Multiaccess, Mobility and Teletraffic For Wireless Communications

MULTIACCESS, MOBILITY
AND TELETRAFFIC FOR

WIRELESS COMMUNICATIONS,
VOLUME 6
Multiaccess, Mobility
and Teletraffic for
Wireless Communications,
volume 6
edited by
Xavier Lagrange
ENST Bretagne, Site de Rennes
and
Bijan Jabbari
George Mason University
SPRINGER SCIENCE+BUSINESS MEDIA, B.V.

A c.I.P. Catalogue record for this book is available from the Library of Congress.
ISBN 978-1-4419-5290-5 ISBN 978-1-4757-5918-1 (eBook)

DOI 10.1007/978-1-4757-5918-1
Printed on acid-free paper
All Rights Reserved

© 2002 Springer Science+Business Media Dordrecht
Originally published by Kluwer Academic Publishers in 2002
No part of the material protected by this copyright notice may be reproduced or
utilized in any form or by any means, electronic or mechanical,
including photocopying, recording or by any information storage and
retrieval system, without written permission from the copyright owner.
Table of Contents
Preface ix
A New Approach for Partitioning the Received SNR Space for Tractable 3
Performance Analysis in Wireless Packet Networks
M. Hassan, M. Krunz, W. Ryan
Capacity Analysis of Voice over IP over GERAN with Statistical 25

Multiplexing
A. Wautier, J. Antoine, L. Husson, J. Brouet, C. Thirouard
Uplink RRM for Conversational and Interactive Services in UTRA-FDD 43

O. Salient, J. Perez-Romero, R. Agusti, J. Sanchez
Rate and Power Adaptation for Downlink Shared Channel in WCDMA 61

S. Akhtar, D. Zeghlache
Capacity and CII Performance of Different Cell Clusters in a Cellular 75

Network
A. Masmoudi, S. Tabbane
Performance Study of Soft Handover with CDMA Heterogeneous Cellular 87

Architectures
L.C. Wang, C.Y. Liao, C.J. Chang
Packet Service in UMTS: Effects of the Radio Interface Parameters on the 103
Performance of the Downlink Shared Channel
F. Borgonovo, A. Capone, M. Cesana, L. Fratta
Cellular Multihop Networks and the Impact of Routing on the SNIR and 115
Total Power Consumption
K. M. Pepe, B. R. Vojcic
Terminal Migration Model in which Cell Dwell Time is Defined by 133

Different Probability Distributions in Different Cells
H. Hidaka, K. Saitoh, N. Shinagawa, T. Kobayashi
Concatenated Location Management 143

K. Sasada, S. Hiyama, M. Yabusaki
v
Vl
Handoff Scheme Improvement in Wireless Networks 155

A. Hac, Y. Zhang
Hierarchical Mobility Controlled by the Network 167

Y. Khouaja, K. Guillouard, J.M. Bonnin, Ph. Bertin
Approximate and Exact ML Detectors for CDMA and MIMO Systems: a 183
Tree Detection Approach
S. Vaton, T. Chonavel, S. Saoudi,
Block Turbo Code with Binary Input for Improving Quality of Service 195
P. Adde, R. Pyndiah, S. Kerouedan
Gallager Codes for Asynchronous Multiple Access 205

A. de Baynast, D. Declercq
Bounding Techniques for the Design of Channel Coding and Modulation 221
Systems
Y. W. Blankenship, B. K. Classon
Quality of Service of Internet Applications over the UMTS Radio Interface 239
S. Heier, A. Kemper, S. Grabner, lO. Rock
Interactions between the TCP and RLC Protocols in UMTS 251

R. Bestak, Ph. Godlewski, Ph. Martins
Impact of SR-ARQ with Finite Buffer on TDD/TDMA Wireless LAN 263

E. Strinati, J. Gosteau, S. Simeons, P. Pellati
Traffic Performance Analysis of Multimedia Applications in Evolved GSM 279

Networks
P. Stuckmann, C. Hoymann
VoicelVideo over IP with Multi-class QoS in 3G Mobile Networks 295

L. Zhang, H. Fang
Establishment of Mobile Extranets through Mobile Ipv6 and GPRS: 309

Enabling Universal Access to Corporate Intranets
K. Koutsopoulos, N. Alexiou, C. Konstantinopoulou, P. Demestichas, M.
Theologou
vii
IP Traffic Control on UMTS Terminal Equipment 323

M. Ricardo, R. Soares, J. Dias, J. Ruela
An Approach for Managing Networks and Services in a Diversified Radio 337

Environment
P. Demestichas, G. Vivier, G. Martinez, F. Galliano, L. Papadopoulou, V.
Stavroulaki, M. Theologou
Preface
Third generation networks have been specified and are now being deployed
in a few countries. They are expected to reach maturity in the next several
years and to provide various services including audio, video, and world wide
web browsing. Furthermore, radio terminals are expected to be integrated in
a number of devices such as personal computers, personal digital assistants,
and even television sets. Such a wide-usage of radio mandates ongoing
research to address design of networks with high capacity while providing
acceptable quality of service.
This volume is the sixth in the edited series Multiaccess, Mobility and
Teletraffic for Wireless Communications. It presents the selected papers for
the proceedings of the Seventh Workshop (MMT'2002) held on this topic in
June 2002 in Rennes, France. The aim of this workshop has been to address
a set of important issues of interest to the wireless communications
community. In particular, the focus of this workshop is to identify, present
and discuss the theoretical and implementation issues critical to the design of
land based mobile cellular and microcellular as well as wireless local area
networks. Included in this book are recent research results on performance
analysis of wireless packet networks, channel coding and receiver design,
radio resource management in third generation systems, mobility
management in cellular and mobile IP networks, performance of transport
protocols (TCP) over radio link control protocols, and ad-hoc networks.
We express our gratitude to the reviewers: Bardia Alavi, Thorsten Benkner,

Fredrik Berggren, Philippe Bertin, Saad Biaz, Ezio Biglieri, Jean-Marie
Bonnin, Jerome Brouet, Prosper Chemouil, Jyh-Cheng Chen, Anthony
Ephremides, David Everitt, Romano Fantacci, Bernard Fino, Luigi Fratta,
Woldemar Fuhrmann, Aura Ganz, Paolo Grazioso, Philippe Jacquet, Markku
Juntti, Kin K. Leung, Seong-Lyun Kim, Takehiko Kobayashi, Zhuyu Lei,
Philippe Martins, Isabelle Moreau, Loutfi Nuaymi, Jae Yoon Park, Ramjee
Prasad, Guy Pujolle, Ramesh Pyndiah, David Ros, Gerardo Rubino, Steven
S. Pietrobon, Stephen S. Rappaport, Zafer Sahinoglu, Gordon StUber, Peter
Stuckmann, Sami Tabbane, Bruno Tuffin, Guillaume Vivier, Branimir
Vojcic, Bernhard Walke, Djamal Zeghlache, and Michele Zorzi.
Weare also grateful to the authors who helped make the development of this
volume possible.
Xavier Lagrange
Bijan Jabbari
ix
MMT
Vol 6
X. Lagrange
B. Jabbari
A NEW APPROACH FOR PARTITIONING
THE RECEIVED SNR SPACE FOR TRACTABLE
PERFORMANCE ANALYSIS IN WIRELESS
PACKET NETWORKS
Mohamed Hassan, Marwan Krunz, and William Ryan

Department of Electrical & Computer Engineering
University of Arizona
Tucson, AZ 85721
[mhassan, krunz, ryan]@ece.arizona.edu
Abstract Successful provisioning of multimedia services over wireless networks hinges

on the ability to guarantee certain levels of quality of service (QoS). Prior
assessment of the QoS performance requires employing realistic channel
models that not only reflect the physical characteristics of the channel,
but that also facilitate analytical investigation of its performance. Finite-
state Markov chain (FSMC) models have often been used to characterize
the wireless channel, whereby the range of the signal-to-noise ratio (SNR)
is partitioned according to some criteria into a set of intervals (states).
Different partitioning criteria have been used in the literature, but none of
them was targeted to facilitating the performance analysis of the packet
delay and loss performance over the wireless link. In this paper, we pro-
pose a new method for partitioning the received SNR space that results
in a FSMC model with tractable queueing performance. We make use
of the level-crossing analysis, the distribution of the received SNR, and
the producer-consumer queueing model of Mitra [14] to arrive at the pro-
posed FSMC model. An algorithm is provided for computing the various
parameters of the model, including the number of states, the partition-
ing thresholds, and the "nominal" bit error rates. The usefulness of the
obtained model is then highlighted by deriving a closed-form expression
for the effective bandwidth (EB) subject to a packet loss constraint. Nu-
merical examples are presented to study the interactions between various
key parameters and the adequacy of the proposed model.
1. Introduction
Wireless networks are characterized by time-varying channels in which
the bit error rate (BER) varies dramatically according to the received
3
X. Lagrance and B. labbari (eds.),
Multiaccess, Mobility and Teletraffic for Wireless Communications, Volume 6, 3-24.
© 2002 Kluwer Academic Publishers.
4
signal-to-noise ratio (SNR). Due to their mathematical tractability, finite-

state Markov chain (FSMC) models have often been used to characterize
the BER behavior [1, 2, 3J. Typically, an FSMC model is constructed by
partitioning the range of the received SNR into a set of non-overlapping
intervals [0,1"1), h, 1"2)"" [1"i' 1"i+l) , ... [1"N, 00). Each interval is repre-
sented by a nominal BER, which in turn represents a certain channel
quality (see Figure 1, where each interval represents one state of the FSMC
model). In our study, the selection of the thresholds 1"i, i = 1, ... ,N, and
their corresponding nominal BERs is done in such a manner that leads
to tractable analysis of the packet loss and delay performance over the
wireless link.
•
Rt!alv~dSNR
,\'f(JteN
.l'll1teN-1
.fIate J
.flareO
~.
time
Figure 1. Partitioning of the received SNR space.
Several researchers have addressed the modeling of wireless links through

partitioning of the received SNR [6, 7, 8, 9, 10J. In [6J the wireless chan-
nel was characterized by a Markov model that was developed based on
experimental measurements of real channels. To maximize the through-
put while minimizing the probability of packet error, Rice and Wicker [7J
represented the time-varying channel by a FSMC model in which the
thresholds 1"i, i = 0, 1, ... N, were obtained by counting the number of re-
transmission requests during the so-called observation or frame intervals.
In [8J a FSMC model was constructed by optimizing the system perfor-
mance in the sense of maximizing the channel capacity. By requiring all
states to have the same mean durations, Zhang and Kassam [9J employed
a FSMC model for the Rayleigh fading channel. Using the Nakagami-m
distributions as the basis for partitioning, the authors in [10J characterized
the dynamics of amplitude variations of time-varying multi-path fading.
They also built a FSMC model whose states represent the different inter-
vals of fading amplitude.
None of the above models was designed to enable tractable analysis
of packet-level performance degradations. More specifically, the use of
5
packet buffering at the transmitter side of a wireless link introduces vari-

able queueing delays and occasional packet loss (due to buffer overflow).
Prior assessment of such degradations is a key to providing QoS guaran-
tees and to the design of online admission control policies. Hence, a good
model should not only reflect the physical characteristics of the channel,
but it should also facilitate analytical investigation of its performance.
In [11] the authors presented a framework for analyzing the packet loss
performance over a wireless link. They used a two-state FSMC model to
characterize the channel and to evaluate the packet loss performance. Ar-
guably, a two-state model provides a coarse approximation of the channel
behavior, and may not always be acceptable. Therefore, the motivation
behind our study is to come up with a multi-state FSMC model that
overcomes the inadequacy of the 2-state Markov model. Our objective
is to design a partitioning methodology of the received SNR space that
will result in a "tractable" FSMC model in the sense di:,;cu:,;:,;ed later. The
resulting channel model not only simplifies the performance analysis of
wireless networks, but it abo accurately models the channel physical char-
acteristics. In this paper, we focus on the partitioning approach. Then
we give an example that shows the efficiency of the proposed model in
obtaining a less conservative estimate of the EB while guaranteeing the
required QoS. In an extended study [12], we investigated the concept of
wireless effective bandwidth, as a metric to meet applications QoS re-
quirements in terms of both packet losses and delays, using the obtained
FSMC channel model.
The paper is organized as follows. Section 2 presents our analytical
framework. In Section 3 we present the adopted SNR partitioning ap-
proach. Section 4 illustrates the applicability of our model, provides the
adopted queueing model and the corresponding packet-loss based wireless
EB analysis. Section 5 reports the numerical results. Finally, Section 6
summarizes the results of this study and outlines our future work.
2. Wireless Link Model

2.1 Preliminaries
Consider Figure 2, which represents a simplified representation of a
wireless link. Arriving packets at the transmitter are stored temporarily
in a FIFO buffer, which is drained at a rate that depends on the state
of the channel at the receiver. Throughout this work, we refer to the
draining (or service) rate when the received SNR is r by c(r). After de-
parting the buffer, a packet undergoes a strong CRC encoding followed
by partial FEC that allows for correcting only a fraction of packet errors.
This hybrid ARQ approach is often used to enhance the efficiency of the
6
wireless link by minimizing the number of retransmissions. In practice,

the transmission path also includes packet and bit interleavers, possibly in
conjunction with multiple FEC encoders (e.g., outer and inner encoders).
For simplicity, we do not directly account for the impact of interleaving,
but assume that such impact has already been incorporated in the FEC
"box" in Figure 2. In other words, this box could consist of multiple stages
of encoding and interleaving. In a wireless packet network, the traffic ar-
onloffsource
[]JJ]]rmm
Incoming trafic
Figure 2. Wireless-link model.
riving at a node can be viewed as an alternating sequence of active and

idle periods. During an active period, one or more network-layer packets
(e.g., IP datagrams) arrive back-to-back, forming a burst (an active pe-
riod). Quite often, a packet is first fragmented into fixed-size link-layer
(LL) packets before being transmitted over the wireless interface. The
sheer difference between the burst and LL-packet time scales makes it
reasonable to separate the two by adopting a fluid approximation of the
arrival traffic. Such a decomposition approach, which has been success-
fully applied in wireline networks [14, 15, 16, 17], allows us to emphasize
the time scale of most relevance to network-layer performance (e.g., packet
loss rate and queueing delay).
Accordingly, an incoming traffic flow is modeled as a fluid source with
exponentially distributed on and off periods. The means of the on and
off periods are 1/0'. and 1//3, respectively. When the source is active, it
transmits at a peak rate (J. The channel is modeled by the (N + I)-state
FSMC model shown in Figure 3. This particular structure of the Markov
chain is chosen because it lends itself to the queueing analysis of Mitra's
producer-consumer model [14] (to be described later), which allows us to
evaluate the packet loss and delay performance over the wireless channel
and compute closed-form expressions for the Effective Bandiwth. Let 7ri be
the steady-state probability that the channel is in state i, i = 0, ... , N.
The FSMC stays in state i for an exponentially distributed time with
mean Ti . It is assumed that bit errors within any given state are mutually
independent. For an FEC code with a correction capability of T bits per
7
eN-I)}. II
N-l
2u rN-llu No
Figure 3. Markov chain in Mitra's producer-consumer model.
code block (packet), the probability of an uncorrectable error in a received

packet when the channel is in state i is given by:
PCi = t
j=r+l
(r:) (Pe(i\))j (1 -
J
Pe(ri))n- j (1)
where n is the number of bits in a code block, including the FEC bits,
Pe(r) is the BER when the instantaneous SNR is r (the form of Pe(r)
depends in the underlying modulation scheme), and f\ is the "nominal"
SNR value in state i (its calculation will be discussed later). The packet
transmission/retransmission process can be approximated by a Bernoulli
process [1]. We assume that the transmitter always gets the feedback
message from the receiver before the next transmission slot, and a packet
is retransmitted persistently until it is successfully received. The nominal
service rate in state i, denoted by Ci, can be approximated by the inverse
of the mean of the geometrically distributed retransmission process:
Ci = c.e.(l - Pc;}, i = 0,1,2, ... , N (2)
where C is the error-free service rate, e = kin is the FEC overhead, and
k is the number of information bits in a code block.
Wireless transmission of continuous waveforms in obstacle environ-
ments is prone to multi-path fading, which results in randomly varying
envelope for the received signal. This randomness has been shown to fol-
Iowa Rayleigh distribution. In the presence of additive Gaussian noise,
the instantaneous received SNR r is proportional to the square of the sig-
nal envelope [5]. Accordingly, the SNR is exponentially distributed with
pdf:
1 -r
PR(r)=-e p , r>O (3)
p
where p ~ E[r]. An important parameter that reflects the behavior of

the SNR process at the receiver is the level-crossing rate (LCR), defined
as the average rate at which the signal envelope crosses a given SNR level
r. For the Rayleigh fading channel, the LCR at an instantaneous SNR r
is given by [5]:
(4)
8
where f m = ~ is the Doppler frequency, w is the speed of the electro-

magnetic wave, v is the speed of the mobile user, and fo is the carrier
frequency. We assume slow fading with respect to symbol transmission
time. Furthermore, we assume that transitions between channel states
take place only at the end of a packet transmission. As mentioned before,
the FSMC model that represent the time-varying behavior of the Rayleigh
fading channel will be obtained by partitioning the received SNR into
N + 1 intervals (states). Let TO = 0 < TI < T2 < ... < TN-I < TN <
TN+1 = 00 be the N + 2 thresholds that define the partitioning. The
steady-state probability that the FSMC is in state i is given by:
7ri = l ri
ri +1 1 =
-e p dT
P
=e
::!:i
P - e
-ritl
P (5)
2.2 Mitra's Producers-Consumers Fluid Model

In this section, we briefly describe Mitra's producers-consumers fluid
queueing model [14J. This model facilitates the analytical investigation
of communication systems possessing randomly varying statistical prop-
erties. According to this model, the fluid produced by m producers is
supplied to a FIFO buffer that is drained by n consumers. Note that
m = 1 in our wireless model. Each producer and consumer alternates
between independent and exponentially distributed active and idle peri-
ods. Let A-I and f-L- I denote the mean of the idle and active periods of
a consumer, respectively. When active, a consumer drains fluid from the
buffer at a constant rate, which is the same for all consumers. It is easy
to see that the number of active consumers fluctuates in time according
to the Markov chain in Figure 3. In [14J Mitra analyzed the steady-state
behavior of this queueing system. We will use his analysis as the basis to
derive the wireless Effective Bandiwth. But first we need to partition the
wireless channel in a manner that produces the same Markovian structure
in Figure 3. In other words, we match the service rate at the transmitter
buffer to the total instantaneous consumption rate in Mitra's model. This
requires that we choose the partitioning thresholds such that each state
corresponds to a given number of active consumers. Let O(T) ~ c'(<;J;) be
the ratio between the service rate at a received SNR T and its value when
the channel is error-free (note that c(oo) < c due to the FEC overhead).
We form a FSMC model based on the requirement that in each SNR in-
terval h, Ti+I), there exists a point ri, Ti :S ri < Ti+1' that satisfies the
following relationship:
O(T")
, ~ cC((ooTi)) = Ni , i = 0,1, ... , N - 1 (6)
9
Note that in the producer-consumer model, the consumption rate in state

i, i = 0,1, ... ,N, is i times the consumption rate in state 1. The BER
that corresponds to f'i is called the nominal BER associated with state i.
We set e(ro) = e(TO) = 0 and e(rN) = e(oo) = 1 (Le., rN = TN+! = (0).
3. SNR Partitioning
It is easy to see that in the producer-consumer model, the steady-state
probability distribution is binomial, i.e.,
(7)
For our wireless channel to fit the Markov chain in [14], the partitioning
must be done so that no more than one state falls in the "good" (BER
close to zero) and "bad" (BER close to one) regions of the SNR space,
since either scenario will lead to an unrealistic number of states.
To completely specify the underlying Markov chain, we need to deter-
mine N, ~, and the thresholds Tl, T2, ... , TN. Our procedure for comput-
ing these quantities is outlined as follows. First, by equating (5) and (7),
we get an expression for N in terms of Tl and ~. Using the level-crossing
analysis and the structure of the embedded Markov chain, we then ob-
tain an expression for ~ in terms of Tl and TN. Combining the obtained
expressions, we can eipress N in terms of Tl and TN only. Then, by
selecting an appropriate value for TN, the value of the threshold Tl can
be obtained. After obtaining Tl, the other thresholds can be obtained
recursively, as follows: The steady-state probability that the channel is in
state i is given by (5), from which and after obtaining the ith threshold
Ti, the (i + l)th threshold can be obtained using the following expression:
Ti+l = -pln(e -;; -7ri), i = 1,2, ... , N - 2 (8)
The details of the above procedure are now presented. Let {X (t) : t 2
O} be the (irreducible) Markov process that represents the state of the
channel. The time spent in any state Si is exponentially distributed with
mean Ti = t. The parameter qi represents the total rate out of state Si·
Consider states 0 and N in Figure 3. The total rate out of state 0 can be
approximated by the level-crossing rate (LCR) at Tl, i.e.,
(9)
10
Similarly, the total rate out of state N can be approximated by the LCR
at rN:
(10)
Dividing (9) by (10), we get the following expression for ~:
(11)
The above equation relates *'

rl, and r N. We still need to relate N to
these parameters. This can be done as follows. From (7), the steady-state
probability that the channel is in state 0 is given by:
1
11'0 = N (12)
(1 + *)
From (5) 11'0 is also given by:
.=!l.
11'0 = 1- e p (13)
(recall that ro ~ 0). From (12) and (13), we have
(14)
Similarly, the steady-state probability that the channel is in state N is

given by:
1I'N= (~)N (l-e 7 ) (15)
From (5), and noting that rN+1 ~ 00, 1I'N can also be written as
:::!:.J:i..
1I'N = e p (16)
From (15) and (16), an expression for the ratio ~ in terms of rI, rN, and
N can be obtained: J.L
>. e:::!:.J:i..
p ) 11
( (17)
~= 1-e7
11
From (15), (16), and (11), we can write the following equation for 7fN:
Taking the logarithm of both sides, we end up with:
(18)
To explicitly express N in terms of rl and rN, we replace * in (14)

by its expression in (11). Equation (18) can now be solved to obtain rl
provided that a suitable r N is chosen. Recall that by a suitable r N, we
mean a value that will not lead to the situation of two states both being
in the good or in the bad regions. By arranging the terms in (18), we
arrive at the following nonlinear equation that can be solved numerically
to obtain rl:
(19)
where N is given by
-In ( l-e .::.!:l.)

p
N = -~-'---""':""":- (20)
In (1 + V7f :~ )
In our analysis, we choose r N such that the service rate at state N is
almost equal to the error-free service rate (this depends on the relation-
ship between the SNR and the BER, which in turn is dependent on the
modulation scheme). This is done by solving for rN in
where 0 < E* « 1 is a predefined control parameter. In (21) the expression

for Pe(r) depends on the deployed modulation scheme. Using the obtained
rN, we numerically solve (19) for rl. Then from rl and rN we get N
using (20).
The algorithm in Figure 4 summarizes the steps that are needed to
obtain the parameters of the FSMC model. Note that the procedure
12
Parameterize-FSMC takes as inputs the parameters of the coding

scheme (n, T), the modulation-dependent BER function Pe (.), p, and €*.
It returns the number of states N, the partitioning thresholds Tl, ... , TN,
and the nominal SNR values h, ... ,rN-l (recall that TO = ro = 0 and
TN+l = rN = (0).
Parameterize-FSMC(p, Pe(.), EO, n, T)

1. Choose an appropriate TN by solving (21)
2. Solve (19) numerically for TI with N replaced by (20)
3. Calculate N using (20)
4. Set N = [Nl
5. Recompute TI using (19)
6. Recompute AI jJ, using (17)
7. Solve for the remaining thresholds as follows:
for i=l, ... , N - 2
7ri =( ~ ) (~)i(l_ e7)
Ti+l = -p In(e -;i -
7ri)
end for
8. Check for the existence of appropriate nominal SNR values:
for i=l, ... , N - 1
if ()(Ti) :S ilN, continue
else /* partitioning is not appropriate *1
setN=N-l
goto step 5
end if-else
end for
9. Compute the nominal SNR values:
for i=l, ... , N - 1
Solve ()(h) = it numerically for Ti
end for
1O.return(N, Tl, ... ,TN, TI, ... , TN-I)
Figure 4. Algorithm to calculate the parameters of the FSMC model.
Note that in step 9 of the algorithm if ()(Ti) > i/N for some state
i E {i, 2, ... , N -i}, then there is no Ti E [Ti' Ti+l) for which ()(ri) = i/N,
and the partitioning does not fulfill the requirements of the producer-
consumer model. If that happens, we decrement the value of Nand
repeat the computations (intuitively, decrementing N increases the ranges
of the various states, which improves the likelihood of finding appropriate
nominal SNR values). The recursion is step 8 is obtained from 7ri =
-r- -r-+l
e --t - e ~. Note that the algorithm is guaranteed to return a solution,
13
since for N = 1 the two nominal SNR values, 1'0 and 1'1, are given (the
partitioning reduces to the two-state Gilbert-Elliot model).
4. Performance Analysis
4.1 Queueing Model
Once the parameters of the FSMC are obtained, we proceed to com-
pute the queueing performance. The underlying queueing model can be
described by 2(N + I)-state Markov chain, as shown in Figure 5. The state
space of this chain is given by S = {(i, j) : i = 0,1 and j = 0,1,2, ... , N},
where a state (i, j) indicates that the source is in state i (on or off) and
the channel is in state j. Recall that the transition rates for the FSMC
are chosen such that the analysis in [13, 14] stays applicable.
"',I , ~,2
, )..i-I,i , , ,
.
N-2.N-I IN_I,N
n,n 0,1 O,H
.. O,i O,N-l
.. ',N
, • "'.I ,
... J.IN'-1,N-2 JlN,N-I
, , ,
~,o J.li,i-l
a a a a a p a
1,0
"',I ~
1,1
~,2
,
!.i-I
,-l,i ~
I.i
AN-2,N-l • l.N-1 "'-I,N ~
1,:-1
~ ~,o , 112,1
, I\,i-I
... )1N-I,N-2 ~ J.1N,N-l
Figure 5. State transition diagram of the Markov chain that governs the queueing
model.
The evolution of the buffer content follows the following first-order

differential equation [13]:
dIT(x)D = IT(x)M (22)

dx
where M is the generator matrix of the underlying Markov chain , D is
the diagonal drift matrix, and IT is the probability distribution vector for
the buffer content. These quantities are defined as follows:
D = diag[-co, ... , -CN, r - co, ... , r - CN] (23)
IT (x ) = [7fo,o ... 7fO,N 7f1,0 ... 7f1,N] (24)
7f s ,i(X) = Pr{ Q :S x, system is in state (8, in

where Q is the steady-state queue length. Throughout the paper, bold-
faced notation is used to indicate matrices and vectors. The solution
of (22) is generally given by the following spectral decomposition:
IT(x) = L aie-ZiXc/>i (25)

Zi2:0
14
where ai's are constant coefficients, and the pairs (Zi' ¢i), i = 1,2,3, ... ,
are the eigenvalues and left eigenvectors of the matrix MD- 1 (see [13, 14]
for details).
4.2 Wireless Effective Bandwidth

