RATE TRAFFIC

Let us now take a closer look at what happens when we have constant-bitrate traffic multiplexed together. Figure 8.2 shows, for a simple situation,
how repeating patterns develop in the arrival process patterns which
depend on the relative phases of the sources.

Queue
size

(a) All streams out of phase

Figure 8.2. Repeating Patterns in the Size of the Queue when Constant-Bit-Rate
Traffic Is Multiplexed

ANALYSIS OF AN INFINITE QUEUE WITH MULTIPLEXED CBR INPUT: THE ND/D/1

115

Queue
size

(b) Two streams in phase

Queue
size

(c) All streams in phase

Figure 8.2.

(continued)

It is clear from this picture that there are going to be circumstances

where a simple classical queueing system like the M/D/1 will not
adequately model superposed CBR traffic; in particular, the arrival
process is not well modelled by a Poisson process when the number
of sources is small. At this point we need a fresh start with a new
approach to the analysis.

ANALYSIS OF AN INFINITE QUEUE WITH MULTIPLEXED CBR

INPUT: THE ND/D/1
The ND/D/1 queue is a basic model for CBR traffic where the input
process comprises N independent periodic sources, each source with
the same period D. If we take our collection of 1000 CBR sources, then
N D 1000, and D D 2115 time slots. The queueing analysis caters for
all possible repeating patterns and their effect on the queue size. The
buffer capacity is assumed to be infinite, and the cell loss probability is
approximated by the probability that the queue exceeds a certain size x,

116

CELL-SCALE QUEUEING

i.e. Qx. Details of the derivation can be found in [8.1].



 


N


N!
nx n
n  x Nn
CLP Qx D
1

n! N  n!
D
D
nDxC1

DNCx

DnCx
Lets put some numbers in, and see how the cell loss varies with different
parameters and their values. The distribution of Qx for a fixed load of
Buffer capacity
0

10

20

30

40

100
101
102
103

Q(x)

104
105
106
107
108
109

N = 1000
N = 500
N = 200
N = 50

1010



N
NDD1Q x, N,  :D  D

 




 N
nx n
n  x Nn D  N C x

combin
(N,
n)

1



D
D
DnCx
nDxC1
xk :D k
y1k :D NDD1Q k, 1000, 0.95
y2k :D NDD1Q k, 500, 0.95
y3k :D NDD1Q k, 200, 0.95
y4k :D NDD1Q k, 50, 0.95
k :D 0.. 40

Figure 8.3. Results for the ND/D/1 Queue with a Load of 95%, and the Mathcad
Code to Generate (x, y) Values for Plotting the Graph

117

HEAVY-TRAFFIC APPROXIMATION FOR THE M/D/1 QUEUE

 D N/D D 0.95 with numbers of sources ranging from 50 up to 1000 is

given in Figure 8.3. Note how the number of inputs (sources) has such
a significant impact on the results. Remember that the traffic is periodic,
and the utilization is less than 1, so the maximum number of arrivals
in any one period of the constant-bit-rate sources (as well as in any one
time slot) is limited to one from each source, i.e. N. The value of N limits
the maximum size of the queue if we provide N waiting spaces there
would be no loss at all.
The ND/D/1 result can be simplified when the applied traffic is close
to the service rate; this is called a heavy traffic theorem. But lets first
look at a useful heavy traffic result for a queueing system we already
know the M/D/1.

HEAVY-TRAFFIC APPROXIMATION FOR THE M/D/1 QUEUE

An approximate analysis of the M/D/1 system produces the following
equation:

Qx D e

2x

1


Details of the derivation can be found in [8.2]. The result amounts to

approximating the queue length by an exponential distribution: Qx is
the probability that the queue size exceeds x, and  is the utilization. At
first sight, this does not seem to be reasonable; the number in the queue
is always an integer, whereas the exponential distribution applies to a
continuous variable x; and although x can vary from zero up to infinity,
we are using it to represent a finite buffer size. However, it does work:
Qx is a good approximation for the cell loss probability for a finite
buffer of size x. In later chapters we will develop equations for Qx for
discrete distributions.
For this equation to be accurate, the utilization must be high. Figure 8.4
shows how it compares with our exact analysis from Chapter 7, with
Poisson input traffic at different values of load. The approximate results
are shown as lines through the origin. It is apparent that although the
cell loss approximation safely overestimates at high utilization, it can
significantly underestimate when the utilization is low. But in spite of
this weakness, the major contribution that this analysis makes is to show
that there is a loglinear relationship between cell loss probability and
buffer capacity.
Why is this heavy-traffic approximation so useful? We can rearrange
the equation to specify any one variable in terms of the other two.
Recalling the conceptual framework of the trafficcapacityperformance
model from Chapter 3, we can see that the traffic is represented by 
(the utilization), the capacity is x (the buffer size), and the performance

118

CELL-SCALE QUEUEING
Buffer capacity
0

10

20

30

10 0
101
102
103

CLP

104
105
106
107

= 0.95

108

= 0.75

109

= 0.55

1010

k :D 0.. 30
ap95k :D Poisson k, 0.95
ap75k :D Poisson k, 0.75
ap55k :D Poisson k, 0.55
2x

1


MD1Qheavy x,  :D e
i :D 2.. 30
xk :D k
Y1i :D finiteQloss xi , ap95, 0.95
Y2k :D MD1Qheavy xk , 0.95
Y3i :D finiteQloss xi , ap75, 0.75
Y4k :D MD1Qheavy xk , 0.75
Y5i :D finiteQloss xi , ap55, 0.55
Y6k :D MD1Qheavy xk , 0.55
Figure 8.4. Comparing the Heavy-Traffic Results for the M/D/1 with Exact Analysis
of the M/D/1/K, and the Mathcad Code to Generate (x, y) Values for Plotting the
Graph

is Qx (the approximation to the cell loss probability). Taking natural

logarithms of both sides of the equation gives
lnQx D 2x

1  


This can be rearranged to give



1

x D  lnQx
2
1

119

HEAVY-TRAFFIC APPROXIMATION FOR THE ND/D/1 QUEUE

and
D

2x
2x  lnQx

We will not investigate how to use these equations just yet. The first
relates to buffer dimensioning, and the second to admission control, and
both these topics are dealt with in later chapters.

HEAVY-TRAFFIC APPROXIMATION FOR THE ND/D/1 QUEUE

Although the exact solution for the ND/D/1 queue is relatively straightforward, the following heavy-traffic approximation for the ND/D/1
Buffer capacity
10 0

10

20

30

40

101
102

Q(x)

103
104
105
106
107
108
109

N = 1000
N = 500
N = 200
N = 50

1010

k :D 0.. 40
2x

x 1
C
N


NDD1Qheavy x, N,  :D e
xk :D k
y1k :D NDD1Q k, 1000, 0.95
y2k :D NDD1Q k, 500, 0.95
y3k :D NDD1Q k, 200, 0.95
y4k :D NDD1Q k, 50, 0.95
y5k :D NDD1Qheavy k, 1000, 0.95
y6k :D NDD1Qheavy k, 500, 0.95
y7k :D NDD1Qheavy k, 200, 0.95
y8k :D NDD1Qheavy k, 50, 0.95

Figure 8.5. Comparison of Exact and Approximate Results for ND/D/1 at a Load
of 95%, and the Mathcad Code to Generate (x, y) Values for Plotting the Graph

120

CELL-SCALE QUEUEING

[8.2] helps to identify explicitly the effect of the parameters:

Qx D e

x 1
2x N C 

Figure 8.5 shows how the approximation compares with exact results
from the ND/D/1 analysis for a load of 95%. The approximate results
are shown as lines, and the exact results as markers. In this case the
approximation is in very good agreement. Figure 8.6 shows how the
Buffer capacity
0

10

20

30

40

100
= 0.95

101

= 0.95

102

= 0.95

103

CLP

104
105
106
107
108
109
1010

k :D 0.. 40
xk :D k
y1k :D NDD1Q k, 200, 0.95
y2k :D NDD1Qheavy k, 200, 0.95
y3k :D NDD1Q k, 200, 0.75
y4k :D NDD1Qheavy k, 200, 0.75
y5k :D NDD1Q k, 200, 0.55
y6k :D NDD1Qheavy k, 200, 0.55
Figure 8.6. Comparison of Exact and Approximate Results for ND/D/1 for a
variety of Loads, with N D 200, and the Mathcad Code to Generate (x, y) Values for
Plotting the Graph

CELL-SCALE QUEUEING IN SWITCHES

121

approximation compares for three different loads. For low utilizations,

the approximate method underestimates the cell loss.
Note that the form of the equation is similar to the approximation for
the M/D/1 queue, with the addition of a quadratic term in x, the queue
size. So, for small values of x, ND/D/1 queues behave in a manner
similar to M/D/1 queues with the same utilization. But for larger values
of x the quadratic term dominates; this reduces the probability of larger
queues occurring in the ND/D/1, compared to the same size queue in
the M/D/1 system. Thus we can see how the Poisson process is a useful
approximation for N CBR sources, particularly for large N: as N ! 1,
the quadratic term disappears and the heavy traffic approximation to the
ND/D/1 becomes the same as that for the M/D/1. In Chapter 14 we
revisit the M/D/1 to develop a more accurate formula for the overflow
probability that both complements and extends the analysis presented in
this chapter (see also [8.3]).

CELL-SCALE QUEUEING IN SWITCHES

It is important not to assume that cell-scale queueing arises only as a
result of source multiplexing. If we now take a look at switching, we will
find that the same effect arises. Consider the simple output buffered 2 2
switching element shown in Figure 8.7.
Here we can see a situation analogous to that of multiplexing the
CBR sources. Both of the input ports into the switch carry cells coming
from any number of previously multiplexed sources. Figure 8.8 shows a
typical scenario; the cell streams on the input to the switching element
are the output of another buffer, closer to the sources. The same queueing
principle applies at the switch output buffer as at the source multiplexor:
the sources may all be CBR, and the individual input ports to the switch
may contain cells such that their aggregate arrival rate is less than the

Figure 8.7. An Output Buffered 2 2 Switching Element

122

CELL-SCALE QUEUEING

Source
1

Source
multiplexer

Source
2
.
.
.
Source
N
Source
1

Source
multiplexer

Source
2
.
.
.

2 2 ATM
switching
element

Source
N

Figure 8.8.

Cell-Scale Queueing in Switch Output Buffers

Buffer capacity

10

15

20

1E+00

Cell loss probability

1E01
1E02
1E03
1E04
1E05
1E06
1E07
1E08
1E09
1E10

Figure 8.9. Cell Loss at the Switch Output Buffer

output rate of either of the switch output ports, but still there can be cell
loss in the switch. Figure 8.9 shows an example of the cell loss probabilities
for either of the output buffers in the switch for the scenario illustrated in
Figure 8.8. This assumes that the output from each source multiplexor is
a Bernoulli process, with parameter p0 D 0.5, and that the cells are routed

CELL-SCALE QUEUEING IN SWITCHES

123

in equal proportions to the output buffers of the switching element. Thus

the cell-scale queueing in each of the output buffers can be modelled with
binomial input, where M D 2 and p D 0.25.
So, even if the whole of the ATM network is dedicated to carrying only
CBR traffic, there is a need for buffers in the switches to cope with the
cell-scale queueing behaviour. This is inherent to ATM; it applies even if
the network allocates the peak rate to variable-bit-rate sources. Buffering
is required, because multiple streams of cells are multiplexed together. It
is worth noting, however, that the cell-scale queueing effect (measured
by the CLP against the buffer capacity) falls away very rapidly with
increasing buffer length so we only need short buffers to cope with it,
and to provide a cell loss performance in accord with traffic requirements.
This is not the case with the burst-scale queueing behaviour, as we will
see in Chapter 9.

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

Burst-Scale Queueing
information overload!

ATM QUEUEING BEHAVIOUR

We have seen in the previous chapter that queueing occurs with CBR
traffic when two or more cells arrive during a time slot. If a particular
source is CBR, we know that the next cell from it is going to arrive after
a fixed duration given by the period, D, of the source, and this gives the
ATM buffer some time to recover from multiple arrivals in any time slot
when a number of sources are multiplexed together (hence the result that
Poisson arrivals are a worst-case model for cell-scale queueing).
Consider the arrivals from all the CBR sources as a rate of flow of cells.
Over the time interval of a single slot, the input rate varies in integer
multiples of the cell slot rate (353 208 cell/s) according to the number of
arrivals in the slot. But that input rate is very likely to change to a different
value at the next cell slot; and the value will often be zero. It makes more
sense to define the input rate in terms of the cycle time, D, of the CBR
sources, i.e. 353 208/D cell/s. For the buffer to be able to recover from
multiple arrivals in a slot, the number of CBR sources, N, must be less
than the inter-arrival time D, so the total input rate 353 208 N/D cell/s
is less than the cell slot rate.
Cell-scale queueing analysis quantifies the effect of having simultaneous arrivals according to the relative phasing of the CBR streams, so
we define simultaneity as being within the period of one cell slot.
Lets relax our definition of simultaneity, so that the time duration
is a number of cell slots, somewhat larger than one. We will also alter
our definition of an arrival from a single source; no longer is it a
single cell, but a burst of cells during the defined period. Queueing
occurs when the total number of cells arriving from simultaneous (or
overlapping) bursts exceeds the number of cell slots in that simultaneous
period.

126

BURST-SCALE QUEUEING

But how do we define the length of the simultaneous period? Well, we

dont: we define the source traffic using cell rates, and assume that these
rates are on for long enough such that each source contributes rather
more than one cell. Originally we considered CBR source traffic, whose
behaviour was characterized by a fixed-length inactive state followed by
the arrival of a single cell. For variable bit-rate (VBR), we redefine this
behaviour as a long inactive state followed by an active state producing
a burst of cells (where burst is defined as a cell arrival rate over a period
of time). The state-based sources in Chapter 6 are examples of models for
VBR traffic.
With these definitions the condition for queueing is that the total input
rate of simultaneous bursts must exceed the cell slot rate of the ATM buffer.
This is called burst-scale queueing. For the N CBR sources there is no
burst-scale queueing because the total input rate of the simultaneous and
continuous bursts of rate 353 208/D cell/s is less than the cell slot rate.
Lets take a specific example, as shown in Figure 9.1. Here we have two
VCs with fixed rates of 50% and 25% of the cell slot rate. In the first 12
time slots, the cells of the 25% VC do not coincide with those of the 50%
VC and every cell can enter service immediately (for simplicity, we show
this as happening in the same slot). In the second set of 12 time slots, the
cells of the 25% VC do arrive at the same time as some of those in the 50%
VC, and so some cells have to wait before being served. This is cell-scale
queueing; the number of cells waiting is shown in the graph.
Now, lets add in a third VC with a rate of 33% of the cell slot rate
(Figure 9.2). The total rate exceeds the queue service rate and over a
period of time the number of cells waiting builds up: in this case there are
two more arrivals than available service slots over the period shown in
50%
25%
75%
2

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Figure 9.1.

Cell Scale Queueing Behaviour

BURST-SCALE QUEUEING BEHAVIOUR

127

50%
25%
33%
108%
3

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Figure 9.2. Burst-Scale and Cell-Scale Queueing Behaviour

the diagram. This longer-term queueing is the burst-scale queueing and

is shown as a solid line in the graph. There is still the short-term cell-scale
queueing, represented by the fluctuations in the number in the queue.
ATM queueing comprises both types of behaviour.

BURST-SCALE QUEUEING BEHAVIOUR

The previous example showed that an input rate exceeding the service
capacity by 8%, i.e. by 0.08 cells per time slot, would build up over a
period of 24 time slots to a queue size of 0.08 24 2 cells. During this
period (of about 68 s) there were 26 arriving cells, but would be only
24 time slots in which to serve them: i.e. an excess of 2 cells. These two
cells are called excess-rate cells because they arise from excess-rate
bursts. Typical bursts can last for durations of milliseconds, rather than
microseconds. So, in our example, if the excess rate lasts for 2400 time
slots (6.8 ms) then there would be about 200 excess-rate cells that must
be held in a buffer, or lost.
We can now distinguish between buffer storage requirements for cellscale queueing (of the order of tens of cells) and for burst-scale queueing
(of the order of hundreds of cells). Of course, there is only one buffer,
through which all the cells must pass: what we are doing is identifying
the two components of demand for temporary storage space. Burstscale queueing analyses the demand for the temporary storage of these
excess-rate cells.

128

BURST-SCALE QUEUEING

We can identify two parts to this excess-rate demand, and analyse the
parts separately. First, what is the probability that an arriving cell is an
excess-rate cell? This is the same as saying that the cell needs burst-scale
buffer storage. Then, secondly, what is the probability that such a cell is
lost, i.e. the probability that a cell is lost, given that it is an excess-rate
cell? We can then calculate the overall cell loss probability arising from
burst-scale queueing as:
Prfcell is lostg Prfcell is lostjcell needs bufferg Prfcell needs bufferg
The probability that a cell needs the buffer is called the burst-scale loss
factor; this is found by considering how the input rate compares with the
service rate of the queue. A cell needs to be stored in the buffer if the
total input rate exceeds the queues service rate. If there is no burst-scale
buffer storage, these cells are lost, and
Prfcell is lostg Prfcell needs bufferg
The probability that a cell is lost given that it needs the buffer is called
the burst-scale delay factor; this is the probability that an excess-rate cell
is lost. If the burst-scale buffer size is 0, then this probability is 1, i.e. all
excess-rate cells are lost. However, if there is some buffer storage, then
only some of the excess-rate cells will be lost (when this buffer storage
is full).
Figure 9.3 shows how these two factors combine on a graph of cell
loss probability against the buffer capacity. The burst-scale delay factor is
shown as a straight line with the cell loss decreasing as the buffer capacity
increases. The burst-scale loss factor is the intersection of the straight line
with the zero buffer axis.
Buffer capacity
0

10

20

30

1
0.1
CLP

Burst scale loss

0.01
Burst scale delay
0.001
0.0001
0.00001
0.000001
Figure 9.3.

The Two Factors of Burst-Scale Queueing Behaviour

40

129

CONTINUOUS FLUID-FLOW APPROACH

Cell rate

R
Excess cell rate
C
Service for
excess-rate
cells

Time
Figure 9.4.

Burst-Scale Queueing with a Single ON/OFF Source

FLUID-FLOW ANALYSIS OF A SINGLE SOURCE PER-VC

QUEUEING
The simplest of all burst-scale models is the single ON/OFF source
feeding an ATM buffer. When the source is ON, it produces cells at
a rate, R, overloading the service capacity, C, and causing burst-scale
queueing; when OFF, the source sends no cells, and the buffer can
recover from this queueing by serving excess-rate cells (Figure 9.4). In
this very simple case, there is no cell-scale queueing because only one
source is present. This situation is essentially that of per-VC queueing: an
output port is divided into multiple virtual buffers, each being allocated
a share of the service capacity and buffer space available at the output
port. Thus, in the following analysis, C can be thought of as the share
of service capacity allocated to a virtual buffer for this particular VC
connection. We revisit this in Chapter 16 when we consider per-flow
queueing in IP.
There are two main approaches to this analysis. The historical approach
is to model the flow of cells into the buffer as though it were a continuous
fluid; this ignores the structure of the flow (e.g. bits, octets, or cells). The
alternative is the discrete approach, which actually models the individual
excess-rate cells.

CONTINUOUS FLUID-FLOW APPROACH

The source model for this approach was summarized in Chapter 6; the
state durations are assumed to be exponentially distributed. A diagram
of the system is shown in Figure 9.5. Analysis requires the use of partial
differential equations and the derivation is rather too complex in detail
to merit inclusion here (see [9.1] for details). However, the equation for
the excess-rate loss probability is

130

BURST-SCALE QUEUEING

Ton

X
Toff
Figure 9.5.
Analysis

Source Model and Buffer Diagram for the Continuous Fluid-Flow

CLPexcess-rate D

C  R e

XCR
Ton 1RCC

1   C  R  C e

XCR
Ton 1RCC

where
R = ON rate
C = service rate of queue
X = buffer capacity of queue
Ton = mean duration in ON state
Toff = mean duration in OFF state
and
=

Ton
D probability that the source is active
Ton C Toff

Note that CLPexcess-rate is the probability that a cell is lost given that it is an
excess-rate cell. The probability that a cell is an excess-rate cell is simply
the proportion of excess-rate cells to all arriving cells, i.e. R  C/R. Thus
the overall cell loss probability is
CLP D

RC
CLPexcess-rate
R

Another way of looking at this is to consider the number of cells lost in a

time period, T.
R T CLP D R  C T CLPexcess-rate
The mean number of cells arriving in one ON/OFF cycle is R Ton , so
the mean arrival rate is simply R . The mean number of cells arriving
during the time period is R T. Thus the number of cells actually
lost (on average) during time period T is given by the left-hand side
of the equation. But cells are only lost when the source is in the ON

DISCRETE FLUID-FLOW APPROACH

131

state, i.e. when there are excess-rate arrivals. Thus the mean number of
excess-rate arrivals in one ON/OFF cycle is R  C Ton , and so the mean
excess rate is simply R  C . The number of excess-rate cells arriving
during the time period is R  C T, and so the number of excess-rate
cells actually lost during a time period T is given by the right-hand
side of the equation. There is no other way of losing cells, so the two
sides of the equation are indeed equal, and the result for CLP follows
directly.
We will take an example and put numbers into the formula later on,
when we can compare with the results for the discrete approach.

DISCRETE FLUID-FLOW APPROACH

This form of analysis sees each of the excess-rate arrivals [9.2]. The
derivation is simpler than that for the continuous case, and the approach
to deriving the balance equations is a useful alternative to that described
in Chapter 7. Instead of finding the state probabilities at the end of a time
slot, we find the probability that an arriving excess-rate cell finds k cells
in the buffer. If an arriving excess-rate cell finds the buffer full, it is lost,
and so CLPexcess-rate is simply the probability of this event occurring.
We start with the same system model and parameters as for the
continuous case, shown in Figure 9.5. The system operation is as follows:
IF the source is in the OFF state AND
a)
b)

the buffer is empty, THEN it remains empty

the buffer is not empty, THEN it empties at a constant rate C
IF the source is in the ON state AND

a)
b)

the buffer is not full, THEN it fills at a constant rate R  C

the buffer is full, THEN cells are lost at a constant rate R  C

As was discussed in Chapter 6, in the sources OFF state no cells are

generated, and the OFF period lasts for a geometrically distributed
number of time slots. In the ON state, cells are generated at a rate of R.
But for this analysis we are only interested in the excess-rate arrivals, so
in the ON state we say that excess-rate cells are generated at a rate of
R  C and the ON period lasts for a geometrically distributed number of
excess-rate arrivals. In each state there is a Bernoulli process: in the OFF
state, the probability of being silent for another time slot is s; in the ON
state, the probability of generating another excess-rate arrival is a. The
model is shown in Figure 9.6.
Once the source has entered the OFF state, it remains there for at least
one time slot; after each time slot in the OFF state the source remains in the
OFF state with probability s, or enters the ON state with probability 1  s.

132

BURST-SCALE QUEUEING

Pr{no} = 1a
Silent for
another
time slot?

Generate
another excessrate arrival?

Pr{yes} = s

Figure 9.6.

Pr{yes} = a

Pr{no} = 1s

The ON/OFF Source Model for the Discrete Fluid-Flow Approach

On entry into the ON state, the model generates an excess-rate arrival;

after each arrival the source remains in the ON state and generates another
arrival with probability a, or enters the OFF state with probability 1  a.
This process of arrivals and time slots is shown in Figure 9.7.
Now we need to find a and s in terms of the system parameters, R, C,
Ton and Toff . From the geometric process we know that the mean number
of excess-rate cells in an ON period is given by
E[on] D

1
1a

But this is simply the mean duration in the ON state multiplied by the
excess rate, so
1
D Ton R  C
E[on] D
1a
giving
aD1

1
Ton R  C

In a similar manner, the mean number of empty time slots in the OFF
state is
1
D Toff C
E[off] D
1s
Cell rate
a

R
1a

1-a

C
1s

s s s s

s
Time

Figure 9.7.

The Process of Arrivals and Time Slots for the ON/OFF Source Model

133

DISCRETE FLUID-FLOW APPROACH

giving
sD1

1
Toff C

In Chapter 7, we developed balance equations that related the state of

the buffer at the end of the current time slot with its state at the end
of the previous time slot. This required knowledge of all the possible
previous states and how the presence or absence of arrivals could achieve
a transition to the current state. For this discrete fluid-flow approach, we
use a slightly different form of balance equation, developed according to
the so-called line crossing method.
Consider the contents of a queue varying over time, as shown in
Figure 9.8. If we draw a line between states of the queue (in the
figure we have drawn one between state hthere are 3 in the queuei and
state hthere are 4 in the queuei) then for every up-crossing through
this line, there will also be a down-crossing (otherwise the queue
contents would increase for ever). Since we know that a probability
value can be represented as a proportion, we can equate the proportion of transitions that cause the queue to cross up through the line
(probability of crossing up) with the proportion of transitions that cause
it to cross down through the line (probability of crossing down). This
will work for a line drawn through any adjacent pair of states of the
queue.
We define the state probability as
pk D Prfan arriving excess-rate cell finds k cells in the bufferg
Number in the queue
5
Pr{crossing up}
4
3
Pr{crossing down}
2
1
0

Time
Figure 9.8.

The Line Crossing Method

134

BURST-SCALE QUEUEING

An excess-rate cell which arrives to find X cells in the buffer, where X is

the buffer capacity, is lost, so
CLPexcess-rate D pX
The analysis begins by considering the line between states X and X  1.
This is shown in Figure 9.9.
Since we are concerned with the state that an arriving excess-rate cell
sees, we must consider arrivals one at a time. Thus the state can only ever
increase by one. This happens when an arriving excess-rate cell sees X  1
in the queue, taking the queue state up to X, and another excess-rate cell
follows immediately (without any intervening empty time slots) to see
the queue in state X. So, the probability of going up is
Prfgoing upg D a pX  1
To go down, an arriving excess-rate cell sees X in the queue and is lost
(because the queue is full), and then there is a gap of at least one empty
time slot, so that the next arrival sees fewer than X in the queue. (If there
is no gap, then the queue will remain full and the next arrival will see X
as well.) So, the probability of going down is
Prfgoing downg D 1  a pX
Equating the probabilities of going up and down, and rearranging gives
pX  1 D

1a
pX
a

We can do the same for a line between states X  1 and X  2. Equating

probabilities gives
a pX  2 D 1  a s pX C 1  a s pX  1
The left-hand side is the probability of going up, and is essentially the
same as before. The probability of going down, on the right-hand side of
the equation, contains two possibilities. The first term is for an arriving
excess-rate cell which sees X in the queue and is lost (because the queue
a.p(X1)
Figure 9.9.
X1

X
X1

(1-a).p(X )

Equating Up- and Down-Crossing Probabilities between States X and

135

DISCRETE FLUID-FLOW APPROACH

is full), and then there is a gap of at least two empty time slots, so that
the next arrival sees fewer than X  1 in the queue. The second term is
for an arriving excess-rate cell which sees X  1 in the queue, taking the
state of the queue up to X, and then there is a gap of at least two empty
time slots, so that the next arrival sees fewer than X  1 in the queue.
Rearranging, and substituting for pX, gives
pX  2 D

s
pX  1
a

In the general case, for a line between X  i C 1 and X  i, the probability

of going up remains the same as before, i.e. the only way to go up is for
an arrival to see X  i, and to be followed immediately by another arrival
which sees X  i C 1. The probability of going down consists of many
components, one for each state above X  i, but they can be arranged in
two groups: the probability of coming down from X  i C 1 itself; and
the probability of coming down to below X  i C 1 from above X  i C 1.
This latter is just the probability of going down between X  i C 2 and
X  i C 1 multiplied by s, which is the same as going up from X  i C 1
multiplied by s. This is illustrated in Figure 9.10.
The general equation then is
pX  i D

s
pX  i C 1
a

The state probabilities form a geometric progression, which can be

expressed in terms of pX, a and s, for i > 0:
 i

pX  i D

s
a

1a
pX
s

The probabilities must sum to 1, so

X


pX  i D pX C

iD0

a.p(Xi+1)

iD1

pX D 1

Xi+2
Xi+1

a.p(Xi)

X  i

1a
s

Xi

(1a) .p(Xi+1) + a.p(X-i+1)

Multiply by s

Figure 9.10. Equating Up- and Down-Crossing Probabilities in the General Case

136

BURST-SCALE QUEUEING

which can be rearranged to give the probability that an excess-rate arrival

sees a full queue, i.e. the excess-rate cell loss probability
pX D
1C

X 

iD1

1
s
a

i

1a
s

This can be rewritten as

 
X

pX D
1C

s
a

 

1

1a
sa

which is valid except when a D s (in which case the previous formula
must be used). As in the case of the continuous fluid-flow analysis, the
overall cell loss probability is given by
CLP D

RC
RC
CLPexcess-rate D
pX
R
R

COMPARING THE DISCRETE AND CONTINUOUS FLUID-FLOW

APPROACHES
Lets use the practical example of silence-suppressed telephony, with the
following parameter values:
R D 167 cell/s
Ton D 0.96 seconds
Toff D 1.69 seconds
Thus
D

0.96
D 0.362
0.96 C 1.69

and the mean arrival rate

 D R D 60.5 cell/s
In order to have burst-scale queueing, the service capacity, C, must be less
than the cell rate in the active state, R. Obviously this does not correspond
to a normal ATM buffer operating at 353 208 cell/s (although it does if we
are considering a virtual buffer with per-VC queueing). We will see, also,
that one application of this analysis is in connection admission control
for estimating a bandwidth value, C, to allocate to a source in order to
meet a cell loss probability requirement. Figure 9.11 shows the overall

137

COMPARING THE DISCRETE AND CONTINUOUS FLUID-FLOW APPROACHES

Buffer capacity
0

10

20

30

40

100

Cell loss probability

101

102

103

104

C = 80
C = 120
C = 150
C = 160

105
k :D 1..

40


CLPcontFF C, , R, X, Ton :D

RC
R

C  R e

XCR
Ton1RCC

1   C  R  C e

1

CLPdiscFFC, , R, X, Ton :D
Toff
Ton

a
1

Ton R  C

1

Toff C

RC




R
s X
1a

1
C

1

a
sa
Xk :D k
y1k :D CLPdiscFF 80, 0.362, 167, k, 0.96
y2k :D CLPdiscFF 120, 0.362, 167, k, 0.96
y3k :D CLPdiscFF 150, 0.362, 167, k, 0.96
y4k :D CLPdiscFF 160, 0.362, 167, k, 0.96
y5k :D CLPcontFF 80, 0.362, 167, k, 0.96
y6k :D CLPcontFF 120, 0.362, 167, k, 0.96
y7k :D CLPcontFF 150, 0.362, 167, k, 0.96
y8k :D CLPcontFF 160, 0.362, 167, k, 0.96
Figure 9.11.

XCR
Ton1RCC

Cell Loss Probability against Buffer Capacity for a Single ON/OFF Source

138

BURST-SCALE QUEUEING

Table 9.1. Average Number of Excess-Rate

Cells in an Active Period

Capacity
(cell/s)

Average number of excess-rate

cells per active state

80
120
150
160

83.52
45.12
16.32
6.72

Buffer capacity
0

10

20

30

40

100

Cell loss probability

101

102

103
* 10
*1
* 0.1

104

105

k :D 1..

40

xk :D k
y1k :D CLPdiscFF 150, 0.362, 167, k, 9.6
y2k :D CLPdiscFF 150, 0.362, 167, k, 0.96
y3k :D CLPdiscFF 150, 0.362, 167, k, 0.096
y4k :D CLPcontFF 150, 0.362, 167, k, 9.6
y5k :D CLPcontFF 150, 0.362, 167, k, 0.96
y6k :D CLPcontFF 150, 0.362, 167, k, 0.096
Figure 9.12. The Effect of Scaling the Mean State Durations, Ton and Toff , when
C D 150 cell/s

139

MULTIPLE ON/OFF SOURCES OF THE SAME TYPE

cell loss probability plotted against the buffer capacity, X, as the service
capacity is varied between the mean and peak cell rates of the source.
The discrete fluid-flow results are shown as markers, and the continuous
as solid lines.
As the service capacity gets closer to the ON rate, R, the gradient
steepens. This means that the buffer is better able to cope with the bursts
of excess-rate cells. We can see more clearly why this is so by looking
at Table 9.1, which shows the average number of excess-rate cells in an
active period. When this number is large relative to the capacity of the
buffer, then the buffer does not cope very well because it only takes
a fraction of an average burst to fill it up. It would not make much
difference if there was no buffer spacethere is so little difference to the
cell loss over the range of buffer capacity shown. The buffer only makes
a difference to the cell loss if the average excess-rate burst length is less
than the buffer capacity, i.e. when it would take a number of bursts
to fill the buffer. Notice that it is only in these circumstances that the
discrete and continuous fluid-flow results show any difference; and then
the discrete approach is more accurate because it does not include the
fractions of cells allowed by the continuous fluid-flow analysis. These
small amounts actually represent quite a large proportion of an excessrate burst when the average number of excess-rate cells in the burst is
small.
Figure 9.12 illustrates the strong influence of the average state durations
on the results for cell loss probability. Here, C D 150 cell/s, with other
parameter values as before, and the Ton and Toff values have been scaled
by 0.1, 1 and 10. In each case the load on the buffer remains constant at a
value of
167
R
D 0.403
D 0.362
C
150

MULTIPLE ON/OFF SOURCES OF THE SAME TYPE

Lets now consider burst-scale queueing when there are multiple ON/OFF
sources being fed through an ATM buffer. Figure 9.13 shows a diagram
of the system, with the relevant source and buffer parameters. There are
N identical sources, each operating independently, sending cells into an
ATM buffer of service capacity C cell/s and finite size X cells. The average
ON and OFF durations are denoted Ton and Toff , as before; the cell rate in
the active state is h cell/s, so the mean cell rate for each source is
mDh

Ton
Ton C Toff

140

BURST-SCALE QUEUEING

Ton

Ton

Toff

Source 1

Ton

Toff

X
Mean rate of source
m = hTon/(Ton + Toff)

Source 2
Toff
Source N
Figure 9.13.

Multiple ON/OFF Sources Feeding an ATM Buffer

and the probability that the source is active is

D

Ton
m
D
h
Ton C Toff

which is also called the activity factor.

The condition for burst-scale queueing is that the total input rate from
active sources must be greater than the service rate of the buffer. An
important parameter, then, is how many times the peak rate, h, fits into
the service capacity, C, denoted by N0 :
N0 D

C
h

This may well not be an integer value. If we round the value up, to
dN0 e (this notation means take the first integer above N0 ), this gives the
minimum number of sources required for burst-scale queueing to take
place. If we round the value down, to bN0 c (this notation means take the
first integer below N0 ), this gives the maximum number of sources we
can have in the system without having burst-scale queueing.
We saw earlier in the chapter that the burst-scale queueing behaviour
can be separated into two components: the burst-scale loss factor, which
is the probability that a cell is an excess-rate cell; and the burst-scale delay
factor, which is the probability that a cell is lost given that it is an excessrate cell. Both factors contribute to quantifying the cell loss: the burst-scale
loss factor gives the cell loss probability if we assume there is no buffer;

141

THE BUFFERLESS APPROACH

this value is multiplied by the burst-scale delay factor to give the cell loss
probability if we assume there is a buffer of some finite capacity.

THE BUFFERLESS APPROACH

For multiple ON/OFF sources, we start by assuming there is no buffer
and calculate the burst-scale loss factor. For the single source, this is
simply the proportion of cells that are excess-rate cells, i.e. R  C/R, or
with the new parameters, h  C/h. Another way of looking at this is the
mean excess rate divided by the mean arrival rate:
Prfcell needs bufferg D

h  C
h  C
hC
D
D
m
h
h

The probability of an excess rate of h  C is the same as the probability

that the source is active, i.e. , hence the mean excess rate is just h  C.
In the case of multiple sources, we need to calculate the probability that
n sources are active, where N0 < n
N and multiply by the excess rate,
n h  C. This probability is given by the binomial distribution
pn D

N!
n 1  Nn
n! N  n!

and so the mean excess rate is

N


pn n h  C

nDdN0 e

The mean arrival rate is simply N m, so the probability that a cell needs
the buffer is given by the ratio of the mean excess rate to the mean
arrival rate:
N


Prfcell needs bufferg D

pn n h  C

nDdN0 e

Nm

which, if we substitute for C D N0 h and D m/h, gives

N


Prfcell needs bufferg D

pn n  N0 

nDdN0 e

Lets put some numbers into this formula, using the example of two
different types of video source, each with a mean bit-rate of 768 kbit/s and
peak bit-rates of either 4.608 Mbit/s or 9.216 Mbit/s. The corresponding

142

BURST-SCALE QUEUEING

Table 9.2.

h (cell/s)

Parameter Values

2000
D
h

12 000
24 000

0.167
0.083

N0 D

353 207.55
h
29.43
14.72

cell rates are m D 2000 cell/s and h D 12 000 cell/s or 24 000 cell/s, and
the other parameter values are shown in Table 9.2.
Figure 9.14 shows how the probability that a cell needs the buffer
increases with the number of video sources being multiplexed through
the buffer. The minimum number of sources needed to produce burstscale queueing is 30 (for h D 12 000) or 15 (for h D 24 000). The results
show that about twice these values (60 and 30, respectively) produce
loss probabilities of about 1010 , increasing to between 101 and 102
for 150 of either source (see Figure 9.14). For both types of source the
mean rate, m, is 2000 cell/s, so the average load offered to the buffer, as a
fraction of its service capacity, ranges from 30 2000/353 208 17% up
to 150 2000/353 208 85%.
We know from Chapter 6 that the binomial distribution can be approximated by the Poisson distribution when the number of sources, N,
becomes large. This can be used to provide an approximate result for
Prfcell needs bufferg, the burst-scale loss factor, and it has the advantage
of being less demanding computationally because there is no summation [9.3].
Prfcell needs bufferg


N0 bN0 c 
N0
1
e

1 
2 N0
bN0 c!

where the offered load,

, is given by

m h
N
Nm
DN D
C
h C
N0

Figure 9.15 shows results for ON/OFF sources with peak rate h D
12 000 cell/s, and mean rates varying from m D 2000 cell/s ( D 0.167)
down to 500 cell/s ( D 0.042). N0 is fixed at 29.43, and the graph plots
the loss probability varying with the offered load,
. We can see that
for any particular value of
the burst-scale loss factor increases, as the
activity factor, , decreases, towards an upper limit given by the approximate result. The approximation thus gives a conservative estimate of the
probability that a cell needs the buffer. Note that as the activity factor
decreases, the number of sources must increase to maintain the constant
load, taking it into the region for which the Poisson approximation is
valid.

143

THE BUFFERLESS APPROACH

Number of sources, N

30

60

90

120

150

10 0
h = 24000 cell/s
h = 12000 cell/s

10 1

Pr{cell needs a buffer}

10 2
10 3
10 4
10 5
10 6
10 7
10 8
10 9
10 10

k :D 0.. 120
N


BSLexact , N, NO :D
xk :D k C 30

dbinom n, N,  n  NO

nDceil NO


2000
353207.55
, xk ,
24000
24000


2000
353207.55
y2k :D BSLexact
, xk ,
12000
12000


y1k :D BSLexact

Figure 9.14. The Bufferless ApproachResults for Multiple ON/OFF Sources

How does this Poisson approximation change our view of the source
process? Instead of considering N identical ON/OFF sources each with
probability of being in the active state and producing a burst of fixed
rate h, we are modelling the traffic as just one Poisson source which
produces overlapping bursts. The approximation equates the average
number of active sources with the average number of bursts in progress.
Its similar to our definition of traffic intensity, but at the burst level.
The average number of active sources is simply N ; now, recalling
that the probability of being active is related to the average durations in

144

BURST-SCALE QUEUEING

Offered load,
0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

10 0
101

Pr{cell needs buffer}

102
103
104
105
106
107
approx.
=0.042
=0.083
=0.167

108
109
1010

k :D 0..

130

BSLapprox 
, NO :D
N2k :D k C 30
N2k 2000
x1k :D
353207.5
y1k :D BSLapprox
x2k :D x1k

y2k :D BSLexact
N3k :D 2 k C 60
N3k 1000
x3k :D
353207.5 
y3k :D BSLexact

1

NOfloor NO 
NO

e
1 
2 NO
floor NO!


x1k ,

353207.55
12000

2000
353207.55
, N2k ,
12000
12000

1000
353207.55
, N3k ,
12000
12000

N4k :D 4 k C 120
N4k 500
x4k :D
353207.5


500
353207.55
y4k :D BSLexact
, N4k ,
12000
12000
Figure 9.15.

An Approximation to the Burst-Scale Loss Factor

1.0

145

THE BURST-SCALE DELAY MODEL

the ON and OFF states:

D

Ton
Ton C Toff

we can substitute for to obtain

average number of active sources D Ton

N
D Ton 
Ton C Toff

which is the average burst duration multiplied by the burst rate,  (each
source produces one burst every cycle time, Ton C Toff ). This is the average
number of bursts in progress.

THE BURST-SCALE DELAY MODEL

We are now in a position to extend the burst-scale analysis to finding the
probability that an excess-rate cell is lost given that it is an excess-rate
cell. With the bufferless approach, this probability is 1; every excess-rate
cell is lost because we assume there is no buffer in which to store it
temporarily. Now we assume that there is a finite amount of buffer space,
X, as shown in Figure 9.13.
We will view the N ON/OFF sources as a single Poisson source
producing bursts of cell rate h and duration Ton at a rate of  bursts per
second, where
N
D
Ton C Toff
Note that there is now no limit to the number of overlapping bursts; the
Poisson model can exceed N simultaneous bursts. But if N is sufficiently
large, the approximation to a Poisson source is reasonable. The average
number of cells per burst, b, is given by:
b D Ton h
so the load offered to the queue, as a fraction of the service capacity is

b
C

If we substitute for b and  (just to check) we obtain

Ton h

N
Ton C Toff
Nm
D
C
C

which is what we had for the bufferless model.

146

BURST-SCALE QUEUEING

An approximate analysis of this burst-scale delay model uses the

M/M/N queueing system (where the number of parallel servers, N, is
taken to be the maximum number of bursts which can fit into the service
capacity of the ATM buffer, N0 ) to give the following estimate for the
probability of loss [9.3]:



CLPexcess-rate D e

X 1 
3
N0
b 4
C1

This is similar in form to the heavy traffic approximations of Chapter 8:

an exponential function of buffer capacity and utilization. Note that the
buffer capacity here can be considered in units of the average burst
length, i.e. as X/b.
Recall that for the ND/D/1 approximation, N, the number of CBR
sources, is in the denominator of the exponential function. With a constant
load, as N increases, the gradient on the graph of cell loss against buffer
capacity decreases, i.e. the buffer is less able to cope with more sources of
smaller fixed cell rate. In contrast, this burst-scale delay result has N0 in
the numerator of the exponential function. N0 is the minimum number of
overlapping bursts required for burst-scale queueing. As N0 increases, so
does the gradient, and the buffer is better able to cope with more sources
of smaller ON rates. Why? If it takes more bursts to achieve queueing in
the buffer then the period of overlap will be smaller, reducing the effective
size of the excess-rate burst. An intuitive way of viewing this is to think of
b/N0 as the average excess-rate burst length; then N0 X/b can be considered as the buffer capacity in units of the average excess-rate burst length.
Lets continue with the example of the video sources we used earlier.
The mean cell rate for both types of source is m D 2000 cell/s and the
peak cell rates are equal to either 12 000 cell/s or 24 000 cell/s. What we
still need to specify are the state durations. If we assume that the ON
state is equivalent to a highly active video frame, then we can use a value
of 40 ms for Ton , which means the average number of cells per burst is
0.04 12 000 D 480 cells or 960 cells respectively. Toff is given by
Toff D Ton

hm
m

so Toff takes values of 0.2 second or 0.44 second respectively. The ON/OFF
source cycle times Ton C Toff  are 0.24 s and 0.48 s, so the burst rates for
the equivalent Poisson source of bursts are 4.167 (i.e. 1/0.24) or 2.083
times the number of sources, N, respectively.
Figure 9.16 shows the effect of the buffer capacity on the excess-rate
cell loss when there are 60 sources, giving an offered load of 0.34. The
results for three types of source are shown: the two just described, and

147

THE BURST-SCALE DELAY MODEL

Buffer capacity, X

Excess-rate cell loss probability

100

100

200

300

400

101

102

h = 24000 cell/s, b = 960 cells

h = 24000 cell/s, b = 480 cells
h = 12000 cell/s, b = 480 cells

k :D 1.. 40

BSDapprox NO, X, b,
 :D e
xk :D k 10

 NO

X 1
3

b 4
C1


353207.5
y1k :D BSDapprox
, xk , 960, 0.34
24000


353207.5
, xk , 480, 0.34
y2k :D BSDapprox
24000


353207.5
, xk , 480, 0.34
y3k :D BSDapprox
12000

Figure 9.16.

Probability of Excess-Rate Cell Loss

then the higher peak-rate source with an average active state duration
of half the original. This makes the average burst length, b, the same as
that for the lower-rate source. We can then make a fair assessment of the
impact of N0 , with b and
kept constant. It is clear then that as the peak
rate decreases, and therefore N0 increases, the buffer is better able to cope
with the excess-rate bursts.
Figure 9.17 shows how the two factors which make up the overall
cell loss probability are combined. The buffer capacity value was set at
400 cells. This corresponds to a maximum waiting time of 1.1 ms. The
burst-scale delay factor is shown for the two different peak rates as the

148

BURST-SCALE QUEUEING

Number of sources, N
30

60

90

120

150

100

Estimate of cell loss probability

101
102
103
104
105
106
burst scale delay
burst scale loss
combined (h = 24000 cell/s)
burst scale loss
burst scale delay
combined (h = 12000 cell/s)

107
108
109
1010

k :D 0..

120

xk :D k C 30



353207.55
xk 2000
y1k :D BSDapprox
, 400, 960,
24000
353207.5


2000
353207.55
y2k :D BSLexact
, xk ,
24000
24000
y3k :D y1k y2k


353207.55
xk 2000
y4k :D BSDapprox
, 400, 480,
12000
353207.5


2000
353207.55
y5k :D BSLexact
, xk ,
12000
12000
y6k :D y4k y5k
Figure 9.17.
Loss Factor

Combining Results for Burst-Scale Delay Factor with the Burst-Scale

curves with markers only. These results tend to an excess-rate loss probability of 1 as the number of sources, and hence the offered load increases.
The burst-scale loss results from Figure 9.14 are shown as the lines without
markers. The overall cell loss probabilities are the product of the two
factors and are the results shown with both lines and markers. Notice that
the extra benefit gained by having a large buffer for burst-scale queueing
does not appear to be that significant, for the situation considered here.

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

10

Connection Admission
Control
the net that likes to say YES!

No network operator likes to turn away business; if it does so too

often customers are likely to take their business elsewhere. Yet if the
operator always accepts any connection request, the network may become
congested, unable to meet the negotiated performance objectives for the
connections already established, with the likely outcome that many
customers will take their business elsewhere.
Connection admission control (CAC) is the name for that mechanism
which has to decide whether or not the bandwidth and performance
requirements of a new connection can be supported by the network,
in addition to those of the connections already established. If the new
connection is accepted, then the bandwidth and performance requirements form a traffic contract between the user and the network. We have
seen in Chapter 9 the impact that changes in traffic parameter values have
on performance, whether it is the duration of a peak-rate burst, or the
actual cell rate of a state. It is important then for the network to be able
to ensure that the traffic does not exceed its negotiated parameter values.
This is the function of usage parameter control. This in turn ensures
that the network meets the performance requirements for all the connections it has admitted. Together, connection admission control and usage
parameter control (UPC) are the main components in a traffic control
framework which aims to prevent congestion occurring. Congestion is
defined as a state of network elements (such as switching nodes and
transmission links) in which the network is not able to meet the negotiated performance objectives. Note that congestion is to be distinguished
from queue saturation, which may happen while still remaining within
the negotiated performance objective.

150

CONNECTION ADMISSION CONTROL

In a digital circuit-switched telephone network the admission control

problem is to find an unused circuit on a route from source to destination
for a single type of traffic. If a 64 kbit/s circuit is not available, then the
connection is blocked. In ATM the problem is rather more complicated:
not only must the route be found, but also a check must be made at
each link on a proposed route to ensure that the new connection, with
whatever traffic characteristics, can be supported without violating the
negotiated performance requirements of connections established over
each link.
In this chapter we focus on how we may make the check on each
link, by making use of the cell-scale and burst-scale queueing analysis of
previous chapters.

THE TRAFFIC CONTRACT

How are the bandwidth and performance requirements of the traffic
contract specified? In our burst-scale analysis so far, we have seen that
there are three traffic parameters which are important in determining
the type of queueing behaviour: peak cell rate, mean cell rate, and
the average active state duration. For the performance requirement,
we have concentrated on cell loss probability, but cell delay and CDV
(cell-delay variation) can also be important, particularly for interactive
services.
The number of bandwidth parameters in the traffic contract is closely
related to the complexity of the CAC algorithm and the type of queueing
behaviour that is being permitted on the network. The simplest approach
is CAC based on peak cell rate only: this limits the combined peak cell rate
of all VCs through a buffer to less than or equal to the service capacity of
the buffer. In this case there is never any burst-scale queueing, so the CAC
algorithm is based on cell-scale queueing analysis. The ITU Standards
terminology for a traffic control framework based on peak cell rate only
is deterministic bit-rate (DBR) transfer capability [10.1]. The equivalent
to this in ATM Forum terminology is constant bit-rate (CBR) service
category [10.2]. If we add another bandwidth parameter, the mean cell
rate, to the traffic contract and allow the peak cell rate to exceed the service
capacity, this is one form of what is called the statistical bit-rate (SBR)
transfer capability. In this case the CAC algorithm is based on both cellscale queueing analysis and burst-scale loss factor analysis (for reasons
explained in the previous chapter), with buffers dimensioned to cope
with cell-scale queueing behaviour only. The ATM Forum equivalent is
the variable bit-rate (VBR) service category.
Adding a third bandwidth parameter to quantify the burst length
allows another form of statistical bit-rate capability. This assumes buffers
are large enough to cope with burst-scale queueing, and the CAC

151

ADMISSIBLE LOAD: THE CELL-SCALE CONSTRAINT

algorithm is additionally based on analysis of the burst-scale delay

factor. In ATM Forum terminology this is the non-real-time (nrt) VBR
service category. However, if the burst length is relatively small, the
delays may be small enough to support real-time services.
Note that specifying SBR (VBR) or DBR (CBR) capability does not
imply a particular choice of queueing analysis; it just means that the
CAC algorithm is required to address both burst-scale and cell-scale
queueing components (in the case of SBR/VBR) or just the cell-scale
queueing component (in the case of DBR/CBR). Likewise, the bandwidth
parameters required in the traffic contract may depend on what analysis
is employed (particularly for burst-scale queueing).

ADMISSIBLE LOAD: THE CELL-SCALE CONSTRAINT

Lets say we have dimensioned a buffer to be 40 cells capacity for a cell
loss limit of 10
10 and a load of 75% (see Table 10.1). We could make our
maximum admissible load 75%, and not accept any more traffic if the
extra load would increase the total beyond 75%. But what if the cell loss
requirement is not so stringent? In this case the admissible load could be
greater than 75%. Some straightforward manipulation of the heavy load
Table 10.1. CAC Look-up Table for Finite M/D/1: Admissible Load, Given Buffer Capacity and
Cell Loss Probability

cell loss probability

x
(cells)

10
1

10
2

10
3

10
4

10
5

10
6

10
7

10
8

10
9

10
10

10
11

10
12

5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100

96.3%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%

59.7%
85.2%
92.4%
95.6%
97.2%
98.2%
98.9%
99.4%
99.7%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%
99.9%

41.9%
71.2%
82.4%
87.7%
90.7%
92.7%
94.0%
95.0%
95.7%
96.3%
96.7%
97.1%
97.4%
97.7%
97.9%
98.1%
98.2%
98.4%
98.5%
98.6%

16.6%
60.1%
74.2%
81.3%
85.4%
88.2%
90.1%
91.5%
92.6%
93.5%
94.2%
94.7%
95.2%
95.6%
95.9%
96.2%
96.5%
96.7%
96.9%
97.1%

6.6%
50.7%
66.9%
75.5%
80.7%
84.1%
86.6%
88.4%
89.8%
90.9%
91.8%
92.6%
93.2%
93.7%
94.2%
94.6%
95.0%
95.3%
95.5%
95.8%

2.9%
42.7%
60.4%
70.2%
76.2%
80.3%
83.2%
85.4%
87.1%
88.5%
89.6%
90.5%
91.3%
92.0%
92.5%
93.0%
93.5%
93.9%
94.2%
94.5%

1.35%
35.8%
54.4%
65.2%
72.0%
76.7%
80.0%
82.6%
84.6%
86.2%
87.5%
88.6%
89.5%
90.3%
91.0%
91.5%
92.1%
92.5%
92.9%
93.3%

0.62%
29.9%
49.0%
60.5%
68.0%
73.2%
77.0%
79.8%
82.1%
83.9%
85.4%
86.7%
87.7%
88.6%
89.4%
90.1%
90.7%
91.2%
91.7%
92.1%

0.28%
24.9%
44.0%
56.2%
64.2%
69.9%
74.0%
77.2%
79.7%
81.7%
83.4%
84.8%
86.0%
87.0%
87.9%
88.6%
89.3%
89.9%
90.5%
91.0%

0.13%
20.7%
39.5%
52.1%
60.6%
66.7%
71.2%
74.6%
77.4%
79.6%
81.4%
83.0%
84.3%
85.4%
86.4%
87.2%
88.0%
88.7%
89.3%
89.8%

0.06%
17.1%
35.4%
48.2%
57.2%
63.6%
68.4%
72.1%
75.1%
77.5%
79.5%
81.2%
82.6%
83.8%
84.9%
85.9%
86.7%
87.4%
88.1%
88.7%

0.03%
14.2%
31.6%
44.6%
53.9%
60.7%
65.8%
69.7%
72.9%
75.5%
77.6%
79.4%
81.0%
82.3%
83.5%
84.5%
85.4%
86.2%
86.9%
87.6%

152

CONNECTION ADMISSION CONTROL

approximation for the M/D/1 system (see Chapter 8) gives:

D

2x
2 x
lnCLP

where we have the maximum admissible load defined in terms of the

buffer capacity and the cell loss probability requirement.

A CAC algorithm based on M/D/1 analysis

How do we use this equation in a CAC algorithm? The traffic contract is
based on just two parameters: the peak cell rate, hi , and the required cell
loss probability CLPi , where i D 1, 2, . . . , n denotes the set of connections
which have already been accepted and are currently in progress, i.e.
they have not yet been cleared. Connection n C 1 is that request which
is currently being tested. This connection is accepted if the following
inequality holds:
hnC1
hi
C


C
C
iD1
n

2 x
ln

2x

min CLPi 

iD1!nC1

where C is the bandwidth capacity of the link. Obviously it is not necessary

to perform a summation of the peak rates every time because this can be
recorded in a current load variable which is modified whenever a new
connection is accepted or an existing connection is cleared. Similarly, a
temporary variable holding the most stringent (i.e. the minimum) cell loss
probability can be updated whenever a newly accepted connection has
a lower CLP. However, care must be taken to ensure that the minimum
CLP is recomputed when calls are cleared, so that the performance
requirements are based on the current set of accepted connections.
It is important to realize that the cell loss probability is suffered by all
admitted connections, because all cells go through the one link in question. Hence the minimum CLP is the one that will give the most stringent
limit on the admitted load, and it is this value that is used in the CAC
formula. (This is in fact an approximation: different VCs passing through
the same first-come first-served link buffer can suffer different cell loss
probabilities depending on their particular traffic characteristics, but the
variation is not large, and the analysis is complicated.) Priority mechanisms can be used to distinguish between levels of CLP requirements; we
deal with this in Chapter 13.
We know that the inequality is based on a heavy traffic approximation.
For a buffer size of 40 cells and a CLP requirement of 10
10 , the equation gives a maximum admissible load of 77.65%, slightly higher than

153

ADMISSIBLE LOAD: THE CELL-SCALE CONSTRAINT

the 74.6% maximum obtained using the exact analysis. An alternative

approach is to use look-up tables based on exact analysis instead of the
expression on the right-hand side of the inequality. Table 10.1 shows such
a table, giving the maximum percentage load that can be admitted for
finite buffer sizes ranging from 5 cells up to 100 cells, and cell loss probabilities ranging from 10
1 down to 10
10 . This table is generated by iteration of the output buffer analysis of Chapter 7 with Poisson input traffic.

A CAC algorithm based on ND/D/1 analysis

But what if all the traffic is CBR and the number of sources is relatively
small? We know from the ND/D/1 analysis that the admissible load
can be greater than that given by the M/D/1 results for a given CLP
requirement. The problem with the ND/D/1 analysis is that it models
a homogeneous source mix, i.e. all sources have the same traffic characteristics. In general, this will not be the case. However, it turns out that
for a fixed load,
, and a constant number of sources, N, the worst-case
situation for cell loss is the homogeneous case. Thus we can use the
ND/D/1 results and apply them in the general situation where there are
N sources of different peak cell rates.
As for the M/D/1 system, we manipulate the heavy load approximation for the ND/D/1 queue by taking logs of both sides, and rearrange
in terms of
:


CLP D e

x 1



2x N C

which gives the formula

D

2xN
2 x N
2 x2 C N lnCLP

It is possible for this formula to return values of admissible load greater

than 100%, specifically when
2 x2 C N lnCLP > 0
Such a load would obviously take the queue into a permanent (burstscale) overload, causing significantly more cell loss than that specified.
However, it does provide us with a first test for a CAC algorithm based
on this analysis, i.e. if


nC1


ln

2 x2


min CLPi 

iD1!nC1

154

CONNECTION ADMISSION CONTROL

then we can load the link up to 100% with any mix of n C 1 CBR sources,
i.e. we can accept the connection provided that
n
hi
hnC1

C

1
C
C
iD1

Otherwise, if

2 x2

nC1>

ln


min CLPi 

iD1!nC1

then we can accept the connection if

hnC1
hi
C


C
C
iD1
n

2 x n C 1

2 x n C 1
2

x2

C n C 1 ln



min CLPi 

iD1!nC1

It is also important to remember that the ND/D/1 analysis is only

required when N > x. If there are fewer sources than buffer places, then
the queue never overflows, and so the admissible load is 100%.
Like the M/D/1 system, this inequality is based on a heavy load
approximation. A look-up table method based on iteration of the equation
CLP

N


nDxC1

N!

n! N
n!

n
x
D

n 

1

n
x
D

N
n

D
NCx

D
nCx

provides a better approximation than the heavy load approximation, but

note that it is not an exact analysis as in Table 10.1 for the finite M/D/1.
The approach is more complicated than for the M/D/1 system because
of the dependence on a third parameter, N. Table 10.2 shows the
maximum number of sources admissible for a load of 100%, for combinations of buffer capacity and cell loss probability. Table 10.3 then shows
the maximum admissible load for combinations of N and cell loss probability, in three parts: (a) for a buffer capacity of 10 cells, (b) for 50 cells,
(c) for 100 cells.
The tables are used as follows: first check if the number of sources is
less than that given by Table 10.2 for a given CLP and buffer capacity; if
so, then the admissible load is 100%. Otherwise, use the appropriate part
of Table 10.3, with the given number of sources and CLP requirement,
to find the maximum admissible load. Note that when the maximum
admissible load is less than 100% of the cell rate capacity of the link, the
bandwidth that is effectively being allocated to each source is greater than
the sources peak cell rate, hi . This allocated bandwidth is found simply

155

ADMISSIBLE LOAD: THE CELL-SCALE CONSTRAINT

Table 10.2. CAC Look-up Table for Deterministic Bit-Rate Transfer Capability:
Maximum Number of Sources for 100% Loading, Given Buffer Capacity and Cell
Loss Probability

cell loss probability

x
(cells) 10
1 10
2 10
3 10
4 10
5 10
6 10
7 10
8 10
9 10
10 10
11 10
12
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100

23
89
200
353
550
790
1064
1389
1758
2171
2627
3126
3669
4256
4885
5558
6275
7035
7839
8685

11
45
100
176
275
395
537
701
886
1085
1313
1563
1834
2128
2442
2779
3137
3517
3919
4342

8
30
67
118
183
264
358
467
591
729
881
1042
1223
1418
1628
1852
2091
2345
2613
2895

6
23
50
89
138
198
269
351
443
547
661
786
922
1064
1221
1389
1568
1758
1959
2171

5
19
41
71
111
159
215
281
355
438
529
629
738
856
982
1111
1255
1407
1567
1737

5
5
5
16
14
13
34
30
26
60
52
45
92
80
70
133 114 100
180 155 136
234 201 176
296 254 223
365 313 275
441 379 332
525 450 394
616 528 462
714 612 536
819 702 615
931 799 699
1045 901 789
1172 1005 884
1306 1119 985
1447 1240 1085

5
12
24
41
63
89
121
157
198
244
295
351
411
477
547
622
702
786
876
970

5
11
22
37
57
81
109
142
179
220
266
316
371
429
493
560
632
708
788
873

5
11
20
34
52
74
100
129
163
201
242
288
337
391
448
510
575
644
717
794

5
10
19
32
48
68
92
119
150
185
223
264
310
359
411
468
527
591
658
729

by dividing the peak cell rate of a source by the maximum admissible

load (expressed as a fraction, not as a percentage).
This CAC algorithm, based on either the ND/D/1 approximate analysis or the associated tables, is appropriate for the deterministic bit-rate
capability. The parameters required are just the peak (cell) rate hi , and
the required cell loss probability, CLPi , for each source i, along with the
buffer capacity x, the cell rate capacity C, and the number of connections currently in progress, n. Note that it is acceptable when using
the deterministic bit-rate capability to mix variable and constant bitrate sources, provided that the peak cell rate of a source is used in
calculating the allocated load. The important point is that it is only
the peak cell rate which is used to characterize the sources traffic
behaviour.

The cell-scale constraint in statistical-bit-rate transfer capability,

based on M/D/1 analysis
A cell-scale constraint is also a component of the CAC algorithm for the
statistical bit-rate transfer capability. Here, the M/D/1 system is more

156

CONNECTION ADMISSION CONTROL

Table 10.3. (a) Maximum Admissible Load for a Buffer Capacity of 10 Cells, Given Number of
Sources and Cell Loss Probability

cell loss probability

N
10
11
12
13
14
15
16
17
18
19
20
30
40
50
60
70
80
90
100
200
300
400
500
600
700
800
900
1000

1

10

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
98.9%
98.0%
93.5%
92.0%
91.3%
90.9%
90.6%
90.4%
90.3%
90.2%
90.1%

2

10

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
96.2%
93.8%
90.9%
88.9%
88.2%
87.0%
83.0%
81.7%
81.1%
80.8%
80.5%
80.4%
80.2%
80.1%
80.1%

3

10

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
88.9%
84.8%
82.2%
80.5%
78.4%
77.6%
76.9%
73.5%
72.3%
71.8%
71.6%
71.5%
71.4%
71.2%
71.1%
71.0%

4

10

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
85.7%
78.4%
74.6%
72.3%
70.7%
69.0%
68.2%
67.6%
64.7%
63.7%
63.3%
63.1%
62.8%
62.7%
62.6%
62.5%
62.5%

5

10

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
95.2%
75.0%
69.0%
64.9%
63.2%
61.4%
60.6%
60.0%
59.2%
56.7%
56.0%
55.6%
55.3%
55.2%
55.1%
55.0%
54.9%
54.9%

10
6

10
7

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
94.4%
85.7%
82.6%
80.0%
65.2%
59.7%
56.8%
55.1%
53.9%
53.0%
52.3%
51.8%
49.5%
48.9%
48.5%
48.4%
48.2%
48.1%
48.1%
48.0%
48.0%

100.0%
100.0%
100.0%
100.0%
100.0%
93.8%
84.2%
81.0%
75.0%
73.1%
69.0%
56.6%
51.3%
49.0%
47.6%
46.7%
46.0%
45.5%
44.8%
43.1%
42.6%
42.3%
42.1%
42.0%
41.9%
41.9%
41.8%
41.8%

10
8

10
9

10
10

10
11

10
12

100.0% 100.0% 100.0% 100.0% 100.0%

100.0% 100.0% 100.0% 100.0% 84.6%
100.0% 100.0% 85.7% 70.6% 57.1%
100.0% 81.3% 68.4% 59.1% 48.2%
87.5% 73.7% 60.9% 51.9% 42.4%
79.0% 65.2% 55.6% 46.9% 39.5%
72.7% 61.5% 51.6% 43.2% 36.4%
68.0% 56.7% 48.6% 41.5% 34.7%
64.3% 54.6% 46.2% 39.1% 33.3%
61.3% 52.8% 44.2% 38.0% 32.2%
58.8% 50.0% 42.6% 36.4% 30.8%
48.4% 41.7% 35.7% 30.3% 25.9%
44.4% 38.1% 32.8% 28.0% 24.0%
42.4% 36.5% 31.5% 26.9% 22.9%
41.1% 35.5% 30.5% 26.1% 22.3%
40.5% 34.8% 29.9% 25.6% 21.9%
39.8% 34.3% 29.5% 25.3% 21.6%
39.3% 34.0% 29.2% 25.0% 21.3%
38.9% 33.7% 28.9% 24.8% 21.2%
37.4% 32.3% 27.9% 23.9% 20.5%
37.0% 31.9% 27.5% 23.6% 20.2%
36.7% 31.7% 27.3% 23.5% 20.1%
36.6% 31.6% 27.3% 23.4% 20.0%
36.5% 31.6% 27.2% 23.3% 20.0%
36.4% 31.5% 27.1% 23.3% 20.0%
36.4% 31.5% 27.1% 23.3% 19.9%
36.3% 31.4% 27.1% 23.3% 19.9%
36.3% 31.4% 27.1% 23.2% 19.9%

appropriate, using the mean cell rate, mi , instead of the peak cell rate hi ,
to calculate the load in the inequality test; i.e. if
n
mnC1

mi
C


C
C
iD1

2 x
ln

2x

min CLPi 

iD1!nC1

is satisfied, then the cell-scale behaviour is within the required cell

loss probability limits, and the CAC algorithm must then check the
burst-scale constraint before making an accept/reject decision. If the
inequality is not satisfied, then the connection can immediately be
rejected. For a more accurate test, values from the look-up table in
Table 10.1 can be used instead of the expression on the right-hand side of
the inequality.

157

ADMISSIBLE LOAD: THE BURST SCALE

Table 10.3. (b) Maximum Admissible Load for a Buffer Capacity of 50 Cells, Given Number of
Sources and Cell Loss Probability

cell loss probability

N
180
190
200
210
220
240
260
280
300
350
400
450
500
550
600
700
800
900
1000

1

10

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%

2

10

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%

3

10

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
99.3%
98.6%
98.1%

4

5

10

10

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
99.8%
99.0%
97.9%
97.0%
96.3%
95.7%

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
99.6%
98.4%
97.5%
96.6%
95.5%
94.7%
94.0%
93.5%

10
6

10
7

10
8

10
9

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
98.5%
97.2%
96.0%
95.2%
94.3%
93.2%
92.4%
91.7%
91.2%

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
98.0%
96.2%
94.7%
93.6%
92.8%
92.0%
90.9%
90.2%
89.6%
89.1%

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
99.3%
98.0%
95.6%
93.9%
92.4%
91.4%
90.5%
89.8%
88.7%
87.9%
87.4%
86.9%

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
98.5%
96.9%
95.5%
93.1%
91.5%
90.2%
89.1%
88.3%
87.6%
86.5%
85.8%
85.2%
84.8%

10
10

10
11

10
12

100.0% 100.0% 100.0%

100.0% 100.0% 99.0%
100.0% 100.0% 97.1%
100.0% 98.6% 95.9%
100.0% 97.4% 94.8%
97.6% 95.2% 92.7%
95.9% 93.5% 90.9%
94.6% 92.1% 89.7%
93.2% 90.9% 88.5%
90.9% 88.6% 86.2%
89.1% 87.0% 84.8%
87.9% 85.7% 83.5%
87.0% 84.8% 82.5%
86.1% 84.0% 81.9%
85.5% 83.3% 81.2%
84.4% 82.4% 80.3%
83.7% 81.6% 79.6%
83.1% 81.1% 79.0%
82.6% 80.7% 78.6%

Table 10.3. (c) Maximum Admissible Load for a Buffer Capacity of

100 Cells

cell loss probability

N
700
750
800
850
900
950
1000

8

10

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%

9

10

100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
99.6%

10
10

10
11

10
12

100.0%
100.0%
100.0%
100.0%
99.6%
99.0%
98.4%

100.0%
100.0%
99.9%
99.1%
98.4%
97.7%
97.2%

100.0%
99.5%
98.6%
97.8%
97.2%
96.5%
96.0%

ADMISSIBLE LOAD: THE BURST SCALE

Lets now look at the loads that can be accepted for bursty sources. For
this we will use the burst-scale loss analysis of the previous chapter,
i.e. assume that the buffer is of zero size at the burst scale. Remember
that each source has an average rate of m cell/s; so, with N sources, the
utilization is given by

D

Nm
C

158

CONNECTION ADMISSION CONTROL

Unfortunately we do not have a simple approximate formula that can

be manipulated to give the admissible load as an explicit function of
the traffic contract parameters. The best we can do to simplify the
situation is to use the approximate formula for the burst-scale loss
factor:

N0 bN0 c

N0
1

e
CLP
1

2 N0
bN0 c!
How can we use this formula in a connection admission control algorithm? In a similar manner to Erlangs lost call formula, we must use the
formula to produce a table which allows us, in this case, to specify the
required cell loss probability and the source peak cell rate and find out
the maximum allowed utilization. We can then calculate the maximum
number of sources of this type (with mean cell rate m) that can be accepted
using the formula

C
ND
m
Table 10.4 does not directly use the peak cell rate, but, rather, the number
of peak cell rates which fit into the service capacity, i.e. the parameter N0 .
Example peak rates for the standard service capacity of 353 208 cell/s are
shown.
So, if we have a source with a peak cell rate of 8830.19 cell/s (i.e.
3.39 Mbit/s) and a mean cell rate of 2000 cell/s (i.e. 768 kbit/s), and we
want the CLP to be no more than 10
10 , then we can accept
Table 10.4. Maximum Admissible Load for Burst-Scale Constraint

cell loss probability

h
(cell/s)

N0

10
1

10
2

10
3

10
4

10
5

10
6

10
7

10
8

10
9

10
10 10
11 10
12

35 320.76
17 660.38
11 773.59
8 830.19
7 064.15
5 886.79
5 045.82
4 415.09
3 924.53
3 532.08
1 766.04
1 177.36
883.02
706.42
588.68
504.58

10
20
30
40
50
60
70
80
90
100
200
300
400
500
600
700

72.1%
82.3%
86.5%
88.9%
90.5%
91.7%
92.5%
93.2%
93.7%
94.2%
96.4%
97.3%
97.8%
98.1%
98.4%
98.5%

52.3%
67.0%
73.7%
77.8%
80.5%
82.5%
84.1%
85.3%
86.3%
87.2%
91.7%
93.6%
94.7%
95.4%
95.9%
96.3%

37.9%
54.3%
62.5%
67.5%
71.1%
73.7%
75.8%
77.4%
78.8%
80.0%
86.4%
89.2%
90.8%
91.9%
92.7%
93.3%

28.1%
44.9%
53.8%
59.5%
63.5%
66.6%
69.0%
71.0%
72.6%
74.0%
81.8%
85.3%
87.4%
88.8%
89.9%
90.7%

21.2%
37.7%
46.9%
53.0%
57.4%
60.8%
63.5%
65.7%
67.5%
69.1%
78.0%
82.0%
84.5%
86.2%
87.4%
88.4%

16.2%
32.0%
41.4%
47.7%
52.4%
55.9%
58.8%
61.2%
63.2%
64.9%
74.7%
79.2%
82.0%
83.9%
85.3%
86.4%

12.5%
27.4%
36.8%
43.3%
48.1%
51.8%
54.8%
57.3%
59.5%
61.3%
71.8%
76.8%
79.8%
81.9%
83.4%
84.7%

9.7%
23.6%
32.9%
39.4%
44.3%
48.2%
51.3%
54.0%
56.2%
58.1%
69.3%
74.6%
77.8%
80.1%
81.8%
83.1%

7.6%
20.5%
29.6%
36.1%
41.1%
45.0%
48.3%
51.0%
53.3%
55.3%
67.0%
72.6%
76.0%
78.4%
80.2%
81.7%

5.9%
17.8%
26.7%
33.2%
38.2%
42.2%
45.5%
48.3%
50.6%
52.7%
64.9%
70.7%
74.4%
76.9%
78.8%
80.3%

4.7%
15.6%
24.1%
30.6%
35.6%
39.6%
43.0%
45.8%
48.2%
50.4%
62.9%
69.0%
72.8%
75.5%
77.5%
79.1%

3.7%
13.6%
21.9%
28.2%
33.2%
37.3%
40.7%
43.6%
46.0%
48.2%
61.1%
67.5%
71.4%
74.2%
76.3%
78.0%

159

ADMISSIBLE LOAD: THE BURST SCALE

ND

0.332 353 208

D 58.63
2000

i.e. 58 connections of this type. This is 18 more connections than if we

had used the deterministic bit-rate capability (assuming 100% allocation
of peak rates, which is possible if the buffer capacity is 25 cells or more).
The ratio
N
GD
N0
is called the statistical multiplexing gain. This is the actual number
accepted, N, divided by the number N0 if we were to allocate on the peak
rate only. It gives an indication of how much better the utilization is when
using SBR capability compared with using DBR capability. If peak rate
allocation is used, then there is no statistical multiplexing gain, and G is 1.
But what happens if there are different types of source? If all the sources
have the same peak cell rate, then the actual mean rates of individual
sources do not matter, so long as the total mean cell rate is less than

C, i.e.


mi

C
i

So, the connection is accepted if the following inequality holds:

n
mi
mnC1

C


CLP, N0 
C
C
iD1

where
is chosen (as a function of CLP and N0 ) from Table 10.4 in the
manner described previously.