Determining the effective bandwidth amounts to determining the min-
imum value of c (the error-free service rate before accounting for the FEe
overhead) that guarantees a desired QoS requirement. In this section, we
determine the wireless EB subject to packet loss constraint. The compu-
tation of this EB has been performed in [11] for the case of a two-state
channel model. As we show later on, this results in an unnecessarily con-
servative allocation of network bandwidth. For the packet loss case, the
QoS requirement is given by the pair (x,p) where Pr[Q > x] = p. The cor-
responding EB is defined by: c* ~ min{c that results in Pr[Q > x] = p}
We obtain this quantity in terms of the source, the channel, and the
error control parameters. A common approximation for the EB, which
becomes exact as the buffer size goes to infinity, is based on the dominant
eigenvalue, z*, in (25). Essentially, z* is the eigenvalue with the small-
est positive real part. In [17] it is shown that z* is the unique positive
solution for the following equation:
(26)
where AA(Z) and Ac(z) are the Gartner-Ellis limits for the accumulative
arrival process {A(t) : t ~ O} and the accumulative service process {C(t) :
t ~ O}, respectively, defined as
AA(Z) ~ lim sup C 1 log E[exp(zA(t))]

t-+oo
and similarly for Ac. Once z* is obtained, the packet loss performance
can be approximately given by:
G(x) ~Pr[Q > x] = e-z*x (27)
For a two-state arrival process, AA(Z) is given by [17]:
AA(Z) ~~ (-a - {3 + az + v(a + (3 - az)2 + 4{3az) (28)
As for the multi-state service process, it can be viewed as a "negative-

flow" arrival process, where each consumer represents a fictitious flow with
arrival rate that alternates between zero and -Cl. Since the N negative-
flow sources are independent, Ac(z) for the total consumption process is
15
given by (Lemma 9.2.1 [17]):

N
Ac(-z) ~ L ACi(-Z) = NAc;(-z) (29)
i=l
where ACi(-Z) is the Gartner-Ellis limit for one consumer, and is given
by:
Ac; ( - z) ~ ~ (-A - J.L - Cl Z + J(A + J.L + Cl z)2 - 4ACI z) (30)
Using (26), (28), and (30), one can solve for z* that is found to be one
of the roots of a quadratic polynomial in z (hence, it is given in closed
form). Next, we compute c*. Let 'f7 ~ cdc = e/N = k/(nN). From (27),
z* can also be written in terms of the packet loss requirement (p, x) as
z* = _IO!p ~~. Substituting (28) and (30) in (26) and setting z = C we
arrive at the following equation:
-0: - j3 + a~ + J(o: + j3 - a~)2 + 4j3a~+

(31)
N (- A- J.L - 'f7~C + J (,\ + J.L + 'f7c~F - 4'f7c'\~) = 0
Let n be defined as follows:

n ~ -0: - j3 + a~ + J(o: + j3 - a~)2 + 4j3a~ - N (,\ + f-t)
N
Note that n is not a function of c, hence can be treated as a constant
with respect to c. Equation (31) can now be written as:
Rearranging and squaring both sides of the above equation, we get
(32)
Solving for c, we obtain a closed-form expression for the wireless EB c*
subject to a packet loss constraint:
c* = n2 - (A + f-t)2 (33)
2'f7~(f-t -,\ + n)
5. Numerical Results
In this section, we provide numerical results obtained based on the
previously presented analysis. A modulation scheme needs to be specified
16
in order to obtain the BER curve and the corresponding service ratios. In
our examples, we mostly focus on the binary phase-shift keying (BPSK)
and the differential phase-shift keying (DPSK) modulation schemes. For
BPSK modulation with coherent demodulation, the BER is given by:
Pb(r) = Q(~) (34)
where r is the instantaneous received SNR and Q(.) is the complementary

error function. For DPSK modulation, the BER is given by:
(35)
In all of our examples, we let n = 424 and use Reed-Solomon (RS) code
for error correction. We also take E* = 10- 10 and 1m = 50 Hz. Unless
specified otherwise, we take p = 4 and T = 5 (hence k = 414 for the RS
code). The first two examples are intended to illustrate the partitioning
approach. The first example uses BPSK modulation. Following the al-
gorithm in Figure 4, we first obtain rN i=:::j 6.970 by solving (21) for rN.
From this value, we obtain rl = 0.396 and N = 2.9397, which we trun-
cate to three, i.e., the channel is represented by a 4-state FSMC. We then
recompute the values of rl and AI fJ,. The resulting partitioning, shown in
Table 1, is found adequate (i.e., satisfies the service-ratio criterion).
I State I Partition I Service rate ratio I

0 [0,0.396) ~O
1 [0.396,2.2284) 0.33
2 [2.2284,6.970) 0.66
3 [6.970 ,(0) ~l
Table 1. Channel Partitioning in the case of BPSK (N = 3 and p = 4).
For the second example, we consider DPSK modulation. In the first

pass, we obtain rN = 6.6520, rl = 0.4425, and N = 2.7935. Setting
N = 3, we find that the resulting partition does not satisfy the required
service ratios. Hence, we decrement N and recompute the calculations.
For N = 2 the partitioning, shown in Table 2, is found to be appropri-
ate. Figure 6 depicts O(r) versus r for three different modulation schemes
(BPSK, DPSK, and FSK). Their contrasting behaviors indicates that the
channel partitioning is dependent on the modulation scheme. Note that
the faster B(r) reaches the saturation region, the smaller is the resulting
value of N. So one should always try to use the maximum possible N sub-
ject to the ability to enforce the requirements of the producer-consumer
17
State Partition Service rate ratio I
0 [0,0.4975) ~O
1 [0.4975,6.6520) 0.5
2 [6.6520,00 ) ~ 1
Table 2. Channel Partitioning in the case of DPSK (N = 2 and p = 4).
ReceivedSNR(r)
Figure 6. Service ratio (J(r) versus r for three modulation schemes (7 = 5).
model. Figure 7 depicts O(r) versus r for different values of T with BPSK
modulation. It is worth noting that the higher the capability of the FEC
code, the lower the expected number of states. This can be seen from the
sharper slopes of the service ratios in Figure 7. This can be explained
by the fact that the larger the value of T (stronger FEC code), the faster
O(r) approaches its asymptotic value, leading to a smaller value of N.
Figure 8 shows the solution for rl as a function of r N. This figure shows
that higher values of rN will result in smaller values for rl. The impact
of the choice of rN on N is shown in Figure 9 for different values of p (rl
fixed at 1.0). It can be seen that there exists an "optimal" value for rN
at which N is maximized, thus reflecting the channel characteristics more
accurately. Figure 10 shows the variation of N as a function of rl for
different values of p (rN fixed at 9.70). Thus, it is clear that the value of
the parameter N depends on the separation between the thresholds r N
and rl. Figure 11 shows the effect of p on N. It is observed that as p
increases, so does N, suggesting the possibility of using p as a means to
control N (since p can be controlled by adjusting the signal power at the
transmitter). Finally, from Figures 9 , 10, and 11 it is clear that a desired
value of the parameter N can be obtained by the appropriate choice of p
18
OB
i'
0.6
04
0.2
0
0 20
Figure 7. {}(r) versus r for BPSK with different error correction capabilities.
2.5,---,---,--.,..--.,---.,---,----,
1.5
Figure 8. Solution for rl as a function of TN, and N.
'.
Figure 9. N versus rN for different values of p (rl = 1.0).
19
and the thresholds rl and r N that satisfy the service ratios as discussed
before. Next, we study the impact of our channel partitioning approach
Figure 10. N versus rl for different values of p (rN=9.7).
Average Received SNA (P)
Figure 11. N versus the average received SNR p (rN=9.7).
on EB-based allocation subject to a packet loss constraint. For brevity,

we only show the results for BPSK modulation. The source parameters
are set to (J = 2604.1667 packets/second (about 1.1 Mbps when using
424-bit packets), a-I = 0.02304 seconds, and /3-1 = 0.2304 seconds. Un-
less noted otherwise, the results are obtained based on the previously
described BPSK example with default parameter values for 1m, T, and
p. We vary most of the remaining parameters and examine their impact
on the EB. Figure 12 depicts the impact of N on the EB. In theory, the
channel partition, and subsequently the values of A and /-l, depend on N.
20
However, in order to isolate the effect of N and gauge its impact on the
EB separately from other factors, we fix ,\ and f-l at their values in the
BPSK example (with N = 3). For the packet loss case, it is clear that for
a given QoS constraint the EB decreases dramatically with an increase
in N. This corroborates our intuition of the conservative nature of the
popular two-state Markov channel model. For example, by using four
states (N = 3) instead of two, c* subject to a PLR of 10- 6 is reduced by
almost 40% (and by 46% from the source peak rate). The reduction in the
EB can be explained by the fact that characterizing the channel with a
larger N reflects its error characteristics more accurately, and hence leads
to more efficient allocation. Figure 13 depicts the effective bandwidth as
a function of the PLR constraint p using different buffer sizes (x) with
N = 3. The figure shows that even with a small buffer, typical PLR
requirements (e.g., 10- 6 to 10- 3 ) can be guaranteed using an amount of
bandwidth that is less than the source peak rate. The significance of our
EB analysis is that it allows the network operator to decide beforehand
the amount of resources (buffer and bandwidth) needed to provide certain
QoS guarantees. A reduction in the per-connection allocated bandwidth
translates into an increase in the network capacity (measured in the num-
ber of concurrently active mobile users).
Figure 12. c' versus N (Buffer size = 50).
Figure 14 depicts the effective bandwidth as a function of the buffer

size using different PLR's with N = 3. The figure coincides with the
previous figure and shows that with relatively small buffer sizes stringent
PLR requirement can still be guaranteed. Figures 15 depicts the wireless
effective bandwidth versus the number of correctable bits (T) for the PLR
case. As T increases, the EB decreases up to a certain point T = T*, which
21
'~,--.------.------.----~.-----.
PLR
Figure 13. c· versus the PLR p (N = 3).
one may call the "optimal" FEC. Beyond r* the trend is reversed (Le.,
the overhead of FEC starts to outweigh its benefits).
I
~ 2000
!
~
~ 1500
~
1000
BufferSiz8
Figure 14. c· versus the buffer size (N = 3).
6. Conclusion
In this paper, we presented a new approach for partitioning the re-
ceived SNR range that enables tractable analysis of the packet loss and
delay performance over a time-varying wireless channel. This was done by
adapting the wireless channel to Mitra's producer-consumer fluid model,
which has known queueing performance. Our analysis exploited several
properties of a slowly-varying wireless channel, including its level-crossing
analysis and the Rayleigh distribution of the signal envelope. We provided
22
Figure 15. c* versus T (buffer size = 50).
an algorithm to iteratively obtain the various thresholds of the partitioned

SNR space. Numerical examples showed that the number of states de-
pends on the underlying modulation scheme, the average SNR p, and
the separation between the thresholds rl and rN. Finally, we presented
how tractable our proposed FSMC model in obtaining a less conservative
closed-form expression for the wireless effective bandwidth. Our future
work will focus on the use of the presented channel model on designing
efficient adaptive coding schemes through predicting channel states and
picking the optimal code rate for the probable state.
23
References
[1] M. Zorzi, R. R. Rao, and L. B. Milstein, "Error statistics in data transmission

over fading channels," IEEE Trans. Commun., vol. 46, pp. 1468-1477, 1998.
[2] H. Bischl and E. Lutz, "Packet error rate in the non-interleaved Rayleigh chan-
nel," IEEE Trans. Commun., vol. 43, pp. 1375-1382, 1995.
[3] H. Steffan, "Adaptive generative radio channel models," In Proceedings of the 5th
IEEE International Symposium on Personal, Indoor and Mobile Radio Commu-
nications., vol. 1, pp. 268-273, 1994.
[4] S. Lin and D. J. Costello, Error Coding: Fundamentals and Applications, Engle-
wood Cliffs, NJ: Printce Hall, 1984.
[5] W. C. Jakes, Microwave Mobile Communications, New York, NY: John Wiley &
Sons, INC., 1977.
[6] B. Vucetic, "An adaptive coding scheme for time-varying channels," IEEE Trans.
Commun., vol. 39, pp. 653-663, May 1991.
[7] M. Rice and S. B. Wicker, "Adaptive error control for slowly-varying channels,"
IEEE Trans. Commun., vol. 42, pp. 917-925,1994.
[8] H.S. Wang and N. Moayeri, "Finite-state Markov Channel- a useful model for
radio communication channels," IEEE Trans. Vech. technol., vol. 44, pp. 163-
171, 1995.
[9] Q. Zhang and Salem A. Kassam, "Finite-state Markov model for Rayleigh fading
channels," IEEE Trans. Commun., vol. 47, Nov. 1999.
[10] Y. L. Guan and L. F. Turner, "Generalized FSMC model for radio channels with
corrlated fading," IEEE Proc. Commun., vol. 146, April. 1999.
[11] J. Kim, and M. Krunz, "Bandwidth allocation in wireless networks with guaran-
teed packet-loss performance," IEEEIACM Transactions on Networking, vol. 8,
pp. 337-349, 2000.
[12] M. Hassan, M. Krunz, and I. Matta "Markov-based Channel Characterization
for Tractable Performance Analysis in Wireless Packet Networks," Submitted to
IEEE Transactions on Wireless Communications.
[13) D. Anick, D. Mitra, and M. M. Sondhi, "Stochastic theory of a data-handling
system with multiple sources," Bell Syst. Tech. J., vol. 61, pp. 1871-1894, Feb.
1982.
24
[14] D. Mitra, "Stochastic theory of a fluid model of producers and consumers coupled
by a buffer," Adv. Appl. Prob., vol. 20, pp. 646-676, 1988.
[15] A. I. Elwalid and D. Mitra, "Effective bandwidth of general Markovian traffic

sources and admission control of high speed networks, "IEEE/ACM Trans. Net-
working, vol. 1, pp. 329-343, 1993.
[16] A. I. Elwalid, and D.Mitra, "Statistical multiplexing with loss priorities in rate-
based congestion control of high-speed networks," IEEE Trans. Commun., vol. 42,
pp. 2989-3002,1994.
[17] C.-S. Chang, Performance Guarantees in Communication Netwoks, London

Berlin Heidelberg: Springer-Verlag, 2000.
[18] P G. Hoel, S. Port, and C. Stone, Introduction to Stochastic Processes, Prospect

Heights, IL: Waveland Press, Inc., 1972.
Capacity analysis of voice over IP over GERAN with
statistical multiplexing
A. WautierO, J. AntoineO, L. HussonO, J. BrouetOO, C. ThirouardOO

o Supelec, Dpt. Radioelectricite et Electronique, Plateau de Moulon, 91192 Gif-sur-Yvette,
France, 00 Alcatel CIT, Research & Innovation, Route de Nozay, 91460 Marcoussis, France
Abstract: The requirements in terms of service flexibility, spectrum efficiency and

speech quality introduce new challenges when voice is transmitted over packet
and over wireless. This paper analyses the perceived voice quality when voice
frames are transported on packet radio bearers of GSM/EDGE Radio Access
Networks (GERAN) and are statistically multiplexed. The resulting quality
depends on the link layer quality, on the scheduling and on the header
compression algorithms. This paper identifies the range of capacity gain
obtained with statistical multiplexing for a given speech quality considering
the different radio bearers ofGERAN.
Key words: Voice over IP over Wireless, wireless Internet, statistical multiplexing.
1. INTRODUCTION
Today, circuit-switched radio cellular systems like GSM offer good

service quality and spectrum efficiency, but provide very little service
flexibility. Recently, GPRS (General Packet Radio Service) and its enhanced
version EGPRS, which makes use of a modified physical layer EDGE
(Enhanced Data Rates for GSM Evolution), have been introduced to support
efficient data transmission (e.g. interactive IP -Internet Protocol- services
like web browsing or W AP) in GSM wireless access.
In future all-IP cellular networks, all types of services (real-time or not),
will be carried by a unique network infrastructure from the core to the access
networks serving the end-users. In particular, in the future releases of
GERAN (GSMIEDGE Radio Access Network), end-to-end packet
transmission of real-time IP applications is planned to support, for instance,
25
26 A. Wautier, J. Antoine, L. Husson, J. Brouet , C. Thirouard
IMS (Internet Multimedia Services) using GPRSIEGPRS radio bearers. In

this context, voice packets of different users can be dynamically multiplexed
on the same packet data radio channels (PDCH). They may also be
multiplexed with packets coming from other services.
When voice is transmitted over IP and over wireless, the requirements in
terms of service flexibility, spectrum efficiency and speech quality introduce
additional challenges. This paper deals with VoIP over GERAN. In
particular, it evaluates the benefit of statistical multiplexing, which exploits
silence periods in speech, in the evaluation of the capacity for a quality of
service measured by a speech quality estimator and by considering the
different packet radio bearers ofGERAN. The paper is organized as follows.
Section 2 gives an overview of the different aspects of the transmission of
voice over IP in the GERAN context. Section 3 presents the simulation
models we used for capacity evaluation. Section 4 presents some results
followed by concluding remarks.
2. TRANSMISSION OF PACKETIZED VOICE
Key features, when designing voice services over IP for cellular radio
links, are spectrum efficiency and robustness to transmission errors. This
section provides a short overview ofVoIP over GERAN specific issues.
2.1 Header compression
One of the problems encountered with IP over wireless is the large

overheads introduced by IP and other higher layer protocol headers such as
UDP (User Datagram Protocol) and RTP (Real-time Protocol). These
headers are used to transmit each packet to the correct host with the
appropriate application in the correct order and at the right time. In case of
real time IP services, IPIUDPIRTP protocol stack is used to convey the
media frames. When transporting packet voice frames, the length of an
IPIUDPIRTP header (40 bytes for IPv4 and 60 bytes for IPv6) is larger than
the payload. Namely, for the GSM Enhanced Full Rate (GSM EFR) codec
this corresponds to a payload of 30.5 bytes every 20 ms. Header
compression is essential for spectrally efficient transmission of VoIP and is
implemented as a three phase process: initial context establishment by
exchange of uncompressed header, regular context updates by transmission
of compressed headers, context restoration in case of excessive errors in the
header decompression. The compression algorithms must be efficient and
Capacity analysis ofVoIP over GERANwith statistical multiplexing 27
robust against errors to be used on the air interface. The CRTP (Compressed
RTP) algorithm [I], which is used for wireline VoIP, is not robust enough
for wireless links: if a compressed header is lost, the decompressor is not
able to reconstruct the subsequent headers (error propagation). Then, a single
packet error causes several consecutive lost packets (headers + voice
payloads). Some more adequate algorithms as ROHC (RObust Header
Compression) have been proposed [2]. ROHC is significantly less sensitive
to radio link errors thanks to repair mechanisms in the decompressor (no
error propagation) and reduces IPIUDPIRTP-packet header sizes down to
only 2 bytes most of the time [3]. In this paper, it is assumed that the
additional frame error rate due to header compression is negligible compared
to the other sources of errors further discussed in the following.
2.2 EGPRS Radio link layer
EGPRS is an evolution of GSM. Its air interface uses main basic physical
layer parameters of GSM (carrier spacing, TDMA frame and burst
structure), with the additional possibility of adaptive modulation: 8PSK
modulation is used instead of GMSK modulation when the radio link
conditions are favorable, which significantly increases the throughput.
Besides, different transmission rates are available. Depending on the chosen
modulation and coding scheme (MCS), the data rates range from 8.8 kbitls
(MCS-I) to 59.2 kbit/s (MCS-9) per time-slot (cf. Table I).
Modulation R£ Data Ulit MaJilnlm Code Rate Numberof30.5

(h)1esl bytHpeech
I MCS9 8·PS< 2X74 =~

59.2 1
frames per block
4
I MCS8 2Jd18 54.4 0.92 4
I MCS7 2lC56 44.8 0.76 3
MCSS 74 29.6 0.49 2
I MCS5 56 22.4 0.37 1
MCS4 GMS< 44 17.6 1 1
37 14.8 0.85 1
: MCS3
28 11.2 0.66 .
I MCS2
MCS1 22 8.8 0.53 .
Table 1 - GERAN: EGPRS radio bearers
The channel coding schemes are derived from the same convolutional
code, having a rate of 113 and constraint length of 7 by applying different
puncturing schemes [4]. It should be noted that all information bits are
equally protected. The transmission on the radio interface in EGPRS is based
on radio blocks transported by 4 bursts (time-slots) in 4 consecutive TDMA
frames. The interleaving scheme is rectangular and limited to a depth of 4
frames. This structure in radio blocks enables a highly dynamic resource
28 A. Wautier, J Antoine, L. Husson, J Brouet , C. Thirouard
sharing on each PDCH: from one radio block to the next one, the resource
can be used by different users.
2.3 Mapping voice on GERAN bearers
GSM voice codecs implement variants of LPC (Linear Predictive Coding).

It consists in analyzing voice frames of 20 ms in order to identify the
parameters of the model (e.g. autoregressive filter coefficients and source
samples), which are themselves coded in a very efficient way to reduce the
bit rate. The bits of a coded frame have different degrees of importance and
are divided in 2 classes, namely class I and class II. Class I is further divided
into two sub-categories: class Ia and class lb. The error free reception of
class Ia bits is essential for reconstructing the original speech frames.
Class Ib bits can tolerate some residual errors, and class II bits tolerate
higher error rate.
In a circuit-switched GSM voice transmission, the protection provided by
channel coding is hierarchical and adapted for each bit class. If GSM EFR
voice frames are transported by an EGPRS bearer, the same protection is
provided for all the bits (MCS-3 provides a 0.85 code rate for example).
Then, class Ia bits may not be enough protected and class II bits may be
over-protected. An improvement may consist in splitting the bits over
different radio blocks having different levels of protection [5]. Actually, the
mapping strategy has a significant impact on perceived speech quality as
exposed in [5].
EGPRS radio bearers provide a wide range of data rates on the air
interface. Therefore, several voice frames from the same user can be
transported on the same radio bearer. This number is a function of the MCS
and of the codec as indicated in Table 1 for HR (Half Rate), FR and EFR
codecs. Obviously, this enables to increase the number of serving channels
on the same time slot. For instance, MCS6 bearer conveys 2 EFR voice
frames (equivalent to 40 ms) per radio block (i.e. 20 ms) and then, the
available number of serving channel is 2.
2.4 Bursty nature of voice and statistical multiplexing
It is well-known that voice consists of a succession of talk-spurts and

silences. Ifthe voice codec has means to quickly detect transitions from talk-
spurts to silences and vice versa, then, voice packet streams become bursty.
GSM vocoders deliver 20 ms frames regularly during talk-spurts and no
frames (or few silence description frames) during silences. In a wireless
Capacity analysis of VoIP over GERAN with statistical multiplexing 29
packet transmission, a traffic channel can be assigned to a mobile user only

during the talk-spurts. The bursty nature of voice can either be used to pack
additional users through statistical multiplexing [6-8], or to convey other
data information streams, which can be either real-time or not, from same or
other users [8].
In this paper only statistical multiplexing of voice is addressed as in [6].
Actually, the scheduler exploits the silence periods to pack in more users in
average than the number of available serving channels, which enhances the
offered traffic per cell. However, it may happen that the number of voice
frames coming from several users is higher than the available number of
serving channels. Those extra frames can then be buffered and transmitted
later. Nevertheless, as voice service has very a stringent constraint on the
end-to-end transmission delay, the size of the buffer must be limited. If the
buffer overflows, packets that could not be delivered in time have to be
discarded.
With statistical multiplexing, the increase of the number of admitted calls
(which gives the offered traffic gain) is then limited by the acceptable speech
quality of those calls, which is sensitive to the number of shared serving
channels, to the selected MCS, to the chosen buffer size, and to the voice
activity factor. Consequently, a tradeoff between the increase of delay and
the packet loss rate has to be optimized Finally, overall capacity gain should
also take into account the reuse factor, which can be deduced from system-
level signal-to-noise ratio (SNR) simulations [9].
2.5 Voice quality evaluation
In a GSM circuit-switched transmission, the speech quality is only

related to the link level performance, which can measured by the following
metrics: the frame error rate (FER), the residual bit error rate (RBER) on
class Ib bits, and the BER of class II bits [10].
In a GERAN packet-switched transmission, these metrics cannot be
reused for several reasons. First, the physical link configuration is different
and is characterized by a plurality of transmission options. Secondly,
statistical multiplexing (and header compression to a lower extent) causes
additional packet loss and increased delays. Finally, the resulting speech
quality also depends on the missing frames processing [11]. We therefore
recommend to come back to the actual evaluation of perception of the
speech quality to carefully analyze all the impacting parameters in VoIP over
GERAN.
Speech perception is a complex process. Subjective and objective

methods have been developed. Subjective tests are realized with listening
tests, and the most well-known is the MOS (Mean Opinion Score) scale
ranging from 1 to 5. Recently, objective methods have also been introduced.
They can be divided in three groups. Comparative methods are based on a
comparison between the original signal and the delivered signal. Absolute
methods are based on the analysis of the delivered signal only. Finally,
parametric methods exploit the network transmission parameters to evaluate
the quality.
The comparative methods require the simulation of the whole
transmission link from "mouth" to "ear", whereas the parametric methods
allow faster evaluation assuming that the degradations due to the diverse
alterations (codec, BER, FER, delay, echo, ... ) are additive. These alterations
can therefore be studied apart and added in the degradation factor. In our
study, we resorted to two complementary methods: the PESQ (Perceptual
Evaluation of Speech Quality) that belongs to the first category [12], and the
E-model [13] that belongs to the last one. They are further detailed in the
following section.
3. SYSTEM MODELS
The models developed to determine the capacity gain due to statistical

multiplexing of voice over GERAN are described below.
3.1 Transmission model
In this study, the radio link including modulation, radio channel,

demodulation and equalization (DFSE, Decision Feedback Sequence Estimator
[14]) has been modeled and replaced by a two-layer hierarchical error-event
model using Markov chains. This kind of flexible model permits extensive
and fast characterization of wireless channels. With such a model, it is
possible to analyze or to simulate burst errors and therefore to compute
analytically or by simulation the bit error rate for any channel
coding/interleaving scheme [15]. The first model dedicated to channels with
memory has been suggested by Gilbert-Eliot, which is a two-state Markov
model where the states are called good and bad states. Fritchman extended
this model to an M-state Markov model with K good states (with error-free
events) and M-K bad states (with error events), where there is no transition
neither between good states nor between bad states (cf figure 1) [16].
Fig. 1 - Fritchman channel model
A two-layer hierarchical model is more relevant for wireless channels. It

consists of an external and of an internal chain sub-models. The external chain
gives the time variations of the mean energy per bit for each time-slot (in GSM,
the energy is assumed to be constant on a GSM burst). It can either be modeled
by a Markov chain [17] or be simply obtained from the random multi-path
Rayleigh fading impulse response model. The internal chain models the
distribution of the bit errors at the output of the equalizer over a time-slot. This
chain can be modeled by a Fritchman Markov model [16], which is
characterized by the number of bad/good states associated to error/error-free
events and by the matrix [pij] giving the probabilities of transitions between
states. Such a transition probability matrix must be computed for each state
of external chain. Different matrix sets are determined for each channel type
with the considered mobile speed (e.g. TU50), and for both GERAN
modulations. Then, error events over a burst can be easily determined by
generating state occurrences and the associated events and included in a
complete transmission link.
TV 50, GSMJEFR
100
--TCHlEFR
--.\-MCS 3
--MCS6
10 1:-----4_-+-A-----It-----I ---- MCS 1+5
0,1
o 5 10 15 20 25 30 35
SNR(d8)
Fig. 2 - FER due to transmission channel
Figure 2 illustrates the results obtained with a TU50 channel model with
different coding schemes encountered in GSM (TCHlEFR), in GERAN
(MCS3, MCS6), and with the coding scheme (MCS1+5) proposed in [5].
The Markov model represents a transmission including modulation, radio
channel, DFSE receiver structure with perfect channel impulse response
estimation. The Fritchman model has 2 bad states and 2 good states. Besides,
soft decisions at the output of the demodulator are considered [18, 19].
Moreover, perfect detection of erroneous frames is assumed.
3.2 Voice activity model
While it is difficult to model the voice activity ofa single user, Weinstein
found that the number of active lines could be modeled by a continuous-time
birth-death process [20]. He showed that this model is quite valid when the
number of users is superior to 25. The parameters that govern the transition
rates are the mean talk-spurt duration a-I and the mean silence duration p-I.
The voice activity factor II is defined by the ratio of mean talk-spurt duration
to the sum of mean talk-spurt and mean silence duration:
11 = n. -1 -1
p +a
Typical values encountered in the literature for the voice activity factor
are 0.445 (obtained with a-I = 1.41 sand p-I = 1.74 s ) or 0.36 (obtained
with a-I = 0.96 sand p-I = 1.69 s).
If btlt) is the probability of having i active lines at time t assuming that
we have} admitted voice communications, the steady-state probability of
having i active lines among} admitted lines is given by:
bijj -- (i}
i i (1- 11 )j-i fior 0 <
- l. <
- )
.
3.3 Traffic model
Speech traffic is modeled by a Poisson process for call arrival and by an

exponential distribution for call duration. The parameters of the model are
the call rate A and the mean call duration !l0l. The mean offered traffic
intensity is equal to the product A!l-l, which is denoted by p in the following.
3.4 Statistical multiplexing model
The statistical multiplexing model considered in this paper, is assumed to

be UAS (Uniform Arrival and Service) as in [5]. With such a model, an
analytical study can be conducted as proposed in [5, 6]. The parameters of
the scheduler is the number of shared serving channels c and the size of the
buffer m (introducing a maximal delay D).
In GERAN, the number of serving channels shared in a cell by a set of
active users is restricted. Indeed, the capacity of processing of the terminals
is limited. One terminal can only handle a maximum number of time-slots
per TDMA frame. Besides, if the terminal has a multi-slot capability, those
slots must be on the same radio frequency if multiple slots are actually used
in the same TDMA frame. Therefore, depending on the terminal capabilities,
either it has access to a single physical channel for the whole duration of the
call or to a pool of physical channels but located on the same transceiver
(TRx). The maximum feasible values for the number of serving channels c
per TRx is given in Table 2 for various MCS of GERAN with GSM FR or
EFR codecs assuming that all communications use the same physical link
configuration.
MCS type bearer C/TRx