A practical CAC scheme

Notice in the table that the value of
decreases as the peak cell rate
increases. We could therefore use this approach in a more conservative
way by choosing
according to the most stringent (i.e. highest) peak-rate
source in the mix. This is effectively assuming that all sources, whatever
their mean rate, have a peak rate equal to the highest in the traffic mix. The
CAC algorithm would need to keep track of this maximum peak rate (as
well as the minimum CLP requirement), and update the admissible load
accordingly. The inequality test for this scheme is therefore written as:

n
C
mi
mnC1

C


min CLPi ,
iD1!nC1
C
C
max
h

i
iD1
iD1!nC1

160

CONNECTION ADMISSION CONTROL

Equivalent cell rate and linear CAC

A different approach is to think in terms of the cell rate allocated to
a source. For the DBR capability, a CAC algorithm allocates either the
sources peak cell rate or a value greater than this, because cell-scale
queueing limits the admissible load. This keeps the utilization, defined in
terms of peak rates, at or below 100%. With SBR capability, the total peak
rate allocated can be in excess of 100%, so the actual portion of service
capacity allocated to a source is below the peak cell rate (and, necessarily,
above the mean cell rate). This allocation is called the equivalent cell rate.
Other terms have been used to describe essentially the same concept:
effective bandwidth and equivalent capacity are the most common
terms used, but the precise definition is usually associated with a particular analytical method. Equivalent cell rate is the term used in the ITU
Standards documents.
The key contribution made by the concept of equivalent cell rate is the
idea of a single value to represent the amount of resource required for
a single source in a traffic mix at a given CLP requirement. This makes
the admission control process simply a matter of adding the equivalent
cell rate of the requested connection to the currently allocated value. If it
exceeds the service rate available then the request is rejected. This is an
attractive approach for traffic mixes of different types of sources because
of its apparent simplicity. It is known as linear CAC.
The difficulty lies in defining the equivalent cell rate for a particular
source type. The issue rests on how well different types of sources are
able to mix when multiplexed through the same buffer. The analysis we
have used so far is for a traffic mix of sources of the same type. In this
case, the equivalent cell rate can be defined as
ECR D

C h
N0
h
C
D
D
hD
N
h N
N
G

When the statistical multiplexing gain, G, is low (i.e. approaching a value

of 1), the equivalent cell rate approaches the peak rate of the source and
the cell loss probability will be low. Conversely, when the gain is high,
the equivalent cell rate approaches the mean rate of the source, and the
cell loss probability is high.
Equivalent cell rate based on a traffic mix of sources of the same type may
underestimate the resources required when sources of very different characteristics are present. The exact analysis of heterogeneous source multiplexing is beyond the scope of this book, but there are other approaches.

Two-level CAC
One of these approaches, aimed at simplifying CAC, is to divide the
sources into classes and partition the service capacity so that each source

ADMISSIBLE LOAD: THE BURST SCALE

161

class is allocated a proportion of it. The homogeneous source analysis can

be justified in this case because the fraction of service rate allocated to the
class is used instead of the total service capacity (within the fraction, all
the sources being the same). This has the effect of reducing N0 , and hence
reducing the admissible load per class. The problem with this approach
is that a connection of one type may be rejected if its allocation is full
even though there is unused capacity because other service classes are
underused.
A solution is to divide the CAC algorithm into two levels. The first level
makes accept/reject decisions by comparing the current service-class
allocations with the maximum number allowed. But this is supported by
a second-level back-room task which redistributes unused capacity to
service classes that need it. The second level is computationally intensive
because it must ensure that the allocations it proposes conform to the
required cell loss probability. This takes time, and so the allocations are
updated on a (relatively) longer time scale. However, the first level is
a very simple comparison and so a connection request can be assessed
immediately.
The basic principle of the two-level scheme is to have a first level which
can make an instant decision on a connection request, and a second level
which can perform detailed traffic calculations in the background to keep
the scheme as accurate as possible. The service class approach is just
one of many: other algorithms for the first and second levels have been
proposed in the literature.

Accounting for the burst-scale delay factor

Whatever the size of buffer, large or small, the actual burst-scale loss
depends on the two burst-scale factors: the loss factor assumes there is
no buffer, and the delay factor quantifies how much less is the loss if we
incorporate buffering. Thus if we use the loss factor only, we will tend to
overestimate the cell loss; or for a fixed CLP, we will underestimate the
admissible load.
So, for small buffer capacities, just using the loss factor is a good
starting point for admission control at the burst scale. But we have
already incorporated some conservative assumptions into our practical scheme, and even small buffers can produce some useful gains
under certain circumstances. How can the scheme be modified to
account for the burst-scale delay factor, and hence increase the admissible load?
Lets use our previous example of 58 connections (peak cell rate
8830.19 cell/s, mean cell rate 2000 cell/s, N0 D 40, and a CLP of 10
10 )
and see how many more connections can be accepted if the average burst
duration is 40 ms and the buffer capacity is 475 cells. First, we need to

162

CONNECTION ADMISSION CONTROL

calculate:

40 475
N0 x
D
D 53.79
b
0.04 8830.19

and the admissible load (from the burst-scale loss analysis) is

D

58 2000
D 0.328
353 208

So we can calculate the CLP gain due to the burst-scale delay factor:


CLPexcess-rate D e

X 1

3

N0 b 4
C1

D 8.58 10
4

Thus there is a further CLP gain of about 10
3 , i.e. an overall CLP of
about 10
13 .
Although the excess-rate cell loss is an exponential function, which can
thus be rearranged fairly easily, we will use a tabular approach because
it clearly illustrates the process required. Table 10.5 specifies the CLP
and the admissible load in order to find a value for N0 x/b (this was
introduced in Chapter 9 as the size of a buffer in units of excess-rate
bursts). The CLP target is 10
10 . By how much can the load be increased
so that the overall CLP meets this target? Looking down the 10
2 column
of Table 10.5, we find that the admissible load could increase to a value
of nearly 0.4. Then, we check in Table 10.4 to see that the burst-scale loss
contribution for a load of 0.394 is 10
8 . Thus the overall CLP meets our
target of 10
10 .
The number of connections that can be accepted is now
ND

0.394 353 208

D 69.58
2000

i.e. 69 connections of this type. This is a further 11 connections more

than if we had just used the burst-scale loss factor as the basis for the
CAC algorithm. The penalty is the increased size of the buffer, and
the correspondingly greater delays incurred (about 1.3 ms maximum,
for a buffer capacity of 475 cells). However, the example illustrates the
principle, and even with buffers of less than 100 cells, worthwhile gains
in admissible load are possible. The main difficulty with the process is
in selecting a load to provide cell loss factors from Tables 10.4 and 10.5
which combine to the required target cell loss. The target cell loss can
be found by trial and error, gradually reducing the excess rate CLP by
taking the next column to the left in Table 10.5.

0.02
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0.18
0.20
0.22
0.24
0.26
0.28
0.30
0.32
0.34
0.36
0.38
0.40
0.42
0.44
0.46
0.48
0.50
0.52
0.54

load,

10

2.6
3.0
3.4
3.9
4.4
5.0
5.6
6.4
7.2
8.1
9.1
10.3
11.6
13.1
14.8
16.7
18.9
21.4
24.3
27.7
31.6
36.2
41.5
47.8
55.3
64.1
74.8

1

5.3
6.0
6.9
7.8
8.8
10.0
11.3
12.7
14.4
16.2
18.2
20.6
23.2
26.2
29.5
33.4
37.8
42.9
48.7
55.4
63.3
72.4
83.1
95.6
110.5
128.3
149.5

10

2

7.9
9.1
10.3
11.7
13.3
15.0
16.9
19.1
21.5
24.3
27.4
30.8
34.8
39.2
44.3
50.1
56.7
64.3
73.0
83.1
94.9
108.6
124.6
143.5
165.8
192.4
224.3

10

3

10.6
12.1
13.8
15.6
17.7
20.0
22.6
25.5
28.7
32.4
36.5
41.1
46.4
52.3
59.1
66.8
75.6
85.7
97.4
110.9
126.5
144.8
166.1
191.3
221.0
256.5
299.0

10

4

13.2
15.1
17.2
19.5
22.1
25.0
28.2
31.9
35.9
40.5
45.6
51.4
58.0
65.4
73.8
83.5
94.5
107.2
121.7
138.6
158.1
180.9
207.6
239.1
276.3
320.6
373.8

10

5

15.9
18.1
20.6
23.4
26.5
30.0
33.9
38.2
43.1
48.6
54.7
61.7
69.6
78.5
88.6
100.2
113.4
128.6
146.1
166.3
189.8
217.1
249.2
286.9
331.6
384.8
448.5

10
6
18.5
21.1
24.1
27.3
31.0
35.0
39.5
44.6
50.3
56.7
63.9
72.0
81.1
91.5
103.4
116.9
132.3
150.0
170.4
194.0
221.4
253.3
290.7
334.7
386.8
448.9
523.3

10
7

cell loss probability

21.1
24.2
27.5
31.2
35.4
40.0
45.2
51.0
57.5
64.8
73.0
82.2
92.7
104.6
118.2
133.6
151.2
171.5
194.8
221.7
253.0
289.5
332.2
382.5
442.1
513.0
598.0

10
8
23.8
27.2
30.9
35.1
39.8
45.0
50.8
57.3
64.6
72.9
82.1
92.5
104.3
117.7
132.9
150.3
170.1
192.9
219.1
249.4
284.6
325.7
373.8
430.4
497.4
577.1
672.8

10
9
26.4
30.2
34.4
39.0
44.2
50.0
56.5
63.7
71.8
81.0
91.2
102.8
115.9
130.8
147.7
167.0
189.0
214.3
243.5
277.2
316.3
361.9
415.3
478.2
552.6
641.3
747.5

10
10

Table 10.5. Burst-Scale Delay Factor Table for Values of N0 x/b, Given Admissible Load and CLP

31.7
36.2
41.3
46.8
53.1
60.0
67.8
76.5
86.2
97.1
109.5
123.4
139.1
156.9
177.2
200.4
226.8
257.2
292.2
332.6
379.5
434.3
498.3
573.8
663.1
769.5
897.0

10
12

(continued overleaf )

29.1
33.2
37.8
42.9
48.6
55.0
62.1
70.1
79.0
89.0
100.3
113.1
127.5
143.9
162.5
183.7
207.9
235.8
267.8
304.9
347.9
398.1
456.8
526.0
607.9
705.4
822.3

10
11

ADMISSIBLE LOAD: THE BURST SCALE

163

0.56
0.58
0.60
0.62
0.64
0.66
0.68
0.70
0.72
0.74
0.76
0.78
0.80
0.82
0.84
0.86
0.88
0.90
0.92
0.94
0.96
0.98

load,

10

2

10

3

10

4

10

5

10
6

10
7

cell loss probability

10
8

10
9

10
10

10
11

10
12

87.6
175.2
262.7
350.3
437.9
525.5
613.1
700.6
788.2
875.8
963.4
1 051.0
103.2
206.4
309.5
412.7
515.9
619.1
722.3
825.5
928.6
1 031.8
1 135.0
1 238.2
122.3
244.6
367.0
489.3
611.6
733.9
856.3
978.6
1 100.9
1 223.2
1 345.6
1 467.9
146.0
292.1
438.1
584.1
730.2
876.2
1 022.2
1 168.2
1 314.3
1 460.3
1 606.3
1 752.4
175.7
351.4
527.1
702.8
878.5
1 054.2
1 229.9
1 405.6
1 581.3
1 756.9
1 932.6
2 108.3
213.2
426.5
639.7
853.0
1 066.2
1 279.5
1 492.7
1 706.0
1 919.2
2 132.5
2 345.7
2 558.9
261.4
522.8
784.2
1 045.6
1 307.0
1 568.4
1 829.8
2 091.2
2 352.6
2 614.0
2 875.4
3 136.8
324.1
648.1
972.2
1 296.3
1 620.3
1 944.4
2 268.5
2 592.5
2 916.6
3 240.7
3 564.7
3 888.8
407.0
814.0
1 220.9
1 627.9
2 034.9
2 441.9
2 848.9
3 255.8
3 662.8
4 069.8
4 476.8
4 883.8
518.8
1 037.6
1 556.4
2 075.2
2 593.9
3 112.7
3 631.5
4 150.3
4 669.1
5 187.9
5 706.7
6 225.5
672.9
1 345.8
2 018.8
2 691.7
3 364.6
4 037.5
4 710.4
5 383.4
6 056.3
6 729.2
7 402.1
8 075.0
890.9
1 781.9
2 672.8
3 563.7
4 454.7
5 345.6
6 236.5
7 127.5
8 018.4
8 909.3
9 800.3
10 691.2
1 208.9
2 417.7
3 626.6
4 835.4
6 044.3
7 253.1
8 462.0
9 670.9
10 879.7
12 088.6
13 297.4
14 506.3
1 689.8
3 379.7
5 069.5
6 759.3
8 449.1
10 139.0
11 828.8
13 518.6
15 208.4
16 898.3
18 588.1
20 277.9
2 451.0
4 902.0
7 353.0
9 804.0
12 255.0
14 706.0
17 157.0
19 608.0
22 058.9
24 509.9
26 960.9
29 411.9
3 725.8
7 451.5
11 177.3
14 903.0
18 628.8
22 354.5
26 080.3
29 806.1
33 531.8
37 257.6
40 983.3
44 709.1
6 023.0
12 045.9
18 068.9
24 091.9
30 114.8
36 137.8
42 160.8
48 183.7
54 206.7
60 229.7
66 252.6
72 275.6
10 591.9
21 183.8
31 775.7
42 367.6
52 959.5
63 551.3
74 143.2
84 735.1
95 327.0
1 05 918.9
1 16 510.8
1 27 102.7
21 047.1
42 094.1
63 141.2
84 188.3 1 05 235.3 1 26 282.4 1 47 329.5
1 68 376.5
1 89 423.6
2 10 470.7
2 31 517.7
2 52 564.8
50 742.2 1 01 484.3 1 52 226.5 2 02 968.6 2 53 710.8 3 04 452.9 3 55 195.1
4 05 937.2
4 56 679.4
5 07 421.5
5 58 163.7
6 08 905.8
1 74 133.0 3 48 266.0 5 22 399.0 6 96 532.0 8 70 665.0 10 44 798.0 12 18 931.0 13 93 064.0 15 67 197.0 17 41 330.0 19 15 463.0 20 89 596.0
14 16 089.8 28 32 179.7 42 48 269.5 56 64 359.3 70 80 449.2 84 96 539.0 99 12 628.8 11 328 718.7 12 744 808.5 14 160 898.3 15 576 988.2 16 993 078.0

10

1

Table 10.5. (continued)

164
CONNECTION ADMISSION CONTROL

CAC IN THE STANDARDS

165

CAC IN THE STANDARDS

Connection admission control is defined in ITU Recommendation I.371
[10.1] as the set of actions taken by the network at the call set-up phase
(or during call re-negotiation) to establish whether a connection can
be accepted or whether it must be rejected. The wording in the ATM
Forum Traffic Management Specification 4.1 [10.2] is very similar. We
have seen that the CAC algorithm needs to know the source traffic characteristics and the required performance in order to determine whether
the connection can be accepted or not and, if accepted, the amount of
network resources to allocate. Also it must set the traffic parameters
needed by usage parameter control this will be addressed in the next
chapter.
Neither Recommendation I.371 nor the Traffic Management Specification specifies any particular CAC algorithm; they merely observe that
many CAC policies are possible, and it is up to the network operator
to choose. ITU Recommendation E.736 outlines some possible policies
[10.3]. It distinguishes three different operating principles:
1.

multiplexing of constant-bit-rate streams

2.

rate-envelope multiplexing

3.

rate-sharing statistical multiplexing

The first corresponds to peak rate allocation, i.e. the deterministic bit-rate
transfer capability, and deals with the cell-scale queueing behaviour.
In this book we have considered two different algorithms, based on
either the M/D/1 or ND/D/1 systems. The second and third operating
principles allow for the statistical multiplexing of variable bit-rate streams
and are two approaches to providing the statistical bit-rate transfer
capability. Rate envelope multiplexing is the term for what we have
called the burst-scale loss factor, i.e. it is the bufferless approach. The
term arises because the objective is to keep the total input rate to within
the service rate; any excess rate is assumed to be lost. Rate sharing
corresponds to the combined burst-scale loss and delay factors, i.e. it
assumes there is a large buffer available to cope with the excess cell
rates. It allows higher admissible loads, but the penalty is greater delay.
Thus the objective is not to limit the combined cell rate, but to share
the service capacity by providing sufficient buffer space to absorb the
excess-rate cells.
These three different operating principles require different traffic
parameters to describe the source traffic characteristics. DBR requires
just the peak cell rate of the source. Rate envelope multiplexing additionally needs the mean cell rate, and rate sharing requires peak cell rate,
mean cell rate and some measure of burst length. The actual parameters
depend on the CAC policy and what information it uses. But there is one

166

CONNECTION ADMISSION CONTROL

important principle that applies regardless of the policy: if a CAC policy

depends on a particular traffic parameter, then the network operator
needs to ensure that the value the user has declared for that parameter is
not exceeded during the actual flow of cells from the source. Only then
can the network operator be confident that the performance requirements
will be guaranteed. This is the job of usage parameter control.

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

11

Usage Parameter Control

theres a hole in my bucket

PROTECTING THE NETWORK

We have discussed the statistical multiplexing of traffic through ATM
buffers and connection admission control mechanisms to limit the number
of simultaneous connections, but how do we know that a traffic source is
going to conform to the parameter values used in the admission control
decision? There is nothing to stop a source sending cells over the access
link at a far higher rate. It is the job of usage parameter control to
ensure that any cells over and above the agreed values do not get any
further into the network. These agreed values, including the performance
requirements, are called the traffic contract.
Usage parameter control is defined as the set of actions taken by
the network, at the user access, to monitor and control traffic in terms
of conformity with the agreed traffic contract. The main purpose is to
protect network resources from source traffic misbehaviour that could
affect the quality of service of other established connections. UPC does
this by detecting violations of negotiated parameters and taking appropriate actions, for example discarding or tagging cells, or clearing the
connection.
A specific control algorithm has not been standardized as with CAC
algorithms, the network may use any algorithm for UPC. However, any
such control algorithm should have the following desirable features:

the ability to detect any traffic situation that does not conform to the
traffic contract,
a rapid response to violations of the traffic contract, and

being simple to implement.

168

USAGE PARAMETER CONTROL

But are all these features possible in one algorithm? Lets recall what
parameters we want to check. The most important one is the peak cell
rate; it is needed for both deterministic and statistical bit-rate transfer
capabilities. For SBR, the traffic contract also contains the mean cell
rate (for rate envelope multiplexing). With rate-sharing statistical multiplexing, the burst length is additionally required. Before we look at a
specific algorithm, lets consider the feasibility of controlling the mean
cell rate.

CONTROLLING THE MEAN CELL RATE

Suppose we count the total number of cells being sent in some measurement interval, T, by a Poisson source. The source has a declared mean
cell rate,
, of one cell per time unit. Is it correct to allow no more than
one cell per time unit into the network? We know from Chapter 6 that
the probability of k cells arriving in one time unit from a Poisson source
is given by
Prfk arrivals in one time unitg D


Tk 
T
e
k!

So the probability of more than one arrival per time unit is

D1

10 1 11 1
e 
e D 0.2642
0!
1!

Thus this strict mean cell rate control would reject one or more cells from
a well-behaved Poisson source in 26 out of every 100 time units. What
proportion of the number of cells does this represent? Well, we know that
the mean number of cells per time unit is 1, and this can also be found by
summing the probabilities of there being k cells weighted by the number
of cells, k, i.e.
mean number of cells D 1 D 0

10 1
11 1
12 1
e C1
e C2
e
0!
1!
2!

C C k

1k 1
e C
k!

When there are k
1 cell arrivals in a time unit, then one cell is allowed
on to the network and k  1 are rejected. Thus the proportion of cells
being allowed on to the network is
1

1
1k 1
11 1

1
e C
e
1!
k!
kD2

D 0.6321

169

CONTROLLING THE MEAN CELL RATE

which means that almost 37% of cells are being rejected although the
traffic contract is not being violated.
There are two options open to us: increase the maximum number of
cells allowed into the network per time unit or increase the measurement
interval to many time units. The object is to decrease this proportion of
cells being rejected to an acceptably low level, for example 1 in 1010 .
Lets define j as the maximum number of cells allowed into the network
during time interval T. The first option requires us to find the smallest
value of j for which the following inequality holds:


1



Tk 
T
e
k  j
k!
kDjC1
 1010

T
where, in this case, the mean cell rate of the source,
, is 1 cell per time
unit, and the measurement interval, T, is 1 time unit. Table 11.1 shows
the proportion of cells rejected for a range of values of j.
To meet our requirement of no more than 1 in 1010 cells rejected for
a Poisson source of mean rate 1 cell per time unit, we must accept up
to 12 cells per time unit. If the Poisson source doubles its rate, then our
limit of 12 cells per time unit would result in 1.2 107 of the cells being
rejected. Ideally we would want 50% of the cells to be rejected to keep
the source to its contracted mean of 1 cell per time unit. If the Poisson
source increases its rate to 10 cells per time unit, then 5.3% of the cells are
Table 11.1. Proportion of Cells Rejected when no more than j cells
Are Allowed per Time Unit

proportion of cells rejected for a mean cell rate of

j
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

1 cell/time unit

2 cells/time unit

10 cells/time unit

3.68E-01
1.04E-01
2.33E-02
4.35E-03
6.89E-04
9.47E-05
1.15E-05
1.25E-06
1.22E-07
1.09E-08
9.00E-10
6.84E-11
4.84E-12
3.20E-13
1.98E-14

5.68E-01
2.71E-01
1.09E-01
3.76E-02
1.12E-02
2.96E-03
6.95E-04
1.47E-04
2.82E-05
4.96E-06
8.03E-07
1.21E-07
1.69E-08
2.21E-09
2.71E-10

9.00E-01
8.00E-01
7.00E-01
6.01E-01
5.04E-01
4.11E-01
3.24E-01
2.46E-01
1.79E-01
1.25E-01
8.34E-02
5.31E-02
3.22E-02
1.87E-02
1.03E-02

170

USAGE PARAMETER CONTROL

rejected, and hence over 9 cells per time unit are allowed through. Thus
measurement over a short interval means that either too many legitimate
cells are rejected (if the limit is small) or, for cells which violate the
contract, not enough are rejected (when the limit is large).
Lets now extend the measurement interval. Instead of tabulating for
all values of j, the results are shown in Figure 11.1 for two different time
intervals: 10 time units and 100 time units. For the 1010 requirement, j
is 34 (for T D 10) and 163 (for T D 100), i.e. the rate is limited to 3.4 cells
per time unit, or 1.63 cells per time unit over the respective measurement
intervals. So, as the measurement interval increases, the mean rate is
being more closely controlled. The problem now is that the time taken to
Maximum number of cells, j, allowed in T time units
0
10

20

40

60

80

100

120

140

180

200

T = 100 time units

T = 10 time units

10 1

Proportion of cells rejected

160

10 2
10 3
10 4
10 5
10 6
10 7
10 8
10 9
10 10

k :D 1.. 200

Propreject T,
, j, max j :D

max j

k  j dpoisk,
T

kDjC1

xk :D k
y1k :D Propreject 100, 1, xk , 250
y2k :D Propreject 10, 1, xk , 250

T

Figure 11.1. Proportion of Cells Rejected for Limit of j Cells in T Time Units

171

CONTROLLING THE MEAN CELL RATE

respond to violations of the contract is longer. This can result in action

being taken too late to protect the network from the effect of the contract
violation.
Figure 11.2 shows how the limit on the number of cells allowed per
time unit varies with the measurement interval, for a rejection probability
of 1010 . The shorter the interval, the poorer the control of the mean rate
because of the large safety margin required. The longer the interval, the
slower the response to violations of the contract.
So we see that mean cell rate control requires a safety margin between
the controlled cell rate and the negotiated cell rate to cope with the
15

Number of cells per time unit

Controlled cell rate

Negotiated cell rate
10

0
10 0

10 1

10 2

10 3

log T :D 0.. 30

Findj 
, T, reject, max j :D  j
ceil 
T
 while Propreject T,
, j, max j C j > reject



 j
jC1

j
log T

xlog T :D 10 10
Findj 1, xlog T , 1010 , 500
y1log T :D
xlog T
y2log T :D 1
Figure 11.2.

Controlling the Mean Cell Rate over Different Time Scales

172

USAGE PARAMETER CONTROL

statistical fluctuations of well-behaved traffic streams, but this safety

margin limits the ability of the UPC function to detect violations of the
negotiated mean cell rate. As the measurement interval is extended, the
safety margin required becomes less, but then any action in response to
contract violation may be too late to be an effective protection for network
resources.
Therefore we need to modify how we think of the mean cell rate: it is
necessary to think in terms of a virtual mean defined over some specified
time interval. The compromise is between the accuracy with which the
cell rate is controlled, and the timeliness of any response to violations
of the contract. Lets look at some algorithms which can monitor this
virtual mean.

ALGORITHMS FOR UPC

Methods to control peak cell rate, mean cell rate and different load states
within several time scales have been studied extensively [11.1]. The most
common algorithms involve two basic mechanisms:

the window method, which limits the number of cells in a time

window
the leaky bucket method, which increments a counter for each cell
arrival and decrements this counter periodically

The window method basically corresponds to the description given in the

previous section and involves choosing a time interval and a maximum
number of cells that can be admitted within that interval. We saw, with
the Poisson source example, that the method suffers from either rejecting
too many legitimate cells, or not rejecting enough when the contract is
violated. A number of variations of the method have been studied (the
jumping window, the moving window and the exponentially weighted
moving average), but there is not space to deal with them here.

The leaky bucket

It is generally agreed that the leaky bucket method achieves a better
performance compromise than the window method. Leaky buckets are
simple to understand and to implement, and flexible in application.
(Indeed, the continuous-state version of the leaky bucket algorithm is
used to define the generic cell rate algorithm (GCRA), for traffic contract
conformance see [10.1, 10.2].) Figure 11.3 illustrates the principle. Note
that a separate control function is required for each virtual channel or
virtual path being monitored.
A counter is incremented whenever a cell arrives; this counter, which
is called the bucket, is also decremented at a constant leak rate. If the

173

PEAK CELL RATE CONTROL USING THE LEAKY BUCKET

Cell
stream
Counter
value 5

Bucket
limit

3
2
1
0

Leak
rate

10 11 12

Figure 11.3. The Operation of the Leaky Bucket

traffic source generates a burst of cells at a rate higher than the leak rate,
the bucket begins to fill. Provided that the burst is short, the bucket will
not fill up and no action will be taken against the cell stream. If a long
enough burst of cells arrives at a rate higher than the leak rate, then the
bucket will eventually overflow. In this case, each cell that arrives to find
the counter at its maximum value is deemed to be in violation of the
traffic contract and may be discarded or tagged by changing the CLP
bit in the cell header from high to low priority. Another possible course
of action is for the connection to be released.
In Figure 11.3, the counter has a value of 2 at the start of the sequence.
The leak rate is one every four cell slots and the traffic source being
monitored is in a highly active state sending cells at a rate of 50% of
the cell slot rate. It is not until the tenth cell slot in the sequence that a
cell arrival finds the bucket on its limit. This non-conforming cell is then
subject to discard or tagging. An important point to note is that the cells
do not pass through the bucket, as though queueing in a buffer. Cells do
not queue in the bucket, and therefore there is no variable delay through
a leaky bucket. However, the operation of the bucket can be analysed as
though it were a buffer with cells being served at the leak rate. This then
allows us to find the probability that cells will be discarded or tagged by
the UPC function.

PEAK CELL RATE CONTROL USING THE LEAKY BUCKET

If life were simple, then peak cell rate control would just involve a leaky
bucket with a leak rate equal to the peak rate and a bucket depth of 1. The

174

USAGE PARAMETER CONTROL

problem is the impact of cell-delay variation (CDV), which is introduced

to the cell stream by the access network. Although a source may send
cells with a constant inter-arrival time at the peak rate, those cells have
to go through one or more buffers in the access network before they
are monitored by the UPC algorithm on entry to the public network.
The effect of queueing in those buffers is to vary the amount of delay
experienced by each cell. Thus the time between successive cells from
the same connection may be more than or less than the declared constant
inter-arrival time.
For example, suppose there are 5 CBR sources, each with a peak rate
of 10% of the cell slot rate, i.e. 1 cell every 10 slots, being multiplexed
through an access switch with buffer capacity of 20 cells. If all the sources
are out of phase, then none of the cells suffers any queueing delay in the
access switch. However, if all the sources are in phase, then the worst
delay will be for the last cell in the batch, i.e. a delay of 4 cell slots (the
cell which is first to arrive enters service immediately and experiences no
delay). Thus the maximum variation in delay is 4 cell slots. This worst
case is illustrated in Figure 11.4. At the source, the inter-arrival times
between cells 1 and 2, T12 , and cells 2 and 3, T23 , are both 10 cell slots.
However, cell number 2 experiences the maximum CDV of 4 cell slots,
and so, on entry to the public network, the time between cells 2 and 3,
T23 , is reduced from 10 cell slots to 6 cell slots. This corresponds to a rate
increase from 10% to 16.7% of the cell slot rate, i.e. a 67% increase on the
declared peak cell rate.
It is obvious that the source itself is not to blame for this apparent
increase in its peak cell rate; it is just a consequence of multiplexing in
the access network. However, a strict peak cell rate control, with a leak
rate of 10% of the cell slot rate and a bucket limit of 1, would penalize
the connection by discarding cell number 3. How is this avoided? A
CDV tolerance is needed for the UPC function, and this is achieved by
increasing the leaky bucket limit.
T 12 = 10 cell slots
1

T 23 = 10 cell slots
2

Cell stream at source

max. CDV of 4 cell slots

T 12 = 14 cell slots
1

T 23 = 6 cell slots
2

Cell stream on entry to network

3
Time

Figure 11.4. Effect of CDV in Access Network on Inter-Arrival Times

175

PEAK CELL RATE CONTROL USING THE LEAKY BUCKET

Lets see how the leaky bucket would work in this situation. First, we
must alter slightly our leaky bucket algorithm so that it can deal with any
values of T (the inter-arrival time at the peak cell rate) and  (the CDV
tolerance). The leaky bucket counter works with integers, so we need to
find integers k and n such that
 Dk

T
n

i.e. the inter-arrival time at the peak cell rate is divided into n equal parts,
with n chosen so that the CDV tolerance is an integer multiple, k, of T/n.
Then we operate the leaky bucket in the following way: the counter is
incremented by n (the splash) when a cell arrives, and it is decremented
at a leak rate of n/T. If the addition of a splash takes the counter above its
limit of k C n, then the cell is in violation of the contract and is discarded
or tagged. If the counter value is greater than n but less than or equal to
k C n, then the cell is within the CDV tolerance and is allowed to enter
the network.
Figure 11.5 shows how the counter value changes for the three cell
arrivals of the example of Figure 11.4. In this case, n D 10, k D 4, the leak
rate is equal to the cell slot rate, and the leaky bucket limit is k C n D 14.
We assume that, when a cell arrives at the same time as the counter is
decremented, the decrement takes place first, followed by the addition of
the splash of n. Thus in the example shown the counter reaches, but does
not exceed, its limit at the arrival of cell number 3. This is because the
inter-arrival time between cells 2 and 3 has suffered the maximum CDV
permitted in the traffic contract which the leaky bucket is monitoring.
Figure 11.6 shows what happens for the case when cell number 2 is
delayed by 5 cell slots rather than 4 cell slots. The counter exceeds its
limit when cell number 3 arrives, and so that cell must be discarded
because it has violated the traffic contract.
Bucket
limit

15
Counter
value

10

3
1

0
CDV = 4 cell slots
Figure 11.5. Example of Cell Stream with CDV within the Tolerance

176

USAGE PARAMETER CONTROL

15
Counter
value

Traffic
contract
violation

Bucket
limit
3

10

0
CDV = 5 cell slots
Figure 11.6.

Example of Cell Stream with CDV Exceeding the Allowed Tolerance

The same principle applies if the tolerance, , exceeds the peak rate
inter-arrival time, T, i.e. k > n. In this case it will take a number of
successive cells with inter-arrival times less than T for the bucket to build
up to its limit. Note that this extra parameter, the CDV tolerance, is an
integral part of the traffic contract and must be specified in addition to
the peak cell rate.

The problem of tolerances

When the CDV is greater than or equal to the inter-arrival time at the peak
cell rate the tolerance in the UPC function presents us with a problem. It
is now possible to send multiple cells at the cell slot rate. The length of
this burst is limited by the size of the bucket, but if the bucket is allowed
to recover, i.e. the counter returns to zero, then another burst at the cell
slot rate can be sent, and so on. Thus the consequence of introducing
tolerances is to allow traffic with quite different characteristics to conform
to the traffic contract.
An example of this worst-case traffic is shown in Figure 11.7. The traffic
contract is for a high-bandwidth (1 cell every 5 cell slots) CBR connection.
With a CDV tolerance of 20 cell slots, we have n D 1, k D 4, the leak rate
is the peak cell rate (20% of the cell slot rate), and the leaky bucket limit
is k C n D 5. However, this allows a group of 6 cells to pass unhindered
at the maximum cell rate of the link every 30 cell slots! So this worst-case
traffic is an on/off source of the same mean cell rate but at five times the
peak cell rate.
How do we calculate this maximum burst size (MBS) at the cell slot
rate, and the number of empty cell slots (ECS) between such bursts? We
need to analyse the operation of the leaky bucket as though it were a queue
with cells (sometimes called splashes) arriving and being served. The

177

PEAK CELL RATE CONTROL USING THE LEAKY BUCKET

CBR cell stream

1 cell every 5 slots

Maximum
burst size

Time for leaky bucket to recover

1 2 3 4 5 6

1 2 3 4 5 6

Worst case cell stream

Figure 11.7.

Time

Worst-Case Traffic for Leaky Bucket with CDV Tolerance

first n units in the leaky bucket effectively act as the service space for a
splash. These n units are required for the leaky bucket to operate correctly
on a peak cell rate whether or not there is any CDV tolerance. Thus it is
in the extra k units where a queue forms, and so the leaky bucket limit of
k C n is equivalent to the system capacity.
So, we can analyse the formation of a queue by considering the time
taken, tMBS , for an excess rate to fill up the leaky buckets queueing space, k:
tMBS D

queueing space
k
D
excess rate
n CSR  n PCR

where CSR is the cell slot rate, and PCR is the peak cell rate. We also
know that
n
k D  D  n PCR
T
so, substituting for k gives
tMBS D

 n PCR
 PCR
D
n CSR  PCR
CSR  PCR

The maximum burst size is found by multiplying tMBS by the cell slot rate
and adding one for the first cell in the burst which fills the server space, n:


MBS D 1 C bCSR tMBS c D 1 C

 CSR PCR
CSR  PCR

which, in terms of the inter-arrival time at the peak rate, T, and the cell
slot time, , is



MBS D 1 C
T

178

USAGE PARAMETER CONTROL

We take the integer part of this calculation because there cannot be

fractions of cells in a burst. For the example given in Figure 11.7, we have



20
D 6 cells
MBS D 1 C
51

The time taken, tcycle , for the leaky bucket to go through one complete
cycle of filling, during the maximum burst, and then emptying during
the silence period, is given by
n MBS D n PCR tcycle
where the left-hand side gives the total number of units by which the
leaky bucket is incremented, and the right-hand side gives the total
number by which it is decremented. The total number of cell slots in a
complete cycle is CSR tcycle . It is necessary to round this up to the nearest
integer number of cell slots to ensure that the leaky bucket empties
completely, so the number of empty cell slots is



MBS
 MBS
ECS D CSR
PCR

For the example given in Figure 11.7, we have



6
 6 D 24 cells
ECS D 1
0.2

Resources required for a worst-case ON/OFF cell stream from peak

cell rate UPC
Continuing with this example, suppose that there are five of these CBR
sources each being controlled by its own leaky bucket with the parameter
values calculated. After the UPC function, the cell streams are multiplexed
through an ATM buffer of capacity 20 cells. If the sources do in fact behave
according to their declared contract, i.e. a peak cell rate of 20% of the cell
slot rate, then there is no cell loss. In any five cell slots, we know that
there will be exactly five cells arriving; this can be accommodated in the
ATM buffer without loss.
But if all five sources behave as the worst-case ON/OFF cell stream,
then the situation is different. We know that in any 30 cell slots there
will be exactly 30 cells arriving. Whether the buffer capacity of 20 cells is
sufficient depends on the relative phasing of the ON/OFF cell streams.
If all five sources send a maximum burst at the same time, then 30 cells
will arrive during six cell slots. This is an excess of 24 cells, 20 of which
can be buffered, and 4 of which must be lost.

PEAK CELL RATE CONTROL USING THE LEAKY BUCKET

179

If we reduce the number of sources to four, their worst-case behaviour

will produce an aggregate burst of 24 cells arriving during six cell slots.
This is an excess of 18 cells, all of which can be buffered. Thus the
performance can be maintained only if the number of sources, and hence
the admissible load, is reduced.
The example we have used is a very simple one that demonstrates the
issue. A reduction from five sources admitted to four may not seem to
be a severe consequence of CDV tolerances. In general, each CBR source
of peak cell rate h cell/s is, in the worst case, being considered as an
ON/OFF source with a mean cell rate of h cell/s and a peak cell rate equal
to the cell slot rate of the link. This can reduce the admissible load by a
significant amount.
We can estimate this load reduction by applying the ND/D/1 analysis
to the worst-case traffic streams. The application of this analysis rests on
the observation that the worst-case ON/OFF source is in fact periodic,
with period MBS D. Each arrival is a burst of fixed size, MBS, which
takes MBS cell slots to be served, so the period can be described as one
burst in every D burst slots. A buffer of size X cells can hold X/MBS
bursts, so we can adapt the analysis to deal with bursts rather than cells
just by scaling the buffer capacity. The analysis does not account for the
partial overlapping of bursts, but since we are only after an estimate we
will neglect this detail.
The approximate analysis for the ND/D/1 queue tends to underestimate the loss particularly when the load is not heavy. The effect of this
is to overestimate the admissible load for a fixed CLP requirement. But
the great advantage is that we can manipulate the formula to give the
admissible load,
, as a function of the other parameters, X and D:

2 X C D

lnCLP
D 2
X

with the proviso that the load can never exceed a value of 1. This formula
applies to the CBR cell streams. For the worst-case streams, we just
replace X by X/MBS to give:

X
2
CD
MBS

D
MBS lnCLP
D 2
X
where





/
D1C
MBS D 1 C
T
D1

Note that D is just the inter-arrival time, T, in units of the cell slot time, .

180

USAGE PARAMETER CONTROL

Table 11.2 shows the number of sources (N D

D) that can be
admitted for different CDV values and with a CLP requirement of
1010 . It is assumed that the output cell streams from N UPC functions
are multiplexed together over a 155.52 Mbit/s link (i.e. a cell slot rate of
353 208 cell/s). The link buffer has a capacity of 50 cells. The CDV tolerance allowed by the leaky buckets takes values of 20, 40, 60, 80 and 100 cell
Table 11.2. Number of CBR Sources that can Be Admitted over a 155.52 Mbit/s Link
with Buffer Capacity of 50 Cells, for Different CDV Tolerances and a CLP of 1010

Cell delay variation tolerance

period, D cell rate / D 0 slots 20 slots 40 slots 60 slots 80 slots 100 slots
(slots)
(cell/s)
 D 0 ms
0.057 ms 0.113 ms 0.170 ms 0.226 ms 0.283 ms
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100

35 321
32 110
29 434
27 170
25 229
23 547
22 075
20 777
19 623
18 590
17 660
16 819
16 055
15 357
14 717
14 128
13 585
13 082
12 615
12 180
11 774
10 092
8 830
7 849
7 064
6 422
5 887
5 434
5 046
4 709
4 415
4 155
3 925
3 718
3 532

10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100

10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100

9
9
12
13
13
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100

6
6
8
8
11
11
12
15
15
16
16
17
22
23
24
24
25
25
26
27
27
35
40
45
50
54
58
65
70
75
80
85
90
95
100

5
5
6
7
8
9
10
10
13
13
13
14
17
18
19
19
20
20
26
27
27
30
33
45
50
54
58
61
65
68
71
85
90
95
100

3
4
5
6
7
7
8
9
11
11
11
12
14
15
15
16
16
20
21
21
22
30
33
36
39
54
58
61
65
68
71
75
78
82
85

181

PEAK CELL RATE CONTROL USING THE LEAKY BUCKET

slots (corresponding to time values of 0.057, 0.113, 0.17, 0.226 and 0.283 ms
respectively). The peak cell rates being monitored vary from 1% up to 10%
of the cell slot rate. If the CDV tolerance is zero, then in this case the link
can be loaded to 100% of capacity for each of the peak cell rates shown.
Figure 11.8 plots the data of Table 11.2 as the admissible load against the
monitored peak cell rate. Note that when the CDV is relatively small (e.g.
40 cell slots or less), then there is little or no reduction in the admissible

Admissible load

100.0
9
8
7
6
5
4
3

0.057 ms
0.113 ms
0.170 ms
0.226 ms
0.283 ms

101.0
0

10000

20000
Peak cell rate

30000

40000

k :D 1.. 35


1
max
X, CSR, D, , CLP :D 
CSR







 MaxBS
1 C floor

D

1





X

2
CD

MaxBS



MaxBS lnCLP

D 2

X


floorD

N

N
D if N > D

N

D
Figure 11.8. Admissible Load for CBR Sources with Different CDV Tolerances

182

USAGE PARAMETER CONTROL


Dk :D  k C 9 if k < 22

 k  21 5 C 30 otherwise
CSR :D 353207.5
CSR
PCRk :D
Dk
CLP :D 1010

20
y1k :D max
50, CSR, Dk ,
, CLP
CSR

40
, CLP
y2k :D max
50, CSR, Dk ,
CSR

60
, CLP
y3k :D max
50, CSR, Dk ,
CSR

80
, CLP
y4k :D max
50, CSR, Dk ,
CSR

100
, CLP
y5k :D max
50, CSR, Dk ,
CSR
Figure 11.8. (continued)

load in this example. The CDV in the access network may well be of this
order, particularly if the access network utilization is low and buffers are
dimensioned to cope with only cell-scale and not burst-scale queueing.

Traffic shaping
One solution to the problem of worst-case traffic is to introduce a spacer
after the leaky bucket in order to enforce a minimum time between
cells, corresponding to the particular peak cell-rate being monitored by
the leaky bucket. Alternatively, this spacer could be implemented before
the leaky bucket as per-VC queueing in the access network. Spacing
is performed only on those cells that conform to the traffic contract;
this prevents the bunching together of cells (whether of the worst-case
traffic or caused by variation in cell delay within the CDV tolerance of the
traffic contract). However, spacing introduces extra complexity, which
is required on a per-connection basis. The leaky bucket is just a simple
countera spacer requires buffer storage and introduces delay.

DUAL LEAKY BUCKETS: THE LEAKY CUP AND SAUCER

Consider the situation for a variable-rate source described by a peak cell
rate and a mean cell rate. This can be monitored by two leaky buckets:

183

DUAL LEAKY BUCKETS: THE LEAKY CUP AND SAUCER

one to control the peak cell rate, the other to control a virtual mean
cell rate. In ITU Recommendation I.371 the term used for this virtual
mean is sustainable cell rate (SCR). With two leaky buckets, the effect
of the CDV tolerance on the peak-cell-rate leaky bucket is not so severe.
The reason is that the leaky bucket for the sustainable cell rate limits
the number of worst-case bursts that can pass through the peak-cell-rate
leaky bucket. For each ON/OFF cycle at the cell slot rate the SCR leakybucket level increases by a certain amount. When the SCR leaky bucket
reaches its limit, the ON/OFF cycles must stop until the SCR counter
has returned to zero. So the maximum burst size is still determined by
the PCR leaky-bucket parameter values, but the overall mean cell rate
allowed onto the network is limited by the sustainable cell rate rather
than the peak cell rate.
This dual leaky bucket arrangement is called the leaky cup and
saucer. The cup is the leaky bucket for the sustainable cell rate: it is a
deep container with a base of relatively small diameter. The saucer is
the leaky bucket for the peak cell rate: it is a shallow container with a
large-diameter base. The depth corresponds to the bucket limit and the
diameter of the base to the cell rate being controlled.
The worst-case traffic is shown in Figure 11.9(a). The effect of the leaky
buckets is to limit the number of cells over different time periods. For
the example in the figure, the saucer limit is 2 cells in 4 cell slots and
the cup limit is 6 cells in 24 cell slots. An alternative worst-case traffic
which is adopted in ITU Recommendation E.736 [10.3] is an ON/OFF
source with maximum-length bursts at the peak cell rate rather than at
the cell slot rate. An example of this type of worst-case traffic is shown
in Figure 11.9(b). Note that the time axis is in cell slots, so the area under
the curve is equal to the number of cells sent.
The maximum burst size at the peak cell rate is obtained in a similar
way to that at the cell slot rate, i.e.


MBS D 1 C

IBT
TSCR  TPCR

where IBT is called the intrinsic burst tolerance. This is another important
parameter in the traffic contract (in addition to the inter-arrival times TSCR
and TPCR for the sustainable and peak cell rates respectively). The purpose
of the intrinsic burst tolerance is in fact to specify the burst length limit
in the traffic contract.
Two CDV tolerances are specified in the traffic contract. We are already
familiar with the CDV tolerance, , for the peak cell rate. From now on we
call this PCR to distinguish it from the CDV tolerance for the sustainable
0
. This latter has to be added to the intrinsic burst tolerance
cell rate, SCR
in order to determine the counter limit for the cup. As before, we need to

184

USAGE PARAMETER CONTROL

Saucer reaches limit Cup reaches limit

1

Cell slot rate

Cell rate

0.75
0.5

Peak cell rate

0.25
0

Sustainable cell rate

0

(a)

5
10
Saucer empties

15

Cell rate

30
Time

Cell slot rate

0.75

Cup reaches limit

0.5

Peak cell rate

0.25
0

20
25
Cup empties

Sustainable cell rate

0

10

15

(b)

20
25
Cup empties

30
Time

Figure 11.9. Worst-Case Traffic through Leaky Cup and Saucer

find integers, k and n, such that

0
Dk
IBT C SCR

TSCR
n

In most cases, n can be set to 1 because the intrinsic burst tolerance will
be many times larger than TSCR .

Resources required for a worst-case ON/OFF cell stream from

sustainable-cell-rate UPC
Neither type of worst-case traffic shown in Figure 11.9 easy to analyse. In
the following analysis we use the maximum burst size for the sustainable
cell rate, and assume that that burst actually arrives at the cell slot rate.
Whether or not this is possible depends on the size of the saucer, and
hence on PCR . It is likely to be the worst of all possible traffic streams
because it generates the largest burst size.

185

DUAL LEAKY BUCKETS: THE LEAKY CUP AND SAUCER

The same approximate analytical approach is taken as before. In this

case D is the inter-arrival time, TSCR , in units of the cell slot time, .

X
CD
2
MBS

D
MBS lnCLP
D 2
X
A graph of utilization against the maximum burst size is shown in
Figure 11.10. The CLP requirement varies from 104 down to 1010 . The
1.0

Admissible load

CLP = 1e04
CLP = 1e06
CLP = 1e08
CLP = 1e10

0.5

0.0
0

10

20
30
Maximum burst size

k :D 1.. 50
2
max
CS X, D, MaxBS, CLP :D

40

X
CD
MaxBS

50

MaxBS lnCLP
D 2
X

MBSk :D k
y1k :D max
CS 50, 100, MBSk , 104 
y2k :D max
CS 50, 100, MBSk , 106 
y3k :D max
CS 50, 100, MBSk , 108 
y4k :D max
CS 50, 100, MBSk , 1010 
Figure 11.10. Admissible Load for Worst-Case Traffic through Leaky Cup and
Saucer

186

USAGE PARAMETER CONTROL

link buffer has a capacity of 50 cells, the cell slot rate is 353 208 cell/s,
and the sustainable cell rate is chosen to be 3532 cell/s, i.e. D D 100. The
maximum burst size allowed by the leaky cup and saucer is varied from
1 up to 50 cells. The peak cell rate and intrinsic burst tolerance are not
specified explicitly; different combinations can be calculated from the
maximum burst size and the sustainable cell rate.
It is important to use the correct value of MBS because this obviously
can have a significant effect on the admissible load. Suppose that the
peak cell rate is twice the sustainable cell rate, i.e. TPCR D TSCR /2. The
maximum burst size at the peak cell rate is

MBSPCR

D1C




IBT
D 1 C 2 IBT
TSCR
T

TSCR 

SCR

and the maximum burst size at the cell slot rate is





IBT
IBT
1C
MBSCSR D 1 C
TSCR 
TSCR
The difference between these two maximum size bursts is almost a factor
of two (for reasonable values of the intrinsic burst tolerance), and this
corresponds to a difference in the admissible load of a factor of roughly
0.6 across the range of burst sizes in the graph. So the assumption that
the worst-case traffic is based on the maximum burst size at the peak cell
rate carries with it a 40% penalty on the admissible load.

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

12

Dimensioning
real networks dont lose cells?

COMBINING THE BURST AND CELL SCALES

The finite-capacity buffer is a fundamental element of ATM where cells
multiplexed from a number of different input streams are temporarily
stored awaiting onward transmission. The flow of cells from the different
inputs, the number of inputs, and the rate at which cells are served
determine the occupancy of the buffer and hence the cell delay and cell
loss experienced. So, how large should this finite buffer be?
In Chapters 8 and 9 we have seen that there are two elements of
queueing behaviour: the cell-scale and burst-scale components. We evaluated the loss from a finite buffer for constant bit-rate, variable bit-rate
and random traffic sources. For random traffic, or for a mix of CBR traffic,
only the cell-scale component is present. But when the traffic mix includes
bursty sources, such that combinations of the active states can exceed the
cell slot rate, then both components of queueing are present.
Lets look at each type of traffic and see how the loss varies with the
buffer size for different offered loads. We can then develop strategies
for buffer dimensioning based on an understanding of this behaviour.
First, we consider VBR traffic; this combines the cell-scale component
of queueing with both the loss and delay factors of the burst-scale
component of queueing.
Figure 9.14 shows how the burst-scale loss factor varies with the
number of sources, N, where each source has a peak cell rate of
24 000 cell/s and a mean cell rate of 2000 cell/s. From Table 9.2 we
find that the minimum number of these sources required for burst-scale
queueing is N0 D 14.72. Table 12.1 gives the burst-scale loss factor, CLPbsl ,
at three different values of N (30, 60 and 90 sources) as well as the offered
load as a fraction of the cell slot rate (calculated using the bufferless
analysis in Chapter 9). These values of load are used to calculate both the

188

DIMENSIONING

Table 12.1. Burst-Scale Loss

Factor for N VBR Sources

CLPbsl

load

30
60
90

4.46E-10
1.11E-05
9.10E-04

0.17
0.34
0.51

Buffer capacity, X
0

10

20

30

40

50

60

70

80

90

100

100
101

Cell loss probability

102

N = 90
N = 60
N = 30

103
104
105
106
107
108
109
1010

Nm
OverallCLP X, N, m, h, C, b :D


C




C

N0


h






m




h



for i 2 0 .. X



ai
Poisson i, 




finiteQloss X, a, 

csloss



bsloss
BSLexact , N, N0 BSDapprox N0, X, b, 



csloss C bsloss
xk :D k
y1k :D OverallCLP k , 90 , 2000 , 24000 , 353207.5 , 480 
y2k :D OverallCLP k , 60 , 2000 , 24000 , 353207.5 , 480 
y3k :D OverallCLP k , 30 , 2000 , 24000 , 353207.5 , 480 
k :D 2.. 100

Figure 12.1.

Overall Cell Loss Probability against Buffer Capacity for N VBR Sources

189

COMBINING THE BURST AND CELL SCALES

cell-scale queueing component, CLPcs , and the burst-scale delay factor,

CLPbsd , varying with buffer capacity.
The combined results are plotted in Figure 12.1. The cell-scale component is obtained using the exact analysis of the finite M/D/1 described
in Chapter 7. The burst-scale delay factor uses the same approach as that
for calculating the values in Figure 9.16. For Figure 12.1, an average burst
length, b, of 480 cells is used. The overall cell loss shown in Figure 12.1 is
calculated by summing the burst- and cell-scale components of cell loss,
where the burst-scale component is the product of the loss and delay
factors, i.e.
CLP D CLPcs C CLPbsl CLPbsd
Now, consider N CBR sources where each source has a constant cell rate
of 2000 cell/s. Figure 12.2 shows how the cell loss varies with the buffer
Buffer capacity, X
0

10

20

30

40

50

60

70

80

90

100

100
101

Cell loss probability

102
103
104
105
106
107
108

N = 170
N = 150
N = 120

109
1010

k :D 0.. 50
xk :D k


170 2000
353207. 5 

150 2000
y2k :D NDD1Q k , 150 ,
353207. 5 

120 2000
y3k :D NDD1Q k , 120 ,
353207. 5

y1k :D NDD1Q

k , 170 ,

Figure 12.2. Cell Loss Probability against Buffer Capacity for N CBR Sources

190

DIMENSIONING

Buffer capacity, X
0

10

20

30

40

50

60

70

80

90

100

100
load = 0.96
load = 0.85
load = 0.68

101

Cell loss probability

102
103
104
105
106
107
108
109
1010

k :D 0.. 100
aP68k :D Poisson k , 0 . 68
aP85k :D Poisson k , 0 . 85
aP96k :D Poisson k , 0 . 96
i :D 2.. 100
xi :D i
y1i :D finiteQloss xi , aP68 , 0 . 68
y2i :D finiteQloss xi , aP85 , 0 . 85
y3i :D finiteQloss xi , aP96 , 0 . 96
Figure 12.3. Cell Loss Probability against Buffer Capacity for Random Traffic

capacity for 120, 150 and 170 sources. The corresponding values for the
offered load are 0.68, 0.85, and 0.96 respectively. Figure 12.3 takes the load
values used for the CBR traffic and assumes that the traffic is random.
The cell loss results are found using the exact analysis for the finite
M/D/1 system. A summary of the three different situations is depicted
in Figure 12.4, comparing 30 VBR sources, 150 CBR sources, and an
offered load of 0.85 of random traffic (the same load as 150 CBR sources).

DIMENSIONING THE BUFFER

Figure 12.4 shows three very different curves, depending on the characteristics of each different type of source. There is no question that the

191

DIMENSIONING THE BUFFER

Buffer capacity, X
0

10

20

30

40

50

60

70

80

90

100

100
VBR
random
CBR

101

Cell loss probability

102
103
104
105
106
107
108
109
1010

k :D 0.. 100
aP85k :D Poisson k , 0 . 85
i :D 2.. 100
xi :D i


150 2000
y1i :D NDD1Q i , 150 ,
353207. 5
y2i :D finiteQloss xi , aP85 , 0 . 85
y3i :D OverallCLP i , 30 , 2000 , 24000 , 353207. 5 , 480
Figure 12.4. Comparison of VBR, CBR and Random Traffic through a Finite Buffer

buffer must be able to cope with the cell-scale component of queueing

since this is always present when a number of traffic streams are merged.
But we have two options when it comes to the burst-scale component, as
analysed in Chapter 9:
1.

Restrict the number of bursty sources so that the total input rate only
rarely exceeds the cell slot rate, and assume that all excess-rate cells
are lost. This is the bufferless or burst-scale loss option (also known
as rate envelope multiplexing).

2.

Assume that we have a big enough buffer to cope with excessrate cells, so only a proportion are lost; the other excess-rate cells are
delayed in the buffer. This is the burst-scale delay option (rate-sharing
statistical multiplexing).

192

DIMENSIONING

It is important to notice that how big we make the buffer depends on

how we intend to accept traffic onto the network (or vice versa). Also a
dimensioning choice has an impact on a control mechanism (connection
admission control).
For the first option, the buffer is dimensioned according to cell-scale
constraints. The amount of bursty traffic is not the limiting factor in
choosing the buffer capacity because the CAC restrictions on accepting
bursty traffic automatically limit the burst-scale component to a value
below the CLP requirement, and the CAC algorithm assumes that the
buffer size makes no difference. Thus for bursty traffic the mean utilization is low and the gradient of its cell-scale component is steep (see
Figure 12.1). However, for either constant-bit-rate or random traffic the
cell-scale component is much more significant (there is no burst-scale
component), and it is a realistic maximum load of this traffic that determines the buffer capacity. The limiting factor here is the delay through
the buffer, particularly for interactive services.
If we choose the second option, the amount of bursty traffic can be
increased to the same levels of utilization as for either constant-bit-rate
or random traffic the price to pay is in the size of the buffer which
must be significantly larger. The disadvantage with buffering the excess
(burst-scale) cells is that the delay through a large buffer can be too
great for services like telephony and interactive video, which negates
the aims of having an integrated approach to all telecommunications
services. There are ways around the problem segregation of traffic
through separate buffers and the use of time priority servers but this
does introduce further complexity into the network, see Figure 12.5.
We will look at traffic segregation and priorities in more detail in
Chapter 13.
Delay
sensitive
cells

Short buffer
Single
server

Loss
sensitive
cells

...
Long

Buffer
...

A time priority scheme would involve serving cells

in the short buffer before cells in the long buffer
Figure 12.5.

Time Priorities and Segregation of Traffic

193

DIMENSIONING THE BUFFER

Small buffers for cell-scale queueing

A comparison of random traffic and CBR traffic (see Figure 12.4) shows
that the cell-scale component of the random traffic gives a worse CLP
for the same load. Even with 1000 CBR sources, each of 300 cell/s (to
keep the load constant at 0.85), Table 10.3(b) shows that the cell loss is
about 109 for a buffer capacity of 50 cells. This is a factor of 10 lower
than for random traffic through the same size buffer.
So, to dimension buffers for cell-scale queueing we use a realistic
maximum load of random traffic. Table 12.2 uses the exact analysis for
Table 12.2

load

155.52 Mbit/s link

622.08 Mbit/s link

buffer
capacity

mean

maximum

mean

maximum

(cells)

delay s

delay s

delay s

delay s

(a) Buffer Dimensioning for Cell-Scale Queueing: Buffer Capacity, Mean and Maximum
Delay, Given the Offered Load and a Cell Loss Probability of 108
0.50
0.51
0.52
0.53
0.54
0.55
0.56
0.57
0.58
0.59
0.60
0.61
0.62
0.63
0.64
0.65
0.66
0.67
0.68
0.69
0.70
0.71
0.72
0.73
0.74
0.75
0.76
0.77

16
16
17
17
17
18
18
19
19
20
20
21
21
22
23
23
24
25
25
26
27
28
29
30
31
33
34
35

4.2
4.3
4.4
4.4
4.5
4.6
4.6
4.7
4.8
4.9
5.0
5.0
5.1
5.2
5.3
5.5
5.6
5.7
5.8
6.0
6.1
6.3
6.5
6.7
6.9
7.1
7.3
7.6

45.3
45.3
48.1
48.1
48.1
51.0
51.0
53.8
53.8
56.6
56.6
59.5
59.5
62.3
65.1
65.1
67.9
70.8
70.8
73.6
76.4
79.3
82.1
84.9
87.8
93.4
96.3
99.1

1.1
1.1
1.1
1.1
1.1
1.1
1.2
1.2
1.2
1.2
1.2
1.3
1.3
1.3
1.3
1.4
1.4
1.4
1.5
1.5
1.5
1.6
1.6
1.7
1.7
1.8
1.8
1.9

11.3
11.3
12.0
12.0
12.0
12.7
12.7
13.4
13.4
14.2
14.2
14.9
14.9
15.6
16.3
16.3
17.0
17.7
17.7
18.4
19.1
19.8
20.5
21.2
21.9
23.4
24.1
24.8
(continued overleaf )

194

DIMENSIONING

Table 12.2. (continued)

155.52 Mbit/s link

622.08 Mbit/s link

buffer
capacity

mean

maximum

mean

maximum

load

(cells)

delay s

delay s

delay s

delay s

0.78
0.79
0.80
0.81
0.82
0.83
0.84
0.85
0.86
0.87
0.88
0.89
0.90
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98

37
39
41
43
45
48
51
54
58
62
67
73
79
88
98
112
129
153
189
248
362

7.9
8.2
8.5
8.9
9.3
9.7
10.3
10.9
11.5
12.3
13.2
14.3
15.6
17.1
19.1
21.6
25.0
29.7
36.8
48.6
72.2

104.8
110.4
116.1
121.7
127.4
135.9
144.4
152.9
164.2
175.5
189.7
206.7
223.7
249.1
277.5
317.1
365.2
433.2
535.1
702.1
1024.9

2.0
2.0
2.1
2.2
2.3
2.4
2.6
2.7
2.9
3.1
3.3
3.6
3.9
4.3
4.8
5.4
6.3
7.4
9.2
12.2
18.0

26.2
27.6
29.0
30.4
31.9
34.0
36.1
38.2
41.1
43.9
47.4
51.7
55.9
62.3
69.4
79.3
91.3
108.3
133.8
175.5
256.2

(b) Buffer Dimensioning for Cell-Scale Queueing: Buffer Capacity, Mean and Maximum
Delay, Given the Offered Load and a Cell Loss Probability of 1010
0.50
0.51
0.52
0.53
0.54
0.55
0.56
0.57
0.58
0.59
0.60
0.61
0.62
0.63
0.64
0.65
0.66
0.67
0.68
0.69

19
20
20
21
21
22
23
23
24
24
25
26
26
27
28
29
30
31
32
33

4.2
4.3
4.4
4.4
4.5
4.6
4.6
4.7
4.8
4.9
5.0
5.0
5.1
5.2
5.3
5.5
5.6
5.7
5.8
6.0

53.8
56.6
56.6
59.5
59.5
62.3
65.1
65.1
67.9
67.9
70.8
73.6
73.6
76.4
79.3
82.1
84.9
87.8
90.6
93.4

1.1
1.1
1.1
1.1
1.1
1.1
1.2
1.2
1.2
1.2
1.2
1.3
1.3
1.3
1.3
1.4
1.4
1.4
1.5
1.5

13.4
14.2
14.2
14.9
14.9
15.6
16.3
16.3
17.0
17.0
17.7
18.4
18.4
19.1
19.8
20.5
21.2
21.9
22.6
23.4

195

DIMENSIONING THE BUFFER

Table 12.2. (continued)

155.52 Mbit/s link

622.08 Mbit/s link

buffer
capacity

mean

maximum

mean

maximum

load

(cells)

delay s

delay s

delay s

delay s

0.70
0.71
0.72
0.73
0.74
0.75
0.76
0.77
0.78
0.79
0.80
0.81
0.82
0.83
0.84
0.85
0.86
0.87
0.88
0.89
0.90
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98

34
35
37
38
39
41
43
45
47
49
51
54
57
60
64
68
73
79
85
93
102
113
126
144
167
199
246
324
476

6.1
6.3
6.5
6.7
6.9
7.1
7.3
7.6
7.9
8.2
8.5
8.9
9.3
9.7
10.3
10.9
11.5
12.3
13.2
14.3
15.6
17.1
19.1
21.6
25.0
29.7
36.8
48.6
72.2

96.3
99.1
104.8
107.6
110.4
116.1
121.7
127.4
133.1
138.7
144.4
152.9
161.4
169.9
181.2
192.5
206.7
223.7
240.7
263.3
288.8
319.9
356.7
407.7
472.8
563.4
696.5
917.3
1347.6

1.5
1.6
1.6
1.7
1.7
1.8
1.8
1.9
2.0
2.0
2.1
2.2
2.3
2.4
2.6
2.7
2.9
3.1
3.3
3.6
3.9
4.3
4.8
5.4
6.3
7.4
9.2
12.2
18.0

24.1
24.8
26.2
26.9
27.6
29.0
30.4
31.9
33.3
34.7
36.1
38.2
40.3
42.5
45.3
48.1
51.7
55.9
60.2
65.8
72.2
80.0
89.2
101.9
118.2
140.9
174.1
229.3
336.9

(c) Buffer Dimensioning for Cell-Scale Queueing: Buffer Capacity, Mean and Maximum
Delay, Given the Offered Load and a Cell Loss Probability of 1012
0.50
0.51
0.52
0.53
0.54
0.55
0.56
0.57
0.58
0.59
0.60
0.61

23
24
24
25
26
26
27
28
28
29
30
31

4.2
4.3
4.4
4.4
4.5
4.6
4.6
4.7
4.8
4.9
5.0
5.0

65.1
67.9
67.9
70.8
73.6
73.6
76.4
79.3
79.3
82.1
84.9
87.8

1.1
1.1
1.1
1.1
1.1
1.1
1.2
1.2
1.2
1.2
1.2
1.3

16.3
17.0
17.0
17.7
18.4
18.4
19.1
19.8
19.8
20.5
21.2
21.9
(continued overleaf )

196

DIMENSIONING

Table 12.2. (continued)

155.52 Mbit/s link

622.08 Mbit/s link

buffer
capacity

mean

maximum

mean

maximum

load

(cells)

delay s

delay s

delay s

delay s

0.62
0.63
0.64
0.65
0.66
0.67
0.68
0.69
0.70
0.71
0.72
0.73
0.74
0.75
0.76
0.77
0.78
0.79
0.80
0.81
0.82
0.83
0.84
0.85
0.86
0.87
0.88
0.89
0.90
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98

32
33
34
35
36
37
38
39
41
42
44
46
47
49
52
54
56
59
62
65
69
73
78
83
89
96
104
113
124
138
154
176
204
244
303
400
592

5.1
5.2
5.3
5.5
5.6
5.7
5.8
6.0
6.1
6.3
6.5
6.7
6.9
7.1
7.3
7.6
7.9
8.2
8.5
8.9
9.3
9.7
10.3
10.9
11.5
12.3
13.2
14.3
15.6
17.1
19.1
21.6
25.0
29.7
36.8
48.6
72.2

90.6
93.4
96.3
99.1
101.9
104.8
107.6
110.4
116.1
118.9
124.6
130.2
133.1
138.7
147.2
152.9
158.5
167.0
175.5
184.0
195.4
206.7
220.8
235.0
252.0
271.8
294.4
319.9
351.1
390.7
436.0
498.3
577.6
690.8
857.9
1132.5
1676.1

1.3
1.3
1.3
1.4
1.4
1.4
1.5
1.5
1.5
1.6
1.6
1.7
1.7
1.8
1.8
1.9
2.0
2.0
2.1
2.2
2.3
2.4
2.6
2.7
2.9
3.1
3.3
3.6
3.9
4.3
4.8
5.4
6.3
7.4
9.2
12.2
18.0

22.6
23.4
24.1
24.8
25.5
26.2
26.9
27.6
29.0
29.7
31.1
32.6
33.3
34.7
36.8
38.2
39.6
41.8
43.9
46.0
48.8
51.7
55.2
58.7
63.0
67.9
73.6
80.0
87.8
97.7
109.0
124.6
144.4
172.7
214.5
283.1
419.0

the finite M/D/1 queue to show the buffer capacity for a given load and
cell loss probability. The first column is the load, varying from 50% up to
98%, and the second column gives the buffer size for a particular cell loss
probability requirement (Table 12.2 part (a) is for a CLP of 108 , part (b)
is for 1010 , and part (c) is for 1012 ). Then there are extra columns which

197

DIMENSIONING THE BUFFER

give the mean delay and maximum delay through the buffer for link
rates of 155.52 Mbit/s and 622.08 Mbit/s. The maximum delay is just the
buffer capacity multiplied by the time per cell slot, s, at the appropriate
link rate. The mean delay depends on the load, , and is calculated using
the formula for an infinite M/D/1 system:
tq D s C

s
2 1  

This is very close to the mean delay through a finite M/D/1 because the
loss is extremely low (mean delays only differ noticeably when the loss
from the finite system is high).
Figure 12.6 presents the mean and maximum delay values from
Table 12.2 in the form of a graph and clearly shows how the delays
increase substantially above a load of about 80%. This graph can be used
to select a maximum load according to the cell loss and delay constraints,
and the buffers link rate; the required buffer size can then be read from
the appropriate part of Table 12.2.
So, to summarize, we dimension short buffers to cope with cell-scale
queueing behaviour using Table 12.2 and Figure 12.6. This approach is
applicable to networks which offer the deterministic bit-rate transfer
capability and the statistical bit-rate transfer capability based on rate
envelope multiplexing. For SBR based on rate sharing, buffer dimensioning requires a different approach, based on the burst-scale queueing
behaviour.