MCS-9 32
MCS-6 16
MCS-3 8
MCS-1IMCS-5 8
Table 2 - Maximum values for the number of serving channel per bearer
The size of the pool and the number of pools per TRx will impact the
statistical multiplexing capacity gain. Table 3 summarizes the number of
pools p of size c per TRx (p x c) for different codecs and physical link
configurations versus the mobile station (MS) multi-slot capability.
MS multi-slot capability I 2 4 8
GSMFRorEFR MCS3 8xl 4x2 2x4 lx8
MCS6 8x2 4x4 2x8 lx16
GSMHR MCS3 8x2 4x4 2x8 lx16
MCS6 8x4 4x8 2x16 lx32
Table 3 - Possible values for the number serving channels per TRx (pxc)
The packet loss rate can be analytically obtained for a given number of
active users denoted by i assuming j admitted users. The considered voice
activity model is the one described by Weinstein [20]. A packet is lost when
the buffer is full (the actual number of packets in the buffer is q = m), this
can of course occur only when i is greater than c. Then, the packet loss rate
can be expressed as [5]:
j .
Pd(J) = I l-.C [P(n = ilN = j) -P(n = i,q < ml N = j)]

i:c TV
Considering the traffic model seen in 3.3 and assuming that the
maximum number of admitted users is N, the relationship between the
blocking probability and the offered traffic p is given by the classical Erlang
B formula [21]:
pN
P(N)= N!
I-k!
N pk
k:O
The worst case for the packet loss rate is obtained for j=N, which is given
by p ~N). Voice activity model can be combined with the traffic model: the
voice activity Markov chain can be viewed as a sub-chain of the traffic chain
[6]. Then, assuming that the two models are independent, the mean packet
loss rate is given by:
pj
I
N
Pd = Pd(J)P(J) where P(J·) = N
j! k
I~
j:c
k:O k!
The criterion to evaluate the capacity (possible offered traffic) must take
into account not only the blocking probability threshold of 2% but also a
criterion about the packet loss probability, which can be either a worst
packet loss rate threshold (1 % for example) or a packet loss rate threshold
for a given percentage of time (e. g. a threshold of 1% for 95% of time). In
the latter case, the distribution of the packet loss has to be considered.
Capacity analysis of VolP over GERAN with statistical multiplexing 35
3.5 Speech quality evaluation models
Two models are used to evaluate the quality of speech: the PESQ model
and the E-model.
The PESQ method (Perceptual Evaluation of Speech Quality) is dedicated
to end-to-end speech quality assessment of narrow band telephone networks
and speech codecs [12]. The original speech samples and the same but
degraded samples passed through a communication system are compared
using a psycho-acoustic model of the human ear. This method is also valid
for communication systems introducing distortions (e.g. time misalignment,
transmission errors, ... ). It is thus relevant for assessing the impact of the
radio link on voice quality perception.
The PESQ delivers MOS scores which is convenient to make the link with
other methods. In particular, the PESQ permits to calibrate some factors used
in the E-model and both models are thus complementary.
The E-model is a parametric model, which assumes that the degradations
due to different factors are cumulative and therefore studied separately [13].
The quality measurement R uses a scale between 0 (poor) and 100 (good)
which is linked to the MOS scale by the following transformation rule (it
should be noted that the degradations are only cumulative in the R scale)
[13]:
MOS = 1 + 0.035 R + 710-6 R (R -60) (100 -R)
In this paper, we consider two terms in the quality evaluation: ldd and Ie.
Original speech quality

(Backfound noise effect)
/)e-~
Degradation due to equipments (Codec. FER) Degradation due to delay
The parameter ldd depends on the transmission delay Ta due to the size of
the buffer and to the interleaving depth of the coding scheme. Figure 3-a
plots ldd whose expression is [13]:
with X = log(Ta 1100) if Ta > lOOms

log2
The parameter Ie depends on the intrinsic quality of the codec and on the
sensitivity to the FER (which includes the processing made to restore lost
frames). This degradation does not come from a closed-form formula but is
determined through listening tests or objective methods like the PESQ
method. For example, the GSM EFR codec has an intrinsic quality of 4,32 in
the MOS scale. In this paper, we used the PESQ method to determine Ie.
Figure 3-b gives Ie in case of uniformly distributed packet loss for GSM EFR
codec obtained with PESQ simulation and it is compared to the normalized
MOS evaluation [13]. This actually reflects the packet dropping due to
statistical multiplexing.
GSM EFR + lost frames processing, random FER
30 30
ri--Mosnonn V
25 25 _ _ PESQnote
20
/
V/V
20
J
15
~
15
~V
~
10 10
~~
~
o 100 200 300 400 500 0,1 10
Delay (ms) FER(%)
(a) (b)
Fig. 3 - Influence of delay and FER on speech quality degradation with E-model
4. SIMULATION AND RESULTS
4.1 Capacity gain due to statistical multiplexing
Achievable capacity gain due to statistical multiplexing is illustrated in

figure 4 for a given size of the pool corresponding to MCS-6 radio bearer as
depicted in table 3. It gives the packet loss rate Pij) as a function of the
number of admitted users j, when the size of the buffer is set for a delay from
oto 120 ms (6 speech frame duration). For a given target of the packet loss
rate, the capacity can be optimised by resorting to an increase of the buffer
delay.
E(talk]= 1.415 ; E[silence1= 1.745 ; c=16; Delay between 0 and 6 speech frames
5%
4%
~
j 3% I ................. , ..........................,..... +. . . . . . + ;············I+IIIH,r.····j
~
~
as 2%
t
1% I·························,··························· ........•........ ,············//AijH > ••..••••••.•••.•..• ~
oL---~---=~~~~~-L--~
16 18 20 22 24 26 28 30 32 34
number of users J
Fig. 4 - Influence of statistical mUltiplexing
Figure 5 illustrates the relative capacity gain as a function of the size of

the channel pool for packet loss targets of 1% and 3%. It shows that the
benefits of statistical multiplexing strongly depend on the size of the pool.
For values of c lower than 4, no gain can be expected (critical size effect).
Delay = 0 s. 'l = 0.445. E[talk]=I.4 s

120%.---~~--"r-~..:.....:.~~'-.'-..:.....:.-.--, 400%,
packet loss rate Delay ~ 100 ms
Packet loss rate =1 %
100% 300%'
~
~ 80% ~
packet loss rate =1% ~ 200%~
·5
"
" 60% ~
100%'
40%
O%:L'_ _ _ _ _ _ _==~~
20%0 5 10 15 20 25 30 35 40 45 50 0.2 0.3 0.4 0.5 0.6 0.7 0.8
C
Speech activity factor 11
Fig. 5 - Influence of the size of the channel Fig. 6 - Influence of the speech activity
pool on capacity gain for a given FER factor on the capacity gain
Figure 6 depicts the relative capacity gain as a function of the voice

activity factor for given sizes of the pool (c = 4, 8, 16, 32). As expected the
gain is higher for lower activity factors. Moreover, there are noticeable
variations when considering values belonging to the typical agreed scale (i.e.
between 0.36 and 0.45, see part 3.2). In order to reduce the impact of voice
activity factor in practical implementations, statistical multiplexing could be
combined with a real time estimation of the activity factor.
······..···.......!:[talk]=1.41 s, E[sil.]=1.74 S,'1=0.44

5 ....
c~ 4
,.-.,
.........
..~.." ..
.......
... ........ ........
..............................
]'" 3
....,
~ 2· ...
·..····· ........~[talk]=0.96 s, E[sil.]=1.69 s, ,'1 =0.36
.............................................................................................
00 20 40 60 80 100 120 140 160 180 200

delay (ms)
Fig. 7 - Influence of delay on FER
The delay and the packet loss rate are two parameters that induce
different effects on the perceived quality. The resulting degradations are
separately evaluated in the E-model (refer to figure 3). However, there is a
dependency between delay and FER. This relationship is illustrated in
figure 7, which confirms that the packet loss rate decreases as the delay is
increased. Considering both influences on the quality measurement leads to a
tradeoff as shown in figure 8 for the EFR codec.
70 L -_ _---L_ _ _--L.._-'-:-:-_.L..._,.-:----'
o m ~ ~ M m m ~ ~
Buffer delay ( ms)

Fig. 8 -Trade-off between degradation due to FER and delay on speech quality (EFR codec)
4.2 Incidence of radio bearers
The radio link behavior impact on the perceived speech quality has been
evaluated by means of simulation. The simulator models a complete
transmission chain comprising voice frame generation, coding, channel error
model, decoding, and voice frame reconstruction. The channel model that is
included in the simulator is a two-layer model replacing the modulation-
channel-demodulation as described in 3.1. The output parameters of the
simulator are the FER at the output of the decoder and the voice quality
measure obtained with the PESQ method. The MOS obtained with PESQ
method is converted in the R scale to be further exploited in the E-model
through Ie factor. Speech quality evaluation with E-model is illustrated in
figure 9 for the GSM EFR codec for three different MCS as a function of the
SNR.
TUSO, GSMlEFR + lost frames processing
100
90
80
..IV 0
.1 -;
70
I . . . I-r"'"
...~
0
60
I / /'
/'
,
50
~=: ./ I /
,
40 --TCHlEFRR
/ I I / --MCS3 R
/ /
30
--MCS6R
"/
20
10 I' -.l-MCS9 R
--MCS1+5R
0
o 10 15 20 25 30 35
SNR(dB)
Fig. 9 - Influence of radio link performance on speech quality (no statistical multiplexing)
4.3 Capacity evaluation
Offered traffic for a given quality of service defined by the blocking

probability of incoming calls and by the perceived speech quality of
admitted calls is calculated. This capacity can be given in a first step in
Erlang per TRx, then in Erlang per cell when the number of TRx is fixed and
the reuse factor can be deduced with the help of interference propagation
model. Finally, once the cell size has been calculated, capacity in
ErlangIMHzlkm2 can be evaluated. The design parameters are the MCS, the
maximum number of admitted calls per TRx N, and the buffer size
parameterized by the delay D. The choice of a radio bearer is directly linked

to the size of the serving channel pool (ef table 2 & 3). It is indirectly related
to the frequency reuse factor.
Without statistical multiplexing (i.e. the maximal number of admitted
calls per TRx is equal to the size of channel pool per TRx: N = e), the
offered traffic is obtained by the Erlang B model and it depends on the target
blocking rate. For example, for e = 16, the offered traffic is 9.8 Erlang per
TRx for a blocking rate of 2%. Then, the reuse factor can be taken into
account by considering the perceived quality related to a SNR threshold (ef
figure 9). For example, if we consider an acceptable speech quality threshold
of 60 in the E-model scale (3 in the MOS scale), the SNR threshold is 16 dB
with MCS-6 for the GSM EFR codec (ef figure 9). In a typical cellular
network with a reuse factor of 3x4, this minimum SNR value is reached for
80% of a cell. Here a propagation attenuation factor of 3.5 and a standard
deviation for the shadowing equal to 7 dB are assumed [20].
Shared channel per TRx: c 4 8 16 32

Offered traffic per TRx: 1.1 3.6 9.8 23.7
peN = c) for 2% blocking probability
Table 4 - Offered traffic per TRx wIthout statIstIcal multlplexmg
With statistical multiplexing, the offered traffic is a function of N, which

in tum is related to the perceived speech quality tradeoff. Table 5 gives the
relative capacity gain expressed in Erlang per TRx when considering that the
blocking probability is 2% and that the worst packet loss rate induced by
statistical multiplexing equals 1% (i.e. P(P~l%) = 100%). The capacity
gain given in table 5 can be further enhanced by tolerating a delay of 100 ms
(i.e. 40 ms for interleaving and 60 ms for the buffer) and by accepting
variable quality of service between different users. However, in order to
compensate for the effect of increased transmission delay and additional
packet loss due to statistical multiplexing, the SNR threshold must be
increased to ensure that the global perceived speech quality is maintained.
For example, for e equal to 16, tables 4 and 5 show that statistical
multiplexing is helpful in increasing the offered traffic from 9.8 Erlang per
TRx to 20 Erlang per TRx (MCS-6 and voice activity factor of 0.44). Also,
1 % FER for GSM EFR codec (ef figure 3) corresponds to a degradation of
10 points in E-model. So, if we consider that the global perceived quality
must be above 60, it is mandatory that the operating SNR be increased from
16 dB to 18 dB. For a given reuse factor, this will reduce the percentage of
satisfied users (70% of the cell instead of 80%). However, it is shown in [5],
Capacity analysis ofVoIP over GERANwith statistical multiplexing 41
that smart organization of transport of voice frames can compensate this

phenomenon thus making the statistical multiplexing gain more relevant.
Shared channel per TRx: c 4 8 16 32

Number of admitted calls per TRx: 5 12 28 62
N (P deN) = 1% with D= 0 ms, ,,=0.44)
Offered traffic per TRx: 1.7 6.5 20 51.5
peN) for 2% blocking rate
Relative gain in admitted calls: (N - c)/c +25% +50% +75% +94%
Relative gain in offered traffic: rp(N)-p(c)lIp(c) +54% +80% +106% +117%
Table 5 - Offered traffic per TRx with statistical multlplexmg
5. CONCLUSION
Capacity evaluation for voice service over IP over GERAN packet radio
bearers based on speech quality estimated through E-model and PESQ
methods, is presented. After an overview of main aspects of VoIP over
GERAN, we have evaluated the impact of statistical multiplexing on
capacity by combining closed-form analytical studies and simulations.
Substantial capacity gain can be obtained through dimensioning of
system parameters (e.g. buffer size, frequency reuse factor, ... ). Besides,
straightforward transmission of voice frames on PDCH can lead to high
requirements in terms of SNR. It has been shown however, that those
requirements can be easily mitigated, resulting in more typical SNR target
[5].
An exhaustive study of VoIP over GERAN should ideally consider
second order impacts resulting from header compression mechanisms as well
as the associated signaling channel overheads. Finally, system complexity
should be addressed and compared with more conventional solutions.
References
[1] "RFC 2508, Compressing IPIUDP/RTP Headers for Low-Speed Serial
Links", IETF, February 1999.
[2] "RFC 3095, RObust Header Compression (ROHC): Framework and four
profiles: RTP, UDP, ESP, and uncompressed", IETF, July 2001.
[3] L. Larzon et al. "Efficient transport of voice over IP over cellular links",
Proceedings ofPIMRC'OO, London, Sept. 2000.
[4] "3GPP TS 05.30: Channel coding", v. 8.6.1, Release 1999, January 2001.
[5] N. Paul et ai., "Efficient Evaluation of Voice Quality in GERAN", Proc. of

VTC'OI Fall, Atlantic City, September 200!.
[6] R. Tucker, "Accurate Method for analysis of a packet-speech Multiplexer
with limited delay", IEEE Trans. on Comm., Vol. 36, pp. 479-483, April
1988.
[7] K. Samaras, et ai." "Capacity calculation of a packet switched voice
cellular network", Proceedings ofVTC'OO Spring, Tokyo, May 2000.
[8] S. Fabri et ai., "Proposed evolutions of GPRS for the support of voice
services", lEE Proc. Commun., vol. 146, n05, pp.325-330, October 1999.
[9] M. Eriksson et ai., "The GSM/EDGE Radio Access Network -GERAN-
System Overview and Performance Evaluation", Proceedings of VTC'OO,
Tokyo, May 2000.
[10] "3GPP TS 45.005: Radio transmission and reception (Release 4)", v. 404.0,
June 2001.
[11] C. Perkins et al., "A survey of packet loss recovery for streaming audio",
IEEE Network Magazine, ppo40-48, 1998.
[12]"ITU-T P.862: Perceptual evaluation of speech quality (PESQ), an
objective method for end-to-end speech quality assessment of narrow-band
telephone networks and speech codecs", pre-published 0212001.
[13]"ITU-T G.107: The E-model, a computational model for use in
transmission planning", May 2000.
[14]A. Wautier, J-C. Dany, C. Mourot, "Phase correcting filter for sub-optimal
equalizers", Proceedings of the 1994 International. Zurich seminar on
Digital Mobile communications, Springer Verlag Lecture notes in computer
science, Vol. 783, March 1994.
[15]M. Zorzi, R.R. Rao, "Impact of burst errors on framing", PIMRC'98,
Boston, Sept. 98.
[16] B.D. Fritchman, "A Binary Channel Characterization Using Partitioned
Markov Chains", IEEE Transaction of Information Theory, Vol. IT-13,
n02, Apri11967.
[17] Babich G. Lombardi, "On verifying a first-order Markovian model for the
multi-threshold success/failure process for Rayleigh channel", VTC'97,
1997.
[18]N. Nefedov, "Generative Markov models for discrete channel modelling",
VTC '97,1997.
[19]N. Nefedov, "Discrete channel models for wireless communications", VTC
'98, May 1998.
[20] C. Weinstein, "Fractional Speech Loss and Talker Activity Model for T ASI
for Packet-Switched Speech", IEEE Trans. on Comm., Vol. 26, pp. 1253-
1257, Aug. 1978.
[21]X. Lagrange, P. Godlewski, S. Tabbane, "Reseaux GSM-DCS" Third
edition, chapter 6, Hermes, 1997.
UPLINK RRM FOR CONVERSATIONAL AND
INTERACTIVE SERVICES IN UTRA-FDD
O. SalIent, J. Perez-Romero, R. Agnsti, J. Sanchez

Universitat Politecnica de Catalunya
cI Jordi Girona 1-3, 08034 Barcelona, Spain
email: [sallent,jorperez, ramon@tsc.upc.esJ
Abstract The definition and assessment of suitable Radio Resource

Management (RRM) strategies able to provide a required QoS in the framework of
the UTRA segment of UMTS is a key issue for achieving the expectations created
on 3G technology. This paper evaluates specific algorithms for the admission
control of a new connection and UE-MAC strategies for the dynamic management
of the transmission parameters. Conversational and interactive-like services are
studied. Results reveal that the behaviour of the different proposed UE-MAC
algorithms in the uplink has an impact on the admission phase and, consequently,
they should be taken into account for a proper admission control algorithm design.
1. INTRODUCTION
W-CDMA access networks, such as the considered in UTRA-FDD proposal [1],

provide an inherent flexibility to handle the provision of future 3G mobile
multimedia services. 3G will offer an optimization of capacity in the air interface by
means of efficient algorithms for Radio Resource and QoS Management. RRM
entity is responsible for utilization of the air interface resources and covers power
control, handover, admission control, congestion control and packet scheduling [2].
These functionalities are very important in the framework of 3G systems because the
system relies on them to guarantee a certain target QoS, to maintain the planned
coverage area and to offer a high capacity. RRM functions are crucial because in W-
CDMA based systems there is not a constant value for the maximum available
capacity, since it is tightly coupled to the amount of interference in the air interface.
Moreover, RRM functions can be implemented in many different ways, this having
an impact on the overall system efficiency and on the operator infrastructure cost, so
that definitively RRM strategies will play an important role in a mature UMTS
scenario. Additionally, RRM strategies are not subject of standardisation, so that
they can be a differentiation issue among manufacturers and operators.
43
Multiaccess, Mobility and Teletraffic for Wireless Communications, Volume 6, 43--60.
44
To cope with a certain QoS a bearer service with clearly defmed characteristics
and functionality must be set up from the source to the destination of the service,
maybe including not only the UMTS network but also external networks. Within the
UMTS bearer service, the role of the Radio Bearer Service is to cover all the aspects
of the radio interface transport and this will be the focus of the present paper.
Control Plane l.!1£r.l.!A.ru! ConlTol PI.ne
L2
U!\rTS RADIO ACCESS NETWORK USER TERMINAL
Figure 1. UTRA radio interface protocol stack.
The radio interface of the UTRA is layered into three protocol layers: the
Physical Layer (Ll), the Data link Layer (L2) and the Network Layer (L3).
Additionally, the layer 2 is split into two sub-layers, the Radio Link Control (RLC)
and the Medium Access Control (MAC). On the other hand, the RLC and layer 3
protocols are partitioned in two planes, namely the User plane and the Control plane.
In the Control plane, Layer 3 is partitioned into sub layers where only the lowest
sublayer, denoted as Radio Resource Control (RRC), terminates in the UTRAN, as
Figure 1 shows.
Connections between RRC and MAC as well as RRC and Ll provide local
inter-layer control services and allow the RRC to control the configuration of the
lower layers. In the MAC layer, logical channels are mapped to transport channels.
A transport channel defmes the way in which traffic from logical channels is
processed and sent to the physical layer. The smallest entity of traffic that can be
transmitted through a transport channel is a Transport Block (TB). Once in a certain
period of time, called Transmission Time Interval (TTl), a given number of TB will
be delivered to the physical layer in order to introduce some coding characteristics,
interleaving and rate matching to the radio frame. The set of specific attributes are
referred as the Transport Format (TF) of the considered transport channel. Note that
the different number of TB transmitted in a TTl indicates that different bit rates are
associated to different TF. As the UE may have more than one transport channel
simultaneously, the Transport Format Combination (TFC) refers to the selected
combination of TF. The network assigns a list of allowed TFC to be used by the UE
in what is referred as Transport Format Combination Set (TFCS).
45
It is worth mentioning that for the optimisation of the radio interface utilisation,
RRM functions should consider the differences among the different services, not
only in terms of QoS requirements but also in terms of the nature of the offered
traffic, bit rates, etc. The RRM functions include:
1. Admission control: it controls requests for setup and reconfiguration of radio
bearers.
2. Congestion control: it faces situations in which the system has reached a
congestion status and therefore the QoS guarantees are at risk due to the
evolution of system dynamics (mobility aspects, increase in interference, etc.).
3. Mechanisms for the management of transmission parameters: are devoted to
decide the suitable radio transmission parameters for each connection (i.e. TF,
target quality, power, etc.).
4. Code management: for the downlink it is devoted to manage the OVSF code
tree used to allocate physical channel orthogonality among different users.
Within the UMTS architecture, RRM algorithms will be carried out in the Radio
Network Controller (RNC). Decisions taken by RRM algorithms are executed
through Radio Bearer Control Procedures (a subset of Radio Resource Control
Procedures) such as [3]:
1. Radio Bearer Set-up.
2. Physical Channel Reconfiguration.
3. Transport Channel Reconfiguration.
3GPP has provided a high degree of flexibility to carry out the RRM functions,
so that the parameters that can be managed are mainly:
1. TFCS (Transport Format Combination Set), which is network controlled
and used for Admission Control and Congestion Control.
2. TFC (Transport Format Combination), which in the case of the uplink is
controlled by the UE-MAC
3. Power, as the fundamental physical parameter that must be set according to
a certain quality target (defined in terms of a SIRtarget) and taking into
consideration the spreading factor used and the impact of all other users in
the system and their respective quality targets.
4. OVSF (Orthogonal Variable Spreading Factor) code
In the above framework, this paper focuses on the admission control and the
mechanisms for the management of transmission parameters for conversational and
interactive services carried out at UE-MAC in the uplink direction. It is worth
mentioning that the problem of QoS provisioning for multimedia traffic has gained
interest in the literature in recent years, as the problem arises in the context of 2.5G
and 3G systems and is not present in 2G systems. Thus, Naghshineh and Acampora
[4] introduced resource sharing schemes for QoS guarantee to different service
classes in microcellular networks. Akyildiz et al. [5] proposed the so-called
WISPER protocol, scheduling the transmissions according to their BER
requirements. Das et al. [6] developed a general framework for QoS provisioning by
combining call admission control, channel reservation, bandwidth reservation and
bandwidth compaction. Lately, Dixit et al. [7] among others have discussed the
evolution scenarios from 2G to 3G networks and the QoS network architecture
46
proposal by 3GPP for UMTS. For the decentralized uplink RRM component,
proposals such as the ones presented in [8] could be adapted to the VTRA-FDD
framework. In this respect, few studies aligned to the 3GPP specifications are
available in the open literature [9-11]. For the admission control several schemes
have been suggested for the uplink [12, 13] under different conditions and at a lower
extent for the downlink [14]. More recently, Ho et a1. [15] have built mathematical
models for various call admission schemes and have proposed an effective linear
programming technique for searching a better admission control scheme. The
admission approach presented in this paper is innovative in the sense that the
admission procedure in the uplink is related to the decentralised algorithm applied at
VE-MAC level and the relevance to take this fact into account is pointed out.
The rest of the paper is organised as follows. Section 2 details the uplink RRM
approach by proposing two different VE-MAC algorithms and a statistical-based
admission control strategy. Section 3 details the simulation model used to evaluate
the strategies through system level simulation in Section 4, where some basic
assumptions concerning congestion control are introduced. Finally, Section 5
summarises the obtained results.
2. UPLINK RRM
RRM strategies have to be applied in a consistent way to both uplink and

downlink. Focusing in the uplink direction, centralized solutions (i.e. RRM
algorithms located at the RNC) may provide better performance compared to a
distributed solution (i.e. RRM algorithms located at the UE) because much more
RRM relevant information related to all users involved in the process may be
available at the RNC. In return, executing decisions taken by RRM algorithms
would be much more costly in terms of control signalling because in this case VE
must be informed about how to operate. Consequently, strategies face with the
performance/complexity trade-off, which usually fmds a good solution in an
intermediate state where both centralized and decentralized components are present.
3GPP approach for the uplink could be included in this category, as it can be divided
in two parts:
1. Centralized component (located at RNC). Admission and congestion
control are carried out.
2. Decentralized part (located at VE-MAC). This algorithm autonomously
decides a TF (or TFC if combination of RABs exists) within the allowed
TFCS for each TTl, and thus operates at a "short" term in order to take full
advantage of the time varying conditions.
The expected effects of applying RRM strategies can be better explained by
comparison with the situation where there is not a tight control of the use of radio
resources, for example in a W-CDMA packet network in the uplink direction such as
the ones considered in [16, 17]. The typical uplink behavior of such a network
expressed in terms of throughput and delay is shown in Figure 2. Two regions can
be distinguished: in region A the offered load is low and the interference is also low,
so that packets are correctly transmitted, whereas in region B the offered load is high
47
and the interference is also high, so that packets are incorrectly transmitted and the
throughput decreases at the time that delay increases due to retransmissions. This
behaviour is due to the lack of coordination among mobile terminals. Despite in
strict sense the W-CDMA networks considered in [16, 17] are inherently unstable
due to the random access mechanism, in practice the system operation point may
provide a controlled performance.
A B A B
<:==:c> <!===:>
OFFERED LOAD OFFERED LOAD
Figure 2. Operation with no RRM (i.e. S-ALOHA W-CDMA network).
:'
.--~---t-··"-·/
OFFERED LOAD
Figure 3. Operation when RRM strategies are applied to a W -CDMA network.
In case of applying RRM, the purpose of admission and congestion control

would be to keep the system operation point in the region A, otherwise the system
becomes unstable and no QoS can be guaranteed. Smart admission and congestion
control strategies will shift region A at some extent to the right side, so the system
capacity is increased. Additionally, the performance achieved under region A is
dependant on the access mechanism and in some cases it could happen that the
system operation is access-limited instead of the more efficient case, which is
interference-limited. A suitable UE-MAC strategy should try to take full advantage
of the load conditions by pushing the system into a interference-limited situation,
48
which in turns provides a performance improvement in terms of delay (see Figure 3)

because active users could transmit at a higher rate. The challenge is to achieve a
good balance between improving the performance (for example in terms of
decreasing the delay under low load situations by increasing the transmission rate)
and maintaining the interference level manageable by the congestion and admission
control algorithms. Moreover, RRM can control and exchange the gain levels
between capacity and delay: if desired the admission region can be extended at the
expense of some reduction in the delay gain or the reverse, the delay gain can be
increased at the expense of some reduction in the admission region.
2.1. Admission control
Within a CDMA cell, all users share the common bandwidth and each new
connection increases the interference level of other connections, affecting their
quality expressed in terms of a certain (E,/N,). For n users transmitting
simultaneously at a given cell, the following inequality must be satisfied [18]:
PI X SF.I > (Eb

__ J i=1..n (1)
PN + X + [PR - p;] - No i
(2)
where Pi is the k-th user received power at the base station, SFi is the i-th user
spreading factor, PN is the thermal noise power, X is the intercell interference and
(E,/N,)i stands for the i-th user requirement. P R is the total received own-cell power
at the base station. Implicitly in the above inequalities a certain received power level
is assumed in each case:
i=1..n (3)
Adding all n inequalities it holds:
(4)
49
[~J
PR ~ _ _ _ _ _N_O---'-i_ _ => PR =---------'----
[~J (5)
1 1
1-2::--- 1-2::---
n n
SF;
i=1 SF;
i=1
+1 +1
Claiming in (5) for the inherent positivity of PR (i.e. PR>O) leads to:
n 1 (6)
I---<1
i=1 SF; + 1
Claiming in (5) for the inherent positivity of % (i.e. %>0) leads to:
(1+;'Jt. SF,1 +1<1

(7)
[!:J,
Claiming in (5) for the inherent positivity of PN (i.e. PN >0) leads to:
:J~
(8)
( 1+ SF;l + 1< 1
(~:l
The later expression is commonly known as the load factor [19]. The load factor
measures the theoretical spectral efficiency ofa W-CDMA cell:
(9)
50
Notice that 7]<1 is equivalent to claim that PN>O and so 7]<1 is the same
expression as (8). Introducing the definition of the load factor 17, (3) can be
expressed as:
P _1_
(PN + X +PR ) N 1-1"] i=l..n (10)
P> =---~
,- SF SF
---=-'-+1 j +1
where it can be observed that as the load factor increases the power demands also
increase. Consequently, and due to the limited power available at mobile terminals
and also for efficiency reasons the cell load factor must be controlled. Admission
control is one of the RRM strategies devoted to achieve such an objective.
The admission control procedure is used to decide whether to accept or reject a
new connection depending on the interference (or load) it adds to the existing
connections. Therefore, it is responsible for deciding whether a new RAB (Radio
Access Bearer) can be set-up and which is its allowed TFCS. Admission control
principles make use of the load factor and the estimate of the load increase that the
establishment of the bearer request would cause in the radio network. From the
implementation point of view, admission control policies can be divided into
modeling-based and measurement-based policies [20]. In case the air interface load
estimation is based on measurements and assuming that K users are already admitted
in the system, the (K+ 1)th request should verify:
< . _ PR +X l+l
11 + ~11 - 11 max With 11 - ---'-'---'-'-- and PR (11)
PR + X + PN ~ll = --S-'F---"----
_ _---"K:...c+-=--l_ _ +1
VK+l.(~b) o K+l
VK+l being the activity factor of the (K + l)th traffic source. In the case of the voice
service this factor is typically set to 0.67. For interactive services, like www surfing,
this factor should be estimated on a service by service basis.
Capacity and coverage are closely related in W -CDMA networks, and therefore
both must be considered simultaneously. In the measurement based approach llrnax is
obtained from radio network planning so that coverage can be maintained. The
coverage problem is directly related to the power availability, so that the power
demands deriving from the system load level should be in accordance with the
planned coverage. So, it must be satisfied that the required transmitted power will be
lower than P Tmax allowed and high enough to be able to get the required (EbINo)
target even at the cell edge:
51
P _1_
P =L (d) (pN + X + PR ) =L (d) N 1-11 i=1..n (12)
T., p' SF. p , SF
----"-+ 1 ; +1
( ~)
No ;
P 1
P . =L (R) N 1 -11 max i=1..n (13)
T.max p SF.
----'-,- + 1
( ~)
No ;
PT.; being the power transmitted by the i-th user, Lp (d;) the path loss (including
shadowing effects) at distance d; R the cell radii and T]max the maximum allowable
load factor for assuring coverage. The term 1I( 1 -17 ) is known as the interference
margin.
In case the air interface load is estimated in statistical terms it is the cell
throughput which is maintained and cell breathing effects may arise due to the fact
that intercell interference can not be directly and precisely included. For the
statistical-based approach and assuming that K users are already admitted in the
system, the (K +1)th request should verify:
K 1 1
+ (1 + f) (14)
(1+ f)I :<;;11 max
;=1 SF; SFK+1
+1 +1
v, {!:} VK+1(!b )
o K+1
where other-cell interference power is modeled as a fraction of the own-cell received
power (X=jxPR). According to (14) different admission strategies arise by balancing
the following parameters:
>- The spreading factor: by setting SFj as an estimated average value the user
will adopt along its connection time the assumed load will be closer to the
real situation at the expense of relying on the statistical traffic multiplexing.
In turns, considering SFj as the lowest SF in the defmed RAB covers the
worst case at the expense of overestimating the impact of every individual
user and, consequently, reducing the capacity.
>- The activity factor of the traffic source: by setting V; <1 the admission
procedure can be closer to the real situation of discontinuous activity
(typical in interactive-like services) at the expense of relying on the
statistical traffic multiplexing. In turns, V; =1 covers the worst case at the
52
expense of overestimating the impact of every individual user and,

consequently, reducing the capacity.
~ The overall load level: by setting 1]max the admission procedure allows for
some protection against traffic mUltiplexing situations above the average
(for example having more active connections than the expected average
number, or having more users making use of low SF than the expected
number).
2.2. UE-MAC strategy
For the conversational service, the DE-MAC strategy is straightforward because

the service is of constant bit rate nature, so that the DE will transmit at every frame
at the same bit rate. For interactive-like services (e.g. WWW browsing), two specific
algorithms are proposed:
2.2.1. Maximum Rate (MR) algorithm

It consists on selecting the TF that allows the highest transmission bit rate
according to the amount of bits Lb to be transmitted. Thus, the number of transport
blocks to be transmitted in a TTl would be:
numTB = min(TBmax,1 Lb.

TBslze I 1J (15)
TBmax being the maximum number of Transport Blocks that can be transmitted per
TTl and TBsize being the number of bits per Transport Block.
2.2.2. Service credit (Ser) algorithm

When a certain mean bit rate should be guaranteed, a new possibility arises that
makes use of the "service credit" (SCr) concept. The SCr of a connection accounts
for the difference between the obtained bit rate (measured in TB per TTl) and the
expected bit rate for this connection. Essentially, if SCr < 0 the connection has
obtained a higher bit rate than expected, if SCr > 0 the connection has obtained a
lower bit rate than expected. At the beginning of the connection: SCr(O)=O. In each
TTl, the SCr for a connection should be updated as follows:
SCr(n) = SCr(n-l) +( GuaranteedJate I TB_size) - Transmitted_TB(n-l) (16)
where SCr(n) is the Service Credit for TTI=n, SCr(n-l) is the Service Credit in the
previous TTl, GuaranteedJate is the number of bits per TTl that would be
transmitted at the guaranteed bit rate, TB_size is the number of bits of the Transport
Block for the considered RAB, Transmitted_TB(n-l) is the number of successfully
transmitted Transport Blocks in the previous TTL
The quotient GuaranteedJatelTB _size reflects the mean number of transport
blocks that should be transmitted per TTl in order to keep the guaranteed mean bit
rate. As a result, SCr(n) is a measure of the number of Transport Blocks that the
53
connection should transmit in the current TTl to keep the guaranteed bit rate. For
example, if TB_size=240 bits, GuaranteedJate=24 Kb/s, and TTI=20 rns, the UE
adds 2 service credits each TTL
Then, assume that in the buffer there are Lb bits, the number Transport Blocks to
be transmitted in the current TTI=n would be:
numTB = min(f Lb. 1, SCr{n), TBmaxJ

I TBslze (17)
Once numTB is calculated, the determination ofTF is straightforward.
3.- SYSTEM MODEL
The radio access bearer considered for supporting the interactive service has a
maximum bit rate of 64 Kbps in the uplink and an associated 3.4 Kbps signalling
radio bearer [21]. The radio access bearer selected for videophone service has a
constant bit rate of 64 Kbps when transmitting [21]. TB error rate target is 0.5%.
Possible transport formats are detailed in Table 1.
Table 1 Transport formats for the considered RABs

Service WWW VIDEOPHONE
TrCHtype DCH DCH
TB sizes, bit 336 (320 payload, 16 640
MACIRLC header)
TFS TFO, bits Ox336 Ox640
TFI, bits lx336 (16 Kb/s, SF=64) 2x640 (64 Kb/s, SF=16)
TF2, bits 2x336 (32 Kb/s, SF=32) -
TF3, bits 3x336 (48 Kb/s, SF=16) -
TF4, bits 4x336 (64 Kb/s, SF=16) -
TTl, rns 20 20
The interactive traffic model considers the generation of activity periods (i.e.
pages for www browsing), where several information packets are generated, and a
certain thinking time between them, reflecting the service interactivity. The specific
parameters are: average thinking time between pages 30 s, average number of packet
arrivals per page: 25, number of bytes per packet: average 366 bytes, maximum
6000 bytes (truncated Pareto distribution), time between packet arrivals: average
0.125 s, exponential distribution. The videophone traffic model is a constant bit rate
source of 64 Kbps with average duration 120s. As the interest of the present paper in
what admission control concerns is on the statistical terms in (11), the simulation
model includes a cell with radii 0.5 km and intercell interference is represented by
f=0.6. Physical layer performance, including the rate 1/3 turbo code effect and the
1500 Hz closed loop power control, is taken from [22] to feed the system level
simulator presented here with BLER (BLock Error Rate) statistics. The mobility
54
model and propagation models are defined in [23], taking a mobile speed of 50 km/h
and a standard deviation for shadowing fading of 10 dB.
4.- RESULTS
As a previous result and for a better understanding of the admission control

phase, several figures regarding UE-MAC algorithms should be detailed. One
important measurement to understand the behaviour of the different UE-MAC
strategies is the transport format distribution used. Referring to Table 1, UE-MAC
has the freedom to choose among TFO (when the buffer is empty or when SCr<O),
TF1, TF2, TF3 and TF4. For SCr24 (see Figure 4, SCrX standing for a service credit
strategy with a guaranteed rate X Kb/s) it can be observed that most of the time TFI
and TF2 are used because the UE buffer queues several packets and so it tends to
transmit the information at 24 Kb/s. In turns, in the periods that the UE buffer is
empty the UE is gaining service credits and, when a new packet arrives the
transmission rate is increased over the guaranteed one (i.e. TF3 and TF4 are used).
For MR strategy, as it chooses the TF according to the buffer occupancy and tries to
transmit the information as fast as possible, most of the time TF4 is being used (see
Figure 2). Additionally, Table 2 shows the average delay performance for both MR
and SCr strategies. It can be seen that MR provides a lower delay because the
strategy tends to maximise the transmission rate according to the buffer occupancy.
Table 2. Avera~e de1ay perfiormance fior UE -MAC algorithms.

UE-MAC Average packet
strategy delay (s)
SCr24 0.54
MR 0.12
MR
SCr24 0.9 r - - - - - -- - - -- - - -
G . r - - - -- - - - - ---- 0.8 t - - - - - - -- - --
el.lS 0.7t - - - -- - - -- - -
OJ o.&t----- - - -- - -
02' 0.5t - - - - - -- - - --
0.• t - - -- - - - -- - -
0.150 0.3 t_- - - - -- - - - -
G.' O~ t_----------
0.050 0.1 t - - - - -- - -
O~. .~~--. .~-
IF. TF. T.2 TFJ TF ..
Figure 4. TF distribution for SCr. Figure 5. TF distribution for MR.
According to (13), Figure 6 plots the maximum cell radii for a 95% coverage
probability as a function of the load factor for the considered services with a cr= 10
55
dB shadow fading and unity antenna gains. Thus, for a cell range of 500m, the load
factor must be below 75%.
95% coverage probability
Figure 6. Cell range for different load factors.
Admission probabilities for videophone service with 11 max =0.75 are shown in
Table 3. Since videophone is of constant bit rate nature, the admission procedure is
quite easy to handle. For 15 users in the cell, the system is still below the planned
load factor, all users can be admitted and the performance is good, as shown in
Figure 7 and Figure 8. In particular, Figure 7 shows the power limitation probability
(i.e. the probability that the required transmitted power is above the maximum
value) and Figure 8 the TB error rate also as a function of the distance to the cell
site. If the offered load is above the planned load factor, as it is the case for 30 users,
the admission procedure is able to assure the system stability, planned coverage and
planned quality by rejecting connection requests. Figures 7 and 8 show a little
increase in the power limitation probability as well as the TB error rate due to the
increase in the load factor. In turns, Figure 9 shows the load factor distribution in the
system.
c: 0,05
0
E :g~ 0,04
. .
0,03 __ 15 users
~ ftI
,g
0,02 __ 30 users
G) 0
~ CL
0,01
0
CL
0
o II) 0 II) 0 II) 0
I'- II) N 0 I'- II)
~ N C") ('I') 'V
Distance to cell site (m)
Figure 7. Dynamic power limitation probability.

56
-
a~4
GI 3 l
1--15
....
1U
0:::
2
/, users
w 1
--30 users
0
/~
...-..-
10
l- 0
\)
'\~ ,,'0\) ~ s::,\) (I..~ 0\)
'1,. ~ '!l l>i
Distance to cell site (m)
Figure 8. Transport Block error rate.
40~--------------------------,
~ 30+--------------T~----------~
~ 20+-------------~~----------~
.c
£ 10 +------------"7....------'1,.----------1
o N M ro
a a a a a a a a a
~ ~ ~ ~ ~ ~
Load factor
Figure 9. Load factor distribution for 20 videophone users.
Admi'
T,ahie3. SSlOn probabT'
I It1es fIor VI'deoplhone serviceo
Number of videophone users Admission
probability
15 1
20 0.98
25 0.79
30 0.59
For the interactive service the situation is not so easy to handle because of two
dynamic issues affecting the system behaviour that are difficult to predict: the
statistical traffic multiplexing (the interactive service is of discontinuous nature and,
consequently, the number of simultaneous users in a given frame in principle is not
known in advance) and the TF used in uplink transmissions (it is decided in a
decentralized way by UE-MAC and, consequently, the set of SF used by
simultaneous users in a given frame in principle is not known in advance). Table 4
shows the admission probabilities for different values of the admission TF
57
(equivalent to SF) and 11 max for both MR and SCr strategies. The activity factor is
assumed to be the average value coming from the traffic model. The criterion for
considering the system under a congested situation is when (18) holds for more than
90 out of 100 consecutive frames, revealing that the CDMA capacity has been
overcome. Note that depending on the specific congestion detection and congestion
resolution algorithms, the system could continue operating under normal conditions
or not and the interest of the present criterion is only for establishing a basis for
comparison purposes since we are not dealing with congestion control algorithms in
this paper.
n
1 (18)
(1+ /)"L >l1th
SF;
(~)
;;1
+1
No ;
T,a hie 4.AdmiSSlOn

. probabTf
Illes uor d'fu
I eren cases.
Number Admission Admission Admission
ofwww probability probability probability
users TF211max =0.75 TF411max =0.75 TF4 11 max =0.9
MR SCr MR SCr MR SCr
450 1 1 0.98 0.98 1 1
500 Congo 1 0.93 0.91 1 1
550 Congo 1 0.84 0.82 Congo 0.98
600 Congo Congo 0.76 0.74 Congo 0.93
It can be observed from Table 4 that for a proper admission procedure the
characteristics of the decentralized algorithm being applied at DE-MAC layer should
be taken into account. For example, if TF2 and 11 max =0.75 are considered in the
admission phase for MR strategy, and since the dynamic behavior of this algorithm
tends to use TF4 in most cases, the system enters in congestion with 500 users
because the admission is too soft. In turns, for SCr the TF considered for admission
purposes is much better adjusted to the real dynamic value, so that admission allows
for more than 550 users to enter in the system while maintaining a controlled
performance. On the other hand, if TF4 is considered for admission purposes,
congestion is avoided because from the transmission rate point of view the worst
case is considered and from the traffic multiplexing point of view 11 max =0.75 is low
enough to absorb traffic fluctuations without causing congestion. Nevertheless, for
SCr strategy this is not so suitable because the admission is too strict. It is worth
noting that the value for 11 max eventually allows for a softer or stricter admission as
shown in the example in Table 4, where increasing the value up to 11 max =0.9
58
improves the percentage of admitted users for SCr strategy compared to the TF4 and
11 max =0.75 case. For this later case, Figure 10 plots the statistical load factor
distribution, showing that it tends to be quite low because of the still strict admission
procedure. For 550 wwwusers, 2% of the requests are already rejected (see Table 4)
while the performance in the system is still very good. As a matter of fact, Figure
11 shows that the power limitation probability as a function of the distance to the
cell site (i.e., the probability that a given user requires more power than the
maximum allowed for achieving the target EbINo) is within the coverage probability
design even at the cell edge. Also, Figure 12 plots the average packet delay again as
a function of the distance to the cell site, and reveals that no performance
degradations are observed as one moves far from the cell site.
100
10
~
~ 0.'
:;;
"
~
0,01
Q. 0,001
0.0001
0,00001
,
'" "'. ",'!- ",,,
"'.'" ",'? "'.
~
Load factor
Figure 10. Load factor distribution for 550www, SCr,

admission TF411max =0.9.
l>
~ 0,025 ~---';";"'---:--...."......,..,.___;;"'::;:"";''''';'''--4-I
~" 0,02 ~---~:""""-;:'"...--::-"";:'''''''::'':'':'''--:-''':'''';:'''--I---l

Q.
c
g 0.015 ~~~~':';""--:-::"""';''''''''~=:':'''''-'----r'----l
!
!!
0,01 ~"";:,~=-=,..,-~.,,...;,.....,..,=,::.,-..-...,--'-..r---"-;""'--l
t 0.005~;""'~~~'--~;""'-~--:-:~~--:-:~___--l
o
Q.
~ ~ &# $ # $ # ~
Olsta nee to ce II site (m)
Figure 11. Power limitation probability for 550www, SCr,

admission TF411max =0.9.
59
~ # $ # # # $ # ~
DISIa nee to cell site (m)
Figure 12. Average packet delay for 550www, SCr,

adnrission TF411max =0.9.
5. CONCLUSIONS
3G will offer different QoS guarantees and an optimization of capacity in the air
interface by means of efficient RRM algorithms, which should consider the
differences among the different services, not only in terms of QoS requirements but
also in terms of the nature of the offered traffic, bit rates, etc. In the framework of
VTRA-FDD, this paper has focused on the admission control and the mechanisms
for the management of transmission parameters for conversational and interactive
services in uplink direction. Results for VE-MAC strategies show that for SCr
strategy most of the time TFI and TF2 are used while for MR strategy most of the
time TF4 is used This different behaviour of the VE-MAC algorithms impacts on
the admission control process, which should take this fact into account for avoiding
either too strict or too soft policies.
6. ACKNOWLEDGMENTS
This work is part of the ARROWS project, partially funded by the European
Commission under the 1ST framework (1ST 2000-25133) and by the Spanish
Research Council under grant TIC2000-2813-CE.
7. REFERENCES
[I] 3GPP TS 25.211, "Physical channels and mapping of transport channels onto physical
channels (FOO)"
[2] 3GPP TR 25.922 v4.0.0, "Radio resource management strategies"
[3] 3GPP TS 25.331 v4.0.0, "RRC protocol specification"
60
[4] M. Naghshineh, A. S. Acampora, "Design and control of micro cellular networks with QoS
provisioning for data traffic", Wireless Networks 3 (1997) , pp. 249-256.
[5] I. F. Akyldiz, D. A. Levine, I. Joe, "A Slotted CDMA Protocol with BER Scheduling for
Wireless Multimedia Networks", IEEE/ACM Transactions On Networking, Vol. 7, No2,
April 1999, pp. 146-158.
[6] S. K. Das et ai, "A call admission and control scheme for QoS provisioning in next
generation wireless networks", Wireless Networks 6 (2000) ,pp. 17-30.
[7] S. Dixit, Y. Guo, Z. Antoniou, "Resorce Management and Quality of Service in Third
Generation Wireless Networks", IEEE Communications Magazine, Vol. 39, No.2, February
2001.
[8] O. Sallent, R. Agusti, "A Proposal for an Adaptive S-ALOHA Access System for a
Mobile CDMA Environment", IEEE Transactions on Vehicular Technology, Vol. 47, n° 3,
August 1998.
[9] K. Dimou, P. Godlewski, "MAC Scheduling for Uplink Transmission in UMTS W-
CDMA", Proceeding of the IEEE Vehicular Technology Conference, VTC 2001 Spring.
[10] R. De Bernardi et aI., "Load Control Strategies for Mixed Services in WCDMA",
Proceedings of the Vehicular Technology Conference 2000, VTC'OO, pp. 825-829.
[II] W. Rave et aI., "Evaluation of Load Control Strategies in an UTRAlFDD Network",
Proceedings of the Vehicular Technology Conference 200 I, VTC'O I, pp. 2710-2714.
[12] H. Holma, J. Laakso, "Uplink Admision Control and Soft Capacity with MUD in
CDMA", Proceedings of Vehicular Technology Conference, VTC'99 Fall.
[13] Z. Lui, M. El Zarki, "SIR Based Call Admission Control for DS-CDMA Cellular
Systems", IEEE Journal on Selected Areas in Communications, Vol. 12, 1994.
[14] J. Knutsson et aI., "Downlink Admission Control Strategies for CDMA Systems in a
Manhattan Environment", Proceedings of Vehicular Technology Conference, VTC'98.
[15] C. J. Ho et aI., "On Call Admision Control in DS/CDMA Cellular Networks", IEEE
Transactions on Vehicular Technology, Vol. 50, n° 6, November 2001.
[16] R. K. Morrow, J.S. Lehnert "Packet Throughput in Slotted ALOHA DS/SSMA Radio
Systems with Random Signature Sequences", IEEE Transactions on Communications, Vol.
Com. 40, No.4, July 1992, pp. 1223-1230.
[17] N. Abramson, "Multiple Access in Wireless Digital Networks", Proceedings of the IEEE,
Vol. 82, No.9, September 1994, pp. 1360-1369.
[18] O. Sallent, J. Perez-Romero, F. Casadevall, R. Agusti, "An Emulator Framework for a
New Radio Resource Management for QoS Guaranteed Services in W-CDMA Systems",
IEEE Journal on Selected Areas in Communications, Vol. 19, No. 10, October 2001.
[19] H. Holma, A. Toskala (editors), W-CDMAfor UMTS, John Wiley and Sons, 2000.
[20] V. Phan-Van, S. Glisic, "Radio Resource Management in CDMA Cellular Segments of
Multimedia Wireless IP Networks", WPMC'OI Conference Proceedings.
[21] 3G TS 34.108 v.3.2.0, "Common Test Environment for User Equipment. Conformance
Testing"
[22] J. Olmos, S. Ruiz, "UTRA-FDD Link Level Simulator for the ARROWS Project",
IST'OI Conference Proceedings, pp. 782-787.
[23] 3GPP TR 25.942 v.2.1.3, "RF System Scenarios"
Rate and Power Adaptation for Downlink Shared
Channel in WCDMA
Saleem Akhtar and Djamal Zeghlache