900
Delay in microseconds through
155.52 Mbit/s link buffer

250

Maximum delay

225

800

200

700

175

600

150

CLP = 108

500

125

CLP = 1010

400

100

CLP = 1012

300

75

200

50

100
0

25

Mean delay
0.5

0.6

0.7

0.8

0.9

Delay in microseconds through

622.08 Mbit/s link buffer

1000

Load

Figure 12.6. Mean and Maximum Delays for a Buffer with Link Rate of either
155.52 Mbit/s or 622.08 Mbit/s for Cell Loss Probabilities of 108 , 1010 and 1012

198

DIMENSIONING

Large buffers for burst-scale queueing

A buffer-dimensioning method for large buffers and burst-scale queueing
is rather more complicated than for short buffers and cell-scale queueing
because the traffic characterization has more parameters. For the cellscale queueing case, random traffic is a very good upper bound and it has
just the one parameter: arrival rate. In the burst-scale queueing case, we
must assume a traffic mix of many on/off sources, each source having the
same traffic characteristics (peak cell rate, mean cell rate, and the mean
burst length in the active state). For the burst-scale analytical approach
we described in Chapter 9, the key parameters are the minimum number
of peak rates required for burst-scale queueing, N0 , the ratio of buffer
capacity to mean burst length, X/b, the mean load, , and the cell loss
probability.
We have seen in Chapter 10 that the overall cell loss target can be
obtained by trial and error with tables; combining the burst-scale loss
and burst-scale delay factors from Table 10.4 and Table 10.5 respectively.
Here, we present buffer-dimensioning data in two alternative graphical
forms: the variation of X/b with load for a fixed value of N0 and different
overall CLP values (Figure 12.7); and the variation of X/b with load for a
fixed overall CLP value and different values of N0 (Figure 12.8).
To produce a graph like that of Figure 12.7, we take one row from
Table 10.4, for a particular value of N0 . This gives the cell loss contribution, CLPbsl , from the burst-scale loss factor varying with offered load

X/b, buffer capacity in units of

average burst length

10000

1000

100
CLP = 1012

CLP = 1010

10
CLP = 108
1

0.5

0.6

0.7

0.8

0.9

1.0

Load

Figure 12.7. Buffer Capacity in Units of Mean Burst Length, Given Load and Cell
Loss Probability, for Traffic with a Peak Cell Rate of 1/100 of the Cell Slot Rate (i.e.
N0 D 100)

199

COMBINING THE CONNECTION, BURST AND CELL SCALES

0.0

0.1

0.2

0.3

0.4

Load
0.5

0.6

0.7

0.8

0.9

1.0

X/b, buffer capacity in units of

average burst length

1000

100

10

1
N0 = 10, 20, ... 100, 200, ... 700
Figure 12.8. Buffer Capacity in Units of Mean Burst Length, Given Load and the
Ratio of Cell Slot Rate to Peak Cell Rate, for a Cell Loss Probability of 1010

(Table 12.3). We then need to find the value of X/b which gives a cell
loss contribution, CLPbsd , from the burst-scale delay factor, to meet the
overall CLP target. This is found by rearranging the equation


X 13
 N0 b 4C1
CLPtarget
D CLPbsd D e
CLPbsl

to give

4C1
X

D
b
1  3

ln

CLPtarget
CLPbsl
N0

Table 12.3 gives the figures for an overall CLP target of 1010 , and
Figure 12.7 shows results for three different CLP targets: 108 , 1010 and
1012 . Figure 12.8 shows results for a range of values of N0 for an overall
CLP target of 1010 .
Table 12.3. Burst-Scale Parameter Values for N0 D 100 and a CLP Target of 1010

CLPbsl

101

102

103

104

105

106

107

108

109

1010

load, 

0.94

0.87

0.80

0.74

0.69

0.65

0.61

0.58

0.55

0.53

CLPbsd

109

108

107

106

105

104

103

102

101

100

5064.2

394.2

84.6

31.1

14.7

7.7

4.1

2.1

0.8

X/b

200

DIMENSIONING

COMBINING THE CONNECTION, BURST AND CELL SCALES

We have seen in Chapter 10 that connection admission control can be
based on a variety of different algorithms. An important grade of service
parameter, in addition to cell loss and cell delay, is the probability of
a connection being blocked. This is very much dependent on the CAC
algorithm and the characteristics of the offered traffic types, and in general
it is a difficult task to evaluate the connection blocking probability.
However, if we restrict the CAC algorithm to one that is based on
limiting the number of connections admitted then we can apply erlangs
lost call formula to the situation. The service capacity of an ATM link
is effectively being divided into N circuits. If all of these circuits are
occupied, then the CAC algorithm will reject any further connection
attempts. It is worth noting that the cell loss and cell delay performance
requirements determine the maximum number of connections that can
be admitted. Thus, for much of the time, the traffic mix will have fewer
connections, and the cell loss and cell delay performance will be rather
better than that specified in the traffic contract requirements.
Consider the situation with constant-bit-rate traffic. Figure 12.9 plots
the cell loss from a buffer of capacity 10 cells, for a range of CBR sources
where D is the number of slots between arrivals. Thus, with a particular
CLP requirement, and a constant cell rate given by
hD

1E+00

10

C
cell/s
D

Maximum number of connections

20
30
50
60
70
40
D=60

Cell loss probability

1E02

100
D=100

D=40

1E03
1E05
1E06

90

D=80

1E01

1E04

80

D=20

D=150
D=200

1E07
1E08
1E09

D=300

1E10
1E11
1E12
Figure 12.9. Maximum Number of CBR Connections, Given D Cell Slots between
Arrivals and CLP

201

COMBINING THE CONNECTION, BURST AND CELL SCALES

(where C is the cell slot rate), we can find from Figure 12.9 the maximum
number of connections that can be admitted onto a link of cell rate C.
The link cannot be loaded to more than 100% capacity, so the maximum
possible number of sources of any particular type is numerically equal
to the (constant) number of cell slots between arrivals. Lets take an
example. Suppose we have CBR sources of cell rate 3532 cell/s being
multiplexed over a 155.52 Mbit/s link, with a CLP requirement of 107 .
This gives a value of 100 for D, and from Figure 12.9, the maximum
number of connections is (near enough) 50.
Now that we have a figure for the maximum number of connections,
we can calculate the offered traffic at the connection level for a given
probability of blocking. Figure 12.10 shows how the connection blocking
probability varies with the maximum number of connections for different
offered traffic intensities. With our example, we find that for 50 connections maximum and a connection blocking probability of 0.02, the offered
traffic intensity is 40 erlangs. Note that the mean number of connections
in progress is numerically equal to the offered traffic, i.e. 40 connections.
The cell loss probability for this number of connections can be found
from Figure 12.9: it is 2 109 . This is over an order of magnitude lower
than the CLP requirement in the traffic contract, and therefore provides
a useful safety margin.
For variable-bit-rate traffic, we will only consider rate envelope multiplexing and not rate sharing. Figure 12.11 shows how the cell loss

Probability of connection blocking

10

20

Maximum number of connections

30
40
50
60
70
80

90

0.1

100

100 E
90 E

0.01
80 E
0.001

70 E

0.0001
10 E

20 E

30 E

40 E

50 E

60 E

Figure 12.10. Probability of Blocking, Given Maximum Number of Connections and

Offered Traffic

202

Cell loss probability

DIMENSIONING

1E+00
1E01
1E02
1E03
1E04
1E05
1E06
1E07
1E08
1E09
1E10
1E11
1E12

0.1

0.2

0.3

0.4

Load
0.5
0.6

0.7

0.8

0.9

N0 =10, 20, ... 100

Figure 12.11. Maximum Utilization for VBR Connections, Given N0 and CLP

probability from a (short) buffer varies with the utilization for a range of
VBR sources of different peak cell rates. The key parameter defining this
relationship is N0 , the ratio of the cell slot rate to the peak cell rate. Given
N0 and a CLP requirement, we can read off a value for the utilization.
This then needs to be multiplied by the ratio of the cell slot rate, C, to the
mean cell rate, m, to obtain the maximum number of connections that can
be admitted onto the link.
So, for example, for sources with a peak cell rate of 8830 cell/s and
a mean cell rate of 3532 cell/s being multiplexed onto a 155.52 Mbit/s
link, N0 is 40 and, according to Figure 12.11, the utilization is about
0.4 for a CLP of 108 . This utilization is multiplied by 100 (i.e. C/m)
to give a maximum of 40 connections. From Figure 12.10, an offered
traffic intensity of 30 erlangs gives a connection blocking probability of
just under 0.02 for 40 connections maximum. Thus the mean number
of connections in progress is 30, giving a mean load of 0.3, and from
Figure 12.11 the corresponding cell loss probability is found to be 1011 .
This is three orders of magnitude better than for the maximum number
of connections.
As an alternative to using Figure 12.10, a traffic table based on Erlangs
lost call formula is provided in Table 12.4.
So, we have seen that, for both CBR and VBR traffic, when the connection blocking probability requirements are taken into account, the actual
cell loss probability can be rather lower than that for the maximum
allowed number of connections.

203

COMBINING THE CONNECTION, BURST AND CELL SCALES

Table 12.4.

Traffic Table Based on Erlangs Lost Call Formula

Probability of blocking a connection

N

0.1

0.05

0.02

0.01

0.005

0.001

0.0001

0.02
0.22
0.60
1.09
1.65
2.27
2.93
3.62
4.34
5.08
5.84
6.61
7.40
8.20
9.00
9.82
10.65
11.49
12.33
13.18
14.03
14.89
15.76
16.63
17.50
18.38
19.26
20.15
21.03
21.93
22.82
23.72
24.62
25.52
26.43
27.34
28.25
29.16
30.08
30.99

0.01
0.15
0.45
0.86
1.36
1.90
2.50
3.12
3.78
4.46
5.15
5.87
6.60
7.35
8.10
8.87
9.65
10.43
11.23
12.03
12.83
13.65
14.47
15.29
16.12
16.95
17.79
18.64
19.48
20.33
21.19
22.04
22.90
23.77
24.63
25.50
26.37
27.25
28.12
29.00

0.005
0.10
0.34
0.70
1.13
1.62
2.15
2.72
3.33
3.96
4.61
5.27
5.96
6.66
7.37
8.09
8.83
9.57
10.33
11.09
11.85
12.63
13.41
14.20
14.99
15.79
16.59
17.40
18.21
19.03
19.85
20.67
21.50
22.33
23.16
24.00
24.84
25.68
26.53
27.38

0.001
0.04
0.19
0.43
0.76
1.14
1.57
2.05
2.55
3.09
3.65
4.23
4.83
5.44
6.07
6.72
7.37
8.04
8.72
9.41
10.10
10.81
11.52
12.24
12.96
13.70
14.43
15.18
15.93
16.68
17.44
18.20
18.97
19.74
20.51
21.29
22.07
22.86
23.65
24.44

0.0001
0.03
0.08
0.23
0.45
0.72
1.05
1.42
1.82
2.26
2.72
3.20
3.71
4.23
4.78
5.33
5.91
6.49
7.09
7.70
8.31
8.94
9.58
10.22
10.88
11.53
12.20
12.88
13.56
14.24
14.93
15.63
16.33
17.04
17.75
18.46
19.18
19.91
20.63
21.37

offered traffic in erlangs:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

0.11
0.59
1.27
2.04
2.88
3.75
4.66
5.59
6.54
7.51
8.48
9.47
10.46
11.47
12.48
13.50
14.52
15.54
16.57
17.61
18.65
19.69
20.73
21.78
22.83
23.88
24.93
25.99
27.05
28.11
29.17
30.23
31.30
32.36
33.43
34.50
35.57
36.64
37.71
38.78

0.05
0.38
0.89
1.52
2.21
2.96
3.73
4.54
5.37
6.21
7.07
7.95
8.83
9.72
10.63
11.54
12.46
13.38
14.31
15.24
16.18
17.13
18.07
19.03
19.98
20.94
21.90
22.86
23.83
24.80
25.77
26.74
27.72
28.69
29.67
30.65
31.63
32.62
33.60
34.59

(continued overleaf )

204

DIMENSIONING

Table 12.4. (continued)

Probability of blocking a connection

N

0.1

0.05

0.02

0.01

0.005

0.001

31.91
32.83
33.75
34.68
35.60
36.53
37.46
38.39
39.32
40.25
44.9
49.6
54.3
59.1
63.9
68.6
73.4
78.3
83.1
87.9
92.8
97.6
102.5
107.4
112.3
117.1
122.0
126.9
131.8
136.8
141.7
146.6
151.5
156.4
161.4
166.3
171.3
176.2
181.2
186.1

29.88
30.77
31.65
32.54
33.43
34.32
35.21
36.10
37.00
37.90
42.4
46.9
51.5
56.1
60.7
65.3
70.0
74.6
79.3
84.0
88.7
93.4
98.2
102.9
107.7
112.4
117.2
122.0
126.7
131.5
136.3
141.1
145.9
150.7
155.5
160.4
165.2
170.0
174.9
179.7

28.23
29.08
29.93
30.79
31.65
32.51
33.38
34.24
35.11
35.98
40.3
44.7
49.1
53.6
58.1
62.6
67.2
71.7
76.3
80.9
85.5
90.1
94.7
99.3
104.0
108.6
113.3
118.0
122.7
127.3
132.0
136.7
141.5
146.2
150.9
155.6
160.4
165.1
169.8
174.6

25.23
26.03
26.83
27.64
28.44
29.25
30.06
30.87
31.69
32.51
36.6
40.7
44.9
49.2
53.5
57.8
62.1
66.4
70.8
75.2
79.6
84.0
88.5
92.9
97.4
101.9
106.4
110.9
115.4
119.9
124.4
129.0
133.5
138.1
142.6
147.2
151.8
156.4
161.0
165.6

0.0001

offered traffic in erlangs:

41
42
43
44
45
46
47
48
49
50
55
60
65
70
75
80
85
90
95
100
105
110
115
120
125
130
135
140
145
150
155
160
165
170
175
180
185
190
195
200

39.86
40.93
42.01
43.08
44.16
45.24
46.32
47.40
48.48
49.56
54.9
60.4
65.8
71.2
76.7
82.2
87.6
93.1
98.6
104.1
109.5
115.0
120.5
126.0
131.5
137.0
142.5
148.1
153.6
159.1
164.6
170.1
175.6
181.1
186.7
192.2
197.7
203.2
208.7
214.0

35.58
36.57
37.56
38.55
39.55
40.54
41.54
42.53
43.53
44.53
49.5
54.5
59.6
64.6
69.7
74.8
79.9
85.0
90.1
95.2
100.3
105.4
110.6
115.7
120.9
126.0
131.2
136.3
141.5
146.7
151.8
157.0
162.2
167.3
172.5
177.7
182.9
188.1
193.3
198.5

22.10
22.84
23.58
24.33
25.08
25.83
26.58
27.34
28.10
28.86
32.7
36.6
40.5
44.5
48.6
52.6
56.7
60.9
65.0
69.2
73.4
77.6
81.9
86.2
90.4
94.7
99.0
103.4
107.7
112.1
116.4
120.8
125.2
129.6
134.0
138.4
142.8
147.2
151.7
156.1

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

13

Priority Control
the customer comes first

PRIORITIES
ATM networks can feature two forms of priority mechanism: space and
time. Both forms relate to how an ATM buffer operates, and these are
illustrated in Figure 13.1. Space priority addresses whether or not a cell
is admitted into the finite waiting area of the buffer. Time priority deals
with the order in which cells leave the waiting area and enter the server
for onward transmission. Thus the main focus for the space priority
mechanism is to distinguish different levels of cell loss performance,
whereas for time priority the focus is on the delay performance. For both
forms of priority, the waiting area can be organized in different ways,
depending on the specific priority algorithm being implemented.
The ATM standards explicitly support space priority, by the provision
of a cell loss priority bit in the ATM cell header. High priority is indicated
by the cell loss priority bit having a value of 0, low priority with a value of
1. Different levels of time priority, however, are not explicitly supported
in the standards. One way they can be organized is by assigning different
levels of time priority to particular VPI/VCI values or ranges of values.

SPACE PRIORITY AND THE CELL LOSS PRIORITY BIT

An ATM terminal distinguishes the level of space priority for the traffic
flows it is generating by setting the value of the cell loss priority bit.
Within the network, if buffer overflow occurs, the network elements may
selectively discard cells of the lower-priority flow in order to maintain
the performance objectives required of both the high- and low-priority
traffic. For example, a terminal producing compressed video can use high
priority for the important synchronization information. This then avoids
the need to operate the network elements, through which the video

206

PRIORITY CONTROL

Space priority
mechanism controls
access to buffer
capacity

Time priority
mechanism controls
access to server
capacity

Waiting area

Server
ATM buffer

Figure 13.1. Space and Time Priority Mechanisms

connection is routed, at extremely low levels of cell loss probability for

all the cells in the connection. The priority mechanism is able to achieve a
very low loss probability just for those cells that require it, and this leads
to a significant improvement in the traffic load that can be admitted to
the network.
Two selective cell discarding schemes have been proposed and studied
for ATM buffers: the push-out scheme and partial buffer sharing [13.1].
The push-out scheme is illustrated in Figure 13.2; an arriving cell of high
priority which finds the buffer full replaces a low-priority cell within the
buffer. If the buffer contains only high-priority cells, then the arriving cell
is discarded. A low-priority cell arriving to find a full buffer is always
discarded. The partial buffer sharing scheme (see Figure 13.3), reserves
a part of the buffer for high-priority cells only. If the queue is below a
threshold size, then both low- and high-priority cells are accepted onto
the queue. Above the threshold only high-priority cells are accepted.
The push-out scheme achieves only slightly better performance than
partial buffer sharing. But the buffer management and implementation
The buffer is full with a mix of high
and low priority cells and another
high priority cell arrives
H
6

H
5

H
4

H
3

H
2

H
1

server

ATM buffer
The last low priority cell is pushed
out of the buffer, providing room
for the arriving high priority cell
H
6

H
5

H
4

H
3

H
2

H
1

server

ATM buffer

Figure 13.2.

Space Priority: the Push-out Scheme

207

PARTIAL BUFFER SHARING

Mix of high and

low priority cells

Only high priority cells

server

ATM buffer
Threshold
Figure 13.3. Space Priority: Partial Buffer Sharing

are rather more complex for the push-out mechanism because, when a
high-priority cell arrives at a full buffer, a low-priority cell in the buffer
must be found and discarded. Thus the partial buffer sharing scheme
achieves the best compromise between performance and complexity.
Lets look at how partial buffer sharing can be analysed, so we can
quantify the improvements in admissible load that are possible with
space priorities.

PARTIAL BUFFER SHARING

An analysis of the partial buffer sharing scheme is possible for the sort of
queueing system in Chapter 7: a synchronized server, a finite buffer and
Poisson input (a synchronized M/D/1/X queue). Here, we will use the
line crossing form of analysis (see Chapter 9) as this allows a relatively
simple approach.
In Chapter 7, the input traffic is a batch arrival process, where the size
of a batch can vary from cell slot to cell slot, described by a probability
distribution for the number of cells in the batch. This allows the queue
to be analysed for arbitrary distributions, and in Chapter 7 results are
shown for Poisson and binomial distributions.
For the analysis of an ATM buffer with partial buffer sharing, we restrict
the input to be a Poisson-distributed batch, comprising two streams of
traffic: one for each level of space priority. We define the probability that
there are k arrivals in one slot as
a
k D

ak a
e
k!

where the mean arrival rate (in cells per cell slot) is given by parameter a.
This mean arrival rate is the sum of mean arrival rates, ah and al , for the
high- and low-priority streams respectively:
a D ah C al

208

PRIORITY CONTROL

and so we can define the probability that there are k high-priority arrivals
in one slot as
ak
ah
k D h eah
k!
and the probability that there are k low-priority arrivals in one slot as
al
k D

akl al
e
k!

The probability of the queueing system being in state k is defined as

s
k D Prfthere are k cells, of either priority, in the system
at the end of a slotg
The maximum number of cells in the system, i.e. the waiting area and
the server, is X, and the maximum number of low-priority cells, i.e. the
threshold level, is M, where M < X. Below the threshold level, cells of
either priority are admitted into the buffer.
Equating the probabilities of crossing the line between states 0 and 1
gives
s
1 a
0 D s
0
1  a
0
where the left-hand side gives the probability of crossing down (one cell
in the queue, which is served, and no arrivals), and the right-hand side
gives the probability of crossing up (no cells in the queue, and one or
more cells arrive). Remember that any arrivals during the current time
slot cannot be served during this slot. Rearranging the equation gives
s
1 D

s
0
1  a
0
a
0

In general, equating the probabilities of crossing the line between states

k  1 and k, for k < M, gives
s
k a
0 D s
0 A
k C

k1

s
i A
k  i C 1

iD1

where A
k is the probability that at least k cells arrive during the time
slot, and is expressed simply as the probability that fewer than k cells do
not arrive.
k1


A
k D 1 
a
j
jD0

209

PARTIAL BUFFER SHARING

Ah
k is the probability that at least k high-priority cells arrive during a

time slot, and is defined in a similar manner in terms of ah
j; this is used
later on in the analysis.
So, in general for k < M, we have

s
0 A
k C

k1

s
i A
k  i C 1

iD1

s
k D

a
0

Continuing the analysis for state probabilities s
k at or above k D M is

not so straightforward, because the order in which the cells arrive in
the buffer is important if the system is changing from a state below the
threshold to a state above the threshold.
Consider the case in which a buffer, with a threshold M D 10 cells
and system capacity X D 20 cells, has 8 cells in the system at the end
of a time slot. During the next time slot, 4 low-priority cells and 2
high-priority cells arrive, and one cell is served. If the low-priority cells
arrive first, then 2 low-priority cells are admitted, taking the system up
to the threshold, the other 2 low-priority cells are discarded, and the
2 high-priority cells are admitted, taking the system size to 12. Then
the cell in the server completes service and the system size reduces to
11, which is the system state at the end of this time slot. If the highpriority cells arrive first, then these take the system up to the threshold
size of 10, and so all 4 low-priority cells are discarded. At the end
of the slot the system size is then 9 (the cell in the server completes
service).
To analyse how the system changes from one state to another we need
to know the number of cells that are admitted onto the buffer (at a later
stage we will be interested in the number of cells that are not admitted,
in order to calculate the loss from the system). So, lets say that m C n
cells are admitted out of a total of i cells that arrive during one cell slot.
Of those admitted, the first m are of either high or low priority and take
the system from its current state up to the threshold level, and then the
other n are of high priority. Thus i 
m C n low-priority cells are lost.
We use the following expression for the probability that these m C n cells
are admitted:
0

a
m, n D

1


iDmCn

i  m!
a
i

n!
i  m  n!

ah
a

n  imn 

al
a

The binomial part of the expression determines the probability that, of

the i  m cells to arrive when the queue is at or above the threshold, n
are high-priority cells. Here, the probability that a cell is of high priority

210

PRIORITY CONTROL

is expressed as the proportion of the mean arrival rate that is of high

priority. Note that although this expression is an infinite summation, it
converges rapidly and so needs only a few terms to obtain a value for
a0
m, n.
With the line crossing analysis, we need to express the probability that
m cells of either priority arrive, and then at least n or more high-priority
cells arrive, denoted A0
m, n. This can be expressed as
A0
m, n D

1

a0
m, j

jDn

Another way of expressing this is by working out the probability that

fewer than m C n cells are admitted. This happens in two different ways:
either the total number of cells arriving during a slot is not enough, or
there are enough cells but the order in which they arrive is such that there
are not enough high-priority cells above the threshold.

A0
m, n D 1 

mCn1

a
i 

iD0

ah
a

1

a
m C n C i

iD0

jD0

j  nCij 

n1 

n C i!
j!
n C i  j!

al
a

We can now analyse the system at or above the threshold. Equating

probabilities of crossing the line between M and M  1 gives

s
M ah
0 D s
0 A
M C

M1

s
i A0
M  i, 1

iD1

The left-hand side is the probability of crossing down; to stay at the

threshold level, or to move above it, at least one high-priority cell is
needed, so the state reduces by one only if there are no high-priority
arrivals. The right-hand side is the probability of crossing up from
below the threshold. The first term is for crossing up from the state
when there is nothing in the system; this requires M, or more, cells of
either priority. The second term is for all the non-zero states, i, below
the threshold; in these cases there is always a cell in the server which
leaves the system after any arrivals have been admitted to the queue.
Thus at least one high-priority arrival is required after there have been
sufficient arrivals
M  i of either priority to fill the queue up to the
threshold.

211

PARTIAL BUFFER SHARING

For k > M, we have

0

s
k ah
0 D s
0 A
M, k  M C

M1

fs
i A0
M  i, k  M C 1g

iD1

k1

fs
i Ah
k  i C 1g

iDM

This differs from the situation for k D M in two respects: first, the crossing
up from state 0 requires M cells of either priority and a further k  M of
high-priority; and secondly, it is now possible to cross the line from a state
at or above the threshold this can only be achieved with high-priority
arrivals.
At the buffer limit, k D X, we have only one way of reaching this state:
from state 0, with M cells of either priority followed by at least X  M
cells of high-priority. If there is at least one cell in the queue at the start
of the slot, and enough arrivals fill the queue, then at the end of the slot,
the cell in the server will complete service and take the queue state from
X down to X  1. Thus for k D X we have
s
X ah
0 D s
0 A0
M, X  M
Now, as in Chapter 7, we have no value for s
0, so we cannot evaluate
s
k for k > 0. Therefore we define a new variable, u
k, as
u
k D

s
k
s
0

so
u
0 D 1
Then
u
1 D

1  a
0
a
0

For 1 < k < M

A
k C
u
k D

k1

u
i A
k  i C 1

iD1

a
0

At the threshold
A
M C
u
M D

M1

u
i A0
M  i, 1

iD1

ah
0

212

PRIORITY CONTROL

For M < k < X  1

0

A
M, k  M C

M1

fu
i A0
M  i, k  M C 1g

iD1
k1


C

u
k D

fu
i Ah
k  i C 1g

iDM

ah
0

At the system capacity

u
X D

A0
M, X  M
ah
0

All the values of u
k, 0
k
X, can be evaluated. Then, as in Chapter 7,

we can calculate the probability that the system is empty:
s
0 D

1
X


u
i
iD0

and, from that, find the rest of the state probability distribution:
s
k D s
0 u
k
Before we go on to calculate the cell loss probability for the high-and
low-priority cell streams, lets first show an example state probability
distribution for an ATM buffer implementing the partial buffer sharing
scheme. Figure 13.4 shows the state probabilities when the buffer capacity
is 20 cells, and the threshold level is 15 cells, for three different loads:
(i) the low priority load, al , is 0.7 and the high-priority load, ah , is 0.175 of
the cell slot rate; (ii) al D 0.6 and ah D 0.15; and (iii) al D 0.5 and ah D 0.125.
The graph shows a clear distinction between the gradients of the state
probability distribution below and above the threshold level. Below the
threshold, the queue behaves like an ordinary M/D/1 with a gradient
corresponding to the combined high- and low-priority load. Above the
threshold, only the high-priority cell stream has any effect, and so the
gradient is much steeper because the load on this part of the queue is
much less.
In Chapter 7, the loss probability was found by comparing the offered
and the carried traffic at the cell level. But now we have two different
priority streams, and the partial buffer sharing analysis only gives the
combined carried traffic. The overall cell loss probability can be found

213

PARTIAL BUFFER SHARING

Queue size
0

10

15

20

1E+00
1E01
1E02
State probability

(i)
1E03
(ii)

1E04
(iii)

1E05
1E06
1E07
1E08
1E09

Figure 13.4. State Probability Distribution for ATM Buffer with Partial Buffer Sharing (i) al D 0.7,
ah D 0.175; (ii) al D 0.6, ah D 0.15; (iii) al D 0.5, ah D 0.125

using
CLP D

al C ah 
1  s
0
al C ah

But the main objective of having a space priority scheme is to provide

different levels of cell loss. How can we calculate this cell loss probability
for each priority stream? It has to be done by considering the probability
of losing a group of low- or high-priority cells during a cell slot, and
then taking the weighted mean over all the possible group sizes. The
high-priority cell loss probability is given by

CLPh D

j lh
j

ah

where lh
j is the probability that j high-priority cells are lost in a cell slot
and is given by
lh
j D

M1


iD0

s
i a0
M  i, X  M C j C

X

s
i ah
X  i C j

iDM

The first summation on the right-hand side accounts for the different ways
of losing j cells when the state of the system is less than the threshold.

214

PRIORITY CONTROL

This involves filling up to the threshold with either low- or high-priority

cells, followed by X  M high-priority cells to fill the queue and then a
further j high-priority cells which are lost. The second summation deals
with the different ways of losing j cells when the state of the system is
at or above the threshold; X  i high-priority cells are needed to fill the
queue and the other j in the batch are lost.
The low-priority loss is found in a similar way:

CLPl D

j ll
j

al

where ll
j is the probability that j low-priority cells are lost in a cell slot
and is given by

ll
j D

M1


iD0

1

r 
M  i!
s
i

a
r

r 
M  i  j! j!
rDMiCj

X

ah
a

r
Mij  j

al
a

s
i al
j

iDM

The first term on the right-hand side accounts for the different ways of
losing j cells when the state of the system is less than the threshold. This
involves filling up to the threshold with either M  i cells of either low or
high-priority, followed by any number of high-priority cells along with j
low-priority cells (which are lost). The second summation deals with the
different ways of losing j cells when the state of the system is above the
threshold. This is simply the probability of j low-priority cells arriving in
a time slot, for each of the states at or above the threshold.

Increasing the admissible load

Lets now demonstrate the effect of introducing a partial buffer sharing
mechanism to an ATM buffer. Suppose we have a buffer of size X D 20,
and the most stringent cell loss probability requirement for traffic through
the buffer is 1010 . From Table 10.1 we find that the maximum admissible
load is 0.521. Now the traffic mix is such that there is a high-priority
load of 0.125 which requires the CLP of 1010 ; the rest of the traffic can
tolerate a CLP of 103 , a margin of seven orders of magnitude. Without
a space priority mechanism, a maximum load of 0.521  0.125 D 0.396
of this other traffic can be admitted. However, the partial buffer sharing
analysis shows that, with a threshold of M D 15, the low-priority load can

215

PARTIAL BUFFER SHARING

be increased to 0.7 to give a cell loss probability of 1.16 103 , and the
high-priority load of 0.125 has a cell loss probability of 9.36 1011 . The
total admissible load has increased by just over 30% of the cell slot rate,
from 0.521 to 0.825, representing a 75% increase in the low-priority traffic.
If the threshold is set to M D 18, the low-priority load can only be
increased to 0.475 giving a cell loss probability of 5.6 108 , and the
high-priority load of 0.125 has a cell loss probability of 8.8 1011 . But
even this is an extra 8% of the cell slot rate, representing an increase in
20% for the low-priority traffic, for a cell loss margin of between two and
three orders of magnitude. Thus a substantial increase in load is possible,
particularly if the difference in cell loss probability requirement is large.

Dimensioning buffers for partial buffer sharing

Figures 13.5 and 13.6 show interesting results from the partial buffer
sharing analysis. In both cases, the high-priority load is fixed at 0.125, and
the space above the threshold is held constant at 5 cells. In Figure 13.5, the
low-priority load is varied from 0.4 up to 0.8, and the cell loss probability
results are plotted for the high- and low-priority traffic against the
combined load. This is done for three different buffer capacities. The
results show that the margin in the cell loss probabilities is almost
constant, at seven orders of magnitude. Figure 13.6 shows the same
margin in the cell loss probabilities for a total load of 0.925
ah D 0.125, al D
0.8 as the buffer capacity is varied from 10 cells up to 50 cells.

Cell loss probability

Combined high and low priority load

0.5
0.6
1E+00
1E01
1E02
X =10, M=5
1E03
1E04
1E05
X =15, M =10
1E06
1E07
1E08 X =20, M =15
1E09
1E10
1E11
1E12
1E13
1E14
1E15

Figure 13.5.
ah D 0.125

0.7

0.8

0.9

Low priority

High priority

Low and High-Priority Cell Loss against Load, for X  M D 5 and

216

PRIORITY CONTROL

Buffer capacity, X
0

10

20

30

40

50

1E+00
1E01

Low priority

1E02
Cell loss probability

1E03
1E04
1E05
1E06
1E07
1E08
1E09

High priority

1E10
1E11
1E12

Figure 13.6. Low- and High-Priority Cell Loss against Buffer Capacity, for a D 0.925 and
XMD5

The difference between high- and low-priority cell loss is almost

invariant to the buffer capacity and the total load, provided that the
space above the threshold, and the high-priority load, are kept constant.
Table 13.1 shows how the margin varies with the space above the
threshold, and the high-priority load (note that margins greater than 11
orders of magnitude are not included these are unlikely to be required
in practice). The values are also plotted in Figure 13.7.
With this information, buffers can be dimensioned using the following
procedure:
1.

Set the threshold by using Table 10.1 based on the M/D/1/X analysis
(without priorities) for the combined load and the combined cell loss
probability requirement. The latter is found using the following
relationship (which is based on equating the average number of cells

Table 13.1. Cell Loss Probability Margin between Low- and High-Priority Traffic

High-priority traffic load (as a fraction of the cell slot rate)

X M
1
2
3
4
5

0.01

0.02

0.03

0.04

0.05

0.1

0.15

0.2

0.25

2.7E-03
6.5E-06
1.4E-08
3.0E-11

5.6E-03
2.7E-05
1.2E-07
5.4E-10

8.4E-03
6.3E-05
4.5E-07
2.9E-09
1.8E-11

1.1E-02
1.2E-04
1.2E-06
1.0E-08
9.0E-11

1.4E-02
1.8E-04
2.5E-06
2.8E-08
3.3E-10

2.8E-02
8.4E-04
2.6E-05
6.7E-07
1.9E-08

4.4E-02
2.2E-03
1.1E-04
4.9E-06
2.5E-07

5.9E-02
4.3E-03
3.2E-04
2.2E-05
1.6E-06

7.6E-02
7.5E-03
7.4E-04
7.0E-05
7.0E-06

217

PARTIAL BUFFER SHARING

High priority load

0

0.05

0.1

0.15

0.2

0.25

Cell loss probability margin

1E+00
1E01

XM=1

1E02

XM=2

1E03

XM=3

1E04

XM=4

1E05

XM=5

1E06
1E07
1E08
1E09
1E10

Figure 13.7. Cell Loss Probability Margin against High-Priority Load for Different
Values of X  M

lost per cell slot):

CLP D
2.

al CLPl C ah CLPh
a

Add buffer space above the threshold determined by the high-priority

load and the additional cell loss probability margin, from Table 13.1.

Lets take an example. We have a requirement for a buffer to carry a

total load of 0.7, with low-priority CLP of 106 and high-priority CLP of
1010 . The high-priority load is 0.15. Thus the overall CLP is given by
CLP D

0.55 106 C 0.15 1010

D 7.86 107
0.7

From Table 10.1 we find that the threshold is between 20 and 25 cells,
but closer to 20; we will use M D 21. Table 13.1 gives an additional buffer
space of 3 cells for a margin of 104 and high-priority load of 0.15.
Thus the total buffer capacity is 24. If we put these values of X D 24
and M D 21, al D 0.55 and ah D 0.15 back into the analysis, the results
are CLPl D 5.5 107 and CLPh D 6.3 1011 . For a buffer size of 23, a
threshold of 20, and the same load, the results are CLPl D 1.1 106 and
CLPh D 1.2 1010 .
Two of the values for high-priority load in Table 13.1 are of particular
interest in the development of a useful dimensioning rule; these values
are 0.04 and 0.25. In Figure 13.8, the CLP margin is plotted against the

218

PRIORITY CONTROL
Cuffer space above threshold, X M
0

1E+00
1E01
Cell loss probability margin

1E02
1E03

ah = 0.25

1E04
1E05
1E06

al = 0.04

1E07
1E08
1E09
1E10
1E11

Figure 13.8. Cell Loss Probability Margin against Buffer Space Reserved for
High-Priority Traffic, X  M

buffer space above the threshold (this is shown as a continuous line to

illustrate the log-linear relationshipthe buffer space of course varies in
integer values). At the 25% load, each cell space reserved for high-priority
traffic is worth one order of magnitude on the CLP margin. At the 4%
load, it is two orders of magnitude. We can express this as
CLPmargin D 10
XM
for a high-priority load of 25% of the cell slot rate, and
CLPmargin D 102
XM
for a high-priority load of 4% of the cell slot rate.

TIME PRIORITY IN ATM

In order to demonstrate the operation of time priorities, lets define
two traffic classes, of high and low time priority. In a practical system,
there may be rather more levels, according to the perceived traffic
requirements. The ATM buffer in Figure 13.9 operates in such a way that
any high-priority cells are always served before any low-priority. Thus a
high-priority cell arriving at a buffer with only low-priority cells currently
in the queue will go straight to the head of the queue. Note that at the
beginning of time slot n C 1 the low-priority cell currently at the head of

219

TIME PRIORITY IN ATM

server
low

low

low

State of the buffer at end of time slot n

server
low

low

low

low

high

low

State of the buffer at end of time slot n+1 after 3 cells

have arrived - one of high priority and two of low priority
Figure 13.9. Time Priorities in ATM

the queue goes into service. It is only during time slot n C 1 that the highpriority cell arrives and is then placed at the head of the queue. The same
principle can be applied with many levels of priority. Note that any cell
arriving to find the buffer full is lost, regardless of the level of time priority.
The effect of time priorities is to decrease the delay for the higherpriority traffic at the expense of increasing the delays for the lowerpriority traffic. As far as ATM is concerned, this means that real-time
connections (e.g. voice and interactive video) can be speeded on their
way at the expense of delaying the cells of connections which do not have
real-time constraints, e.g. email data.
To analyse the delay performance for a system with two levels of time
priority, we will assume an M/D/1 system, with infinite buffer length.
Although time priorities do affect the cell loss performance, we will
concentrate on those analytical results that apply to delay.