Telecommunication Networks and Services Department
Institut National des Telecommunications
9, rue Charles Fourier
91011 Evry Cedex, France
Tel: (+33) 1 60 764584, Fax: (+33) 1 60 764291
saleem.akhtar@int-evryJr, djamal.zeghlache@int-evryJr
Abstract- Rate adaptation scheduling with prioritized call admission control

for UMTS WCDMA networks is analyzed. Admission is based on maximum
base station transmit power and service class priority. Conversational and
interactive services are considered. Real Time services with stringent QoS
requirements are operated in circuit switched mode over the dedicated radio
bearers. Downlink shared channels are used to transport Non Real Time data
handled in packet switched mode. Rate adaptation based upon link quality for
NRT services and system load is introduced to improve performance and
increase capacity. Results indicate that rate adaptation scheduling uses radio
resource efficiently and improves system stability. Compared with fixed rate
scheduling, rate adaptation scheduling results in better overall QoS.
1. INTRODUCTION
3G wireless mobile cellular networks have the ability to accommodate

heterogeneous multimedia traffic, composed of voice, video and data. The
mixture of services, data rates and QoS needs in 3G systems require advanced
radio resource allocation algorithms to benefit fully from this enabling
technology. Proper and efficient radio bearer allocation and management is
imperative for respecting QoS and maximizing system throughput.
Adaptive rate transmission can be applied to interactive services that are
bursty in nature and have very flexible delay and throughput requirements [1].
Most of the published work [2-3] for rate adaptation in CDMA focuses on the
uplink dedicated channel for rate adaptation. The downlink is rarely addressed
especially in the multiple services context. In WCDMA, if rate adaptation is
applied to downlink dedicated channels, the dedicated channel spreading
factor can not be varied on a frame by frame basis. Mobile stations are
61
62
allocated channelization codes corresponding to the highest data rate. Data
rate variation is achieved by rate matching operation (selecting a new leaf in
the allocated code tree branch) or discontinuous transmission. Few high speed
data users with low activity factor can make the code tree run out of
channelization codes. For interactive and background service classes
resources need not be permanently allocated. Shared channels combined with
time division scheduling can be used to reduce the downlink code resource
consumption. Over WCDMA Downlink Shared CHannels (DSCH) variable
spreading factors can be allocated to the mobile stations on a frame by frame
basis. Rate adaptation schemes with packet scheduling can be introduced to
improve capacity. Scheduling targets preferably high bit rates because
transmission at high rate requires less energy per transmitted bit and incurs
shorter delays. Transmission at higher bit rates reduces interference in the
network and results in better statistical multiplexing. Transmission times are
shorter and the number of concurrent packet calls is lower. Time division
scheduling should be applied to few very high rate users at a time. However,
higher bit rates can not be assigned under all channel and network load
conditions. In the downlink, users experience different radio link quality
depending upon their location and channel conditions. Increasing transmission
power to overcome bad channel conditions can result in rapid increase of base
station power and interference to other mobile stations. Even if a base station
has enough power budget, mobiles having bad channels may require a large
amount of power. Limiting transmission power for a user is not a proper
solution either as it suppresses both interference and signal. A better approach
is to reduce transmission rate to accommodate users having bad channel
conditions. Users in the best radio conditions get high bit rates. Transmission
rates for users in poor radio conditions are lowered. The resulting reduced
transmission power and interference leads to improved performance and
capacity under poor channel conditions.
2. SCHEDULING AND RATE ADAPTATION
There are three types of channels available for transmissions on downlink

WCDMA: dedicated, shared and control channels [4]. Dedicated channels are
used to transmit conversational and streaming classes to meet stringent delay
requirements. For interactive and background classes, the resources need not
be permanently allocated as data is bursty in nature. Shared channels can be
assigned to these services and scheduling used to exploit service burstiness to
accommodate more users and to maximize resource utilization.
A Physical Downlink Shared CHannel (PDSCH) corresponds to a
channelization code below or at a PDSCH root channelization code within the
code tree (leaves of a code tree branch). A high rate PDSCH can be allocated
to a single user. Alternatively under the same PDSCH root channelization
63
code, several lower bit rate users can be allocated lower rate physical
downlink shared channels on a code multiplexing principle. Downlink shared
channels allocated to users on different radio frames may have different
spreading factors. Each PDSCH is associated with a downlink Dedicated
Physical CHannel (DPCH) to achieve power control and signaling. All
relevant Layer 1 control information is transmitted on the associated
Dedicated Physical Control Channel (DPCCH); i.e. the PDSCH does not
carry any Layer 1 information.
In the downlink, each base station has a maximum power budget that is to
be shared among users belonging to the different QoS classes. A portion of
this base station power is allocated to common control channels such as base
station specific beacon channel and pilot channel. The remaining power is
available for traffic or information channels.
The users in a cell are allocated power levels according to their QoS
requirements and their relative locations. The number of simultaneous users
served by a base station is limited by its maximum radiated power (Pmax).
The received average signal-to-interference (SIR) by a given mobile can be
expressed as:
SIR = W. ~ (1)
R Idownlink
Where ~ is the received average signal power, Idownlink is the total
downlink interference, W is the chip rate and R is the transmission rate.
Since the downlink codes are not completely orthogonal within a cell due to
multipath propagation, the total downlink interference is given as:
Idownlink =y . I intra + linter (2)
I intra is the interference from within the home cell, I inter is the
interference from surrounding cells and y is known as the orthogonality

factor.
In order to attain the required QoS, the mobile station must ensure a
minimum SIR value Ai,req for the service of QoS class i. Assuming that
base station k transmits power PTk =~kPmax' a fraction of the maximum
power (Pmax )' we can rewrite equation 1 for mobile user j in the center cell 0
as:
(3)
64
P;j denotes the required power to be transmitted by the base station to

mobile station j using a service of QoS class i at a data rate of Rij' aij is
the service activity factor for user j while using a service of QoS class i .
Gkj is the path gain from base station k to the mobile station j . Go j is the
path gain to home base station. K is the number of interfering base stations.
From equation 3, we have:
(4)
If there are m services and N; users of service class i in the cell, the
following condition must be satisfied to ensure proper operation:
(5)
The objective of a resource allocation strategy is to provide the best power

allocation P*ij at any given time to achieve highest throughput while
respecting every service class QoS for all on going connections. Equations 4
and 5 indicate that users requiring high power have an important influence on
the overall base station power budget. Poor link quality exacerbates the power
allocation issue by requiring additional resource. Connection Admission
Control should take into account link quality and traffic load and must be
combined with scheduling of services requiring lower grade QoS to achieve
quality control.
The scenario used in this paper to assess performance of rate adaptation for
WCDMA consists of a 256 Kbps interactive service offered simultaneously
with a conversational 64 Kbps RT service as depicted in Figure 1. Each
service is admitted on a different basis according to its priority class and the
radio bearers that are used to convey information. Handover requests have
priority over new connections for all classes.
RT services are operated on a blocked call basis and hold highest priority
in the system. No waiting queue is used because of very tight delay
requirements. New connection requests from the RT service are admitted onto
the system only if two conditions are met. There is enough power remaining
in the base station power budget to compensate for the estimated path loss by
the mobile unit on an open loop basis. The individual traffic channel (DPCH)
power limit is not violated.
65
64 Kbps R T service FIFO Priority 1

--~~TI~~-nll~~~~~------~~~
CAC and serving Queue
Check BS remaining Power Budget,
Power Budget for DPCH
256 Kbps Interactive service
I
FIFO Priority 2
Scheduling Queue
I I I I
r+8
~ II ~ DSCH
CAC and serving Queue Earliest Deadline First
Check BS remaining Power Budget,

Open Loop on DPCCH, Check
Check BS remaining Power
Budget, Quality of higher class
8···8
for growing access delays service, check Link Quality,
Figure 1: Service-handling policies
To admit interactive users, waiting in a CAC queue, the system checks the
associated control channel (DPCCH) power requirements and the mean
transfer delays of the on going connections. Waiting users are not transferred
to the scheduling queue if the delays grow excessively. The average delays
are used as an indirect way to sense high traffic loads and prevent new
requests from entering the serving queue. Upon admission, the users are
transferred to the scheduling queue and allocated a dedicated control channel
for signaling and control purposes.
As indicated, the scheduling queue monitors message delays to assist
CAe. Active connections, in the scheduling queue, are served according to
the Earliest Deadline First policy. The interactive class users are scheduled
only if their link quality is acceptable and higher priority services link quality
is respected as well. By giving precedence to high priority real time users,
only the remaining radio resources are allocated to non real time packet users.
In downlink WCDMA, interference greatly depends on user position.
When a base station operates at low load or the mobile station channel is in
good condition, smaller processing gain and higher transmit power can be
applied to the interactive users. On the contrary, if the base station is
operating at high load or the mobile station experiences poor radio link
quality, the base station decreases the rates of interactive users to stabilize the
system.
66
BEGIN
Prioritize Users on the basis

of P min and EDF Policy
Select Next User j
P minji : Required power at minimum rate for mobile j and service class i
P max,i : Maximum allowed power for mobiles of service class i
R min : Minimum rate for NRT service class i = 32 Kbps
R max,i : Maximum rate for NRT service class i = 256 Kbps
Figure 2: Data rate allocation flow chart in scheduling cycle
Figure 2 shows the flowchart for power and rate adaptation over a
scheduling cycle (a frame). Users are first sorted according to channel
condition between mobile and base station. The channel state information is
provided by the associated DCCHs. The base station can also estimate the
DSCH power level through the associated DCCHs power levels when
deciding transmission rates for mobile stations on DSCHs. Users are then
selected according to the Earliest Deadline First priority policy. The deadline
for each user is calculated according to minimum acceptable throughput and
packet size, assuming all NRT users have the same maximum transmission
rate capability.
NRT users are only scheduled for downlink packets transmissions if their
required transmission power at minimum allowed rate is less than the power
budget for each user. In addition, the base station is not operating at maximum
threshold power. Starting from the maximum allowed rate, transmission rate
67
is decreased to the next lower level if the required mobile transmission power
is greater than the maximum mobile power budget limits for both mobile and
base. In other words, each request is first checked for feasible allocation at
minimum rate. Once this first test is passed, the other rates are checked
sequentially starting for the highest possible rate. The algorithm can be
improved of course but the paper objective is to simply assess the benefits of
rate adaptation combined with CAC and scheduling.
Once the scheduler has decided about user rate and power, the availability
of the PDSCH is checked. If a shared channel of corresponding rate is
available, the mobile gets the reservation for transmission of a Service PDU.
Otherwise, the mobile's packets simply wait in the scheduling queue for the
next scheduling instant. In this way mobiles are segregated on shared
channels of different rates depending upon their locations and channel
conditions.
Rate matching is achieved by mapping NRT users on PDSCHs giving bit
rate of 32, 64, 128 and 256 Kbps at RLC payload. Packets are segmented into
fixed size transport block (RLC PDU) of 320 bits (uncoded). The scheduler
determines the data rate and the number of transport block (RLC PDU) to be
transmitted according to the transmission rate. Therefore, rate adaptation
results in transmission of 1,2,4 or 8 RLC PDUs within a frame.
3. SIMULATION MODEL
Simulation parameters used in an event driven and dynamic simulation,

using OPNET as the network modeling tool, are provided in Table 1. RT
services are represented by a CBR flow with 100% activity factor [5]. A
WWW model described in [5] is used for the interactive service class.
Estimation of the benefits of scheduling combined with rate and power
adaptation for the WCDMA radio network is conducted for a multiple cell
environment consisting of small macro cells [5]. For NRT users, the quality
measures used in this investigation are the percentage of call blocking, the
proportion of satisfied users, system throughput in KbpsIMHz/cell,
normalized SPDU delay in sec/Kbytes and base station and traffic channel
transmission power in dBm. For NRT services user satisfaction corresponds
to throughputs not falling below 10% of the maximum possible service rate.
In addition, the user does not experience session dropping during handoff.
Performance metrics for the R T services at 64 Kbps are the percentage of
users blocked and the percentage of satisfied users. User satisfaction for RT
service users entails two conditions: a) the user has sufficiently good quality
more than 95 % of the session time and b) the user is not dropped due to poor
quality over a continuous time duration, see [5].
68
Table 1: Simulation Parameters
SimulatJo~ Parameter Value Unit

Radio access WCDMA (FDD Downlink)
Chip rate 3.84 Mcps
Hexagonal with omni-
Deployment scheme
directional antennas
Cell radius 500 meters
User speed 0-60 Kmfh
Distance loss exponent 4
Mean: 0
Log normal shadowing dB
Standard deviation: 10
Soft handover margin 3 dB
Max. active set size Real Time: 2
Non Real Time: I
Max BS transmission power 43 dBm
Common channels + Voice service
33 dBm
power
Max. transmit power per traffic
30 dBm
channel
Power control range 25 dB
Power control step size I (400 Hz) dB
Orthogonality factor 'Y 0.4
Scheduling cycle 10 msec
Dedicated channel rate (Inf. bits) 64 (conversational) Kbps
Shared channels rate (Inf. bits) 32,64,128,256
Service activity factor Conversational: 100%
EblNo target RT 64 Kbps service 2.5 dB
256 and 128 Kbps: 2.0
EblNo targets for NRT service 64 Kbps: 2.5 dB
32 Kbps: 3.0
4. SIMULATION RESULTS
Simulation results are reported in Fig. 3 through Fig. 11. Fixed rate
scheduling is used for comparison and serves as a reference. Higher priority is
assigned to the RT type service at 64 Kbps behaving like a conversational
class. The RT service load is held constant at 4 Erlangs. RT traffic flows are
transmitted on dedicated channels while interactive traffic (NRT flows) use
shared channels. For fixed rate scheduling, the two 256 Kbps shared channels
are used. The DSCH code tree allows scheduling of a single user at high bit
rate or several lower bit rate users through code multiplexing.
For rate adaptive scheduling one of the two 256 Kbps branch can be set
aside to provide one 128 Kbps, one 64 Kbps and two 32 Kbps channels. In
this way, these schemes are using the same amount of code tree resources and
can be compared on a fair basis.
69
Fig. 3 through Fig. 9 depict the achieved performance for the interactive
service class. Looking jointly at user blocking, user satisfaction and system
throughput in KbpslMHz/cell for packet users, rate adaptation performs much
better than fixed rate scheduling across all traffic loads. Rate adaptation
achieves more than 90% user satisfaction and less than 1% blocking even at
high loads. Without rate adaptation the users satisfaction degrades to 60%
even if call blocking remains below 1%.
2.0
1.5
1.0
V
Without Rate Adaptation
w".~.~'.~
0.5
\
.. 0.0 ~_~
,,::;;;;;;;~J.;;;;;;;;;:~r.E::=;:::----'==2.~_-JI
I
0.2 O.~ 0.4 0.5
Arrival Rate I Second
Figure 3: % users blocked vs. arrival rate [s-l] of 256 Kbps Interactive Service
110
100
90
80
70
60
50
40r---------~----------~----------~
0.2 0.3 0.4 0.5
Figure 4: % users satisfied vs. arrival rate [s-I] of 256 Kbps Interactive Service
70
30
5r---------~----------~--------~
0.2 0.3 0.4 0.5
Arrival Rate! Second
Figure 5: System throughput [Kbps/MHzlcellJ of256 Kbps Interactive Service
Fig. 6 and Fig. 7 depict the average user throughput in Kbps during the
entire session and the normalized Service PDU (SPDU) transmission delay in
seclKbytes. SPDU delay includes queuing, transmission and retransmission
delays.
200
100
With Rate Adaptation
\
50r---------~----------~----------~
0.2 0.3 0.4 0.5

Arrival Rate! Second
Figure 6: Session average throughput [Kbps J of 256 Kbps Interactive Service

71
.17
.16
.15
.14
.13
.12
.11
.10
.09
.06
.07
.06 \ - - - - - - ; - - - - - - - - , - - - - - ,
0.2 0.3 0.4 0.5
Figure 7: SPDU normalized delay [sec/Kby1es] of256 Kbps Interactive Service
Obviously reducing the rate during rate adaptation increases the

transmission delays and reduces the throughput for users having bad channel
conditions. This expected performance degradation seems reasonable for the
analyzed scenario. The average delay per SPDU increases slightly and
remains acceptable for the interactive service class. The achieved average bit
rates as shown in Fig. 6 are very stable with rate adaptation.
Fixed rate degrades sharply at high traffic loads. The available codes at
256 Kbps can not be used under poor link quality and users must wait for
better radio conditions to enter the system. This results in very inefficient use
of code tree resources, increased delays and blocking at high load.
Base station transmission power [dBm] and traffic channel power, reported
in Fig. 8 and Fig. 9, confirm the benefits of using rate adaptive scheduling.
For fixed rate transmission, 50% of the users operate at maximum power.
Rate adaptation combined with scheduling provides much better stability.
Traffic channel power levels remain strictly below the maximum power limit.
There is even some room left in the power budget.
Fig. 10 and Fig. 11 depict the achieved performance for RT conversational
class in terms of percentage of blocked calls and user satisfaction. Blocking
for RT services with fixed rate transmission is at 3% while rate adaptation
achieves blocking rates lower than 1%. User satisfaction is below 70 % unless
rate adaptation is used to improve the performance to 80%.
72
1.0
With Rate Adaptation
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
RT 4 Erlangs
0.1
0.0
36 38 40 42 44
BS Tx Power [dBm[
Figure 8: CDF of base station Tx power [dBm)
1.0
0.9 With Rate Adaptation
0.8
0.7
0.6
\
0.5
0.4
0.:3
0.2
RT 4 Erlangs
0.1
NRT 0.4 Arrivalslsec
O.O\-.:w.....c=::!:::.----r--------l
10 20 :30
NRT Users Traffic Channel Tx Power [dBm[
Figure 9: CDF of traffic channel power [dBm) ofNRT Interactive class users
73
6
RT Load 4 Erlangs
4
-1r----------.----------.---------~
0.2 0.3 0.4 0.5
Arrival Rate I Second of Interactive Users
Figure 10: % users blocked for 64 Kbps Conversational Service
100
90r--:::r--t--......._
80
70
RT Load 4 Erlangs
60r----------.----------.---------~
0.2 0.3 0.4 0.5
Arrival Rate I Second of Interactive Users
Figure 11: % users satisfied of 64 Kbps Conversational Service

74
5. CONCLUSIONS
Results confirm that high rate users in bad channel conditions cause
increased interference and degrade performance. The introduction of rate
adaptation scheduling for NR T users leads to efficient use of radio resources
and better system stability. By giving priority to the RT conversational class,
only the remaining radio resources are allocated to NRT interactive class data
users. Rate adaptation scheduling results in better QoS for both service classes
and can accommodate more users by allowing bad channel NRT users to
transmit at lower rates. Combining rate adaptation scheduling for delay
tolerant services, prioritized CAC and power control is a promising path to
improve system performance and provide higher capacity for WCDMA
UMTS networks.
REFERENCES
[I] UMTS; QoS concept and architecture, 3GPP TS 23.107
[2] Gyung-Ho Hwang and Dong-Ho Cho, "Dynamic rate control based on interference and
transmission power in 3GPP WCDMA system", IEEE VTC'2000 Fall, vol.6, pp. 2926-
2931
[3] Takumi ITO, Seiichi Sampei and Norihiko Morinaga, "Adaptive transmission rate control
scheme for ABR services in the CBR and ABR services integrated DS/CDMA systems",
IEEE VTC'2000 Fall, vol. 5, pp. 2121-2125
[4] UMTS; Physical channels and mapping of transport channels onto physical channels
(FDD), 3GPP TS 25.21 1 V 3.4.0 Release 1999
[5] UMTS; "Selection procedures for the choice of radio transmission technologies of the
UMTS", (UMTS 30.03 V 3.2.0)
[6] A. lera, S. Marano, and A. Molinaro, "Call level and burst level priorities for effective
management of multimedia services in UMTS", IEEE INFOCOM '96, vol. 3, pp. 1363-
1370
Capacity And CII Performance Of Different Cell
Clusters In A Cellular Network
Anis Masmoudi 1,2, Sami Tabbane2 , Senior Member, IEEE

JCentre d 'Etudes et de Recherche des Telecommunications (CERT)
Anis.Masmoudi@cert.mincom.tn
2 Unite de recherche en Technologies de I 'Information et de la Communication (UTIC)
Ecole Superieure des Communications de Tunis (Sup 'Com)
Sami. Tabbane@supcom.rnu.tn
Abstract: Capacity and interference performance are among the most important issues in
the cellular frequency planning process. The main objective is to reach a
tradeoff between the quality and the offered traffic. In this paper, we study and
compare the CIR and spectral efficiency of different reuse patterns. We also
establish PDF expression of CII assuming one interferer in the serving cell.
The spectral efficiency and trunking efficiency are expressed analytically
versus cluster size, number of sectors and reuse distance. In particular, reuse
partitioning sub-cells sizes are optimized in order to maximize the traffic
capacity parameterized by spectral efficiency; and it is shown that more than
one partition wiI1 decrease offered traffic per area unit, per Hz.
Key words: Cellular Reuse Patterns, Interference, Quality Of Service, Capacity, Offered
Traffic, Spectral Efficiency, Trunking Efficiency, eIR Distribution Function,
Cluster Size, Reuse Distance, Reuse Partitioning (RP).
1. INTRODUCTION
Because of the limited available frequency bandwidth, cellular radio networks
adopt the frequency reuse concept to reuse the same frequency at different locations.
A large reuse distance can enhance the ell level by reducing the interference level
but offers a poor traffic capacity. One challenge for cell engineering is to reach a
tradeoff among channel quality, system capacity, and the costs of infrastructure and
user terminals. Previous works have been focused on frequency assignment
algorithms evaluation and comparison. In this paper, many already existing pattern
75
76
types presenting particular frequency plans are studied and simulated considering
different reuse factors, antenna directivity and sectorisation types. We also consider
some particular cluster plans such as reuse partitioning (RP) and Fractional
Frequency Reuse.
The remaining of this paper is organized as follows: The next section compares
these clusters in terms of eIR performance on the basis of histograms plotted for the
same clusters families. An analytical PDF expression of ell is given in Section 3
assuming one interferer in the serving cell. In Section 4, the spectral efficiency and
the trunking efficiency are both calculated and expressed analytically. SeGtion 5
presents some cluster types examples and especially RP. We exhibit the best cluster
type that realizes the compromise between quality performance and traffic capacity.
Finally, our conclusions are summarized in Section 6.
2. ell SIMULATION RESULTS

All following simulations are performed in a GSM network. The purpose is to
exhibit the performance of different clusters. For this, we calculate the ell for the
whole coverage area of a cell divided in 30.000 square meshes per cell or sector. The
next step is to compute numerical results and statistics, and to plot ell histograms
considering both the first and second tiers of interferers. The front-to-back ratio of
different antennas is supposed infinite for all sectored clusters.
Figure 1 shows plots of eIR histograms for simple clusters with omnidirectional
antennas for different cluster size values (N = 3,4, 7, 9 and 12). Besides, two N = 7
clusters histograms have been depicted (The classical and the trapezoidal cluster
represented respectively in Figures 2 and 3). Each histogram shows a maximum and
the curves don't appear to be as smooth for the lowell values as for the high values.
CIR histograms of simple clusters With omnidirectional antennas
100
E
::>
E 90 , 0000 : K=3
'x
E
OJ
"'~
(J) 80
," ,
,, ' . lOOOO< : K=4
. K=7 simple
::> ________ . : K=7 trapezoidal plan
u
Q) 70 ++++ : K=9
-5
.9 . K=12
2-
Q)
60
>
"m
~ 50
#'
.!: 40
"'
Q)
.c
"'
Q)
E
30
'0
Q)
0)
20
1ll
(J)
~
10
Q)
0..
0
0 10 20 50 60
Figure 1. ell histograms for omnidirectional clusters (Simulation).

77
In fact, the right side of each curve is unlimited whereas the one on the left has a
smaller boundary since we have calculated and represented C/I only within the cell
borders. Concerning the arrangement of the curves, we notice that, for the same
antenna type, the greater the cluster size K, the more the curve position is towards
the higher C/I values. So, C/I increases with cluster size. On the other hand, the
trapezoidal cluster presents the highest CIR values compared to all other patterns
because of its special configuration plotted in Figure 3 presenting only two potential
interferers versus six for the other cluster types. Only horizontal expansion is
possible in this latter plan; for this reason it is suitable for highways and coastal
areas [1].
6 effective interferers
Figure 2. A classical 7-sized cluster with omnidirectional antenna
Figure 3. Trapezoidal frequency plan with a 7-sized cluster
Figure 4 depicts the sectorisation effect through the histograms of clusters with a
size equal to 3. We observe that the more the sectors of a site, the better the quality
for the same reuse factor. In fact, the 6-sectored cluster is more performing than the
tri-sectored ones that also present better CII than the simple omnidirectional cluster.
78
This is, mainly, due to the fact that sectorisation reduces the number of effective
interferers to 2 in the case of tri-sectored patterns and I for 6-sectored ones. Here,
the cyclic and alternate frequency assignments have comparable performance.
CIR histograms of K=3 patterns

100
E
::J
E
'x 90 0000 : simple cluster with omnidirectional antenna
'"E )OOQO( : Tricellular plan with alternate channel assign ent
(f)
80 ++++ : Tricellular plan with cyclic channel assignme
~
::J ...... : K=3/18 (6x600)
U
(]) 70
£;
2
~
ill
60
>
.~
~ 50
#
.~
(f)
40
(])
.<::
(f)
(]) 30
E
'0
(])
0\
20
~ill
~
10
(])
D-
O
0 10 20 30 40 50 60
cn in dB
Figure 4. C/I histograms of clusters with a size 3 (Simulation).
Table I presents numeric results for some clusters. We can conclude that the
cyclic channel assignment provides a better ell quality than the alternate assignment
with a value of more than 4 dB for the cluster 4112. This difference is due to the fact
that, for the alternate assignment case, different cells antenna beams are interfering
by 60° each from the other inducing more interference than the cyclic assignment
cluster where co-channel cell beams are pointing to the back of each other for which
secondary lobes imply less interference.
r.a hie 1 S'Imu IalIon

' numenc resu ts 0 fC/I perfIonnance
Threshold Y 195% of
Cluster Cluster size Mean CII
meshes have CII > Y dB
Sectored (alternate assignment) 4/12 (3x600) 24.62 dB 17.68 dB
Sectored (cyclic assignment) 4/12 (3x600) 30.94 dB 22.04 dB
4112 (3x1200):
Sectored 23.35 dB 16.07 dB
1st model
4112 (120°):
Sectored 26.38 dB 18.70dB
2nd model
Another crucial factor for ell quality is the number of effective interferers. It
appears in the case of two 4/12 clusters that have the same reuse distance but the
fIrst has six potential interferers and the second has only four, which increases ell
by more than 2.5 dB than the fIrst model.
79
3. C/I ANALYTICAL DISTRIBUTION MODEL
In the previous section, we have plotted C/I histograms by simulation. Now, we

will model this law analytically.
Figure 5 shows the general scheme of an omnidirectional serving cell with
several interfering BTSs. The signal power received by MS is
c= A~. (1)
rY ,
with PT is the transmitted power, A a multiplicative constant and r the path-loss

exponent. The total interference assuming M interferers, is
(2)
BTS j
Figure 5. A cell with an MS and an interfering BTS i.
To simplify our study, we suppose the presence of only one interferer (M = 1)

placed at a distance D = J3K.R from the target cell; and we assume that the axis
linking both BTSs is horizontal. After rewriting d l expression, the CIR ratio
expression becomes:
(3)
r
The CDF F is then given by
F(s) = ~r r dr da. (4)

ell <s
Starting with C I 1< s expression, then distinguishing two cases for s: higher or
lower than (3K + 213K
+ 1)112 and applying some integration techni3ues such as
changing successively a variable into u = ...Jcosza + s2Jy - 1, u into x 2 = i y - u2, then x
into rp defmed as x = slly cos rp, yields to the (C/I)dB PDF expression p(x) as follows:
80
p(x) =A(z) - B(z);Vx E]l OyloglJ3K -1),lOylogoC.!3K + 1)],
p(x) =A(z) + B(z);Vx E]lO( log 0 (.!3K + 1),10( logo(& + 1)],

z +1
p(x) = -2az r::;:;;
3 ;VX > lOylogIO(v3K + 1), (5)
y~ (z -1)
with:
J(z) =2 z (11 if -2£(z+l»~ (z+1) ArICO (a+1-z)2 )
(6)
1lYi;J4a-(a+1-z)2 (Z_l)2 Jr(Z_l)3 z 4az '
B( z ) -- a(z+1)3 A rccos (1- (a+1-z)2) , (7)

Jr(z -1) 4a
2x 10
z = exp(-),a = 3K, ~ = - - and X dB = ~ Ln(s).
y~ Ln(10)
The analytical expression established assuming one interferer of the serving cell
is constituted by three parts, each has its own expression and all make up a global
continue curve. We note also that each bound of the three expressions increase with
the cluster size K confirming a previous result in Section 2, Figure 1.
Figure 6 includes both simulated and theoretical histograms for a 7-sized cluster.
The two histograms have the same shape with sharp vertices on their top (not
derivable but continue). The simulated histogram is moved towards lower ell values
due to the consideration of six potential interferers in the first tier and 12 other
interferers in the second one versus only one in the analytical histogram. The
difference between both curves is almost constant equal to about 13 dB since the
interference is approximately the same from each interferer. So, we can approximate
the real simulated expression by the analytical one assuming only one interferer to
which we add the 13 dB constant. I;ksides, the theoretical histogram maximum is
shown to be 10ylogIOh/3K +1).
C/I distribution law for a 7-size cluster (Simulation + Ana~c81 expression)
~ ::: .,t.~.,.,
eoo 0.07 .
_ en probability density (PDF)
Simulation of CIl distribLltion law

o
~ 0.06
LL
·fO.05
.
!
~
~004
.B ;
~ 0.03 :
0: :
0.02 :
0.Q1
15 20 w ~ ~ ~ ~ ~ ~
C/ljndB
Figure 6. Comparison between analytical and simulated CII histograms.

81
If log-normal shadowing is considered, no noticeable modification of the

histogram is observed when simulations have been performed to some pattern types.
In fact, the addition of a random variable having a mean value equal to 0 dB doesn't
introduce considerable alteration of the CIR distribution statistics represented by an
histogram. Furthermore, the propagation conditions variation is sometimes favorable
(in Line Of Sight), sometimes unfavorable because of the presence of obstacles,
which provides an unchanged global distribution.
4. CLUSTER PARAMETERS IMPACT ON CAPACITY

The spectral efficiency as defined in [2] is given by
E=_T_ (8)
B.K.A'
where T denotes a given traffic volume in a given area A, B is the bandwidth
corresponding to a single cell, and K is the cluster size. The traffic law to be
considered in the calculation is based on the Erlang B formula.
Assuming Q the total number of channels available in the network and K the
cluster size, the number of channels per cell is QIK and the traffic per cell is
T = E1,Q' K -I (PB ), (9)
where E1,n-I(PB) denotes the inverse Erlang B formula providing traffic obtained
with n active channels and a blocking probability PB.
Here, A is the cell area equal to lCR 2 , and B.K is the total bandwidth BIOI of the
system. Finally E is written as
E = EI'QIK-IC~B). (10)
BtottrR
By extracting the reuse distance D from its approximation T3K .R valid for
regular clusters, (10) becomes
3fp (K)
E= 8 2' (11)
BtotlrD
where fpo is the characteristic function defined by
!p.(n) = nE1,Qln-1CPB). (12)
Its variation is shown to be decreasing ; since the spectral efficiency E is
proportional to fp (K) , so it decreases with the increase of the cluster size K and is
inversely proportional to the reuse distance squared.
The trunking efficiency expression is
E -I(p')
TE = 100 I,QIK B = 1001; (K). (13)
QIK Q p.
Hence, we see that at constant reuse distance, the trunking efficiency is

proportional to the spectral efficiency, and follows, consequently, the same variation
versus cluster size.
82
At constant reuse distance and fixed total channel number (bandwidth), a pattern
having h sectors has spectral and trunking efficiency proportional to fPR(Kh) ; thus
the capacity decreases when we add sectors with the same cluster size. This result
differs from the one presented in [2] that describes the sectorisation impact on
capacity, treats the uniform traffic case (Constant traffic density), where spectral
efficiency increases with sectorisation for large and medium cells at the same QoS
in terms of CIR value.
The capacity decrease with the cluster size shows a compromise to assume
between interference minimization and network capacity maximization.
5. CLUSTER TYPES EXAMPLES
5.1 RP Quality Performance

In this section, the first example is the Reuse Partitioning pattern. TIlls cluster meets the
C/I objective since 95% of its meshes have a value ofC/I > 14.40 dB.
For the internal region of the cell, the small reuse distance is compensated by the
fact that the mobiles receive a more important signal power because of their
proximity of the base station.
Figure 7 plots the C/I distribution along the cell in a RP cluster showing quality
improvement in the peripheral area.
1QO Different frequencies

130
CD
-0
eo
,5
i5 40
20
0
0.5
OnjiMtes axis (Km)

Abscissas ilns (Kml
Figure 7. erR distribution along RP cell.
83
5.2 Partitioning Impact On Trunking Efficiency

Concerning capacity, RP has a better spectral efficiency than the "Fractional
Frequency Reuse" cluster. It is characterized by a higher trunking efficiency value.
This capacity increase is due to the reuse of cluster with smaller size in the core.
Besides, comparing RP to an ordinary cluster, computation and simulation show that
it enhances capacity but reduces trunking efficiency from 3 to 8%.
Figure 8 shows the trunking efficiency difference variation between a RP and an
ordinary cluster for a total number of channels Q = 416 and for a cluster size varying
from 2 to 6. The reduction is shown to reach its maximum 7.79% for the cluster size
10. In the RP cluster, the radio spectrum must be divided into sub-parts. The number
of channels available per cell partition is smaller than without partitioning, thus
reducing the trunking efficiency [3].
Trunking efficiency difference between an ordinary cluster and a Reuse Partitioning cluster
-3~---,----.----,---,.---,----,----,----,----,----,
-3.5
Q)
(J) -4
co
~
~ -4.5
9-
Q)
u
c -5
Q)
Q5
~ -5.5
>-
u
c
Q)
'u -6
t
Q)
g> -6.5
:2
c
~ -7
-7.5
-8
2 3 4 5 6 7 8 9 10 11 12
Cluster size of the ring
Figure 8. Trunking efficiency reduction effect ofRP cluster.
5.3 Optimization Of RP Sub-Cells Dimensions

Reference [4] states that the optimal bandwidth enhancement in the RP pattern
with one partition is realized when the different sub-cells have equal areas. Our
objective is to determine sub-cells dimensions so as to maximize traffic
improvement. For two partitions, we find the same optimal result as in [4].
However, the result is shown to be different for the 3-partition case. In fact, for the
last case, the calculation of the spectral efficiency gain leads to the (3D)-curve in
84
Figure 9 We note the existence of negative values of the curve especially for Km
:$ 0.53 K where the ordinary cluster is better than RP. This performance reduction
J
reaches its maximum when Km = 0.24 K J and Ks = 0.5 Km = 0.12 K J with a 10-2
approximation where the capacity loss is 19.89%. However, the spectral efficiency
gain is maximum when Km = KJ and Ks = 0.5 Km = 0.5 K J where we have a 29.33
gain. The addition of more than one partition decreases spectral efficiency
performance. This result differs from results obtained in [4] where the bandwidth
increases indefinitely as long as we add partitions. This is due to the fact that for the
case of the addition of more than one partition, the bandwidth gain is more than
compensated by the trunking efficiency loss described in Sub-Section 5.2.
Percentage of spectral efficiency increase for "Reuse Partitioning" in relation to ordinary cluster
30
.s
Q)
II)
20
(II
Q)
g 10
(;'
c:
3 0
~
~
u
-10
'"a.
(f) -20
1
Core surface I middle surface

Middle surface I total surface
Figure 9. Spectral efficiency gain variation of RP cluster versus sub-cells relative dimensions.
5.4 Best Cluster Type Choice

The "Fractional Frequency Reuse" cluster can be considered a best one in terms
of tradeoff between interference performance and capacity supply due to its
particular properties. In fact, it has 95% percentage of the cell area having a ell
> 18.47 dB and 100% of the cell whose ell> 12 dB. Thus, this cluster type offers an
acceptable ell quality. This good performance is due to its sectored site and the
reuse plan based on a back-to-back reuse scheme forming the apex of a triangle and
having more groups of frequencies. Moreover, its spectral efficiency is superior than
the tri-sectored cluster one with K = 7 by 13.55% and the trunking efficiency value
is superior by 8.94%, which affirms partially the result expressed in [5] of 30% more
capacity than tri-sectored with K = 7 without loss of ell performance.
85
However, the last cluster is better than the other one by about 3 dB due to its
bigger reuse distance. In fact, the tri-sectored cluster has a 7-sized cluster, which is
bigger than that of Fractional Frequency Reuse (5.333).
6. CONCLUSIONS AND FUTURE WORKS

In this paper, we have evaluated the capacity and quality perfonnance of different
pattern types in a cellular network by both simulation and analytical study. The quality has
been expressed in tenn of CIR ratio, whereas spectral efficiency and trunking efficiency
have represented capacity.
The comparison of some clusters samples have led to choose Fractional Frequency
Reuse among those which give the best tradeoffbetween interference and traffic supply.
However, each cluster type is suitable for a specific environment and could be
combined with others.
Our work is limited mainly to TDMAlFDMA systems, but the results may be useful
even for CDMA systems since they require a Pseudo-Noise (PN) sequence Offset planning
instead of frequency planning in traditional ones. Moreover CDMA standards such as
WCDMA include sectored clusters and are based on hierarchical cells and different cell
ranges like RP pattern.
Further works can be focused on analyzing CII PDF distribution expression with more
than one interferer as well as for sectored pattems. Other pattern types can also be studied
in the future such as Hierarchical Cell Structures (HCS) in the fields of both quality and
capacity. Adjacent Channel Interference can be treated and studied as an additional quality
criterion to compare pattern types (by computing CIA ratio).
We will also look at the clusters configurations in TDMA third generation systems
such as IS-136 in USA, and try to generalize analytical expressions for irregular clusters
using more realistic hypothesis.
REFERENCES
[1] S. Faruque, "Cellular Mobile Systems Engineering," Artech House, Boston,

Mass., USA, Ch. 7, pp. 199-245, 1996.
[2] Rupert Rheinschmitt, and Michael Tangemann, "Performance of Sectored
Spatial Multiplex Systems," Vehicular IEEE Transactions On Communications
(VTC), Atlanta (Georgia), April 28 - May 1, 1996.
[3] Ulrich Dropmann, "Allocation des ressources dans des systemes radiomobiles a
reservation par paquets," ENST thesis, Paris, 1996.
[4] D. Lucatti, A Pattavina, and V. Trecordi, "Bounds and Performance of Reuse
Partitioning in Cellular Networks," International Journal of Wireless
Information Networks, vol. 4, N° 2, pp. 125-134, 1997.
[5] Saleh Faruque, "High Capacity Cell Planning Based on Fractional Frequency
Reuse With Optimum Trunking Efficiency," VTC'98, IEEE, pp. 1458-1460,
1998.
PERFORMANCE STUDY
OF SOFT HAND OVER
WITH CDMA HETEROGENEOUS
CELL ULAR ARCHITECTURES
Li-Chun Wang
Ching-Yu Liao
and Chung-J u Chang
Department of Communication Engineering,
National Chiao Tung University
Tel: +886-3-5712121 ext. 54511
Email: {lichun.cjchang}@cc.nctu.edu.tw.cyliao.cm86g@nctu.edu.tw
Abstract This paper investigates the downlink user capacity of a heterogeneous

cellular CDMA system with soft handover. We consider the scenario
that a hotspot microcell is adjacent to a larger macro cell instead of a
micro cell embedded within a macrocell. Since traditional soft hand-
off algorithms are developed for homogeneous cell structures, the two
serving base stations allocate equal power in the downlink to the soft
handover users [1]. We observe that this kind of equal power allocation
will cause a serious "power exhausting problem" in the microcell when
soft handover is occurred between two cells with different cell radius.
To quantize the impact of this problem, we present an analytic ap-
proximation method for computing the downlink user capacity with soft
handover in heterogeneous cellular structures. We further propose an
improved quality balancing power allocation method with maximum
power constraint to enhance the downlink soft handover performance
when two involving handover base stations have different cell sizes. The
simulation results show that the improved power allocation for soft han-
dover can support a larger range of the ratios of the cell radius between
microcell to macrocell, thus increasing the total system capacity in the
sense that more micro cells can be installed adjacent to a macrocells.
Our numerical results demonstrate that in a heterogeneous CDMA net-
work with the cell radius ratio between microcell and macrocell equal to
0.3, the proposed constrained unequal power allocation (UPA) technique
with soft handover can enhance the total system capacity eight times
higher than the conventional equal power allocation (EPA) method.
Keywords: Heterogeneous cellular, soft handover, forward link, power allocation
87
Multiaccess, Mobility and Teletrafficfor Wireless Communications, Volume 6,87-102.
88
1. Introduction
Soft handover is one of the most important merits of the code division
multiple access (CDMA) cellular system. For user terminals in the soft
handoff process, the original serving base station and the target base
station will maintain two communications links simultaneously over the
same bandwidth to guarantee a smooth transition without the chance
of dropping the ongoing call. Traditional soft handoff algorithms are
developed for homogeneous cell structures, i.e, the involved base stations
have the same cell size. In practice, however, cell coverage area of each
base station in the existing cellular network differs a lot from each other.
First, due to the coverage problem, a small micro cell may be installed
at the boundary of surrounding macro cell base stations. Second, for
increasing system capacity, a cluster of micro cells may be employed by
cell splitting or other techniques. Thus, a heterogeneous cellular network
will be naturally occurred as shown in Figure 1.
Figure 1. The heterogeneous cellular model

To our best knowledge, a complete performance analysis of downlink
soft handover for heterogeneous cellular architecture is not existing in
the literature. The key challenge here is to incorporate all the effects of
path loss, shadowing, multiple access interference, the combining scheme
in the receiver, and the power allocation algorithms in the transmitter
as well. The primary goal of this paper is to develop an analytical
methodology to compute the downlink user capacity in a heterogeneous
cellular network. Next, we will discuss an improved power allocation
algorithms in the transmitter to enhance the soft handover performance
in a heterogeneous cellular network on the downlink.
Traditionally, based on homogeneous cell structures, the soft handoff
algorithms will ask the two serving base stations to allocate equal power
in the downlink to the user terminals in the handoff region [1], this
is because that the maximum soft handover performance gain can be
achieved in the location with equal path loss from the base stations in
the active set[2, 3]. Although the equal power allocation (EPA) method
is effective for downlink handover in homogeneous cellular structure, it
would cause a serious "power exhausting problem" in the microcell when
89
soft handover is occurred between two cells with different cell radius.
According to the EPA method, the handoff mobile terminal originally
in the macro cell will request the same amount of power to be allocated
from both macro cell base station and micro cell base station. Intuitively,
it is easy to see that the handoff users form macrocell will be very likely
to exhaust most of the power budget in the micro cell which usually
has a small amount of total transmission power, thereby decreasing the
regular user capacity of the microcell. However, to what extent the ratio
of cell radius between two serving base stations affects the effectiveness
of the power allocation method in the performance gain of downlink soft
handover is an open issue.
To overcome the power exhausting problem of the EPA method in
downlink soft handover within heterogeneous cellular networks, we pro-
pose a new quality balancing algorithm with power constraints for allo-
cating the base station transmission power to serve each user in the both
macrocell and microcell. We call the new power allocation algorithm
for soft handover the constrained unequal power allocation with (UPA)
method in this paper since the allocated power form the serving base
station will be different. The constrained UPA method is the modified
version of [4]. Unlike [4] suitable for only a homogeneous cell structure,
the proposed constrained UPA method is applicable for the heteroge-
neous cell structures. More importantly, we add a new criterion to limit
the maximum downlink transmission power allocated from base stations
to each handoff user. The new criterion is based upon the link budget
analysis to calculate the transmission power to achieve the required SIR
requirement at the cell boundary. According to the constrained UPA
method, all the allocated power from the base station for any handoff
request should be constrained below this limit. Our numerical results
will demonstrate that the concept of limiting maximum downlink trans-
mission power is very critical to avoid the power exhausting problem
occurred in the micro cell especially when the cell radius of the micro cell
is less than 50% of that of the macrocell.
The literature survey of the previous work related to downlink soft
handover for heterogeneous network can be classified into three cate-
gories. First, the CDMA downlink soft handover issue was first exam-
ined in [1], where the impact of downlink soft handover was discussed
for a homogeneous cell structure. In [1], it is mentioned that the EPA
method is effective for a mobile station during soft handover, but the
detailed algorithm and downlink capacity with soft handover is not ana-
lyzed. The effect of soft and softer handovers on the downlink capacity of
homogeneous CDMA was discussed in [5], but without addressing the is-
sue of power control. Secondly, as for the downlink power control issues,
many downlink power control algorithms for balancing link quality have
been developed in both centralized version and distributed version [6]
90
and [4]. The concept of quality balancing power control is to let all the
mobile stations in the cell maintain equal link quality. On the contrary,
the conventional method will allocate extra transmission power, result-
ing in extra interference. Few downlink power control works have been
published in the context of heterogeneous cellular networks. At last, for
the mixed cellular architecture, many research works have be published
to analyze the performance of a hotspot micro cell embedded within a
larger macro cell [7] and [10]. Few work except [11] has been published
to discuss the performance for the mixed cellular architecture as shown
in Figure 1. However, in [11], only the reverse link capacity without soft
handover in a CDMA cellular system with power control with mixed cell
sizes was analyzed. Thus, to our knowledge, the performance analysis
of downlink soft handoff with power control for heterogeneous cellular
networks is still an open research area.
As for the uplink performance analysis for heterogeneous CDMA net-
works, [12] investigated the interference issue when a mobile terminal
connecting the macrocell moves toward a mobile terminal connecting
the microcell. Although the reverse link is considered to be a limiting
factor for the CDMA system capacity, the forward-link performance is
becoming increasingly important due to the emergence of asymmetric
wireless data services.
In summary, this paper evaluates the capacity of a heterogeneous cel-
lular network with a hotspot micro cell adjacent to a larger macrocell. We
consider the effects of path loss, shadowing, multiple access interference,
downlink power allocation, soft handover in our performance evaluation.
Furthermore, we present a new quality balancing downlink power allo-
cation method with maximum power constraint, the constrained UPA
method, which can solve the power exhausting problem in the micro cell
when executing soft handover. Using the constrained UPA method, a
smaller micro cell can be employed without suffering the power exhaust-
ing problem, thereby increasing the total system capacity.
The remaining parts of this paper are organized as follows. Focusing
on a simplified cell model with a single micro cell and a single macrocell,
Section 2 first discusses the signal model, downlink power control, and
soft handover gain, and then presents an analytical model for comput-
ing the downlink user capacity with soft handover in the heterogeneous
cellular network. The numerical results are shown in Section 3. Section
4 will give the concluding remarks.
91
2. System Model
2.1 Signal Model
Consider a simplified heterogeneous cellular model with a single mi-
crocell adjacent to a macrocell as shown in Figure 2. Let RM and R ft ,
respectively denote the radii of the macro cell M and the micro cell /L;
r M and r ft are the distance for the user terminal at the point H to the
macro cell M and that to the micro cell /L, respectively.
,
"-
\
\
Il I
I
/
/
Figure 2. A simplified heterogeneous cellular model
Let r(qi,j) be the received bit energy-to-noise density ratio Eb/No of

mobile j with the downlink power qi,j transmitted from the base station
z. Then r(qi,j) is given by :
qi,j . Li,j . G
r(qi,j) = (P't _ q .. ) .L·· + L P k . L k · +7] ;::: 'Yreq, (1)
1,,) 't,) ,) 0
k,kfi
N
where G is the processing gain, Pi = L qi,j is the total downlink trans-
j=1
mission power for N users in cell i; Li,j is the link gain from cell i to
mobile j, 7]0 is the background noise, and 'Yreq the required Eb/No. We
include the effect of both path loss and shadowing in the link gain Li,j.
That is,
Li ,J' = L'1,,). X lO~dlO (2)
In (2), L~,j follows a two-slope path loss model as in [7].
L'
A
'"
{ dZ~-:/3
..
,if di,j > > Zi
(3)
i,j = da. (1 + (~){3) , = (~'~{3Si)
'b,) Zi
dO. Zz
, J'f di,j «Zi
where a and f3 are the path loss exponents, di,j is the distance from
mobile station j to the base station i, Zi is the break point in cell i, and
A is a constant. The standard deviation of the shadowing ~i in (2) is
92
also distance dependent [8],
,di,j ::; Zi
(4)
,di,j ~ Zi
In [7], the breakpoint Zi is given by
(5)
where hi is the antenna height for base station i, h ms the mobile antenna
height, and ,\ the wavelength. We define the cell boundary as the point
at which mobile station j receives the same power from both adjacent
cells M and f-l [9]. Then at the cell boundary, we have
(6)
For simplicity, we only consider the effect of path loss in (6). Then,
combining (3) and (6)
P L' + (!lM..)f3)
Rcx (1 R h
~M = ---1!:2i... = M ZM ex (-.!:!... t+f3 x (---1!:....)f3 (7)
PI' LM,j R~(l + (~)f3) RI' hM
Note that (7) is only valid when the micro cell radius is higher than the
break point distance. From (I), where the noise is neglected and only
micro-cell interference is considered, we have
(8)
where
(d-CXJ-L (1 + ~ )-f3J-L)
D. - I' ZJ-L (9)
) - (dA:tM (1 + ~ )-f3M)
To make macro cell users have the required Eb/ No, the maximum al-
locating transmission power qM can be obtained by substituting the
maximum total transmission power PM and PI' in (8). Then
(10)
where D j is given in (9). Note that the total transmission power of the
base station is varied depending on the summation of power allocated
for each user. The hat in qM indicates that this power level allocated
from the base station is for the user at the cell boundary. From (7) and
93
(10), the downlink maximum allocating power for the micro cell can be
obtained as
~ ~
L'M,j ( )
q/1 = qM' ~ 11
/1,)
where L~,j and L~,j are given in (3). In this paper, we adopt maximum
ratio combining for the forward link soft handover. Thus, based on [15],
the optimal received Eb/No for mobile station j during soft handover is
given by
r( qM,j, q/1,j) = r( qM,j ) + r( q/1,j) , (12)
where r( qM,j, q/1,j) denote the Eb/No after the maximum ratio com-
bining for macrocell transmitting at the power level qM,j and micro cell
transmitting at q/1,j, respectively; r( qM,j) and r( q/1,j) is the E b/ No re-
ceived from the macrocell base station and that from the micro cell base
station before combining, respectively.
2.2 Soft Handover Gain

In this section, we discuss two power allocation methods for soft han-
dover mobile stations: (i) the equal power allocation (EPA) and (ii) the
unequal power allocation (UPA). Consider a mobile h at the location H
near the cell boundary in the simplified heterogeneous cellular model as
shown in Figure 2. Assume that the mobile station h originally be served
by macro cell M and performs soft handover with adjacent microcell /-t,
denoted by M - 7 /-t.
For the EPA method, base stations in active set must transmit the
same power level. Thus, the serving base station M will allocate trans-
mission power for the mobile station h according to (7) with an upper
limit specified by (10). The transmitted power during the handover,
denoted by P~,h for macrocell ,and P~,h for micro cell, can be obtained
as
,
qM,h = q/1,h
, = "21 ~)
· ( qM,h, qM,
mm £or M -7 /-t . (13)
Note that the q~ h indicate the base station allocated power during soft
handover, and qt.d,h is that before the soft handoff. Also, it is noteworthy
that the factor of ~ is related to the number of base stations involved in
soft handoff, i.e. two base stations in our case.
For the UPA method, the two serving base stations will transmit at
different power level according to (7) and (10).
q~,h = ~min( qM,h, qM) for M -7 J1,
(14)
q~,h = ~min( q/1,h, q/1 ) for M -7 /-t
For a microcell user moving into macrocell, i.e., /-t -7 M, we can simply
swap M and /-t in (13) and (14) to obtain the transmitted power for the
macrocell and the microcell during handover.
94
The diversity gain is defined as the difference between the received

Eb/ No with handover and that without handover. For the hard handover
case, a mobile station will connect to the cell with better link gain, and
the hard handover gain Chard can be written as
Ghard(M~!,) = max{ f( qM,h)(dB). r( q!"h)(dB)} - f( qM,h) (dB)
Ghard(!,~M) = max{ f( qM,h)(dB). r( q!',h)(dB) } - f( q!"h)(dB) (15)
For the soft handover case, according to (12), the soft handover gain
Csojt can be obtained by
Gsojt(M~I') = f( qM,h )(dB) + f( q~,h )(dB) - f( qM,h )(dB) (16)

Gsojt(!'_M) = f( qM,h )(dB) + r( q~,h )(dB) - f( q!"h )(dB)
2.3 Downlink Power Control Algorithm

In this section, we extend the cell site transmitter power control al-
gorithm in [4] to the case of heterogeneous cell structures with the max-
imum power constraint. Let Ii be the set of mobile stations who are
served by cell i and Ti be the subset of mobile stations requiring trans-
mission at the level of Pi. Denote N1(i) and NT(i) as the number of
mobile stations in Ii and T i , respectively. Given the total allocating
transmission power (Pi) at base stations i, the quality balancing pro-
cedure is performed according to (7). For ease of notation, denote the
right hand side of (7) be ¢i,j, i.e.
,/.. . _ "Yreq' (PM' LM,j
,+,1,,) -
+ Pj.L· Lj.L,j) (17)
(C + "Yreq) . LM,j
Note that only the mobile terminals j E Ii - Ti will involve the quality
balancing power control, whereas the mobiles j E Ti will use the Pi. The
allocated procedures for each mobile station j in cell i is given by
_
qi,j = (Pi -
Q)
i
¢i,j
~,/, .. (18)
L..J '/",J
JEIi
jetTi
where Qi = NT(i) . qi is the total allocating transmission power in T i .