Mean value analysis

We define the mean arrival rate in cells per slot as ai for cells of priority
i. High-priority is indicated by i D 1 and low priority by i D 2. Note that
the following analysis can be extended to many levels if required.
The formulas for the mean waiting time are:
w1 D
and

a1 C a2
2
1  a1 
a1 C a2
2
1  a1  a2

w1 a1 C
w2 D

220

PRIORITY CONTROL

where wi is the mean wait (in time slots) endured by cells of priority i
while in the buffer.
Consider an ATM scenario in which a very small proportion of traffic,
say about 1%, is given high time priority. Figure 13.10 shows the effect
on the mean waiting times. Granting a time priority to a small proportion
of traffic has very little effect on the mean wait for the lower-priority
traffic, which is indistinguishable from the mean wait when there are no
priorities. We can also see from the results that the waiting time for the
high-priority cells is greatly improved.
Figure 13.11 shows what happens if the proportion of high-priority
traffic is significantly increased, to 50% of the combined high- and
low-priority load. Even in this situation, mean waiting times for the
low-priority cells are not severely affected, and waiting times for the
priority traffic have still been noticeably improved. Figure 13.12 illustrates the case when most of the traffic is high-priority and only
1% is of low priority. Here, there is little difference between the nopriority case and the results for the high-priority traffic, but the very
small amount of low-priority traffic has significantly worse waiting
times.
The results so far are for the mean waiting time. Lets now consider
the effect of time priorities on the distribution of the waiting-time (see
also [13.2]). To find the waiting time probabilities for cells in an ATM

Mean delay in time slots

20

15
Priority 2
No priority
Priority 1

10

0
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Total load
Figure 13.10. Mean Waiting Time for High and Low Time-Priority Traffic, where
the Proportion of High-Priority Traffic is 1% of the Total Load

221

TIME PRIORITY IN ATM

Mean delay in time slots

20

15
priority 2
no priority
priority 1

10

0.1

0.2

0.3

0.4 0.5 0.6

Total load

0.7

0.8

0.9

Figure 13.11. Mean Waiting Time for High and Low Time-Priority Traffic, where
the Proportion of High-Priority Traffic is 50% of the Total Load

Mean delay in time slots

20

15
priority 2
no priority
priority 1

10

0.1

0.2

0.3

0.4 0.5
Total load

0.6

0.7

0.8

0.9

Figure 13.12. Mean Waiting Time for High and Low Time-Priority Traffic, where
the Proportion of High-Priority Traffic is 99% of the Total Load

buffer where different levels of time priority are present requires the use
of convolution (as indeed did finding waiting times in a non-priority
buffersee Chapter 7). A cell, say C, arriving in time slot i will wait
behind a number of cells, and this number has four components:

222

PRIORITY CONTROL

1.

the total number of cells of equal or higher priority that are present
in the buffer at the end of time slot i  1
2. the number of cells of the same priority that are ahead of C in the batch
in which C arrives
3. all the cells of higher priority than C that arrive in time slot i
4.

higher-priority cells that arrive subsequently, but before C enters

service.

Again, if we focus on just two levels of priority, we can find the

probability that a cell of low priority (priority 2) has to wait k time slots
before it can enter service, by finding expressions for the four individual
components. Let us define:
component 1

the unfinished workas u
k

component 2

the batch waitas b
k

component 3

the wait caused by priority-1 arrivals in time slot

i as a1
k

Then, excluding the effect of subsequent high-priority arrivals, we

know that our waiting-time distribution must be (in part) the sum of
the three components listed above. Note that to sum random variables,
we must convolve their distributions. We will call the result of this
convolution the virtual waiting-time distribution, v
k, given by:
v
 D u
 b
 a1

where * denotes convolution. We can rewrite this as:

k
i




u
k  i
v
k D
b
j a1
i  j
iD0

jD0

But where do the three distributions, u
k, b
k and a1
k come from?
As we are assuming Poisson arrivals for both priorities, a1
k is simply:
a1
k D

ak1 a1
e
k!

and for the low-priority cells we will have:

a2
k D

ak2 a2
e
k!

where
a1 is the arrival rate (in cells per time slot) of high-priority cells
a2 is the arrival rate (in cells per time slot) of low-priority cells

223

TIME PRIORITY IN ATM

The unfinished work, u
k, is actually found from the state probabilities,
denoted s
k, the formula for which was given in Chapter 7:
u
0 D s
0 C s
1
u
k D s
k C 1

for k > 0

What about the wait caused by other cells of low priority arriving in the
same batch, but in front of C? Well, there is a simple approach here too:
b
k D PrfC is
k C 1th in the batchg
D

E[number of cells that are
k C 1th in their batch]

E[number of cells arriving per slot]
1

k

a2
i

iD0

a2

So now all the parts are assembled, and we need only implement
the convolution to find the virtual waiting-time distribution. However,
this still leaves us with the problem of accounting for subsequent highpriority arrivals. In fact, this is very easy to do using a formula developed
(originally) for an entirely different purpose. The result is that:
w
0 D v
0
k

w
k D

v
i a1
k  i, k i

iD1

for k > 0

where:
w
k D Prfa priority 2 cell must wait k time slots before it enters serviceg
a1
k, x D

axk
1
eka1
x!

So a1
k, x is simply the probability that k priority-1 cells arrive in x time

slots.
Figures 13.13 and 13.14 show the waiting-time distributions for highand low-priority cells when the combined load is 0.8 cells per time slot
and the high-priority proportions are 1% and 50% respectively. From
these results, it is clear that, even for a relatively large proportion of highpriority traffic, the effect on the waiting-time distribution for low-priority
traffic is small, but the benefits to the high-priority traffic are significant.

224

PRIORITY CONTROL

Waiting time (time slots)

0

10

1E+00
1E01

Probability

1E02
priority 2
priority 1

1E03
1E04
1E05
1E06
1E07
1E08

Figure 13.13. Waiting-Time Distribution for High- and Low Time-Priority Traffic,
where the Proportion of High-Priority Traffic Is 1% of a Total Load of 0.8 Cells per
Time Slot
Waiting time (time slots)
0

10

1E+00
1E01

Probability

1E02
1E03

priority 2
priority 1

1E04
1E05
1E06
1E07
1E08

Figure 13.14. Waiting Time Distribution for High and Low Time-Priority Traffic,
where the Proportion of High-Priority Traffic Is 50% of a Total Load of 0.8 Cells per
Time Slot

Before we leave time priorities, it is worth noting that practical systems

for implementing them would probably feature a buffer for each priority
level, as shown in Figure 13.15, rather than one buffer for all priorities,
as in Figure 13.9. Although there is no explicit provision in the Standards
for distinguishing different levels of time-priority, it is possible to use the

225

TIME PRIORITY IN ATM

buffer for cells of priority 1 only

SERVER

buffer for cells of priority 2 only

.
.
.
buffer for cells of priority i only
Figure 13.15. Practical Arrangement of Priority Buffers at an Output Port

VPI/VCI values in the header. On entry to a switch, the VPI/VCI values

are used to determine the outgoing port required, so it is a relatively
simple extension to use these values to choose one of a number of time
priority buffers at that output port.
Using one buffer per priority level (Figure 13.15) would have little
effect on the delays experienced by the cells but it would affect the
CLP. This is because, for a given total capacity of X cells, the CLP is
minimized if the total space is shared amongst the different priorities
(as in Figure 13.9). However, this has to be balanced against considerations of extra complexity (and hence cost) inherent in a buffer-sharing
arrangement.

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

PART III
IP Performance and
Traffic Management

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

14

Basic Packet Queueing

the long and short of it

THE QUEUEING BEHAVIOUR OF PACKETS IN AN IP ROUTER

BUFFER
In Chapters 7 and 8, we investigated the basic queueing behaviour found
in ATM output buffers. This queueing arises because multiple streams of
cells are being multiplexed together; hence the need for (relatively short)
buffers. We developed balance equations for the state of the system at the
end of any time slot, from which we derived cell loss and delay results.
We also looked at heavy-traffic approximations: explicit equations which
could be rearranged to yield expressions for buffer dimensioning and
admission control, as well as performance evaluation.
In essence, packet queueing is very similar. An IP router forwards
arriving packets from input port to output port: the queueing behaviour
arises because multiple streams of packets (from different input ports) are
being multiplexed together (over the same output port). However, a key
difference is that packets do not all have the same length. The minimum
header size in IPv4 is 20 octets, and in IPv6, it is 40 octets; the maximum
packet size depends on the specific sub-network technology (e.g. 1500
octets in Ethernet, and 1000 octets is common in X.25 networks). This
difference has a direct impact on the service time; to take this into account
we need a probabilistic (rather than deterministic) model of service, and
a different approach to the queueing analysis.
As before, there are three different types of behaviour in which we are
interested:

the state probabilities, by which we mean the proportion of time that

a queue is found to be in a particular state (being in state k means the
queue contains k packets at the time at which it is inspected, and measured over a very long period of time, i.e. the steady-state probabilities);

230

BASIC PACKET QUEUEING

the packet loss probability, by which we mean the proportion of

packets lost over a very long period of time;

the packet waiting-time probabilities, by which we mean the probabilities associated with a packet being delayed k time units.

It turns out that accurate evaluation of the state probabilities is paramount

in calculating the waiting times and loss too, and for this reason we focus
on finding accurate and simple-to-use formulas for state probabilities.

BALANCE EQUATIONS FOR PACKET BUFFERING:

THE GEO/GEO/1
To analyse these different types of behaviour, we are going to start by
following the approach developed in Chapter 7, initially for a very simple
queue model called the Geo/Geo/1, which is the discrete-time version
of the classical queue model M/M/1. One way in which this model
differs from that of Chapter 7 is that the fundamental time unit is reduced
from a cell service time to the time to transmit an octet (byte), Toct . Thus
we have a conveyor belt of octets the transmission of each octet of a
packet is synchronized to the start of transmission of the previous octet.
Using this model assumes a geometric distribution as a first attempt at
variable packet sizes:
b
k D Prfpacket size is k octetsg D
1  qk1 q
where
q D Prfa packet completes service at the end of an octet slotg
We use a Bernoulli process for the packet arrivals, i.e. a geometrically
distributed number of slots between arrivals (the first Geo in Geo/Geo/1):
p D Prfa packet arrives in an octet slotg
Thus we have an independent and identically distributed batch of k octets
(k D 0, 1, 2, . . .) arriving in each octet slot:
a
0 D Prfno octets arriving in an octet slotg D 1  p
a
k D Prfk > 0 octets in an octet slotg D p b
k
The mean service time for a packet is simply the mean number of octets
(the inverse of the exit probability for the geometric distribution, i.e. 1/q)
multiplied by the octet transmission time.
sD

Toct
q

BALANCE EQUATIONS FOR PACKET BUFFERING: THE GEO/GEO/1

231

giving a packet service rate of

D

q
1
D
s
Toct

The mean arrival rate is

D

p
Toct

and so the applied load is given by

D


p
D

q

This is also the utilization, assuming an infinite buffer size and, hence, no
packet loss. We define the state probability, i.e. the probability of being
in state k, as
s
k D Prfthere are k octets in the queueing system at the
end of any octet slotg
As before, the utilization is just the steady-state probability that the
system is not empty, so
 D 1  s
0
and therefore
s
0 D 1 

p
q

Calculating the state probability distribution

As in Chapter 7, we can build on this value, s
0, by considering all the
ways in which it is possible to reach the empty state:
s
0 D s
0 a
0 C s
1 a
0
giving


p
p
1  a
0

D 1
s
1 D s
0
a
0
q
1p

Similarly, we find a formula for s
2 by writing the balance equation for
s
1, and rearranging:
s
2 D

s
1  s
0 a
1  s
1 a
1

a
0

232

BASIC PACKET QUEUEING

which, after substituting in

a
0 D 1  p
a
1 D p q
gives

s
2 D s
1

1q
1p


p
p
D 1

q
1p



p
p

D 1
q
1q

1q
1p



1q 2

1p

By induction, we find that

p
s
k D 1 
q

1q

1q
1p

k

for k > 0

As in Chapter 7, the state probabilities refer to the state of the queue at

moments in time that are the end of time unit instants.
We can take the analysis one step further to find an expression for the
probability that the queue exceeds k octets, Q
k:
Q
k D 1  s
0  s
1   s
k
This gives a geometric progression which, after some rearrangement,
yields



p
1q k
Q
k D
q
1p
To express this in terms of packets, x, (recall that it is currently in terms
of octets), we can simply substitute
k D x (mean number of octets per packet) D x

1
q

giving an expression for the probability that the queue exceeds x packets:
Q
x D

1q
1p

x
q

So, what do the results look like? Lets use a load of 80%, for comparison
with the results in Chapter 7, and assume an average packet size of

233

BALANCE EQUATIONS FOR PACKET BUFFERING: THE GEO/GEO/1

500 octets. Thus

D

p
D 0.8
q

1
D 500 ) q D 0.002
q
p D 0.8 0.002 D 0.0016
The results are shown in Figure 14.1, labelled Geo/Geo/1. Those
labelled Poisson and Binomial are the results from Chapter 7
Buffer capacity, X
0

10

15

20

25

100
Geo/Geo/1
Poisson
Binomial

Pr{queue size > X}

101
102
103
104
105
106

1
500
p :D 0.8 q
q :D

p
packetQ
x :D
q
k :D 0.. 30
xk :D k
y1 :D packetQ
x

1q
1p

x

Figure 14.1. Graph of the Probability that the Queue State Exceeds X, and the
Mathcad Code to Generate (x, y) Values for Plotting the Geo/Geo/1 Results. For
Details of how to Generate the Results for Poisson and Binomial Arrivals to a
Deterministic Queue, see Figure 7.6

234

BASIC PACKET QUEUEING

(Figure 7.6) for fixed service times at a load of 80%. Notice that the
variability in the packet sizes (and hence service times) produces a
flatter gradient than the fixed-cell-size analysis for the same load. The
graph shows that, for a given performance requirement (e.g. 0.01), the
buffer needs to be about twice the size (X D 21) of that for fixed-size
packets or cells (X D 10). This corresponds closely with the difference, in
average waiting times, between M/D/1 and M/M/1 queueing systems
mentioned in Chapter 4.

DECAY RATE ANALYSIS

One of the most important effects we have seen so far is that the state
probability values we are calculating tend to form straight lines when the
queue size (state) is plotted on a linear scale, and the state probability is
plotted on a logarithmic scale. This is a very common (almost universal)
feature of queueing systems, and for this reason has become a key result
that we can use to our advantage.
As in the previous section, we define the state probability as
s
k D Prfthere are k units of data packets,
octets in the queueing systemg
We define the decay rate (DR) as the ratio:
s
k C 1
s
k
However, this ratio will not necessarily stay constant until k becomes
large enough, so we should actually say that:
DR D

s
k C 1
s
k

as k ! 1

as illustrated in Figure 14.2.

From the form of the equation, and the example parameter values in
Figure 14.1, we can see that the decay rate for the Geo/Geo/1 model is
constant from the start:





p
p
1  q kC1


1q
s
k C 1
q
1q
1p
D




 D
s
k
1p
p
p
1q k
1

q
1q
1p

1

235

DECAY RATE ANALYSIS

Queue size - linear scale

0

10

15

20

25

30

State probability - logarithmic scale

100
90% load
80% load

101

102

103

104
Here we see a constant decay rate
105

Figure 14.2.
System

The Decay Rate of the State Probabilities for the M/D/1 Queueing

But, as we mentioned previously, this is not true for most queueing

systems. A good example of how the decay rate takes a little while to
settle down can be found in the state probabilities generated using the
analysis, developed in Chapter 7, for an output buffer. Lets take the case
in which the number of arriving cells per time slot is Poisson-distributed,
i.e. the M/D/1, and choose an arrival rate of 0.9 cells per time slot. The
results are shown in Table 14.1.
The focus of buffer analysis in packet-based networks is always to
evaluate probabilities associated with information loss and delay. For
this reason we concentrate on the state probabilities as seen by an arriving
Table 14.1. Change in
Decay Rate for M/D/1
with 90% Load

Radio
s(1)/s(0)
s(2)/s(1)
s(3)/s(2)
s(4)/s(3)
s(5)/s(4)
s(6)/s(5)
s(7)/s(6)

DR
1.4596
0.9430
0.8359
0.8153
0.8129
0.8129
0.8129

236

BASIC PACKET QUEUEING

packet. This is in contrast to those as seen by a departing packet, as in classical

queueing theory, or as left at random instants as we used in the time-slotted
ATM buffer analysis of Chapter 7. The key idea is that, by finding the
probability of what is seen ahead of an arriving packet, we have a very
good indicator of both:

the waiting time i.e. the sum of the service time of all the packets
ahead in the queue

the loss the probability that the buffer overflows a finite length is
often closely approximated by the probability that the infinite buffer
model contains more than would fit in the given finite buffer length.

Using the decay rate to approximate the buffer overflow probability

Having a constant decay rate is just the same as saying that we have a
geometric progression for the state probabilities:
Prfkg D
1  p pk
To find the tail probability, i.e. the probability associated with values
greater than k, we have
Prf>kg D 1  Prf0g  Prf1g   Prfkg
Buffer capacity
0

10

20

30

100

Overflow probability

constant multiplier
Constant
multiplier
101
decay rate
Decay
rate

102

103

Figure 14.3. Decay Rate Offset by a Constant Multiplier

40

237

DECAY RATE ANALYSIS

After substituting in the geometric distribution, and doing some algebraic

manipulation we have
Prf>kg D 1 
1  p 
1  p p  
1  p pk
p Prf>kg D p 
1  p p  
1  p pk 
1  p pkC1

1  p Prf>kg D 1 
1  p  p C
1  p pkC1
Load
0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

100
101

exact packet loss probability

approx. buffer overflow

Loss probability

102
103
104
105
106
107
108
109
1010


Q
X, s :D  qx0
1  s0

 for i 2 1.. X

 qx
qxi1  si

i

 qxX
i :D 10
j :D 0.. 14
k :D 0.. 30
j
loadj :D
C 0.25
20
aPk,j :D Poisson
k, loadj 
xj :D loadj

if X > 0

y1j :D finiteQloss
i, aPhji , xj 

yPj :D infiniteQ
i, aPhji , xj 
y2j :D Q
i, yPj 
Figure 14.4. Comparison of Q
x and Loss Probability for the M/D/1 Queue Model,
with a Finite Buffer Capacity of 10 Packets

238

BASIC PACKET QUEUEING

which, after dividing through by (1  p), yields

Prf>kg D pkC1
However, many of the buffer models we have to deal with are best
modelled by a constant decay rate, dr , that is offset by another constant
multiplier, Cm . This is illustrated in Figure 14.3. If we know both the
value of the decay rate and the constant multiplier then we can estimate
the buffer overflow probability from:
Prfbuffer overflowg Cm dXC1
r
where X is the buffer length in packets.
This ties in very nicely with our earlier work in Chapter 9 on the
burst-scale loss and burst-scale delay models for ATM buffering, where
the value of the constant was evaluated via the probability that a cell
needs a buffer. We use a similar approach in Chapter 15 for evaluating
the resource implications of many ONOFF sources being multiplexed
in an IP router.
At this stage it is worth looking at some numerical comparisons
for typical queueing systems, plotting the probability of buffer overflow against the packet loss probability. Figure 14.4 compares these
two measures for the M/D/1 system. This comparison (i.e. using state
probabilities seen by arrivals) shows that Prfinfinite buffer contains > Xg
is a good approximation for the loss probability. This is the sort of
simplification that is frequently exploited in buffering analyses.

BALANCE EQUATIONS FOR PACKET BUFFERING: EXCESS-RATE

QUEUEING ANALYSIS
The advantage of the Geo/Geo/1 is that it is simple, and the high
variability in its service times have allowed some to claim it is a worstcase model. We need to note two points: the Bernoulli input process
assumes arrivals as an instantaneous batch (which, as we will see in the
next chapter, has a significant effect); and the geometric distribution of
the packet lengths is an overestimation of the amount of variation likely
to be found in real IP networks. The second of these problems, that the
geometric is an unrealistic model of IP packets as it gives no real upper
limit on packet lengths, can be overcome by more realistic packet-size
distributions.
To address this, we develop an analytical result into which a variety
of different packet-size distributions can be substituted relatively simply.
To begin with, we assume fixed-size packets (i.e. the M/D/1 queue) and
derive a formula that is more convenient to use than the recurrence equations of Chapter 7 and significantly more accurate than the heavy-traffic

BALANCE EQUATIONS FOR PACKET BUFFERING: EXCESS-RATE QUEUEING ANALYSIS

239

approximations of Chapter 8. This formula can be applied to cell-scale

queueing in ATM as well as to packet queueing for real-time services
such as voice-over-IP (which have fixed, relatively short, packet lengths).
Then we show how this formula can be applied for variable-size
packets of various distributions. One particular distribution of interest is
the bi-modal case: here, the packet lengths take one of two values, either
the shortest possible or the longest possible. The justification for this is
that in real IP networking situations there is a clear division of packets
along these lines; control packets (e.g. in RSVP and TCP) tend to be very
short, and data packets tend to be the maximum length allowable for the
underlying sub-network technology.

The excess-rate M/D/1, for application to voice-over-IP

We introduced the notion of excess-rate arrivals in Chapter 9 when we
considered burst-scale queueing behaviour. Then, we were looking at
the excess of arrival rate over service rate for durations of milliseconds
or more, i.e. multiple cell slots. In the example of Figure 9.2, the excess
rate was 8% of the service capacity over a period of 24 cell slots, i.e.
approx. 2 cells in 68 s. Typical bursts last for milliseconds, and so if this
excess rate lasts for 2400 time slots, then about 200 cells must be held
temporarily in the output buffer, or they are lost if there is insufficient
buffer space.
Now, suppose we reduce the duration over which we define excessrate arrivals to the time required to serve a fixed-length packet. Let
this duration be our fundamental unit of time. Thus, excess-rate (ER)
packets are those which must be buffered as they represent an excess of
instantaneous arrival rate over the service rate; if N packets arrive in any
time unit, then that time unit experiences N  1 excess packets.
Why should we do this? Well, for two reasons. First, we get a clearer
idea of how the queue changes in size: for every excess-rate packet,
the queue increases by one; a single packet arrival causes no change
in the queue state (because one packet is also served), and the queue
only decreases when there are no packets arriving in any time unit (see
Figure 14.5). We can then focus on analysing the behaviour that causes
the queue to change in size. Instead of connecting end of slot k with
end of slot k C 1 via a balance equation (i.e. using an Imbedded Markov
Chain at ends of slots, as in Chapter 7), we connect the arrival of excessrate packets via the balance equations (in a similar way to the discrete
fluid-flow approach in Chapter 9).
Secondly, it gives us the opportunity to use a form of arrival process
which simplifies the analysis. We alter the Poisson process to produce
a geometric series for the tail of the distribution (Figure 14.6), giving a
geometrically distributed number of ER packets per time unit. We call

240

BASIC PACKET QUEUEING

Excess arrivals cause

queue level to increase

No arrival during time

unit causes queue
Single packet
level to decrease
arrival causes no
change in the
queue state

IMC

Time unit = mean packet service time

Figure 14.5. Excess-Rate Queueing Behaviour

Number of arrivals, k, per time unit

1
2
3
4

101

e
No change

e
Under load

Probability of k arrivals per time unit

100
Excess arrivals
l e e
q
q

q
102
q

...
103

Figure 14.6. The Geometrically Approximated Poisson Process

this the Geometrically Approximated Poisson Process (GAPP) and define

the conditional probability
q D Prfanother ER packet arrives in a time unitjjust had oneg
Thus the mean number of ER packets in an excess-rate batch, E[B], is
given by
1
E[B] D
1q

BALANCE EQUATIONS FOR PACKET BUFFERING: EXCESS-RATE QUEUEING ANALYSIS

241

But we can find an expression for E[B] based on the arrival probabilities:
1


E[B] D

i  1 a
i

iD2

1  a
0  a
1

where
a
k D Prfk arrivals in a packet service timeg
The numerator weights all the probabilities of having i packets arriving
by the number that are actually excess-rate packets, i.e. i  1. This ranges
over all situations in which there is at least one excess-rate packet arrival.
The denominator normalizes the probabilities to this condition (that there
are ER arrivals). A simple rearrangement of the numerator gives
1


E[B] D

i a
i 

iD1

1


a
i

iD1

1  a
0  a
1

E[a] 
1  a
0
1  a
0  a
1

where E[a] is the mean number of packets arriving per unit time. We now
have an expression for the parameter of the geometric ER series:
qD1

1  a
0  a
1
1
D1
E[B]
E[a]  1 C a
0

Consider now how the queue increases in size by one packet. We define
the state probability as
p
k D Prfan arriving excess-rate packet finds k packets in the bufferg
Remember that we are creating an Imbedded Markov Chain at excessrate arrival instants. Thus to move from state k to k C 1 either we need
another ER packet in the same service time interval, with probability q,
or for the queue content to remain unchanged until the next ER packet
arrival. To express this latter probability we need to define
d
k D Prfqueue content decreases by k packets between ER arrivalsg
and
D
k D Prfqueue content decreases by at least k packets
between ER arrivalsg

242

BASIC PACKET QUEUEING

The queue size decreases only when there is no arrival, and it remains
the same only when there is one arrival. For the queue to decrease
by k packets between excess arrivals, there must be k slots with no
arrivals, with probability a
0, any number of slots with one arrival, with
probability a
1, and then a slot with excess arrivals, with probability
1  a
0  a
1. So d
k is given by
d
k D

1


Ck
a
0k
a
1nk
1  a
0  a
1

nDk

In fact, we only need d
0 in the queue analysis, and this is simply
d
0 D

1


a
1n
1  a
0  a
1

nD0

which reduces to
d
0 D

1  a
0  a
1
1  a
1

We also require
D
1 D 1  d
0 D

a
0
1  a
1

Now, for the balance equations: in a similar way to the discrete fluid-flow
analysis in Chapter 9, we develop these by equating the up and down
probabilities of crossing between adjacent queue states. As before, we are
concerned with the state as seen by an excess-rate arrival, so we must
consider arriving packets one at a time. Thus the state can only ever
increase by one.
Initially, let the buffer capacity be X packets. To cross between states
X  1 and X, an arriving excess-rate packet sees X  1, taking the queue
state up to X, and another excess-rate packet follows to see the queue in
state X. This happens either immediately, with probability q, or after any
number of time units in which the queue state stays the same, i.e. with
probability
1  q d
0. So the probability of going up is
Prfgoing upg D
q C
1  q d
0 p
X  1
To go down, an arriving excess-rate packet sees X in the queue and is
lost (because the buffer is full) and then there is a gap containing any
number of time units, at least one of which is empty and the rest in which
the queue state does not change. Then the next excess-rate arrival sees
fewer than X in the queue. For this to happen, it is simply the probability
that the next ER packet does not see X, or, put another way, one minus

BALANCE EQUATIONS FOR PACKET BUFFERING: EXCESS-RATE QUEUEING ANALYSIS

243

the probability that the next ER packet does see X. This latter probability
has the same conditions as the up transition, i.e. either another ER packet
follows immediately, or it follows after any number of time units in which
the queue state stays the same. So
Prfgoing downg D
1 
q C
1  q d
0 p
X
Equating the probabilities of going up and down gives

q C
1  q d
0 p
X  1 D
1 
q C
1  q d
0 p
X
For a line between X  1 and X  2, equating probabilities gives

q C
1  q d
0 p
X  2 D
1 
q C
1  q d
0 D
1 p
X
C
1 
q C
1  q d
0 D
1 p
X  1
The left-hand side is the probability of going up, and has the same
conditions as before. The probability of coming down, on the right-hand
side of the equation, contains two possibilities. The first term is for an
arriving ER packet which sees X in the queue and is lost (because the
buffer is full) and the second term is for an arriving ER packet which
sees X  1 in the queue, taking the state of the queue up to X. Then, in
both cases, there is a period without ER packets during which the queue
content decreases by at least two empty time units, so that the next ER
arrival sees fewer than X  1 in the queue.
Substituting for p
X, and rearranging gives

q C
1  q d
0 p
X  2 D D
1 p
X  1
In the general case, for a line between X  i C 1 and X  i, the probability
of going up remains the same as before, i.e. the only way to go up is
for an ER packet to see X  i, and to be followed (either immediately
or after a period during which the queue state remains the same) by
another ER packet which sees X  i C 1. The probability of going down
consists of many components, one for each state above X  i, but they
can be arranged in two groups: the probability of coming down from
X  i C 1 itself; and the probability of coming down to below X  i C 1
from above X  i C 1. This latter is just the probability of going down
between X  i C 2 and X  i C 1 multiplied by D
1, which is the same
as going up from X  i C 1 multiplied by D
1. This is precisely the same
grouping as illustrated in Figure 9.10 for the discrete fluid-flow approach.
The general equation then is

q C
1  q d
0 p
X  i D D
1 p
X  i C 1

244

BASIC PACKET QUEUEING

so

q C
1  q d
0
p
X D p
0
D
1

X

The state probabilities must sum to 1, so

X


p
i D p
0

iD0


X 


q C
1  q d
0 i
iD0

D
1

D1

and we can find p
0 thus

q C
1  q d
0
1
D
1
p
0 D



q C
1  q d
0 XC1
1
D
1
Now, although we have assumed a finite buffer capacity of X packets, let
us now assume X ! 1. The term in the denominator for p
0 tends to 1,
and so the state probabilities can be written

q C
1  q d
0
p
k D 1 
D
1

 

q C
1  q d
0

D
1

k

Substituting for q, d
0 and D
1 gives



E[a]
1  a
1  1 C a
1 C
a
02
p
k D 1 
a
0
E[a]  1 C a
0

k
E[a]
1  a
1  1 C a
1 C
a
02

a
0
E[a]  1 C a
0

Now, although this expression looks rather messy, its structure is quite
simple:
p
k D
1  decay rate [decay rate]k
The probability that the queue exceeds k packets is then a geometric
progression which, after rearrangement, yields
Q
k D [decay rate]kC1
i.e.

E[a]
1  a
1  1 C a
1 C
a
02

Q
k D
a
0
E[a]  1 C a
0

kC1

BALANCE EQUATIONS FOR PACKET BUFFERING: EXCESS-RATE QUEUEING ANALYSIS

245

These then are the general forms for p
k and Q
k, into which we can
substitute appropriate expressions from the Poisson distribution, i.e.
E[a] D 
a
0 D e
a
1 D  e
to give


 e  e  2 C  C e
p
k D 1 
  1 C e

kC1
 e  e  2 C  C e
Q
k D
  1 C e

 e  e  2 C  C e

  1 C e

Well, was it really worth all the effort? Lets take a look at some results.
Figure 14.7 shows the queue state probabilities for three different values
of  D 0.55, 0.75, 0.95. In the figure, the lines are the exact results found
using the approach developed in Chapter 7, with Poisson input, and the
markers show the results from the excess-rate analysis with GAPP input.
Note that the results from the exact analysis are discrete, not continuous,
but are shown as continuous lines for clarity.
Figure 14.8 shows the results for Q
k, the probability that the queue
exceeds k, comparing exact and excess-rate GAPP analyses. Figure 14.9
compares the excess-rate GAPP results with those from the heavy-traffic
analysis in Chapter 8. It is clear that the excess-rate GAPP provides a
very accurate approximation to the exact results across the full range of
load values, and it is significantly more accurate than the heavy-traffic
approximation.

The excess-rate solution for best-effort traffic

But how can we use the excess-rate analysis if we have variable-length
packets, as is typical with current best-effort IP networks? The key here is
to go back to the definition of a
k, the probability of k arrivals in a packet
service time, because it is from a
0, a
1 and the mean of this distribution,
E[a], that we derive the excess-rate queueing behaviour (see Figures 14.5
and 14.6).
Previously, we assumed a constant service time, T D 1 time unit. Thus,
for packets arriving at an average rate of  packets per time unit, the
probability that there are no packets arriving in one packet service time
is given by

 T0 T
a
0 D
D e
e
0!

246

BASIC PACKET QUEUEING

Queue size
0

10

20

30

40

100
101
102

State probability

103
104
105
106

Exact 95%
ER GAPP 95%
Exact 75%
ER GAPP 75%
Exact 55%
ER GAPP 55%

107
108
109
1010
k :D 0.. 40
aP95k :D Poisson
k, 0.95
aP75k :D Poisson
k, 0.75
aP55k :D Poisson
k, 0.55


k


 e  e  2 C  C e
 e  e  2 C  C e

GAPPMDI
k,  :D 1 
  1 C e
  1 C e
xk :D k
y1 :D infiniteQ
40, aP95, 0.95
y2 :D GAPPMDI
x, 0.95
y3 :D infiniteQ
40, aP75, 0.75
y4 :D GAPPMDI
x, 0.75
y5 :D infiniteQ
40, aP55, 0.55
y6 :D GAPPMDI
x, 0.55

Figure 14.7. State Probability Distributions at Various Load Levels, Comparing the Exact Analysis
and Excess-Rate Analysis Methods

247

BALANCE EQUATIONS FOR PACKET BUFFERING: EXCESS-RATE QUEUEING ANALYSIS

Buffer capacity, X
0

10

20

30

40

100
101

Pr{queue size > X}

102
Exact 95%
ER GAPP 95%
Exact 75%
ER GAPP 75%
Exact 55%
ER GAPP 55%

103
104
105
106
107
108
109
1010

k :D 0..

40

GAPPMD1Q
k,  :D

 e  e  2 C  C e
  1 C e

kC1

xk :D k
yP1 :D infiniteQ
40, aP95, 0.95
y1 :D Q
40, yP1
y2 :D GAPPM1Q
x, 0.95
yP3 :D infiniteQ
40, aP75, 0.75
y3 :D Q
40, yP3
y4 :D GAPPMD1Q
x, 0.75
yP5 :D infiniteQ
40, aP55, 0.55
y5 :D Q
40, yP5
y6 :D GAPPMD1Q
x, 0.55
Figure 14.8. Probability that the Queue Exceeds X for Various Load Levels,
Comparing the Exact Analysis and Excess-Rate Analysis Methods

But what if 50% of the packets are short, say 40 octets, and 50% of the
packets are long, say 960 octets? The average packet length is 500 octets,
equivalent to one time unit. The probability that there are no packets
arriving in one packet service time is now a weighted sum of the two
possible situations, i.e.

248

BASIC PACKET QUEUEING

Buffer capacity, X
0

10

20

30

40

100
101

Pr{queue size > X}

102
Heavy Traffic 95%
ER GAPP 95%
Heavy Traffic 75%
ER GAPP 75%
Heavy Traffic 55%
ER GAPP 55%

103
104
105
106
107
108
109
1010

k :D 0..

40


2k

HeavyTrafficMD1Q
k,  :D e
xk :D k
y1 :D HeavyTrafficMD1Q
x, 0.95
y2 :D GAPPMD1Q
x, 0.95
y3 :D HeavyTrafficMD1Q
x, 0.75
y4 :D GAPPMD1Q
x, 0.75
y5 :D HeavyTrafficMD1Q
x, 0.55
y6 :D GAPPMD1Q
x, 0.55

1


Figure 14.9. Probability that the Queue Exceeds X for Various Load Levels,
Comparing the Excess-Rate and Heavy Traffic Approximation Methods

a
0 D 0.5

40

500
0!

0

40

960

500
0!

48

e 500 C 0.5

0

giving
a
0 D 0.5 e 25 C 0.5 e 25
The same approach applies for a
1, i.e.
a
1 D 0.5 

960

e 500

2
2
48  48
e 25 C 0.5 
e 25
25
25

BALANCE EQUATIONS FOR PACKET BUFFERING: EXCESS-RATE QUEUEING ANALYSIS

249

Lets make this more general. Assume we have short packets of length 1
time unit, and long packets of length n time units. The proportion of short
packets is ps , and so the proportion of long packets is (1  ps ). Packets
arrive at a rate of  packets per time unit. The mean service time (in time
units) is then simply
s D ps 1 C
1  ps  n
and mean number of arrivals per packet service time (i.e. the utilization)
is given by
E[a] D  D  s D  fps C
1  ps  ng
The general form for the probability of k arrivals in a packet service time
is then given by
a
k D ps

k 

n k n
e C
1  ps 
e
k!
k!

So for bi-modal packet-size distributions, we have the following expressions

E[a] D 
ps C
1  ps  n
a
0 D ps e C
1  ps  en
a
1 D ps  e C
1  ps  n  en
which we can substitute into the general forms for p
k and Q
k from the
previous section. Note that n does not have to be integer-valued.
It is not necessary to limit the packet sizes to two distinct values,
though. The above process can be generalized further to the case of the
general service-time distribution, hence providing an excess-rate solution
for the M/G/1 queueing system.
E[a] D 

1


g
i D 

iD1

a
0 D

1


g
i ei

iD1

a
1 D

1


g
i i  ei

iD1

where
g
k D Prfa packet requires k time units to be servedg
But for now, lets keep to the bi-modal case and look at some results
(in Figure 14.10) for different lengths and proportions of short and long
packets. We fix the short packet length to 40 octets, the mean packet

250

BASIC PACKET QUEUEING

Buffer capacity, X
0

10

20

30

40

100
101

Pr{queue size > X}

102
103
104
105
106

Bi-modal 2340
Geo/Geo/1
Bi-modal 960
Bi-modal 540
M/D/1

107

k :D 0.. 40



longandshortQ
k, , ps, n :D  
ps
C

1
 ps n


 Ea



ps e C
1  ps en
 a0

 a1
ps  e C
1  ps n  en


Ea
1  a1  1 C a1 C a02

 decayrate

a0
Ea  1 C a0

 decayratekC1
xk :D k
y1 :D GAPPMD1Q
x, 0.8
y2 :D longandshortQ
x, 0.8, 0.08, 13.5
y3 :D longandshortQ
x, 0.8, 0.5, 24
y4 :D longandshortQ
x, 0.8, 0.8, 58.5
y5 :D packetQ
x
Figure 14.10. Probability that the Queue Exceeds X for Different Service Distributions (Geometric, Deterministic and Bi-modal)

BALANCE EQUATIONS FOR PACKET BUFFERING: EXCESS-RATE QUEUEING ANALYSIS

251

length to 500 octets and the load to 0.8 packets arriving per mean service
time. We set the time unit to be the time to serve a short packet, so
sD

500
D ps 1 C
1  ps  n
40

Figure 14.10 shows results for three different lengths for the long packets:
2340, 960, and 540 octets. These give values of n D 58.5, 24 and 13.5
with corresponding values of ps D 0.8, 0.5 and 0.08 respectively. Also
in the figure are the results for the M/D/1 with the same load of 0.8,
and the Geo/Geo/1 results from Figure 14.1 (a load of 0.8 and mean
packet-size of 500 octets). Note that the M/D/1 gives a lower bound on
the probabilities, but the Geo/Geo/1 is not the worst case. Introducing
a small proportion of short packets, and hence slightly increasing the
length of the long packets (to maintain an average of 500 octets) results
in a decay rate only a little higher than the M/D/1. When there are equal
proportions of short and long packets, the decay rate approaches that for
the Geo/Geo/1. However, when most of the packets are short, and only
20% are (very) long packets, the decay rate is rather worse than that for
the Geo/Geo/1 queue.