Note that calibrating procedure is necessary, because the cell-site total
power tuning are based on the power of (Pi - Qi), we should calibrate
the total allocating power as follows
P'i X "Y~eq = (Pi - Qi) x "Y~eq + Qi. (19)

"Y "Y
Thus, we can find the calibrated total allocating power P'i as
~-1
P'i = Pi - ( 1'~ ) X Qi. (20)
"i
95
According to [4], the balanced link quality 1'i will be
(21)
Set the stop criterion as I1'M - 1'111 c < 10- 3 . The power control
algorithm is described as follows:
• Step 0: [Initialization]
Set cO > > 10-3 , iteration w = O.
For all cell i E M, p,: set Q? = 0, TP = 0, N¥(i) = 0, ~o = N2(i)'(k
• Step 1: [Power allocation for mobiles not in T i ]
Obtain iir,j and the corresponding 1't' by (21) and (18), respectively.
• Step 2: [Decision of power control procedure]
IF cW > 10- 3 , CONTINUE Step 3.
ELSE GOTO Step 4.
• Step 3: [Adjust the total transmission power at each base station]

Pi(w+l) _ . {p(w) b!L NW(') ~. p~.}
-mIn i x =.(w) , I t X qz, z
'"'Ii
calibrate Pi(W+l) as in (20),
set w = w + 1 and GOTO Step 1.
• Step 4: [Update subset of Ti for each cell i]

IF any iir,j > fA exist,
choose one mobile station j, which is with the largest iir,j
and put into subset T i .
GOTO Step l.
ELSE DONE.
In Step 1, the base station will calculate the allocated power for each
mobile based upon the conventional quality balancing procedures. Then
in Step 3 we constrain the total base station transmission power by Pi or
N'f (i) X Pi. In Step 4 we limit the maximum transmission power allocated
for each user to be Pi and put it into Ti . All the mobile stations in Ti will
not participate in the quality balancing procedures. The procedures will
be stopped when the mobiles are either with the balanced link quality
"'Ii or in Ti .
2.4 Capacity Analysis

Capacity analysis is related to the outage performance. We define
the outage as the case when the receivedEb/No at a mobile station j is
below the threshold. Soft handover provided by macro diversity gain can
96
improve the outage performance. The soft handover is initiated when
the following condition is occurred
PM' L·',j. - PJ.L . Lk ,j. < -'f)
'l1 i r-I.. k (22)
where Li,j Lk,j are the link gain for the mobile j at cell i and cell k,
respectively; TJ is the handoff threshold. As in [1], the downlink outage
probability during soft handoff is defined the probability of the sum-
mation of the requested transmission power from all the mobiles being
larger than the total available power at base station. That is,
(23)
Given that number of soft handover users in macrocell M and micro cell
J.L (NJ.2°, N~ho), we can calculate outage probability for macro cell M
from (8)
(N M-N;';O) N;;o N~ho
pit:) Prob { L PM,j +L qM/2 + L PI"/2 > PM} (24)
Note that the first term in (24) can be used to incorporate the effect
of constrained downlink power allocation, and the second term and the
third term incorporate the effect of UPA or EPA. Let
(NM-N!!1°)
YM = L Dj · 10(!:I"-!:M)/10 (25)
j
and
Then (24) becomes
P~::) = Prob {YM > X} =Q(X- my ) (27)

(Jy
where Q(x) = Jxoo ~e-t2 /2dt. Note that since YM is a sum of independent
log-normal random variables, it can be approximated by another log-
normal random variable YM with mean my and standard deviation (Jy
by using the techniques in [16]. The outage calculation is validate both
for micro cell and macro cell in the forward link.
3. Numerical Results and Discussion

The numerical results of above analysis are shown here, and the sim-
ulation results are presented to verify the validity of the analysis.
97
3.1 Simulation Model

In our simulation, we assume that mobile stations are uniformly dis-
tributed in both macro cell and microcell. Let N M and N I-' be the number
of mobiles in a macro cell and a microcell, respectively. The other system
parameters are listed in Table 1.
Table 1. SYSTEM PARAMETERS
Parameters value
macrocell's radius(km), RM 3.0
microcell's radius(km), R/1- 0.3-3.0
cell radius ratio (R/1-/ RM), p 0.1-1.0
mobile's antenna height(m), h ms 1.5
macrocell antenna height(m), hM 20
micro cell antenna height(m), hI' 6-12
macrocell's max. total transmission power(watt), PM 20
macrocell's max. allocating power(watt), qM 1
2 slope path loss exponent, a, (3 2,2
standard deviation of 2 slope shadowing, 0"1,0"2 4.0,8.0
2 slope path loss model parameter, A 0.01
processing gain( dB), G 21
required received Eb/No(dB), received Eb/No(dB), r 5
We enhance the quality balancing power control method with ad-

justable cell-site transmission power [4J to provide allocation power con-
straint. Since base stations allocate different power to each mobile sta-
tion so that practical total transmission power for each base station will
be different. We categorize mobile stations to with and without soft
handover based on soft handover algorithm (22). Since soft handover
region are often located around cell boundary, we assume that both
base stations will allocate one-half maximum transmission power to soft
handover mobile stations. Here, two power allocation methods men-
tioned in Section 2, both the EPA and the UPA methods are examed.
At first, base stations allocate power for soft handover mobile stations,
and then for the other mobile stations without soft handover. The re-
moval strategy here is to remove the mobile with maximum required
allocating power [14J. The preceding process will be repeated until all
the mobile stations without handover can obtain their required signal
quality. At last, the corresponding received signal quality Eb/No can be
obtained by (1) and (12) for no handover mobile stations and for soft
handover mobile stations, respectively. The outage probability also can
be calculated by comparing the received Eb/No and 'Yreq for all mobile
stations. Therefore, we can obtain the results of outage probability ver-
sus different number of mobile stations in the macro cell and micro cell ,
respectively. The capacity of macro cell and micro cell are on condition
that outage probability equals 0.05, and the total capacity is defined as
the capacity of one macrocell and one microecll: NM + NI-'.
98
3.2 Discussion
In this paper, we study the performance of the soft handover in CDMA
cellular systems with mixed cell sizes. At first, we analyze the perfor-
mance of the received Ebj No for a mobile station in the location H as
shown in Figure 2. Figure 3 shows the EbjNo performance in different
cell radius ratio p supporting soft handover mobile H by EPA and UPA
methods, respectively. In the EPA case, according to (13), executing
soft handover can improve received signal quality in homogeneous cellu-
lar environment, i.e., p = 1.0. However, the received signal quality will
be deteriorated for mobile station h moving from microcell to macrocell,
denoted by f.L ---> M in the figure, as soon as the cell radius ratio is get-
ting smaller. More importantly, we find that EPA method will result in
the power exhausting problem when the cell radius ratio is smaller than
0.3, for which the micro cell consumes too much power in supporting
soft handover mobile stations moving from macrocell, thereby leaving
no power budget to support its own users. On the contrary, the UPA
method, based on (14), the EbjNo can be maintained at the satisfactory
level for p = 0.1 rv 1.0. Thus it is demonstrated that the UPA method
can avoid the power exhausting problem in micro cell for smaller cell ra-
dius ratio.
18
16 -e- M-Il , EPA
-(3- /l-M, EPA
14 -e- M-I', UPA
-El-Il - M. UPA
12
co 10
hard
soft handoff
~
0 8 andoff
z
:0
w 6
.-.-.-.~.-.-.-.
cr--
4 ~~
2
-
---
microcell
0 macrocell
no handoff
-2 0
0.2 0.4 0.6 0.8
Cell radius ratio (p)
Figure 3. Eb/No performance in different cell radius ratio for mobile station h
without handover and with soft handover by equall power allocation and unequall
power allocation methods, respectively. J.L ~ M means a user moves from a microcell
f.1 to a macrocell M, and M -> f.1 means a user moves from a macro cell to a microcell.
Figure 4 validates the accuracy of the proposed outage performance

estimation, i.e. (23) in Section 2, by simulation. It is shown that in terms
of 0.05 outage probability, the approximate capacity are very close to
the simulation results. The convergence time by approximation (23) is
99
1
10 -0- Simulation(macro) p =0.7, EPA
-B - Simulation(micro)
- Analysis
p = 1.0, EPA
~
:cI1l
.c 10 ·1
e
D-
Q)
0>
.:9
:::J
0
10 ·2
20 25 30 35 40 20 25 30 35 40
Capacity (number of mobile stations) Capacity (number of mobile stations)
Figure 4- Outage probability of microcell and macro cell using conventional quality
balancing power control method with equal power allocation (EPA) soft handover
algorithm for p = 0.7, 1.0, where p is the cell radius ratio of micro cell and macrocell.
macrocell
~ 36 35
III
'":::>
'0 32 30
iii
.0
.s
E
28 25
o No power constraint
o Power constraint
~
u
~ 24 20
to ONo power constraint
U o Power constraint
0.3 0.5 0.7 O.g 1 15 0!o----::-0.~1~~0~.3:--~-::0"::.5-~0'"'.7~~-0::-'.g:-----"

Cell radius ratio ( p ) Cell radius ratio (p
W~A ~U~
Figure 5. Capacity of (a) equal power allocation (EPA) and (b) unequal power
allocation (UPA) with soft handover against the ratio of radius of the microcell to
that of the macrocell p.
nearly 10 times less than that by simulation. From the figure, one can
see that the errors in capacity approximation are from zero to two users,
which is tolerable for capacity calculation. Thus, this proves the accu-
racy and efficiency of using this analytical approximation in calculating
the user capacity for the heterogeneous cellular systems.
Figure 5a shows the capacity of the EPA method with soft handoff
against the cell radius ratio p, in terms of outage probability equal to
0.05. It is observed that the power exhausting problem occurred in the
micro cell for p < 0.7 without power constraint and p < 0.5 with power
constraint, respectively. One can see that the smaller the value of p,
100
65
60
~ 55
~ 60
'0 50
2i
I.??
45 UPA
55
40
[ 35
u" 50 L:'-c--'--c:'---'--~~~~~-.J
30
0 0.1 0.3 0.5 0.7 0.9 1 0 0.1 0.3 0.5 0.7 0.9 1
Cell radius ratio ( P ) Cell radius ratio (p )
(a) (b)
Figure 6. Total capacity comparison of the equal power allocation (EPA) and
the unequal power allocation (UPA) during soft handover with and without power
constraint.
the higher the macro cell capacity. The increase of macro cell capacity
is mainly because of the decreasing interference from the microcell and
using up all the power budget of the micro cell base station. Hence,
although constraining the maximum power can help solving the power
exhausting problem in the micro cell, the improvement is not significant.
Figure 5b demonstrates the capacity of the UPA method with soft han-
dover against the cell radius ratio in terms of 0.5 of outage probability.
Unlike the EPA method, the UPA can maintain a good capacity for
both micro cell and macrocell from p = 0.5 '" 1. The power exhausting
problem will not occur even with p = 0.1 although with slightly capacity
degradation in microcell. It is also noted that the power constraint can
improve the capacity, especially when the p is small. For p = 0.1 the
capacity for the constrained UPA method increases microcell capacity
about 30%.
Figure 6 shows the total capacity of EPA and UPA methods. The
total capacity here is the summation of the macrocell capacity and the
micro cell capacity in Figure 5. The purpose of plotting Figure 6 is
to calculate the capacity in the case of one macro cell and a cluster of
microcells. We assume the coverage is equal to two macrocell coverage
area. Based on the area's ratio (i? between macro cell and microcell, the
estimated capacity of one macrocell and a cluster of multiple micro cells
can be calculated as (NM + N", x (i)2), where NM and N", is the user
capacity of a macro cell and a micro cell, respectively. As shown in Figure
7, the concept of constraining the maximum transmission power during
hand off is helpful in avoiding the power exhausting problem only for
p> 0.5. As the star marks in the figure, the EPA method suffers from
the power exhausting problem for p = 0.3 and 0.5, where the capacity of
multiple micro cells is sacrificed for only a macro cell capacity. The UPA
method can provide high user capacity regardless of power constraint.
For p = 0.3, it is easily to see that the UPA method can enhance total
101
system capacity eight times higher than the EPA method. Hence, it is
concluded that the UPA method is an inevitable technique in solving the
power exhausting problem during the soft handoff in a CDMA network
with mixed cell size.
400
.::I no power constraint with EPA *
*
3~5
350 • _ consl",int wit/1 EPA
319 ~ no power constrainl wilh UPA
~ 300 • power oonstraint with UPA
Q)
en
::J
'0 250
Q;
.c 200
E
::J
E- 150
~
ro
u 100 84 91 91 94
c..
ro 60 62 6062
u
50
~ 0
0.5 0.7 1.0
Cell radius ratio (p)
Figure 7. Approximate capacity of one macro cell and multiple microcells for power
control algorithms with and without power constraint, combined with EPA and UPA
methods for soft handover power allocation, where there are 2 microcell for p = 0.7,
4 microcell for p = 0.5, and 333 microcells for p = 0.3, respectively.
4. Concluding Remarks
This paper studies the downlink user capacity of a heterogeneous
CDMA cellular system with soft handover. We consider the scenario
that a hotspot microcell is adjacent to a larger macrocell. We ob-
serve the phenomenon of the power exhausting problem happened in
soft handoff between micro cells and macrocell. To quantize the impact
of this problem, we present an analytic approximation method for com-
puting the downlink user capacity with soft handover in heterogeneous
cellular structures. We further propose an improved quality balancing
power control method, the constrained unequal power allocation (UPA)
method to protect micro cell base station from being used up the trans-
mission power by the hand off mobile terminals in macrocell.
Our simulation results demonstrate that the proposed constrained
UPA technique with soft handover can enhance the total system ca-
pacity eight times higher than the conventional equal power allocation
(EPA) method in a heterogeneous CDMA network with the cell radius
ratio between micro cell and macrocell equal to 0.3. Future work in this
area include to extend the analytical capacity estimation technique to
multiple clusters of micro cells with multiple macrocells, and develop an
optimal downlink power allocation algorithms with soft handover for the
CDMA network with heterogeneous cell structures.
102
References
[1] Viterbi A. J., "CDMA:priciples of spread spectrum communication," Addison-

Wesley, pp. 218-224, June 1995.
[2] Salonaho 0., and Padovani R., "Flexible power allocation for physical control
channel in wideband CDMA," IEEE VTC'99 Spring, Houston, TX, 16-19, pp.
1455-1458, May 1999.
[3] Holma H., and Toskala A., "WCDMA for UMTS: radio access for third gener-
ation mobile communications," John Wiley and Sons, ltd., pp. 208-210, 2000.
[4] Kim D., "A simple algorithm for adjusting cell-site transmitter power in CDMA
cellular systems," IEEE Trans. on Veh. Technol., vol. 48, no. 4, pp.1092-1098,
July 1999.
[5] Lee C. C. and Steele R., "Effect of soft and softer handoffs on CDMA system
capacity," IEEE Trans. on Veh. Technol., vol. 47, no. 3, pp. 830-841, Aug. 1998.
[6] Grandhi S. A., Zander J., and Yates R. D., "Constrained power control," Wire-
less Personal Commun., vol. 1, no. 4, pp. 257-270, 1995.
[7] Erceg V., Ghassemzadeh S., Taylor M., Li D., and Schilling D. L., "Ur-
ban/suburban out-of-sight propagation modeling," IEEE Commun. Mag., pp.
56-61, June 1992.
[8] Min S., and Bertoni H. L., "Effect of path loss model on CDMA system design
for highway microcells," IEEE VTC98, pp. 1009-1013, 1998.
[9] Shapira J., "Microcell engineering in CDMA cellular networks," IEEE Trans.
Veh. Technol., vol. 43, no. 4, pp. 817-825, Nov. 1994.
[10] Shalinee K., Greenstein L., Poor H. V., "Capacity tradeoffs between macro-
cell and microcell in a CDMA sysem: exact and approximate analyses," IEEE
Vehicular Technology Conference, VTC'Ol Fall, Atlantic City, pp. 1172-1176,
October, 2001.
[11] Jeon H. G., Shin S. M., Hwang T., and Kang C. E., "Reverse link capacity
analysis of a CDMA cellular system with mixed cell sizes," IEEE Trans. on
Veh. Technol., vol. 49, no. 6, pp. 2158-2163, Nov. 2000.
[12] Lee D. D., Kim D. H., Chung C. Y., Kim H. G., and Whang K. C., "Other-cell
interference with power in macro/microcell CDMA networks," IEEE Vehicular
Technology Conference, pp. 88-92, 1996.
[13] Kim J. Y., Stuber G. L., Akyildiz 1. F., "Macro diversity power control in hier-
archical CDMA cellular systems," IEEE J. Select. Areas Commun., vol. 19, no.
2, pp. 266 V276, Feb 2001.
[14] Andersin M., Rosberg Z., and Zander J., "Gradual removals in cellular PCS
with constrained power control and noise," ACMjBaltzer Wireless Networks J.,
vol. 2, no. 1, pp. 27-43, 1996.
[15] 3GPP technical Specification 25.942, RF System Scenarios, page 26, Dec. 1999.
[16] Schwartz S. and Yeh Y. S., "On the distribution function and moments of power
sums with log-normal components," Bell System Tech. Journal, vol. 61, pp.
1441-1462, Sept. 1982.
Packet service in UMTS: effects of the
radio interface parameters on the
performance of the downlink shared
channel
Flamini Borgonovo, Antonio Capone, Matteo Cesana, Luigi Fratta
DEI, Politecnico di Milano
borgonov,capone,cesana,fratta@elet.polimi.it
Abstract
The UMTS W-CDMA radio interface is characterized by great
flexibility and a variety of different physical and logical channel
types. For example, on the downlink, the DCH offers circuit switch-
ing, the FACH uses packet switching and the DSCH uses packet
switching with closed loop power control. Furthermore, several user
rates and protections are possible, by choosing suitable parameters,
such as spreading factors, code rates and ARQ schemes. In this
paper we present the results, obtained by a detailed simulation,
about the effect of several parameters and system alternatives on
the capacity of the downlink segment of the W-CDMA interface
with packet service. In particular, we investigate the effect of the
spreading factor and the code rate on the DSCH capacity and delay-
throughput performance.
1 Introd uction
The Universal Mobile Telecommunications System (UMTS) [1, 2] is the third
generation mobile communication system developed by ETSI, the European
Telecommunications Standard Institute, which will allow the use of a new fre-
quency spectrum and is expected to extend the present GSM service to include
multimedia.
In UMTS, users will be provided with data rates up to: 144 kb/s, in macro-
cellular environments, 384 kb/s, in micro-cellular environments, and up to 2 Mb/s
in indoor or pico-cellular environments. Due to the effort of the standardization
bodies, the radio interface is characterized by great flexibility and a variety of
different physical and logical channel types. For instance, several user rates
and protections are possible, by choosing suitable parameters, such as spreading
factors, code rates and ARQ (Automatic Repeat request) schemes.
103
X. Lagrance and B. Jabbari (eds.),
104
Among the new services offered by UMTS, the packet data service is prob-
ably one of the most critical from the system parameters setting point of view
mainly because of the characteristics of the code division access scheme adopted
at the radio interface. Up to date no study that thoroughly investigates the
effects of the different possibilities on UMTS data service performance has yet
appeared.
In the downlink, three different transport channel types are available for data
packets transmission, namely the DCH Dedicated Channel, the DSCH Downlink
Shared Channel and the FACH Forward Access Channel.
The DCHs are assigned to single users through set-up and tear down pro-
cedures and are subject to closed loop power control that, if used for circuit
service such as voice, stabilizes the BER (bit error rate) and optimizes CDMA
performance.
The DSCH is a common channel on which several users can be time mul-
tiplexed. No set-up and tear down procedures are required and the physical
channel on which the DSCH is mapped does not carry power control signaling.
However, since the closed loop power control is still required, users that are al-
lowed to access DSCH services must have an associated active DCH. The DCH,
if not already active due to another transport service, must be activated just to
allow the access to the DSCH and to carry physical layer signaling only.
The FACH is shared by several users to transmit short bursts of data, but,
unlike DSCH, no closed-loop power control is exerted and no DCH must be
activated to access this channel.
For each one of the above channels, different combinations of spreading factor
and code rate can provide the bandwidth and the protection required for different
services and environments. However, it is not altogether clear which combination
is the best.
Well known results for real-time circuit traffic show that CDMA with closed-
loop power control can be very effective in spectrum exploitation [3]. Its effi-
ciency can be further enhanced by using powerful codes and FEe codes have
been proved to be more effective than spreading codes [4].
With packet service, the effect of direct sequence spreading, FEC codes and
closed loop power control is not easily predictable and the optimal combination
of codes and spreading factors may be different from circuit service. In fact,
data traffic is bursty in nature, and, depending on the number of interfering
channels and their power levels, errors can be more efficiently obviated by ARQ
techniques than by forward error correcting codes [5, 6]. For the same reason,
the protection obtained with high spreading factors is questionable.
To understand the roles that the many parameters and system features have
on the overall capacity with packet service, we have implemented a detailed
simulator of the UMTS downlink.
In Section 2 we present the system model adopted for simulations and in
Section 3 we discuss the results obtained. Section 4 includes some final remarks
and concludes the paper.
105
2 Simulator description
The simulator reproduces a system composed of 49 hexagonal cells that lay on a
torus surface to avoid border effects. The base stations (BS) are located at the
center of each cell and irradiate with omni-directional antennas with unit gain.
2.1 Propagation model

The propagation model assumed in this work follows the guidelines of ETSI [7].
In particular, the relationship between the received power Pr and the transmitted
power Pt is given by Pr = Pt 0 2 10TI;- L where L is the path loss, lOTI;- accounts
for the loss due to slow shadowing, E being a normal random variable with zero
mean and 17 2 variance, and 0 2 represents the gain, with a negative exponential
distribution of unit mean, due to fast fading.
In the following we refer to a macro-cellular environment, for which the cell
radius is 300 m, and the path loss L is expressed as 10 log L = 128.1 + 37.6 log r
(dB), where r (in meters) represents the distance between the mobile and the
base. Furthermore, we assume no fading and shadowing standard deviation
equal to 5 dB.
When a new user is generated, its position is chosen randomly over the
torus surface and it is assigned to the BS with the minimum attenuation. No
user mobility is considered.
Each cell is assigned a single tree of orthogonal variable spreading factors
(OVSF), so that channels in the same cell are always orthogonal. The loss of
orthogonality of the received signals due to the multipath effect is accounted for
in the receiver model as specified in section 2.4.
2.2 Traffic model

We have adopted a basic traffic model which emulates typical web-browsing
sessions that become active according to a Poisson Point Process of intensity A,
as described in [7]. Each session, upon activation, generates a flow of packets
whose length is negative exponentially distributed with mean 3840 bits. The
packet flow is composed of a number of packets, geometrically distributed with
mean Np = 25, that arrive according to a Poisson Point Process whose intensity
is chosen to match a given source speed. A user leaves the system as soon as
the last packet of the web-browsing session has been successfully received.
2.3 Transmission model

The packets generated by each user are delivered to the RLC (Radio Link Con-
trol) layer [8] where they are subdivided into transmission blocks before being
queued for transmission. Each transmission block includes an RLC header of 16
bit that also accounts for an ARQ mechanism.
At each 10 ms long frame, the MAC (Medium Access Control) layer [9]
chooses an user queue according to the scheduling mechanism and, after adding
the MAC header, sends to the physical layer the blocks up to filling the frame.
106
I.E+OO , - - - - - - - - - - - - - , I.E+OO , - - - - ' " " ' " " '___- " " ' - - - - - - ,
I.E-Ol t---"''''-~*-------I
I.E-01 t-----t--H---1.----j
I.E-02 +-----+1r1-~-----l
n: n:
~1.E-03 t-------1M-\-----''<----I w
...I
III1.E-02
I.E'()4
I.E-05 +-~-_~~---'.,u..~-~.>...,_-l
-604-2024 10 12 -6 -4 ·2 0 2 4 6 8 10 12 14 16
Eb/No [db] SIR [db]
Figure 1: Bit Error Rate of the Figure 2: Block Error Rate of

convolutional codes adopted in the convolutional codes adopted
UMTS as function of the bit nor- in UMTS as function of the Sig-
malized energy. nal to Interference ratio after de-
spreading.
Before transmission, the physical layer adds the redundancy bits according to
the coding scheme adopted. Several coding schemes are supported by UMTS.
Our simulator adds the parity bits required by Convolutional Codes, with
256 states, Constraint Length K = 9 and optimal puncturing, whose Bit Error
Rate (BER), obtained through link level simulations [10], is shown in Figure
1. In particular we have considered code rates, spreading factors and block
sizes such that the bits introduced by rate matching are very few and add an
overhead, without increasing error protection. To avoid throughput differences
due to different mappings of bits from packets to blocks we have used almost
the same transport-blocks size for all codes and spreading factors. A block
length of about 750 bits has been proved optimal with respect to the maximum
throughput.
2.4 Receiver model

At the receiving side the carrier to interference ratio is evaluated, for each trans-
mission, as
c
I
= ~~--~--~
O'.lintra + linter + PN
(1)
where P r is the received signal strength, PN is the thermal noise assumed

equal to -99 dBm, linter is the sum of the signal powers received from the other
cells, lintra is the sum of the signal powers received from users within the cell,
and 0'. is the loss-of-orthogonality factor assumed equal to 0.4 [11].
The relationship between the carrier to interference ratio (fj-) and the energy
per user bit (Eb) is given by:
107
Eb 1 C
-No = -2R xSFx-
I'
(2)
where R is the coding rate and SF the spreading factor. The term ~ comes
from the fact that since QPSK modulation scheme is adopted on the downlink of
UMTS-FDD, each information symbol is composed of 2 bits. The ratio Ebl No
represents, in fact, the entry in Figure 1. From the curves shown in this figure,
BLock Error Rate (BLER) curves have been obtained as BLER = 1- (1- BER)1 ,
1 being the transmission block length. For each transmission, the normalized bit
energy is used to derive the BLER, and the correctness of the transmission is
decided testing the value of a normalized random variable against BLER.
Otherwise specified, in the following we will refer to the SIR after despread-
ing, which is defined as SIR=SF x C/I. Figure 2 reports the BLER of blocks of
750 bits versus SIR.
Our simulator does not implement an explicit ARQ procedure. Instead, at
the end of the transmission, the transmitted block is kept in the transmitting
queue unless no error occurs. After 10 failed transmissions the block is dropped
and the user is declared in outage.
2.5 Power control model

The power control mechanism adopted for DCHs uses two control loops. The
inner loop controls the transmitted power to maintain the SIR at the target value,
whereas the outer loop controls the SIR target value to achieve the target BLER.
The latter control mechanism has been envisioned to provide different qualities
to different services. Since in our simulation we investigate a service at a time,
the corresponding BLER requirement can be assumed constant and therefore
we have implemented the inner loop only, treating the BLER requirement as an
input.
The UMTS specifications require that power-update requests of ±1 dB are
transmitted every time slot (0.666 ms) in order to fight fast fading effects. How-
ever, since we do not consider fast fading in the propagation formula given in
2.1, we have assumed to transmit power updates every frame (10 ms). This
simplifies the simulator and reduce its run time without affecting the generality
of the results. DCH power updates, limited within the ±15 dB range, are re-
quested at each frame based on the difference between the SIR target and the
SIR evaluated on the last frame.
Each channel can not exceed a transmitted power of 30 dBm, whereas the
overall power transmitted by a base station is limited to 43 dBm. [11].
2.6 Flow Control on DSCH

For several values of system parameters we have observed throughput instability.
Instability is said to occur when the average interference level in the system
gets too high and the power control cannot provide the desired SIR to many
connections due to power limitations. In this situation the curve of throughput
versus the traffic on the channel decreases after reaching a maximum. To get
108
No
Pj "'" Pmax ) - - - - - '
hes
I Ni :=Ni+ll
+
Pick Fe_timer,
in
[O:N,]
Figure 3: Flow Control scheme Figure 4: Flow Control scheme

mobile equipment side. mobile base station side.
a stable behavior, it's necessary limit the average interference level. We have
proposed and implemented in our simulator a flow control mechanism based on
the well known back-off (BO) mechanism which dynamically adjusts the load on
the active channels, and therefore the average interference generated, according
to traffic and propagation conditions.
The BO mechanism is based on a feedback provided by mobile terminals.
The basic idea is to reduce the transmission rate on the channel when one or
more consecutive transmissions fails. The mechanism is triggered only when
a transmission performed at the maximum powers. More in details, a flow-
control timer is started so that transmissions are inhibited for a random number
of frames (10 ms long) uniformly distributed in the interval (l,n), where n is
the number of consecutive wrong transmissions performed at maximum power.
In such a way, we control the traffic on the channels (G), and consequently
we limit the mean interference level. In order to let the information on the
transmission result at the transmitting end available, we adopt one of the FBI
(Feedback Information Bits) bits defined in the uplink transport block format
of the dedicated channel (DCH) associated to the DSCH [12]. The flow chart of
the BO mechanism is reported in Figures 3 and 4.
3 Simulation results
The complexity of the overall system and the interaction among system pa-
rameters and performance variables do not allow a single and straightforward
discussion of the system behavior. In our investigation we have been forced to
focus the study into several sub-problems and to take simplifying assumptions.
109
We have at first investigated a simplified system in which the interference

generated by DCHs is not taken into account in SIR calculation. In the real
system, in fact, the use of DCH control channels beside the PDSCH 1 (Physical
DSCH), has effects on the global interference suffered at the receiving end and
can dramatically affect the performance of data transfer over DSCH.
With this assumption we have analyzed the effect of spreading factor and
code rate with the single physical DSCH, and the performance with multiple
physical channels (subsection 3.1). Then, once acquired a clearer understanding
of the complex mechanisms that affect the system behavior, we have investigated
the effect of DCH interference on the DSCH throughput (subsection 3.2).
3.1 Effect of codes and spreading factors