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

15

Resource Reservation
go with the flow

QUALITY OF SERVICE AND TRAFFIC AGGREGATION

In recent years there have been many different proposals (such as Integrated Services [15.1], Differentiated Services [15.2], and RSVP [15.3]) for
adding quality of service (QoS) support to the current best-effort mode of
operation in IP networks. In order to provide guaranteed QoS, a network
must be able to anticipate traffic demands, assess its ability to supply the
necessary resources, and act either to accept or reject these demands for
service. This means that users must state their communications requirements in advance, in some sort of service request mechanism. The details
of the various proposals are outside the scope of this book, but in
this chapter we analyse the key queueing behaviours and performance
characteristics underlying the resource assessment.
To be able to predict the impact of new demands on resources, the
network needs to record state information. Connection-orientated technologies such as ATM record per-connection information in the network
as hard state. This information must be explicitly created for the duration
of the connection, and removed when no longer needed. An alternative
approach (adopted in RSVP) is soft state, where per-flow information is
valid for a pre-defined time interval, after which it needs to be refreshed
or, if not, it lapses.
Both approaches, though, face the challenge of scalability. Per-flow
or per-connection behaviour relates to individual customer needs. With
millions of customers, each one initiating many connections or flows,
it is important that the network can handle these efficiently, whilst still
providing guaranteed QoS. This is where traffic aggregation comes in.
ATM technology introduces the concept of the virtual path a bundle of
virtual channels whose cells are forwarded on the basis of their VPI value

254

RESOURCE RESERVATION

only. In IP, packets are classified into behaviour aggregates, identified by

a field in the IP header, and forwarded and queued on the basis of the
value of that field.
In this chapter, we concentrate on characterizing these traffic aggregates, and analysing their impact on the network to give acceptable
QoS for the end users. Indeed, our approach divides into these two
stages: aggregation, and analysis (using the excess-rate analysis from
Chapter 9).

CHARACTERIZING AN AGGREGATE OF PACKET FLOWS

In the previous chapter, we assumed that the arrival process of packets
could be described by a Poisson distribution (which we modified slightly,
to derive accurate results for both M/D/1 and M/G/1 queueing systems).
This assumption allowed for multiple packets, from different input ports,
to arrive simultaneously (i.e. within one packet service time) at an output
port, and hence require buffering. This is a valid assumption when the
input and output ports are of the same speed (bit-rate) and there is no
correlation between successive arrivals on an input port.
However, if the input ports are substantially slower than the output
port (e.g. in a typical access multiplexing scenario), or packets arrive in
bursts at a rate slower than that allowed by the input port rate (within
the core network), then the Poisson assumption is less valid. Why? Well,
suppose that the output port rate is 1000 packet/s and the traffic on
the input port is limited to 100 packet/s (either because of a physical
bit-rate limit, or because of the packet scheduling at the previous router).
The minimum time between arrivals from any single input port is then
10 ms, during which time the output port could serve up to 10 packets.
The Poisson assumption allows for arrivals during any of the 10 packet
service times, but the actual input process does not.
So, we characterize these packet flows as having a mean duration, Ton ,
and an arrival rate when active, h (packet/s). Thus each flow comprises
Ton h packets, on average. If the overall mean load is Ap packet/s, then
the rate of flows arriving is simply
FD

Ap
Ton h

We can interpret this arrival process in terms of erlangs of offered traffic:

offered traffic D

Ap
D F Ton
h

i.e. the flow attempt rate multiplied by the mean flow duration.

255

PERFORMANCE ANALYSIS OF AGGREGATE PACKET FLOWS

1
2

PI input ports

Output port
of interest

PI

Figure 15.1. Access Multiplexor or Core Router

It may be that there is a limit on the number of input ports, PI , sending

flows to the particular output port of interest (see Figure 15.1). In this
case, the two scenarios (access multiplexor, or core router/switch) differ
in terms of the maximum number of flows, N, at the output port. For the
access multiplexor, with slow speed input ports of rate h packet/s, the
maximum number of simultaneous flows is
N D PI
However, for the core router with input port speeds of C packet/s, the
maximum possible number of simultaneous flows it can support is


C
N D PI
h
i.e. each input port can carry multiple flows, each of rate h, which have
been multiplexed together upstream of this router.

PERFORMANCE ANALYSIS OF AGGREGATE PACKET FLOWS

The first task is to simplify the traffic model, comprising N input sources,
to one in which there is a single aggregate input process to the buffer (see
Figure 15.2), thus reducing the state space from 2N possible states to just
2. This aggregate process is either in an ON state, in which the input rate
exceeds the output rate, or in an OFF state, when the input rate is not
zero, but is less than the output rate.
For the aggregate process, the mean rate in the ON state is denoted
Ron , and in the OFF state is Roff . When the aggregate process is in the ON
state, the total input rate exceeds the service rate, C, of the output port,
and the buffer fills:
rate of increase D Ron  C

256

RESOURCE RESERVATION

N
N1
N2
Exponentially
distributed
OFF period

Exponentially
distributed
ON period

Ron

Roff

2N state process

Figure 15.2.

2 state process

State Space Reduction for Aggregate Traffic Process

The average duration of this period of increase is denoted T
on. To be

in the ON state, more than C/h sources must be active. Otherwise the
aggregate process is in the OFF state. This is illustrated in Figure 15.3.
In the OFF state, the total input rate is less than the service rate of the
output port, so, allowing the buffer to empty,
rate of decrease D C  Roff
The average duration of this period of decrease is denoted T
off.
Reducing the system in this manner has obvious attractions; however,
just having a simplifying proposal does not lead directly to the model in
detail. Specifically, we need to find values for the four parameters in our
two-state model, a process which is called parameterization.

No. sources active

S
.
.
.
C/h
C/h -1
.
.
.
3
2

ON period

ON period
T(on) = mean ON time
Ron = mean ON rate

Channel
capacity = C

OFF period

OFF period
T(off) = mean OFF time
Roff = mean OFF rate

Figure 15.3.

Time

Two-State Model of Aggregate Packet Flows

257

PERFORMANCE ANALYSIS OF AGGREGATE PACKET FLOWS

Parameterizing the two-state aggregate process

Consider the left-hand side of Figure 15.3. Here we show the combined
input rates, depending on how many packet flows are active. The capacity
assigned to this traffic aggregate is C packet/s this may be the total
capacity of the output port, or just a fraction if there is, for example, a
weighted fair queue scheduling scheme in operation. If C/h packet flows
are active, then the input and output rates of the queue are equal, and
the queue size remains constant. From the burst-scale point of view,
the queue is constant, although there will be small-scale fluctuations
due to the precise timing of packet arrival and departure instants. If
more packet flows are active, the queue increases in size because of
the excess rate; with fewer packet flows active, the queue decreases
in size.
Let us now view the queueing system from the point of view of the
arrival and departure of packet flows. The maximum number of packet
flows that can be served simultaneously is
N0 D

C
h

We can therefore think of the output port as having N0 servers and a

buffer for packet flows which are waiting to be served. If we can find the
mean number waiting to be served, given that there are some waiting,
we can then calculate the mean rate in the ON state, Ron , as well as the
mean duration in the ON state, T
on.
Assuming a memoryless process for the arrival of packet flows (a
reasonable assumption, since flows are typically triggered by user
activity), this situation is then equivalent to the system modelled by
Erlangs waiting-call analysis. Packet flows are equivalent to calls, the
output port is equivalent to N0 circuits, and we assume infinite waiting
space. The offered traffic, in terms of packet flows, is given by
AD

Ap
D F Ton
h

Erlangs waiting-call formula gives the probability of a call (packet flow)

being delayed as



N0
N0  A

DD


N
1
0

AN0
Ar
N0
C

r!
N0 !
N0  A
rD0

AN0

N0!

258

RESOURCE RESERVATION

or, alternatively, in terms of Erlangs loss probability, B, we have

DD

N0 B
N0  A C A B

The mean number of calls (packet flows) waiting, averaged over all calls,
is given by
A
wDD
N0  A
But what we need is the mean number waiting, conditioned on there
being some waiting. This is simply given by
A
w
D
D
N0  A
Thus, when the aggregate traffic is in the ON state, i.e. there are some
packet flows waiting, then the mean input rate to the output port exceeds
the service rate. This excess rate is simply the product of the conditional
mean number waiting and the packet rate of a packet flow, h. So
Ron D C C h

Ap
A
DCCh
N0  A
C  Ap

The mean duration in the excess-rate (ON) state is the same as the
conditional mean delay for calls in the waiting-call system. From Littles
formula, we have
A
w D F tw D
tw
Ton
which, on rearranging and substituting for w, gives
tw D

Ton
A
Ton
wD
D
A
A
N0  A

So, the conditional mean delay is

T
on D

tw
Ton
h Ton
D
D
D
N0  A
C  Ap

This completes the parameterization of the ON state. In order to parameterize the OFF state we need to make use of D, the probability that a
packet flow is delayed. This probability is, in fact, the probability that the

259

PERFORMANCE ANALYSIS OF AGGREGATE PACKET FLOWS

aggregate process is in the ON state, which is the long-run proportion of

time in the ON state. So we can write
T
on
DD
T
on C T
off 
which, after rearranging, gives
T
off  D T
on

1D
D

The mean load, in packet/s, is the weighted sum of the rates in the ON
and OFF states, i.e.
Ap D D Ron C
1  D Roff
and so
Roff D

Ap  D Ron
1D

Analysing the queueing behaviour

We have now aggregated the Poisson arrival process of packet flows
into a two-state ONOFF process. This is very similar to the ONOFF
source model in the discrete fluid-flow approach presented in Chapter 9,
except that the OFF state now has a non-zero arrival rate associated with
it. In the ON state, we assume that there are a geometrically distributed
number of excess-rate packet arrivals. In the OFF state, we assume that
there are a geometrically distributed number of free periods in which
to serve excess-rate packets. Thus the geometric parameters a and s are
given by
1
aD1
T
on
Ron  C
and
sD1

1
T
off 
C  Roff 

For a finite buffer size of X, we had the following results from Chapter 9:
p
X  1 D
and
p
X  i D

1a
p
X
a

s
p
X  i C 1
a

260

RESOURCE RESERVATION

The state probabilities, p
k, form a geometric progression which can be

written as
 k

p
0
0<k<X

s
p
k D 
  k

s
a

p
0
kDX
1a
s
These state probabilities must sum to 1, and so, after some rearrangement,
we can find p
0 thus:
a
s
p
0 D

  X
1s
a
1

1a
s
1

Now, although we have assumed a finite buffer capacity of X packets

for this excess-rate analysis, let us now assume X ! 1. The term in the
denominator for p
0 tends to 1, and so the state probabilities can be
written

  k
a
a
p
k D 1 

s
s
As we found in the previous chapter for this form of expression, the
probability that the queue exceeds k packets is then a geometric progression, i.e.
 kC1
a
Q
k D
s
This result is equivalent to the burst-scale delay factor it is the probability that excess-rate packets see more than k in the queue. It is in our,
now familiar, decay rate form, and provides an excellent approximation
to the probability that a finite buffer of length k overflows. This latter is a
good approximation to the loss probability.
However, we have not quite finished. We now need an expression for
the probability that a packet is an excess-rate arrival. In the discrete fluidflow model of Chapter 9, this was simply
R  C/R the proportion of
arrivals that are excess-rate arrivals. This simple expression needs to be
modified because when the aggregate process is in the OFF state, packets
are still arriving at the queue.
We need to find the ratio of the mean excess rate to the mean arrival
rate. If we consider a single ONOFF cycle of the aggregate model, then
this ratio is the mean number of excess packets in an ON period to the

261

VOICE-OVER-IP, REVISITED

mean number of packets arriving in the ONOFF cycle. Thus

Prfpacket is excess-rate arrivalg D

Ron  C T
on
Ap
T
on C T
off 

which, after substituting for Ron , T
on and T
off , gives

Prfpacket is excess-rate arrivalg D

hD
C  Ap

The queue overflow probability is then given by the expression

xC1
1
1

hD
T
on
Ron  C

Q
x D

C  Ap
1
T
off 
C  Roff 

VOICE-OVER-IP, REVISITED
In the last chapter we looked at the excess-rate M/D/1 analysis as
a suitable model for voice-over-IP. The assumption of a deterministic
server is reasonable, given that voice packets tend to be of fixed size,
and the Poisson arrival process is a good limit for N CBR sources when
N is large (as we found in Chapter 8). But if the voice sources are using
activity detection, then they do not send packets during silent periods.
Thus we have ONOFF behaviour, which can be viewed as a series of
overlapping packet flows (see Figure 15.1).
Suppose we have N D 100 packet voice sources, each producing packets
at a rate of h D 167 packet/s, when active, into a buffer of size X D 100
packets and service capacity C D 7302.5 packet/s. The mean time when
active is Ton D 0.35 seconds and when inactive is Toff D 0.65 second, thus
each source has, on average, one active period every Ton C Toff D 1 second.
The rate at which these active periods arrive, from the population of N
packet sources, is then
FD

N
D 100 s1
Ton C Toff

Therefore, we can find the overall mean load, Ap , and the offered traffic,
A, in erlangs.
Ap D F Ton h D 100 0.35 167 D 5845 packet/s

262

RESOURCE RESERVATION

A D F Ton D 100 0.35 D 35 erlangs

and the maximum number of sources that can be served simultaneously,
without exceeding the buffers service rate is
N0 D

C
D 43.728
h

which needs to be rounded down to the nearest integer, i.e. N0 D 43. Lets
now parameterize the two-state excess-rate model.
AN0
N !
B D N 0 D 0.028 14
0

Ar
rD0

DD

r!

N0 B
D 0.134 66
N0  A C A B

Ron D C C h

Ap
D 7972.22
C  Ap

Ap  D Ron
D 5513.98
1D
h Ton
D 0.0401
T
on D
C  Ap
Roff D

T
off  D T
on

1D
D 0.257 71
D

We can now calculate the geometric parameters, a and s, and hence the
decay rate.
aD1

1
D 0.962 77
T
on
Ron  C

sD1

1
D 0.997 83
T
off 
C  Roff 

decay rate D

a
D 0.964 86
s

The probability that a packet is an excess-rate arrival is then

Prfpacket is excess-rate arrivalg D

hD
D 0.015 43
C  Ap

263

VOICE-OVER-IP, REVISITED

and the packet loss probability is estimated by

hD

Q
X D
C  Ap

 XC1

a
s

D 4.161 35 104

Figure 15.4 shows these analytical results on a graph of Q
x against x. The
Mathcad code to generate the analytical results is shown in Figure 15.5.
Also shown, as a dashed line in Figure 15.4, are the results of applying
the burst-scale analysis (both loss and delay factors, from Chapter 9) to
the same scenario. Simulation results for this scenario show a decay rate
of approximately 0.97. The figure of 0.964 86 obtained from the excessrate aggregate flow analysis is very close to these simulation results,
and illustrates the accuracy of the excess-rate technique. In contrast, the
burst-scale delay factor gives a decay rate of 0.998 59. This latter is typical
of other published techniques which tend to overestimate the decay rate
by a significant margin; the interested reader is referred to [15.4] for a
more comprehensive comparison.
If we return to the M/D/1 scenario, where we assume that the voice
sources are of a constant rate, how many sources can be supported
over the same buffer, and with the same packet loss probability? The
excess-rate analysis gives us the following equation:


 e  e  2 C  C e
Q
100 D
  1 C e

101

D 4.161 35 104

Buffer capacity, X
0

10

20

30

40

50

60

70

80

90 100

Pr{queue size > X}

10 0
101
102
103
104
105

Figure 15.4. Packet Loss Probability Estimate for Voice Sources, Based on ExcessRate Aggregate Flow Analysis

264

RESOURCE RESERVATION

k :D 0.. 100


Ap
afQ
k , h , Tflow , Ap, C  :D A
h

 


C
 N0
floor

h




AN0
1
B

r

N0! N0 A


rD0

r!


N0 B
D

N0  A C A B


h Tflow

 Ton

C  Ap




1-D

Ton
 Toff

D


Ap
 Ron
CCh

C  Ap



Ap  D Ron
 Roff

1D



1

1

Ton

Ron
 C
 decayrate

1

1

Toff
C  Roff



hD
 probexcess

C  Ap

probexcess decayratekC1
xk :D k
y :D afQ
x , 167 , 0.35 , 5845 , 7302.5 
Figure 15.5.

Mathcad Code for Excess-Rate Aggregate Flow Analysis

which is plotted in Figure 15.6 for values of load ranging from 0.8 up to
0.99 of the service capacity.
The value of loss probability we require occurs at an offered load of
about 0.96; in fact 0.961 yields a loss probability of 4.164 104 . This
offered load is just the total input rate from all the CBR sources, divided
by the service rate of the output port. So, we have
NCBR h
D 0.961
C
0.961 7302.5
D 42.02
NCBR D
167

265

TRAFFIC CONDITIONING OF AGGREGATE FLOWS

0.80

0.85

Offered load
0.90

0.95

1.00

100
Packet loss probability

101
102
103
104
105
106
107
108
109
1010
Figure 15.6. Packet Loss Probability Estimate for Voice Sources, Based on ExcessRate M/D/1 Analysis

Thus, instead of transporting 100 voice sources, we can only carry 42 if

there is no activity detection. This gives us nearly two and half times
increase in loading efficiency when activity detection is used.

TRAFFIC CONDITIONING OF AGGREGATE FLOWS

Within both the Integrated Services and Differentiated Services architectures [15.1, 15.2], the concept of a token bucket is introduced to describe
the load imposed by either individual, or aggregate, flows. The token
bucket is, in essence, the same as the leaky bucket used in ATM usage
parameter control, which we described in Chapter 11. It is normally
viewed as a pool, of capacity B octet tokens, being filled at a rate of R
octet token/s. If the pool contains enough tokens for the arriving packet,
then the packet is sent on into the network, and the token bucket is
drained of the appropriate number of octet tokens. However, if there
are insufficient tokens, then the packet is either discarded, or marked
as best-effort, or delayed until enough tokens have replenished the
bucket.
In both architectures, the token bucket can be used to define a traffic
profile, and hence police traffic flows (either single or aggregate). As
we found in Chapter 11, we can analyse the leaky (or token) bucket as
a buffer, even though in reality it is not used to delay packets. So, if
we have an aggregate flow, the same analysis used to assess queueing
performance can be used to dimension the token bucket.

266

RESOURCE RESERVATION

Token rate, R packet/s

10000
9500
9000

12

10

8500
8000

7500

7000

6
4

12
10

6
4

6500
0

20

40
60
80
100
Token bucket size, B packets

120

Figure 15.7. Example of Relationship between Token Bucket Parameter Values for
Voice-over-IP Aggregate Traffic

Figure 15.7 shows the relationship between B and R for various values
of the packet loss probability estimate (102 down to 1012 ). The scenario
is the aggregate flow of voice-over-IP traffic, using the parameter values
and formulas in the previous section. The tokens are equivalent to packets,
rather than octets, in this figure. A simple scaling factor (the number of
octets per packet) can be applied to convert to octets. There is a clear
trade-off between rate allocation (R) and burstiness (B) for the aggregate
flow. With a smaller rate allocation, the aggregate flow exceeds this value
more often, and so a larger token bucket is required to accommodate
these bursts.

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

16

IP Buffer Management
packets in the space time continuum

FIRST-IN FIRST-OUT BUFFERING

In the chapters on packet queueing, we have so far only considered
queues with first-in-first-out (FIFO) scheduling. This approach gives all
packets the same treatment: packets arriving to a buffer are placed at the
back of the queue, and have to wait their turn for service, i.e. after all the
other packets already in the queue have been served. If there is insufficient
space in the buffer to hold an arriving packet, then it is discarded.
In Chapter 13, we considered priority control in ATM buffers, in terms
of space priority (access to the waiting space) and time priority (access
to the server). These mechanisms enable end-to-end quality-of-service
guarantees to be provided to different types of traffic in an integrated
way. For IP buffer management, similar mechanisms have been proposed
to provide QoS guarantees, improved end-to-end behaviour, and better
use of resources.

RANDOM EARLY DETECTION PROBABILISTIC PACKET

DISCARD
One particular challenge of forwarding best-effort packet traffic is that
the transport-layer protocols, especially TCP, can introduce unwelcome
behaviour when the network (or part of it) is congested. When a TCP
connection loses a packet in transit (e.g. because of buffer overflow),
it responds by entering the slow-start phase which reduces the load
on the network and hence alleviates the congestion. The unwelcome
behaviour arises when many TCP connections do this at around the
same time. If a buffer is full and has to discard arriving packets from
many TCP connections, they will all enter the slow-start phase. This
significantly reduces the load through the buffer, leading to a period

268

IP BUFFER MANAGEMENT

of under-utilization. Then all those TCP connections will come out of

slow-start at about the same time, leading to a substantial increase in
traffic and causing congestion in the buffer. More packets are discarded,
and the cycle repeats this is called global synchronization.
Random early detection (RED) is a packet discard mechanism that anticipates congestion by discarding packets probabilistically before the buffer
becomes full [16.1]. It does this by monitoring the average queue size,
and discarding packets with increasing probability when this average is
above a configurable threshold,
min . Thus in the early stages of congestion, only a few TCP connections are affected, and this may be sufficient
to reduce the load and avoid any further increase in congestion. If the
average queue size continues to increase, then packets are discarded with
increasing probability, and so more TCP connections are affected. Once
the average queue size exceeds an upper threshold,
max , all arriving
packets are discarded.
Why is the average queue size used why not use the actual queue
size (as with partial buffer sharing (PBS) in ATM)? Well, in ATM we
have two different levels of space priority, and PBS is an algorithm for
providing two distinct levels of cell loss probability. The aim of RED is to
avoid congestion, not to differentiate between priority levels and provide
different loss probability targets. If actual queue sizes are used, then
the scheme becomes sensitive to transient congestion short-lived bursts
which dont need to be avoided, but just require the temporary storage
space of a large buffer. By using average queue size, these short-lived
bursts are filtered out. Of course, the bursts will increase the average
temporarily, but this takes some time to feed through and, if it is not
sustained, the average will remain below the threshold.
The average is calculated using an exponentially weighted moving
average (EWMA) of queue sizes. At each arrival, i, the average queue
size, qi , is updated by applying a weight, w, to the current queue size, ki .
qi D w ki C 1  w qi1
How quickly qi responds to bursts can be adjusted by setting the weight,
w. In [16.1] a value of 0.002 is used for many of the simulation scenarios,
and a value greater than or equal to 0.001 is recommended to ensure
adequate calculation of the average queue size.
Lets take a look at how the EWMA varies for a sample set of packet
arrivals. In Figure 16.1 we have a Poisson arrival process of packets, at a
load of 90% of the server capacity, over a period of 5000 time units. The
thin grey line shows the actual queue state, and the thicker black line
shows the average queue size calculated using the EWMA formula with
w D 0.002. Figure 16.2 shows the same trace with a value of 0.01 for the
weight, w. It is clear that the latter setting is not filtering out much of the
transient behaviour in the queue.

269

RANDOM EARLY DETECTION PROBABILISTIC PACKET DISCARD

30

Queue size

20

10

0
5000

6000

7000

8000

9000

10000

Time
Figure 16.1.
w D 0.002

Sample Trace of Actual Queue Size (Grey) and EWMA (Black) with

30

Queue size

20

10

0
5000

6000

7000

8000

9000

10000

Time
Figure 16.2.
w D 0.01

Sample Trace of Actual Queue Size (Grey) and EWMA (Black) with

270

IP BUFFER MANAGEMENT

Configuring the values of the thresholds,
min and
max , depends on the
target queue size, and hence system load, required. In [16.1] a rule of
thumb is given to set
max > 2
min in order to avoid the synchronization
problems mentioned earlier, but no specific guidance is given on setting

min . Obviously if there is not much difference between the thresholds,
then the mechanism cannot provide sufficient advance warning of potential congestion, and it soon gets into a state where it drops all arriving
packets. Also, if the thresholds are set too low, this will constrain the
normal operation of the buffer, and lead to under-utilization. So, are there
any useful indicators?
From the packet queueing analysis in the previous two chapters, we
know that in general the queue state probabilities can be expressed as
pk D 1  dr  dr k
where dr is the decay rate, k is the queue size and pk is the queue state
probability. The mean queue size can be found from this expression, as
follows:
1
1




k pk D 1  dr 
k dr k
qD
kD1

kD1

Multiplying both sides by the decay rate gives

dr q D 1  dr 

1

k  1 dr k

kD2

If we now subtract this equation from the previous one, we obtain

1  dr  q D 1  dr 

1

dr k

kD1

qD

1

dr k

kD1

Multiplying both sides by the decay rate, again, gives

dr q D

1

dr k

kD2

And, as before, we now subtract this equation from the previous one to
obtain
1  dr  q D dr
qD

dr
1  dr

271

RANDOM EARLY DETECTION PROBABILISTIC PACKET DISCARD

For the example shown in Figures 16.1 and 16.2, assuming a fixed packet
size (i.e. the M/D/1 queue model) and using the GAPP formula with a
load of 0.9 gives a decay rate of

 e  e  2 C  C e 
dr D


  1 C e

D 0.817

D0.9

and a mean queue size of

qD

0.817
D 4.478
1  0.817

which is towards the lower end of the values shown on the EWMA traces.
Figure 16.3 gives some useful indicators to aid the configuration of the
thresholds,
min and
max . These curves are for both the mean queue size
against decay rate, and for various levels of probability of exceeding a
threshold queue size. Recall that the latter is given by
Prfqueue size > kg D Qk D dr kC1

Queue size, or threshold

103

Q(k) = 0.0001
Q(k) = 0.01
Q(k) = 0.1
mean queue size

102

101

100

0.80

Figure 16.3.
Decay Rate

0.85

0.90
Decay rate

0.95

1.00

Design Guide to Aid Configuration of Thresholds, Given Required

272

IP BUFFER MANAGEMENT

So, to find the threshold k, given a specified probability, we just take logs
of both sides and rearrange thus:
threshold D

logPrfthreshold exceededg
1
logdr 

Note that this defines a threshold in terms of the probability that the
actual queue size exceeds the threshold, not the probability that the
EWMA queue size exceeds the threshold. But it does indicate how the
queue behaviour deviates from the mean size in heavily loaded queues.
But what if we want to be sure that the mechanism can cope with a
certain level of bursty traffic, without initiating packet discard? Recall
the scenario in Chapter 15 for multiplexing an aggregate of packet flows.
There, we found that although the queue behaviour did not go into the
excess-rate ON state very often, when it did, the bursts could have a
substantial impact on the queue (producing a decay rate of 0.964 72). It
is thus the conditional behaviour of the queueing above the long-term
average which needs to be taken into account. In this particular case, the
decay rate of 0.964 72 has a mean queue size of
qD

0.964 72
D 27.345 packets
1  0.964 72

The long-term average load for the scenario is

D

5845
D 0.8
7302.5

If we consider this as a Poisson stream of arrivals, and thus neglect the

bursty characteristics, we obtain a decay rate of

 e  e  2 C  C e 
dr D
D 0.659


  1 C e
D0.8

and a long-term average queue size of

qD

0.659
D 1.933 packets
1  0.659

It is clear, then, that the conditional behaviour of bursty traffic dominates

the shorter-term average queue size. This is additional to the longer-term
average, and so the sum of these two averages, i.e. 29.3, gives us a good
indicator for the minimum setting of the threshold,
min .

273

VIRTUAL BUFFERS AND SCHEDULING ALGORITHMS

VIRTUAL BUFFERS AND SCHEDULING ALGORITHMS

The disadvantage of the FIFO buffer is that all the traffic has to share
the buffer space and server capacity, and this can lead to problems
such as global synchronization as we saw in the previous section. The
principle behind the RED algorithm is that it applies the brakes gradually initially affecting only a few end-to-end connections. Another
approach is to partition the buffer space into virtual buffers, and use a
scheduling mechanism to divide up the server capacity between them.
Whether the virtual buffers are for individual flows, aggregates, or
classes of flows, the partitioning enables the delay and loss characteristics
of the individual virtual buffers to be tailored to specific requirements.
This helps to contain any unwanted congestion behaviour, rather than
allowing it to have an impact on all traffic passing through a FIFO output
port. Of course, the two approaches are complementary if more than
one flow shares a virtual buffer, then applying the RED algorithm just to
that virtual buffer can avoid congestion for those particular packet flows.

Precedence queueing
There are a variety of different scheduling algorithms. In Chapter 13, we
looked at time priorities, also called head-of-line (HOL) priorities, or
precedence queueing in IP. This is a static scheme: each arriving packet
has a fixed, previously defined, priority level that it keeps for the whole
of its journey across the network. In IPv4, the Type of Service (TOS) field
can be used to determine the priority level, and in IPv6 the equivalent
field is called the Priority Field. The scheduling operates as follows (see
Figure 16.4): packets of priority 2 will be served only if there are no packets

Inputs

.
.
.

.
.
.
Packet router

Priority 1 buffer

.
.
.

Outputs

server

Priority 2 buffer
..
Priority P buffer

Figure 16.4. HOL Priorities, or Precedence Queueing, in IP

274

IP BUFFER MANAGEMENT

of priorities 1; packets of priority 3 will be served only if there are no

packets of priorities 1 and 2, etc. Any such system, when implemented in
practice, will have to predefine P, the number of different priority classes.
From the point of view of the queueing behaviour, we can state that, in
general, the highest-priority traffic sees the full server capacity, and each
next highest level sees what is left over, etc. In a system with variable
packet lengths, the analysis is more complicated if the lower-priority
traffic streams tend to have larger packet sizes. Suppose a priority-2
packet of 1000 octets has just entered service (because the priority-1
virtual buffer was empty), but a short 40-octet priority-1 packet turns up
immediately after this event. This high-priority packet must now wait
until the lower-priority packet completes service during which time as
many as 25 such short packets could have been served.

Weighted fair queueing

The problem with precedence queueing is that, if the high-priority loading
on the output port is too high, low-priority traffic can be indefinitely
postponed. This is not a problem in ATM because the traffic control
framework requires resources to be reserved and assessed in terms of the
end-to-end quality of service provided. In a best-effort IP environment
the build-up of a low-priority queue will not affect the transfer of
high-priority packets, and therefore will not cause their end-to-end
transport-layer protocols to adjust.
An alternative is round robin scheduling. Here, the scheduler looks at
each virtual buffer in turn, serving one packet from each, and passing over
any empty virtual buffers. This ensures that all virtual buffers get some
share of the server capacity, and that no capacity is wasted. However,
short packets are penalized the end-to-end connections which have
longer packets get a greater proportion of the server capacity because it
is shared out according to the number of packets.
Weighted fair queueing (WFQ) shares out the capacity by assigning
weights to the service of the different virtual buffers. If these weights
are set according to the token rate in the token bucket specifications
for the flows, or flow aggregates, and resource reservation ensures that
the sum of the token rates does not exceed the service capacity, then
WFQ scheduling effectively enables each virtual buffer to be treated
independently with a service rate equal to the token rate.
If we combine WFQ with per-flow queueing (Figure 16.5), then the
buffer space and server capacity can be tailored according to the delay
and loss requirements of each flow. This is optimal in a traffic control
sense because it ensures that badly behaved flows do not cause excessive
delay or loss among well-behaved flows, and hence avoids the global
synchronization problems. However, it is non-optimal in the overall loss

275

BUFFER SPACE PARTITIONING

.
.
.

N IP
flows
entering
a buffer

Single o/p
line

Figure 16.5. Per-flow Queueing, with WFQ Scheduling

sense: it makes far worse use of the available space than would, for
example, complete sharing of a buffer. This can be easily seen when you
realize that a single flows virtual buffer can overflow, so causing loss,
even when there is still plenty of space available in the rest of the buffer.
Each virtual buffer can be treated independently for performance analysis, so any of the previous approaches covered in this book can be re-used.
If we have per-flow queueing, then the input traffic is just a single source.
With a variable-rate flow, the peak rate, mean rate and burst length can be
used to characterize a single ONOFF source for queueing analysis. If we
have per-class queueing, then whatever is appropriate from the M/D/1,
M/G/1 or multiple ONOFF burst-scale analyses can be applied.

BUFFER SPACE PARTITIONING

We have covered a number of techniques for calculating the decay rate,
and hence loss probability, at a buffer, given certain traffic characteristics.
In general, the loss probability can be expressed in terms of the decay
rate, dr , and buffer size, X, thus:
loss probability Prfqueue size > Xg D QX D dr XC1
This general form can easily be rearranged to give a dimensioning formula
for the buffer size:
X

logloss probability
1
logdr 

For realistically sized buffers, one packet space will make little difference,
so we can simplify this equation further to give
X

logloss probability
logdr 

276

IP BUFFER MANAGEMENT

But many manufacturers of switches and routers provide a certain

amount of buffer space, X, at each output port, which can be partitioned
between the virtual buffers according to the requirements of the different
traffic classes/aggregates. The virtual buffer partitions are configurable
under software control, and hence must be set by the network operator in a
way that is consistent with the required loss probability (LP) for each class.
Lets take an example. Recall the scenario for Figure 14.10. There were
three different traffic aggregates, each comprising a certain proportion of
long and short packets, and with a mean packet length of 500 octets. The
various parameters and their values are given in Table 16.1.
Suppose each aggregate flow is assigned a virtual buffer and is served
at one third of the capacity of the output port, as shown in Figure 16.6. If
we want all the loss probabilities to be the same, how do we partition the
available buffer space of 200 packets (i.e. 100 000 octets)? We require
LP dr1 X1 D dr2 X2 D dr3 X3
given that
X1 C X2 C X3 D X D 200 packets
By taking logs, and rearranging, we have
X1 logdr1  D X2 logdr2  D X3 logdr3 
Table 16.1. Parameter Values for Bi-modal Traffic Aggregates

Parameter

Bi-modal 540

Bi-modal 960

Bi-modal 2340

40
540
13.5
0.08
0.064
0.8
0.4628
0.33 982
0.67 541

40
960
24
0.5
0.064
0.8
0.57 662
0.19 532
0.78 997

40
2340
58.5
0.8
0.064
0.8
0.75 514
0.06 574
0.91 454

Short packets (octets)

Long packets (octets)
Ratio of long to short, n
Proportion of short packets, ps
Packet arrival rate, 
E[a]
a(0)
a(1)
Decay rate, dr

X1
X2
X3

C/3
C/3

Service rate
C packet/s

C/3

Figure 16.6. Example of Buffer Space Partitioning

277

BUFFER SPACE PARTITIONING

and so
X2 D

X1 logdr1 
logdr2 

X3 D

X1 logdr1 
logdr3 

Summing up both sides of these equations gives

X2 C X3 D X  X1 D

X1 logdr1  X1 logdr1 
C
logdr2 
logdr3 

To make it a generic formula, we have



X1 logdr1 
X1 logdr1 
X1 logdr1  X1 logdr1 

X  X1 D
C
C
logdr1 
logdr2 
logdr3 
logdr1 


3


X1 logdr1 
 X1
X  X1 D
logdr

j
jD1
X D X1 logdr1 

3


jD1

1
logdrj 

So we can write

X


Xi D
logdri 

3


jD1

1
logdrj 

In our example, we have

X D 200 packets
dr1 D 0.675 41
dr2 D 0.789 97
dr3 D 0.914 54
By applying our partitioning formula, we obtain (to the nearest whole
packet)
X1 D 28 packets
X2 D 47 packets
X3 D 125 packets
This gives loss probabilities for each of the virtual buffers of approximately 1.5 105 .

278

IP BUFFER MANAGEMENT

If we want to achieve different loss probabilities for each of the traffic

classes, we can introduce a scaling factor, Si , associated with each traffic
class. For example, we may require the loss probabilities in our example
to be related as
10 000 LP1 D 100 LP2 D 1 LP3
i.e.
S1 D 10 000
S2 D 100
S3 D 1
Lets generalize to any number, V, of virtual buffers, and modify the
previous approach by allowing for any scaling of loss probability.
S1 dr1 X1 D S2 dr2 X2 D D Sj drj Xj D D SV drV XV
XD

V

Xj

jD1

By taking logs, and rearranging, we have

logS1  C X1 logdr1  D logS2  C X2 logdr2 
D D logSV  C XV logdrV 
and so
X2 D

logS1  C X1 logdr1   logS2 

logdr2 

..
.
XV D

logS1  C X1 logdr1   logSV 

logdrV 

After summing up both sides, and some rearrangement, we have the

general formula

logSj 
XC
logdrj 
logSi 
jD1
Xi D


V
logdr


i
1
logdri 
logdrj 
jD1
V

279

SHARED BUFFER ANALYSIS

This buffer partitioning formula can be used to evaluate the correct size
of the partition allocated to any traffic class depending only on knowing
the total space available, the decay rates and the desired scaling factors.
For our example with three virtual buffers, we have
X D 200 packets
dr1 D 0.675 41
dr2 D 0.789 97
dr3 D 0.914 54
and
S1 D 10 000
S2 D 100
S3 D 1
By applying the general partitioning formula, we obtain (to the nearest
whole packet)
X1 D 46 packets
X2 D 56 packets
X3 D 98 packets
This gives loss probabilities for each of the virtual buffers of
LP1 D 1.446 108
LP2 D 1.846 106
LP3 D 1.577 104

SHARED BUFFER ANALYSIS

Earlier we noted that partitioning a buffer is non-optimal in the overall
loss sense. Indeed, if buffer space is shared between multiple output
ports, much better use can be made of the resource (see Figure 16.7). But
can we quantify this improvement? The conventional approach is to take
the convolution of the individual state probability distributions, in order
to combine all the possibilities of having a certain total number in the
buffer. Assuming the same state probability distributions for each of the
output buffers, and that the arrivals to each buffer are independent of
each other, let
PN k D Prfqueue state for N buffers sharing D kg

280

IP BUFFER MANAGEMENT

N=8 virtual
o/p buffers
Example
switch
element
with 8 i/p
and 8 o/p
lines, and
a single
shared o/p
buffer.

N=8 o/p
lines

Internal switching fabric - details

not important

Shared o/p buffer

Figure 16.7. Example of a Switch/Router with Output Ports Sharing Buffer Space

The autoconvolution for two buffers sharing is given by

P2 k D

k

P1 j P1 k  j

jD0

i.e. to find the probability that the shared buffer is in state k, find all the
different ways in which the two individual queues can have k packets
between them, and sum these probabilities.
The autoconvolution for N buffers sharing the buffer space can then be
constructed recursively, i.e.

PN k D

k

PN1 j P1 k  j

jD0

and the loss probability estimated from

QN k D 1 

k

PN j

jD0

From the packet queueing analysis in the previous two chapters, we

know that in general the state probabilities for an individual queue can
be expressed as
pk D 1  dr  dr k

281

SHARED BUFFER ANALYSIS

where dr is the decay rate, k is the queue size, and pk is the individual queue state probability. The autoconvolution of this geometric
distribution is given by a negative binomial; thus
PN k D kCN1 CN1 dr k 1  dr N
where k is the size of the combined queues in the shared buffer.
The probability that the shared buffer overflows is expressed as

QN k  1 D 1 

k1

PN j D

jD0

1

PN j

jDk

To avoid having to do summations, a geometric approximation can be

applied to the tail of the negative binomial, i.e.
QN k  1 D

1

PN j

jDk

1

PN k qjk

jDk

Note that this is, in essence, the same approach that we used previously in Chapter 14 for the Geometrically Approximated Poisson Process.
However, we cannot parameterize it via the mean of the excess-rate batch
size instead we estimate the geometric parameter, q, from the ratio of
successive queue state probabilities:
qD

PN k C 1
D
PN k

kCN

CN1 dr kC1 1  dr N

kCN1 C
k
N
N1 dr  1  dr 

which, once the combinations have been expanded, reduces to

qD

k C N dr
kC1

For any practical arrangement in IP packet queueing, the buffer capacity

will be large compared to the number of output ports sharing; so
q dr

for k N

So, applying the geometric approximation, we have

QN k  1 PN k

1
1q

282

IP BUFFER MANAGEMENT
Total queue size
0

10

20

30

40

50

60

70

80

90

100

100
101

State probability

102
103
104
105
10

107
108

separate
conv. 2 buffers
neg. bin. 2 buffers
conv. 4 buffers
neg. bin. 4 buffers
conv. 8 buffers
neg. bin. 8 buffers

pk, dr :D 1  dr drk


Autoconv N , P :D  P if N < 2
 otherwise

 Q
p
 
  max k
last (P)
 
 
 i
1
 
  while i < N
  
   for m 2 0..maxk
  
m


  
   convm
Pj Qmj
  
  
jD0


   Q
conv
  
  i
i
C1
 
 Q
NegBinomial N, k, dr :D combin (k C N  1, N  1 drk 1  drN
k :D 0.. 100
dr :D 0.78997
Psinglek :D p k, dr
xk :D k
y1 :D Psingle
y2 :D Autoconv2, Psingle
y3k :D NegBinomial2, k, dr
y4 :D Autoconv4, Psingle
y5k :D NegBinomial4, k, dr
y6 :D Autoconv8, Psingle
y7k :D NegBinomial8, k, dr
Figure 16.8. State Probabilities for Shared Buffers, and Mathcad Code to Generate
(x, y) Values for Plotting the Graph

283

SHARED BUFFER ANALYSIS

Total buffer capacity

100

10

20

30

40

50

60

70

80

90

100

101

Pr{queue size > X}

102
103
104
105
106
107
108

separate
conv. 2 buffers
neg. bin. 2 buffers
conv. 4 buffers
neg. bin. 4 buffers
conv. 8 buffers
neg. bin. 8 buffers

Q(s) :D  qx0
1  s0

 for i 2 1.. last (s)

 qx
qxi1  si

i
 qx
NegBinomialQ N , k , dr :D combin (k C N  1, N  1 drk 1  drN1
k :D 0.. 100
xk :D k
y1 :D Q (Psingle)
y2 :D Q (Autoconv2, Psingle
y3k :D NegBinomialQ 2, k , dr
y4 :D Q (Autoconv 4, Psingle
y5k :D NegBinomialQ 4, k, dr
y6 :D Q (Autoconv 8, Psingle
y7k :D NegBinomialQ 8, k , dr
Figure 16.9. Overflow Probabilities for Shared Buffers, and Mathcad Code to
Generate (x, y) Values for Plotting the Graph

which, after substituting for PN k and q, gives

QN k  1 kCN1 CN1 dr k 1  dr N1
Applying Stirlings approximation, i.e.
N N eN
N! D p
2 N

284

IP BUFFER MANAGEMENT

Buffer capacity per port, X

0

10

100
10

Pr{queue size > X}

102
103

20

30

40

50

separate
simple, 2 buffers
neg. bin. 2 buffers
simple, 4 buffers
neg. bin. 4 buffers
simple, 8 buffers
neg. bin. 8 buffers

104
105
106
107
108

k :D 0 .. 50

NegBinomialQ (N, k, dr) :D  k
kN

 combin (k C N  1 , N  1 drk 1  drN1


SimpleNBQ N , k , dr :D  k
kN

ekCN1lnkCN1klnkN1lnN1
a
b
eklndrCN1ln1dr

a b
xk :D k
y1 :D Q (Psingle)
y2k :D SimpleNBQ 2, k , dr
y3k :D NegBinomialQ 2, k , dr
y4k :D SimpleNBQ 4, k , dr
y5k :D NegBinomialQ 4, k , dr
y6k :D SimpleNBQ 8, k , dr
y7k :D NegBinomialQ 8, k , dr
Figure 16.10. Comparison Showing the Benefits of Sharing Buffer Space Overflow
Probability vs. Buffer Capacity per Output Port

gives
QN k  1

k C N  1kCN1
dr k 1  dr N1
kk N  1N1

which has the distinct advantage of not requiring the user to evaluate
large factorials. Applying logarithms ensures that all the powers can be

SHARED BUFFER ANALYSIS

285

evaluated without number overflow.

QN k  1 efkCN1lnkCN1klnkN1lnN1Cklndr CN1ln1dr g
Thus we have a simplified negative binomial expression for the overflow probability in terms of the decay rates in (conceptually) separate
queues [16.2].
Lets now suppose we have a number of output ports sharing buffer
space, and each output port is loaded to 80% of its server capacity
with a bi-modal traffic aggregate (e.g. column 2 in Table 16.1 bi-modal
960). The decay rate, assuming no buffer sharing, is 0.789 97. Figure 16.8
compares the state probabilities based on exact convolution with those
based on the negative binomial expression, clearly showing the precision
of the latter approach. Figure 16.9 compares the overflow probabilities
based on exact convolution with those based on the negative binomial
expression (using n Cr ). There is an obvious discrepancy for very small
shared buffers, but above this the comparison is very close indeed.
If the loss probability requirement is 104 then having separate buffers
means that
38 500 D 19 000 octets
are required per output port, whereas if the 8 output ports share buffer
space then only
94 500/8 D 5875 octets
are required per output port (recall that the average packet size is 500
octets in this example).
Figure 16.10 shows this very clearly by plotting the overflow probability against the buffer capacity per output port. In this case we compare
the negative binomial expression (using n Cr ) with the simplified negative
binomial expression (using logarithms).

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

17

Self-similar Traffic
play it again, Sam

SELF-SIMILARITY AND LONG-RANGE-DEPENDENT TRAFFIC

The queueing models and solutions we have presented, developed
and applied in this book are very useful and have wide applicability.
However, one of the most significant recent findings for the design
and performance evaluation of networks has been the discovery of selfsimilarity and long-range dependence (LRD) in a variety of traffic types
[17.1]. Why is it significant? Well, the essence of self-similarity is that a
time-varying process behaves in a similar way over all time scales. The
observations made on a variety of traffic types in different network technologies show bursty behaviour over a wide range of time scales. And, as
we have seen in previous chapters, bursty behaviour has a much greater
impact on finite resources.
Lets take a memoryless process first, and see how that scales with time.
Figure 17.1 shows the results of simulating traffic for 10 000 seconds. The
first 100 seconds of the arrival process are shown as a thin grey line, and
here we see typical variable behaviour around a mean value of about
25 arrivals per second. The thick black line shows the process scaled by
100, i.e. the number of arrivals is averaged every 100 seconds and so
the 100 scaled time units cover the full 100 00 seconds of the simulation.
This averaging clearly shows a reduction in the variability of the process
when viewed on the longer time scale the mean value of 25 arrivals per
second is evident. Figure 17.2 takes a self-similar process and plots it in
the same way. In this case we can see the high variability of the process
even after scaling.
However, it is not self-similarity which is the underlying phenomenon,
but rather it is the presence of many basic communications processes
which have heavy-tailed sojourn-time distributions. In these distributions,
the tail probabilities decay as a power law, rather than exponentially.

288

SELF-SIMILAR TRAFFIC
50
100
1

45

Normalized arrival rate

40
35
30
25
20
15
10
5
0

10

20

30

40
50
60
Time (scaled units)

70

80

90

100

Figure 17.1. Scaling Behaviour of a Memoryless Process

50

100
1

45

Normalized arrival rate

40
35
30
25
20
15
10
5
0

10

20

30

40

50

60

70

80

Time (scaled units)

Figure 17.2. Scaling Behaviour of a Self-Similar Process

90

100

289

THE PARETO MODEL OF ACTIVITY

Another way of expressing this is that the process has long-term (slowly
decaying) correlations. So, an individual communications process with
heavy-tailed sojourn times exhibits long-range dependence. And the
aggregation of LRD sources produces a traffic stream with self-similar
characteristics.
So, how do we model and analyse the impact of this traffic? There
have been claims that traditional approaches to teletraffic modelling
no longer apply. Much research effort has been, and is being, spent on
developing new teletraffic models, such as Fractional Brownian Motion
(FBM) processes (e.g. [17.2]) and non-linear chaotic maps (e.g. [17.3]).
However, because of their mathematical complexity, assessing their
impact on network resources is not a simple task, although good progress
is being made.
In this chapter we take a different approach: with a little effort we can
re-apply what we already know about traffic engineering usefully, and
generate results for these new scenarios quickly. Indeed, this is in line
with our approach throughout this book.

THE PARETO MODEL OF ACTIVITY

A distribution is heavy-tailed if
PrfX > xg D 1  Fx

1
x

as x ! 1, and noting that > 0 (usually takes on values in the range

1 ! 2). The Pareto distribution is one of the class of distributions that are
heavy-tailed, and is defined as



PrfX > xg D

where is the parameter which specifies the minimum value that the
distribution can take, i.e. x
. For example, if D 25, then PrfX > 25g D
1, i.e. X cannot be less than or equal to 25. For our purposes it is often
convenient to set D 1.
The cumulative distribution function is



Fx D 1 
x
and the probability density function is given by
f x D

C1

290

SELF-SIMILAR TRAFFIC

The mean value of the Pareto distribution is

E[x] D

1

Note that for this formula to be correct, > 1 is essential; otherwise the
Pareto has an infinite mean.
Lets put some numbers in to get an idea of the effect of moving to
heavy-tailed distributions. Assume that we have a queue with a timeslotted arrival process of packets or cells. The load is 0.5, and we have a
batch arriving as a Bernoulli process, such that
Prfthere is a batch in a time slotg D 0.25
thus the mean number of arrivals in any batch is 2. We calculate the
probability of having more than x arrivals in any time slot, in two cases:
for an exponentially distributed batch size, and for a Pareto-distributed
batch size. In the former case, we have
x

Prfbatch size > xg D e 2

so
Prf> 10 arrivals in any time slotg D Prfbatch size > 10g
Prfthere is a batch in a time slotg
10

D e 2 0.25 D 0.001 684

In the latter case, we have (with D 1)
E[x] D 1
so
D

D2
1

E[x]
D2
E[x]  1

hence

2

Prfbatch size > xg D

1
x

giving

Prf>10 arrivals in any time slotg D

1
10

2

0.25 D 0.0025

Thus for a batch size of greater than 10 arrivals there is not that much
difference between the two distributions the probability is of the same

291

THE PARETO MODEL OF ACTIVITY

order of magnitude. However, if we try again for more than 100 arrivals
we obtain
Prf>100 arrivals in any time slotg D e

100
2

0.25 D 4.822 1023

in the exponential case, and

Prf>100 arrivals in any time slotg D

0
100
101

20

40

60

1
100

2

0.25 D 2.5 105

Batch size, x
80
100
120

140

160

180

200

Pareto, E(x) = 10
Pareto, E(x) = 2
Exponential, E(x) = 10
Exponential, E(x) = 2

102

Pr{X>x}

103
104
105
106
107
108

exponential , t :D
et


k
Pareto k , , x :D
x
i :D 1.. 1000
xi :D i
y1i :D exponential 0 . 5, xi 
y2i :D exponential 0 . 1, xi 



2
y3i :D Pareto 1 ,
, xi
21



10
y4i :D Pareto 1 ,
, xi
10  1
Figure 17.3. Comparison of Exponential and Pareto Distributions, and the Mathcad
Code to Generate x, y Values for Plotting the Graph

292

SELF-SIMILAR TRAFFIC

100
10

3 4 5 67 101

Batch size, x
2 3 4 5 67 102

3 4 5 67 103

101
102

Pareto, E(x) = 10
Pareto, E(x) = 2
Exponential, E(x) = 10
Exponential, E(x) = 2

Pr{X>x}

103
104
105
106
107
108

Figure 17.4.
Scale for x

Comparison of Exponential and Pareto Distributions, with Logarithmic

in the Pareto case. This is a significant difference, and clearly illustrates

the problems associated with highly variable traffic, i.e. non-negligible
probabilities for large batch sizes, or long sojourn times.
Figure 17.3 compares the exponential and Pareto distributions for two
different mean batch sizes, plotting x on a linear scale. For the exponential
distribution (which we have used extensively for sojourn times in statebased models) the logarithm of the probability falls away linearly with
increasing x. But for the Pareto the distribution bends back so that much
longer values have much more significant probability values than they
would otherwise. In fact we can see, in Figure 17.4, that when both axes
have a logarithmic scale, there is a straight-line relationship for the Pareto.
We can see from these figures that the Pareto distribution has
increasing, not constant, decay rate. This is very important for our analysis;
for example, as the ON period continues, the probability of the ON period
coming to an end diminishes. This is completely different from the exponential model, and the effect on buffer content is predictably dramatic.

IMPACT OF LRD TRAFFIC ON QUEUEING BEHAVIOUR

In previous queueing analysis we have been able to use memoryless
distributions such as the exponential or geometric, in the traffic models,
resulting in constant decay rates for the queueing behaviour. The effect
of using a Pareto distribution is that, as the buffer fill becomes very large,
the decay rate of the buffer-state probabilities tends to 1. This has an

293

IMPACT OF LRD TRAFFIC ON QUEUEING BEHAVIOUR

important practical outcome: above a certain level, there is no practical

value in adding more buffer space to that already available. This is clearly
both important and very different from those queueing systems we have
already studied. The queue with Pareto-distributed input is then one
of those examples (referred to previously in Chapter 14) which are not
covered by the rule of asymptotically constant decay rates except that
it will always eventually be the case that the decay rate tends to 1!

The Geo/Pareto/1 queue

In order to explore the effects of introducing heavy-tailed distributions
into the analysis, we can re-use the queueing analysis developed in
Chapter 7. Lets assume a queue model in which batches of packets
arrive at random, i.e. as a Bernoulli process, and the number of packets in
a batch is Pareto-distributed. The Bernoulli process has a basic time unit
(e.g. the time to serve an average-length packet), and a probability, q, that
a batch arrives during the time unit. This is illustrated in Figure 17.5.
In order to use the queueing analysis from Chapter 7, we need to
calculate the batch arrivals distribution. The probability that there are k
arrivals in any time unit is denoted ak. Thus we write
a0 D 1  q
a1 D q b1
a2 D q b2
..
.
ak D q bk
where bk is the probability that an arriving batch has k packets. Note
that this is a discrete distribution, whereas the Pareto, as defined earlier,
Geometrically distributed period of time between arriving batches

.
.
.

Pareto distributed
number of packets
in an arriving batch

.
.
.
Time

Packet departure process

Figure 17.5.

Model of Arriving Batches of Packets

294

SELF-SIMILAR TRAFFIC

is a continuous distribution. We use the cumulative form



Fx D 1 

1
x

to compute a discrete version of the Pareto distribution. In order to

calculate bk, we use the interval [k  0.5, k C 0.5] on the continuous
0

20

40

60

Batch size
80 100 120

140

160

180

200

100
= 1.9
= 1.1

101

Probability

102
103
104
105
106
107
108


BatchPareto q , k ,  :D  1  q if k D 0

 



1
 1
q if k D 1

 
1 . 5 
 




1
1

q


k  0.5
k C 0.5
maxX :D 1000
k :D 0 .. maxX
l :D 0
.. 1
1.9
:D
1.1
l
Bl :D
l  1
:D 0.25

ql :D
Bl
xk :D k
y1k :D BatchPareto q0 , k , 0 
y2k :D BatchPareto q1 , k , 1 
Figure 17.6.

otherwise

Discrete Version of Batch Pareto Input Distributions

295

IMPACT OF LRD TRAFFIC ON QUEUEING BEHAVIOUR

Queue size
100

3 4 5 6 7 101

3 4 5 6 7 102

2 3 4 5 6 7 103

100

= 1.9
= 1.1

State probability

101
102
103
104
105
106

maxX :D 1000
k :D 0.. maxX
1 :D 0..1



1.9
:D
1.1
1
B1 :D
1  1
:D 0.25

q1 :D
B1
aP1k :D Batchpareto q0 , k, 0 
ap2k :D Batchpareto q1 , k, 1 
xk :D k
y1 :D infiniteQmaxX, aP1, 
y2 :D infiniteQmaxX, aP2, 
Figure 17.7.

State Probability Distributions with Pareto Distributed Batch Input

distribution, i.e.

bx D Fx C 0.5  Fx  0.5 D

1
x  0.5

1
x C 0.5

Note that F1 D 0, i.e. the probability that an arriving batch is less than
or (exactly) equal to 1 packet is zero. Remember this is for a continuous
distribution; so, for the discrete case of a batch size of one packet,

296

SELF-SIMILAR TRAFFIC

we have

b1 D F1.5  F1 D 1 

1
1.5

So, bk is the conditional probability distribution for the number of

packets arriving in a time unit, i.e. given that there is a batch; and ak is the
unconditional probability distribution for the number of packets arriving
in a time unit i.e. whether there is an arriving batch or not.
Intuitively, we can see that the probability there are no arrivals at all
will probably be the biggest single value in the distribution most of
the time there will be zero arrivals, but when packets do arrive watch
out because there are likely to be a lot of them! Figure 17.6 shows some
example distributions for batch Pareto input, with D 1.1 and 1.9. The
figure is plotted on a linear axis for the batch size, so that we can see the
probability of no arrivals. Note that the mean batch sizes are 11 and 2.111
packets respectively. The mean number of packets per time unit is set to
0.25; thus the probability of there being a batch is
qD

0.25
D
B
B

giving q D 0.023 and 0.118 respectively.

100
101

100

3 4 5 6 7 101

Size
3 4 5 6 7 102

3 4 5 6 7 103
= 1.9
= 1.1

Probability

102
103
104
105
106
107
108

Figure 17.8. Comparison of Power-Law Decays for Arrival (Thin) and Queue-State
(Thick) Probability Distributions

297

IMPACT OF LRD TRAFFIC ON QUEUEING BEHAVIOUR

100

3 4 5 6 7 101

Size
2 3 4 5 6 7 102

3 4 5 6 7 103

100
101

= 1.9
= 1.1

Probability

102
103
104
105
106
107
108

BatchparetoTrunc q , k , , X :D  1  qif k D 0
 


 
1

 1

1.5

 q


if k D 1

1
 1

XC0.5






 

1
1




k0.5
kC0.5
q




1

1

X C 0.5


0 if k > X
Xlimit :D 500
aP1k :D BatchParetoTrunc q0 , k, 0 , Xlimit
aP2k :D BatchParetoTrunc q1 , k, 1 , Xlimit

alt1 :D
k.aP1k

if

k  X k > 1

alt1 D 0.242

alt2 :D
k.aP2k
k

alt2 D 0.115
y1 :D infiniteQ maxX, aP1, alt1
y2 :D infiniteQ maxX, aP2, alt1

Figure 17.9. Effect of Truncated Power-Law Decays for Arrival (Thin) and Queue-State (Thick)
Probability Distributions

298

SELF-SIMILAR TRAFFIC

Now that we have prepared the arrival distribution, we can put this
directly into the queueing analysis from Chapter 7. Figure 17.7 shows
the resulting queue state probabilities for both D 1.1 and 1.9. Note that
the queue-state probabilities have power-law decay similar to, but not
the same as, the arrival distributions. This is illustrated in Figure 17.8,
which shows the arrival probabilities as thin lines and the queue-state
probabilities as thick lines.
From these results it appears that the advantage of having a large buffer
is somewhat diminished by having to cope with LRD traffic: no buffer
would seem to be large enough! However, in practice there is an upper
limit to the time scales of correlated traffic activity. We can model this by
truncating the Pareto distribution, and simply using the same approach
to the queueing analysis.
Suppose X is the maximum number of packets in a batch. Our truncated, discrete version of the Pareto distribution now looks like




 



 

1
1

1

1

xD1

0.5
X C 0.5



  



 
1
1
1
bx D


1

X
x>1

x  0.5
x C 0.5
X C 0.5

0
x>X
Note that, because of the truncation, the probability density needs to be
conditioned on what remains, i.e.

1

1
X C 0.5

Figure 17.9 shows the result of applying this arrival distribution to the
queueing analysis from Chapter 7. In this case we have the same values
as before for , i.e. D 1.1 and 1.9, and we set X D 500. The load is
reduced because of the truncation to 0.115 and 0.242 respectively. The
figure shows both the truncated arrival distributions and the resulting
queue-state distributions. For the latter, it is clear that the power-law
decay begins to change, even before the truncation limit, towards an
exponential decay.
So, we can see that it is important to know the actual limit of the
ON period activity in the presence of LRD traffic, because it has such a
significant effect on the buffer size needed.

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

References
[1.1] Griffiths, J.M. (ed.), ISDN Explained: Worldwide Network and Applications Technology, John Wiley & Sons, ISBN 0 471 93480 1 (1992)
[1.2] Cuthbert, L.G. and Sapanel, J-C., ATM: the Broadband Telecommunications Solution, The Institution of Electrical Engineers, ISBN 0 85296 815 9 (1993)
[1.3] Thomas, S.A., IPng and the TCP/IP Protocols: Implementing the Next Generation
Internet, John Wiley & Sons, ISBN 0 471 13088 5 (1996)
[3.1] Flood, J.E., Telecommunications Switching, Traffic and Networks, Prentice Hall,
ISBN 0 130 33309 3 (1995)
[3.2] Bear, D., Principles of Telecommunication Traffic Engineering, Peter Peregrinus
(IEE), ISBN 0 86341 108 8 (1988)
[5.1] Law, A.M. and Kelton, W.D., Simulation Modelling and Analysis, McGraw-Hill,
ISBN 0 07 100803 9 (1991)
[5.2] Pitts, J.M., Cell-rate modelling for accelerated simulation of ATM at the burst
level, IEE Proceedings Communications, 142, 6, December 1995
[6.1] Cosmas, J.P., Petit, G., Lehnert, R., Blondia, C., Kontovassilis, K., and
Cassals, O., A review of voice, data and video traffic models for ATM,
European Transactions on Telecommunications, 5, 2, March 1994
[7.1] Pattavina, A., Switching Theory: Architecture and Performance in Broadband ATM
Networks, John Wiley & Sons ISBN 0 471 96338 0 (1998)
[8.1] Roberts, J.W. and Virtamo, J.T., The superposition of periodic cell arrival
processes in an ATM multiplexer, IEEE Trans. Commun., 39, 2, pp. 298303
[8.2] Norros, I., Roberts, J.W., Simonian, A. and Virtamo, J.T., The superposition of
variable bit rate sources in an ATM multiplexer, IEEE JSAC, 9, 3, April 1991,
pp. 378387
[8.3] Schormans, J.A., Pitts, J.M., Clements, B.R. and Scharf, E.M., Approximation
to M/D/1 for ATM CAC, buffer dimensioning and cell loss performance,
Electronics Letters, 32, 3, 1996, pp. 164165
[9.1] Onvural, R., Asynchronous Transfer Mode Networks: Performance Issues, Artech
House (1995)
[9.2] Schormans, J.A., Pitts, J.M. and Cuthbert, L.G., Exact fluid-flow analysis of
single on/off source feeding an ATM buffer, Electronics Letters, 30, 14, 7 July
1994, pp. 11161117
[9.3] Lindberger, K., Analytical methods for the traffical problems with statistical multiplexing in ATM networks, 13th International Teletraffic Congress,
Copenhagen 1991, 14: Teletraffic and datatraffic in a period of change
[10.1] ITU Recommendation I.371 TRAFFIC CONTROL AND CONGESTION CONTROL IN
B-ISDN, August 1996

300

REFERENCES

[10.2] ATM Forum AF-TM-0121.000 TRAFFIC MANAGEMENT SPECIFICATION, Version

4.1, March 1999
[10.3] ITU Recommendation E.736 METHODS FOR CELL LEVEL TRAFFIC CONTROL IN
B-ISDN, May 1997
[11.1] Rathgeb, E.P., Modeling and performance comparison of policing mechanisms for ATM networks, IEEE JSAC, 9, 3, April 1991, pp. 325334
[13.1] Kroner,
H., Hebuterne, G., Boyer, P. and Gravey, A., Priority management in

ATM switching nodes, IEEE JSAC, 9, 3, April 1991, pp. 418428

[13.2] Schormans, J.A., Scharf, E.M. and Pitts, J.M., Waiting time probabilities in a
statistical multiplexer with priorities, IEE Proceedings I, 140, 4, August 1993,
pp. 301307
[15.1] Braden, R., Clark, D. and Shenker, S. Integrated services in the internet architecture: an overview, RFC 1633, IETF, June 1994
[15.2] Nichols, K., Blake, S., Baker, F. and Black, D., Definition of the Differentiated
Services Field (DS Field) in the IPv4 and IPv6 Headers, RFC 2474, IETF,
December 1998
[15.3] Braden, R., Zhang, L., Berson, S., Herzog, S. and Jamin, S., Resource ReSerVation Protocol (RSVP) Version 1 Functional Specification, RFC 2205, IETF,
September 1997
[15.4] Schormans, J.A., Pitts, J.M., Scharf, E.M., Pearmain, A.J. and Phillips, C.I.,
Buffer overflow probability for multiplexed on-off VoIP sources, Electronics
Letters, 16 March 2000, 36, 6
[16.1] Floyd, S. and Jacobson, V., Random early detection gateways for congestion avoidance, IEEE/ACM Transactions on Networking, 1, 4, August 1993,
pp. 397413.
[16.2] Schormans, J.A., and Pitts, J.M., Overflow probability in shared cell switched
buffers, IEEE Communications Letters, May 2000
[17.1] Leland, W.E., Taqqu, M.S., Willinger, W. and Wilson, D., On the self-similar
nature of Ethernet traffic (extended version), IEEE/ACM Transactions on
Networking, 2, 1, February 1994, pp. 115
[17.2] Norros, I., On the use of fractional Brownian motion in the theory of connectionless networks, IEEE JSAC, 13, 6, August 1995, pp. 953962
[17.3] Mondragon R.J., Pitts J.M. and Arrowsmith D.K., Chaotic intermittencysawtooth map model of aggregate self-similar traffic streams, Electronics Letters, 36, 2, January 2000, pp. 184186

Introduction to IP and ATM Design Performance: With Applications Analysis Software,

Second Edition. J M Pitts, J A Schormans
Copyright 2000 John Wiley & Sons Ltd
ISBNs: 0-471-49187-X (Hardback); 0-470-84166-4 (Electronic)

Index
Asynchronous Transfer Mode (ATM)
standards 58
technology of 3, 8
time slotted nature 8
Switches 11
Bernoulli process
with batch arrivals 19, 87, 98
binomial distribution 16, 89
distribution of cell arrivals 16
formula for 16
blocking
connection blocking 47
buffers
balance equations for buffering 99
delays in 110
at output 99
finite 104
sharing and partitioning 273, 279
see also priorities
virtual buffers in IP 273
burst scale queueing 125
by simulation 77
in combination with cell scale 187
using large buffers 198
call
arrival rate 16, 47
average holding time 47
capacity 49
cell 7
loss priority bit 205
loss probability 104, 236, 245
see also queue
switching 97
cell scale queueing 113, 121

by simulation 73
in combination with burst scale 187
channel 5
circuits 5
circuit switching 3
connection admission control (CAC) 149
a practical scheme 159
CAC in the ITU standards 165
equivalent cell rate and linear CAC 160
using 2 level CAC 160
via burst scale analysis 39, 157, 161
via cell scale analysis 37, 152, 153
using M/D/1 152
using N.D/D/1 153
connectionless service 10
constant bitrate sources (CBR) 113, 125, 150
multiplex of 113
cross connect 9
deterministic bitrate transfer capability (DBR)
150
DIFFSERV 12, 253
Differentiated performance 32
Decay rate analysis 234
as an approximation for buffer overflow
236
used in Dimensioning 42, 187
of buffers 192
effective bandwidth see CAC, equivalent cell
rate
equivalent bandwidth see CAC, equivalent
cell rate
Erlang
Erlangs lost call formula 42, 52
traffic table 54

302
Excess Rate analysis (see also GAPP) 22, 27
for multiplex of ON/OFF sources 27
for VoIP 261
for RED 271
exponential distribution
formula for 16
of inter-arrival times 16
Geometrically Approximated Poisson Process
(GAPP) 22, 23, 240
GAPP approximation for M/D/1 queue
239
GAPP approximation for RED in IP 271
GAPP approximation for buffer sharing
279
geometric distribution 86
formula for 86
of inter-arrival times 86
of number of cell arrivals / empty slots 87
with batch arrivals 87
Internet Protocol (IP) 10, 58
IP queueing see Packet Queueing
IP source models, see sources models for IP
IP best effort traffic, ER queueing analysis
245
IP packet flow aggregation 254, 255
IP buffer management 267
IP virtual buffers 272
IP buffer partitioning 275
IP buffer sharing 279
INTSERV 12, 253
label multiplexing see packet switching
load 47
Long Range Dependency in Traffic 287
in queue model 293
mesh networks 45
MPLS 12
multiplexors
octets 5
packet queueing 229
for variable length packets 247
packet switching 5, 229
Pareto distribution 17, 289
in queueing model 30, 293
performance evaluation 57

INDEX

by analysis 58
by measurement 57
by simulation 57
Poisson 16
distribution of number of cell arrivals per
slot 16, 86
distribution of traffic 51, 86
formula for distribution 16
position multiplexing 4
priorities 32, 205
space priority and selective discarding
205
partial buffer sharing 207
via M/D/1 analysis 207
push out mechanism 206
precedence queueing in IP 273
time priority 35, 218
distribution of waiting times 2/22, 220
via mean value analysis 219
via M/D/1 analysis 220
QoS 253
queue
cell delay variation 68
cell delay in ATM switch 68, 108
cell loss probability in 104
via simulation 73
cell waiting time probability 108, 220
customers of 58
deterministic see M/D/1
discard mechanisms 32
fluid-flow queueing model 129
continuous analysis 129
discrete analysis 131
M/D/1 21, 22, 66, 117
delay in 108
heavy traffic approximation 117
mean delay in 66
M/M/1 19, 62
mean delay in 61
mean number in 62
system size distribution 63
summary of formulae 18
N.D/D/1 21, 115
heavy traffic approximation 117
queueing
burst scale queueing behaviour 127
queue state probabilities 98
queueing system, instability of 100
steady state probabilities, meaning of 99

303

INDEX

per VC queueing 11, 129

theory, summary 18, 58
Kendalls notation for 60
rate envelope multiplexing 191
rate sharing statistical multiplexing 191
Random Early Discard (RED) 13, 35, 267
Resource Reservation 253
Routers 9, 229
Self-similar traffic 287
simulation 59
accelerated simulation 77
cell rate simulation 77
speedup obtained 80
confidence intervals 76
discrete time simulation 59
discrete event simulation 59
generation of random numbers 71
via the Wichmann-Hill algorithm 72
method of batch means 75
of the M/D/1 queue 73
steady state simulations 74
validation of simulation model 77
sources (also called Traffic Models) 81
source models 16, 81
as time between arrivals 83
as rates of flow 89
by counting arrivals 86
generally modulated deterministic
processes 93
for IP 81
memoryless property of inter-arrival times
86
multiplexed ON/OFF sources 139, 254
bufferless analysis of 141
burst scale delay model 145
ON/OFF source 90
star networks 45
switches (also called Cell Switches) 7, 97

output buffered 97
statistical bitrate transfer capability (SBR)
10/2
sustainable cell rate 150
Synchronous Digital Hierarchy (SDH) 5
teletraffic engineering 45
time division multiplexing 3
timeslot 3
traffic
busy hour 51
carried 47
conditioning of aggregate flows in IP
intensity 47, 50
levels of behaviour 81
lost 47
models, see sources
offered 47
traffic shaping 182
traffic contract 150
usage parameter control (UPC) 13, 167
by controlling mean cell rate 168
by controlling the peak cell rate 173
by dual leaky buckets (leaky cup and
saucer) 40, 182
by leaky bucket 172
by window method 172
Tolerance problem 176
worst case cell streams 178
variable bitrate services (VBR) 150
virtual channels 9, 205
virtual channel connection 9, 205
virtual paths 9, 205
Voice over IP (VoIP) 239
basic queueing model 239
advanced queueing model 261
Weighted Fair Queueing (WFQ)

274

265