Figure 5 shows the packet average delay versus the throughput when one PDSCH
is adopted with SF=4 and for different codes, namely R= 1, 3/4, 2/3, 1/2. In
all cases we have adopted the linear back-off mechanism described in section
2.6. Furthermore, we have chosen the SIRtar,qet values in order to minimize the
BLER. If a too small SIR target is chosen too many errors occur since the SIR
fluctuations around the target value often drive the system in a condition where
the code protection is useless. On the other side, with a high SIR target the
power requirement increases and too many transmissions tend to be driven in
saturation.
The best performance is obtained with a light codes (R= 3/4 and R=2/3).
Heavier codes (R= 1/2) achieve a poorer performance since the added redun-
dancy bits provide a useless excess protection and negatively affect the through-
put. The low throughput in the case of no error correction (R= 1) shows that
the protection of the spreading process with SF= 4 is not sufficient to fight
interference.
For a deeper understanding of the system behavior, we have considered addi-
tional performance parameters shown in Figures 6, 8, 7. These figures show the
throughput, the BLER and the average fraction of transmission at the maximum
power (saturation fraction) versus the channel traffic G, which is defined as the
normalized value of transmitting frames effectively occupied by data blocks.
In the R = 1 case, we have adopted a SIR target of 13 dB which should
provide a very low BLER (Figure 2). Unfortunately, due to the power limits,
the power control is not able to reach this value at high loads, and the BLER
results much higher. Such an effect can be explained by observing Figure 7). In
fact, as channel traffic G increases, the power control drives many sources into
saturation causing a sharp increase in the BLER (Figure 8). At this point the
BO mechanism intervenes and drastically limits the maximum G as shown in
Figure 6.
Since the high SIR target is responsible for the bad performance at G = 1,
to improve performance we must adopt correcting codes that allow a lower SIR.
Referring again to Figure 2 we see that with an R = 3/4 code, a 9 dB SIR is
enough to guarantee a very low BLER. However, even in this case the perfor-
mance is too bad in G = 1 and the BO intervenes limiting G to 0.955. A similar
1 Physical channel where DSCH is mapped
110
2000 1200
SF=4 SF=4
! iii'
1000
..
:a
1500
R=2/3
BOO
>-
CO ~
..
'ii :::I
"C 1000 C. 600
..c
....
CI
CO
.
CI
:::I
0 400
«> 500 ..c

t-
200
200 400 600 800 1000 12(

0.2 0.4 0.6 0.8
Throughput (kbs) Channel traffic G
Figure 5: Average delay versus Figure 6: Throughput versus the

throughput for SF= 4 and differ- channel traffic G for the cases re-
ent code rates with the basic traf- ported in Figure 4.
fic model.
SF=4
-
c R=1
o
:s 0.1 .
.Il 0.1
/ R-3/4
,!g a: ./
c
o
g
UJ
~
R=3/4
~:::I ""U
0.01 .2 0.01
111 _____ R-1I2
~ SF=4
0.001 -I-----,---.----.--....J<.....,-----l 0.001

0.2 0.4 0.6 0.8 o 0.2 0.4 0.6 O.B
Channel traffic G Channel traffic G
Figure 7: Fraction of user in sat- Figure 8: Block Error Rate ver-

uration versus the channel traffic sus the channel traffic G for the
G for the cases reported in Figure cases reported in Figure 4.
4·
111
behavior is observed with the code R = 2/3, though the G limit is further in-
creased to 0.97. The delay curves for the two last cases almost perfectly overlap,
showing the ability of the BO to keep the system very close to capacity. How-
ever, although the reduced SIR target reduces the saturation fraction and the
BLER, the obtained BLER is still much higher than what predicted by Figure 2.
The reason is that, the power control mechanism is able to track the SIR at the
target value only with slow varying interference, as in the case of circuit service,
but fails its goal with packet service where traffic and interference are bursty.
We have measured SIR standard deviation values in the range 3.7 - 4.3 dB 2.
With more powerful codes (R= 1/2), G = 1 can be reached and, because of
the further reduced SIR target, the saturation fraction and the BLER is further
reduced. However, the benefits of the reduced number of retransmissions, do not
compensate the loss of throughput due to the increased code redundancy and
the maximum observed throughput is remarkably smaller than that obtained
with R= 2/3.
The system parameters configuration with R= 1/2 is very effective in fighting
interference, since it allows transmissions with lower power levels. It's quite clear
from figure 5 that the redundancy introduced by the code R= 1/2 limits the
capacity of the system, when using a single PDSCH. Under these hypothesis we
have studied the performance of the DSCH service when using multiple SF= 8
in a single BS, and a FEC code with R= 1/2.
The performance is significantly improved, as observed in Figure 9, by using
3 and 4 PDSCHs with a SIR target equal to 4 dB, which in this case has shown to
be optimal. The BO mechanism intervenes with 4 PDSCH limiting the maximum
G to 0.855. With 5 PDSCHs, the increase in the interference prevails, G is limited
to 0.622, and a small instability effect is present despite the BO. The case of 4
channels provides the maximum throughput (1240 kb/s) among those examined.
3.2 Effect of control channels

In this section we discuss the system performance when the effect of the DCH
control channels is considered in the interference and SIR evaluations.
It is expected that the increased interference caused by all the active DCH
decreases the throughput of the DSCH and causes instability. In fact, a through-
put decrease increases the number of users waiting for transmission and conse-
quently the number of active DCH. This further increases the interference and
reduces the throughput, which reaches zero if the waiting queue is infinite.
To control this phenomenon we have limited to N the number of users with
an associated DCH. The users that arrive beyond this limit are queued and wait
for an available DCH. In our model we have assumed that DCH control channels
use a spreading factor equals to 512 and require a SIR target equal to 8 dB.
Figure 10 shows the effect of N for the case of four channels, SF= 8 and
R = 1/2. The curve with N= 0 in the figure represents the ideal case when no
DCH interference is considered. It is clear that an increase of N decreases the
achieved throughput. This is due to the added interference generated by the
2Throughout the paper, all interference and SIR statistics are evaluated considering
logarithmic values
112
2000 - . - - - - - - - - - . - - - - ' " ' ,r.-

h --,
4th
SF=8 ,eh b.D.
R=1/2 b.o
'iii 1500 +---------+---++---H

.§.
.""
.
~ 1000 +-----------jI-----1f---+--l
'"f!
3
.. 500
o+---~-~-~-~--~-~
o 200 400 600 800 1000 1200
Throughput (kb/s)
Figure 9: Average delay versus throughput when a different number of

physical channels are used in parallel for SF= 8 R = 1/2.
DCHs which affects the performance of the DSCHs. Note that limiting to N
the number of active DCH has from one side the beneficial effect of reducing
the interference, but from the other it reduces the efficiency of multiplexing and
might penalize the achieved throughput in the case of sources with low speed.
Therefore an optimal choice of N exists for any given traffic characteristics. In
our case, which assumes the basic traffic model described in section 2, N = 10
is an optimal choice since we did not observe any multiplexing inefficiency.
4 Conclusions
In this paper we have investigated the performance of the UMTS radio interface,
with packets service, mainly evaluating the maximum throughput achievable on
the DSCH with different physical channels configurations, traffic dynamics and
power control mechanisms.
The results show that, when the packet service can use only one single phys-
ical channel, the maximum throughput is attained with the smallest available
spreading factor, SF= 4, and a light code, R = 2/3. The other cases character-
ized by a higher channel protection (lower R ) present a lower throughput since
the loss due to the added overhead is not compensated by the reduced BLER.
If the use of multiple physical channels is allowed, the maximum throughput
is attained by using up to four channels with SF= 8 and R = 1/2, despite
the new intra-cell interference introduced. This is mainly due to the improved
efficiency of the closed-loop power control that, taking advantage of the longer
transmission time and of the reduced interference burstiness, better tracks the
SIR target. This confirms the common belief that CDMA characteristics are
better exploited by circuit services with constant interference and, therefore, by
using many small channels rather than a single big one.
Further investigations, still in progress, show that the conclusions drawn
113
2000
N-10
SF=8
R=1/2
r-- 4 channels
..,
N-20
en 1500
.§.
~
~
., 1000
.,E'"
~ 500
__ I.)
o
o 500 1000 1500
Throughput (kb/s)
Figure 10: A vemge delay versus throughput when a different number of

active users is adopted in the case with four physical channels, SF= 8 and
R = 1/2.
here somehow scale to other scenarios, such as micro-cells and/or corner fed
antennas.
References
[1] A. Samukic, UMTS universal mobile telecommunications system: development of
standards for the third generation, IEEE 'Transactions on Vehicular Technology,
vol. 47, no. 4, Nov. 1998, pp. 1099-1104.
[2] K.W. Richardson, UMTS overview, Electronics & Communication Engineering
Journal, vol. 12, no. 3, June 2000, pp. 93-100.
[3] A.M. Viterbi, A.J. Viterbi, Erlang capacity of a power controlled CDMA system,
IEEE Journal on Selected Areas in Communications, vol. 11, no. 6, Aug. 1993,
pp. 892-900.
[4] J. Y. N. Hui, Throughput analysis for Code Division Multiple Access of the Spread
Spectrum Channel, IEEE Journal on Selected Areas in Communications, vol. 2,
no. 4, July 1984.
[5] R.J. McEliece, W.E. Stark, Channels with block interference, IEEE 'Trans. on
Information Theory, Vol. 30, No.1, January 1984.
[6] F. Borgonovo, A. Capone, L. Fratta, Retransmissions Versus FEC Plus Interleav-
ing for Real·Time Applications: A Comparison Between CDPA and MC-TDMA
Cellular Systems, IEEE Journal on Selected Areas in Communications, vol. 17,
no. 11, Nov. 1999.
[7] UMTS 30.03, Annex B: Test environments and deployment models, TR 101 1112
v.3.2.0, April 1998.
[8] 3rd Generation Partnership Project, RLC Protocol Specification, 3G TS 25.322,
December 2001.
114
[9J 3rd Generation Partnership Project, MAC protocol specification, 3G TS 25.321,

December 2001.
[10J A. Bellini, M. Ferrari, Personal communication, Politecnico di Milano, 2000.
[I1J 3rd Generation Partnership Project, RF system scenarios, 3G TR 25.942, Decem-
ber 2001.
[12J 3rd Generation Partnership Project, Physical channels and mapping of transport
channels onto physical channels (FDD), 3G TS 25.211, December 2001.
Cellular Multihop Networks and the Impact of
Routing on the SNIR and Total Power Consumption
Kevin M. Pepe, Branimir R. Vojcic

K. M. Pepe is with SPARTA. Inc, Rosslyn, VA 22209 USA {telephone: 703-797-3087, e-mail:
Kevin _Pepe@Sparta.com}.
B. R. Vojcic is with the Department ofElectrical and Computer Engineering, The George
Washington University, Washington, DC 20052 USA {e-mail: vojcic@seas.gwu.edu}.
Abstract: This paper investigates the impact of routing on achievable SNIR and power
utilization assuming a Cellular Multihop Network (CMHN) architecture with
code division multiple access (CDMA). To support the investigation a new
centralized routing algorithm is developed to combine network layer routing
decisions with physical layer characteristics. Results are presented showing
significant advantage of the algorithm over traditional least-cost routing when
lognormal Shadowing and fading is present, demonstrating the advantage of
more sophisticated routing to maximize Signal-to-Noise-and-Interference
Ratio (SNIR) and minimize power consumption.
Key words: multihop networks, power control, routing.
1. INTRODUCTION
In this paper we consider the impacts of route selection on the achievable

Signal-to-Noise-and-Interference Ratio (SNIR) and total power consumption
for a Cellular Multihop Network (CMHN) utilizing code division multiple
access (CDMA). Traditional multihop (ad hoc) networks assume no
preexisting infrastructure. However, recently there has been interest in
overlaying multihop networks on a cellular structure ([ 1D. The potential
advantages are power savings, increased capacity, and increased robustness
115
116 Kevin M. Pepe, Branimir R. Vojcic
via path diversity. However, there are unique issues associated with the
CMHN architecture. In particular, since all data is flowing to/from a
common Base Station CBS) there is the potential for considerable congestion
near the BS. Also, the required node complexity, in terms of other-user
traffic that must be handled, varies widely as a function of distance from the
BS. Finally, unlike a Single Hop (SH) network, a node must be able to both
transmit and receive on the forward link as well as the reverse link. This
paper concentrates on the impact routing has on the network interference and
power characteristics and develops a heuristic algorithm to select routes that
balance between these characteristics. It achieves this by making network
layer routing decisions based on physical layer conditions. As a result,
considerable power savings are achievable while still maximizing the SNIR.
Section II provides more detail about the CMHN architecture. Section III
analyzes the SNIR and power characteristics of a multihop network. Section
N presents a new O(K4) routing algorithm to balance between maximal
SNIR and minimum total power utilization. Section V presents simulation
results, and Section VI presents conclusions.
2. CMHN SYSTEM MODEL
In this section we describe the CMHN. The network to be considered

consists of numerous (possibly) mobile subscribers communicating with one
or multiple base stations. The base stations provide access to external
networks or systems. It is assumed that each user communicates through a
base station (e.g., even if a user is communicating information to another
nearby user, it will transmit/receive through the nearest base station). 1
Figure 1 shows an example set of users within a cell transmitting
information to a base station. For purposes of this example we consider the
reverse link (node to BS) although our analysis addresses both forward and
reverse link. In Figure la) each user communicates directly to the BS,
whereas in 1b) information is transmitted over multiple hops. By taking
multiple hops the transmit power can be substantially reduced due to reduced
transmit range, and, possibly less path loss for ranges within line-of-sight.
For simplicity, we assume that each user generates the same one unit of data
per unit time. If a node is used as an intermediate hop then it forwards its
own traffic plus that of all node traffic that passes through it. Consider
Cellular Multihop Networks and the Impact of Routing 117
Figure 1b and the nodes labeled A, B, and C connected to the BS. Each of
these nodes carry traffic for themselves and 2, 4, or 5 other nodes,
respectively. Note that the cumulative sum of the traffic on the last hop
necessarily equals that of all the nodes in the network. In the example, we
have 3 packets (A~BS) + 5 (B~BS) + 6 (C~BS) = 14 packets being
transmitted to the BS on the last hop - the same as the total number of non-
BS nodes in the network. Therefore, the interference on the last hop is at
least as great as that experienced at the BS in a SH architecture.
Figure I. Example cell communications for (a) Single Hop (left) and (b) Multihop (right).
We put few restrictions on the nodes' transmit and receive abilities, except
that a node cannot simultaneously transmit and receive. We assume CDMA
as the multiple access method (although many of the results shown are
equally applicable to time or frequency division multiple access) and each
user's packet has a different spreading sequence (see [2] for code assignment
protocols). Each node is capable of simultaneously transmitting to multiple
nodes and can transmit multiple packets from different users on each link,
where the transmit power associated with each link can be separately set.
Each node can also simultaneously receive from multiple nodes. We treat
the interference caused by other users as independent additive noise ([3-5])
and assume a receiver has a required SNIR to meet performance
requirements. The route for a CMHN is a spanning tree so that on the
I The approach taken in [I] allowed intra as well as inter-cell routing. This mayor may not
be justified, depending on the type of traffic supported by the network.
118 Kevin M Pepe, Branimir R. Vojcic
forward link there are I-to-many transmissions and on the reverse link there
are many-to-I transmissions, with a total of K -1 links (K nodes includes
the BS). Importantly, since all traffic in a CMHN is routed to/from the BS,
unlike an ad hoc network, we can apply variants of more traditional wired-
network routing algorithms.
3. ACHIEVABLE SNIR
In this section we describe an approach for determining transmit power in

order to meet or exceed the Signal-to-Interference Ratio (SIR) (or SNIR)
requirements at each receiver in a CMHN. Usually, the cellular CDMA
system capacity is interference limited and power control is essential in the
use of CDMA with conventional detection. Initial solutions focused on use
of the absolute signal strength at each receiver (see [5, 6]). Later it was
recognized that using SIR provided potential performance improvements and
more responsive systems (see [7-10]). For practical reasons these solutions
all have focused on distributed iterative approaches to controlling SIR, via
incremental changes in transmit power that converge to an optimal solution.
However, centralized solutions may be possible in some situations.
Furthermore, in a simulation environment where all performance related
information is known, a direct solution to achieving the desired SIR is useful,
realizing that it presents a best case answer and may provide useful insight
into more practical solutions. This section focuses on deriving a direct
centralized SIR solution for CMHN. While the results are applied to a
CMHN, they are applicable to multihop networks in general.
We first develop the approach assuming no noise and extend the results to
include thermal noise. The definitions and approach follow those used by
Hanly in [11] and [12] for single hop cellular, extended to account for
multiple packets on a link. In the case of no thermal noise the derivation
below essentially follows that of [13]. It is important to note that it can be
shown that the SNIR and total network power solutions are equivalent for the
forward and reverse link, assuming a symmetric channel and equal thermal
noise at all receivers.
3.1 SIR (No Noise)
Let there be M transmitting nodes and L receiving nodes and at any given
time node i is transmitting to receiving node Ci. The routing is assumed to
be already determined, i.e, c i is given. Transmitting node i transmits at

power qi per packet and transmits mi packets. We assume mi is an integer
number of packets, and always at least one to account for self generated
information. Further we denote the mean channel gain experienced between
each transmitter and receiver by the MxL matrix r, which includes radio
attenuation effects due to distance and shadowing and fading. Thus, the
received signal power for transmitter i at receiving node Ci is qir[i, cJ
The total interference power is modeled as the sum of powers of all active
interferers. Then the SIR of a single user associated with link [i, C i] is
Note that we include the (m i -1) other packets on the same link as
interference. While it would seem pessimistic since orthogonal codes could
be used to remove the interference, in general the effects of multipath may
severely reduce orthogonality (see [14] Section 6.7.3). Furthermore, we
have not assumed any multi-user detection capability at the receiver. Thus,
(m i -1) interfering packets is a worst-case assumption.
We also note that while we derive the following equations assuming a
general set of transmitters and receivers, for the CMHN we ar~ imposing a
spanning tree route. Thus, if there are K nodes in the networlC{including the
BS) there are K-l links. Also, we treat the power setting for each link
separately. Therefore, we effectively have K-l transmitters and K-l
receIvers.
For successful communication it is assumed there exists a threshold value
that the SIR must exceed in order to support each user's required data
rate R i • For a COMA system, the SIR is usually quite small due to the
processing gain and powerful coding. To make the role of processing gain G
more explicit we can use the relationship SIR = E b / I 0 (<; W/ R t 1 , E b the
energy per bit, 10 the interference power, <; a constant of order unity which
depends on the cross-correlation properties of the spread-spectrum codes
(assumed to be one hereafter), and Wthe overall available bandwidth.
Thus for successful communication of a particular user on link ~,Ci] we
want
(1)
where G i = W / Ri is the processing gain experienced by the user on link

[i, cJ and Yi is the threshold. For the CMHN we assume that each user
generates the same traffic and has the same requirements R (and therefore G)
and Y , although the results can be easily extended.
The power control problem is then to find a global allocation of powers q
to solve the M equations
Considering Yi = Y for all i gives
or
m.r(j, c.]
where A is the MxM normalized path-gain matrixAi,j == } [. ]' Note
r I,C i
that the diagonal of A; i = mi . We see that the elements of A are
proportional to the ratio of path loss experienced by interfering transmitters
at the desired cell site, to the path loss experienced by the source transmitter
to the desired cell site. Since the transmitted power is related to the inverse
of this ratio, we see that the elements of A quantify the gain transmitter i
Cellular Multihop Networks and the Impact ofRouting 121
experiences relative to interfering transmitter j.

Note that in the case of strict equality we have an eigenvalue equation
with eigenvalue Gjy + 1. In general, the existence of a unique power vector
solution and the resulting SIR is guaranteed by Perron-Frobenius theory. In
particular, the Subinvariance Theorem (Theorem 1.5 in [15]) states that if
sy ~ Ty , T any non-negative irreducible matrix, s a positive number, and y
a vector with all non-negative elements (not equal to the zero vector), then
all of the elements of y must be strictly positive and s ~ r, r the maximal
eigenvalue of T. 2 Equality holds if and only if s=r. Furthermore, the
eigenvector associated with r is unique and is a strictly positive vector. 3
Since A is strictly positive it is clearly non-negative irreducible. Thus, we
know that a maximal positive eigenvalue exists and that the properly scaled
corresponding eigenvector is strictly positive. Therefore, if Gjy + 1 is less
than r, q will not have all positive elements and no global solution exists.
If we view G and y as fixed then it is expected that some percentage of
configurations would not be supportable. We would then reduce traffic
requirements (e.g. drop nodes, or reduce data rate) until a supportable
solution is found. If we view y and traffic requirements as fixed then the
processing gain G can be adjusted until a feasible solution exists. Finally, if
we view G as fixed, we have r defining the maximum achievable SIR for the
given network, traffic requirements, and routing. Note that the maximal
eigenvalue of the matrix A as defined above is greater than or equal to one
since the trace of a matrix equals the sum of the eigenvalues. Thus, we see
that r ~ 1 SInce
M M M
M'r~ IAi =tr(A)=IAii =Imi ~M~r~l,
i=1 i=l ' i=l
assuring a feasible solution. Note that the closer r gets to 1 the larger the
supportabley , assuming fixed G, or the smaller the required G (and hence
bandwidth), assuming fixed y. Thus, small r is desirable.
Finally, for CMHN we must impose the limitation that a node cannot
transmit and receive at the same time. We address this by segregating link
2 A matrix T is non-negative irreducible if all elements are non-negative and there exists a
positive integer n such that each element of Tn is strictly positive (need not be the same
n for each matrix element).
3 See Corollary 4.2 and Theorem 4.4 in Section 1.4 of [16].
transmissions into two time slots based on whether the number of hops to the
BS is even or odd. Then A is similar (and therefore has the same
- [Ao Ae01where Ao
characteristic equation) to a matrix of the form A = 0
and Ae are normalized path-gain matrices associated with links for the odd
and even time slots, respectively. The maximal eigenvalue of A is then the
maximum of the maximal eigenvalue of Ao and A e , where Ao and Ae
are both irreducible matrices.
Note that the segregation of the channel into two time slots imposes a 3 dB
penalty to the CMHN (or any multihop versus single hop approach). That is,
for a multihop architecture, half of the time available to a node for
supporting the forward (or reverse) link is allocated to transmitting, and half
to receiving. Therefore, the power dedicated to moving information towards
its destination is reduced by half.
3.2 SNIR (With Gaussian Noise)
We extend the results of the previous section by representing the noise

spectral density at each receiving node k given by the LxI vector '11k, with a
noise spectral density at each receiving node of'l1J2. In general, we assume
'11k = '11 for all k, although the results can be easily extended.
Extending equation (1), we require for successful communication
Preceding as above for SIR we derive
(13I-A)q = b
where 13 = G/y + 1, I is the MxM identity matrix, A is the same MxM path-
Cellular Multihop Networks and the Impact ofRouting 123
gain matrix as above and b is the Mxl vectorb; == l1[.W].

r l,C;
In [15] it is shown that an all-positive solution for q exists if and only if
13 > r, r being the maximal eigenvalue of A. In that case, q = (131 - At1b .
Thus, we see that the maximum supported y is nearly identical (except for
the strict inequality) to that for no thermal noise since A (and therefore r) is
the same for both solutions. Obviously, the solution for q differs. Note that
as 13 approaches r from above, the components of the vector q become
larger. In a real system there will be some upper limit to the supportable
transmit power of a node. In that case we would increase 13 such that the
transmit powers m;q; are below the necessary threshold. However, we want
13 to be as close to r as feasible in order to achieve something near the
maximal y , since y = G/(13 -1). We do not further explore the best value
for 13 in this paper and simply set it to 1.1. r for all results presented to
achieve near maximum values of SNIR (see [17]).
Note that the SNIR solution derived above is unique to the network
configuration. That is, if all nodes are allowed to be active then the spectral
radius defines the maximum achievable SNIR. Obviously, if this is not
sufficient to meet the needs of the network (on average or individual users)
then allowed active nodes need to be reduced (or we might restrict admission
to the network if SNIR would fall below a threshold [17]). We might ask
however, via route selection, what SNIR can be achieved while still
minimizing the total network power. What we will show in the remaining
sections is that with proper route selection we can achieve near the
maximum achievable SNIR for a given configuration while simultaneously
greatly reducing the total network power utilization.
4. BALANCED SIR/POWER ROUTING

ALGORITHM
The analysis above indicates that for a given set of nodes and traffic
requirements and known path attenuation characteristics we can apply a
given route and determine the achievable SNIR and total network power.
Our goal in this section, and the main result of this paper, is to use the above
analysis as part of an algorithm that finds routes that balance between the
maximum achievable SNIR and the total network power. To achieve this
balance we define the overall network cost function to be minimized as

a . r + (1- a )pIN' where r is the maximal eigenvalue, PIN is the total
network power, and a E [0,1]. An approach similar to Dijkstra's well known
algorithm for generating least-cost routes can be taken (see for example [18,
19]) where the route is grown from a specified node and a labeling scheme is
used to keep track of the least-cost connections examined at each stage.
While the algorithm can be used to find optimal SNIR routes, there are
computationally less expensive approaches. The real advantage of the
algorithm is its ability to trade off between SNIR and power. We first
describe the algorithm, then derive the algorithm complexity, and finally
discuss the algorithm optimality.
It should be noted that a "brute force" method of searching through all
possible routes is not possible for networks with any significant number of
nodes. For example, if we consider a fully connected network (which is
desirable for modeling the mobile environment) it is known that there are
KK-2 routes, where K is the number of nodes ([19]). For K=50 it would
take ~ 1065 years evaluating 1 billion routes per second to examine all
routes.
4.1 Balanced SIR/Power (BSP) Routing Algorithm
The Balanced SIR/Power (BSP) algorithm is defined as follows:

1. Initialize. Identify the BS (node 0) as the starting node, mark it
permanent, set the BS label L(O) = 0, and set p=O (indicating the
most recently added permanent node). Mark all other nodes as
temporary, and set the label value for all temporary nodes to L(i) =
00.
2. Compute Costs. For each temporary node i, consider the cost of
adding it to the permanent network via the most recent addition p, to
the list of permanent nodes. The cost is computed by creating a link
between the temporary node and the permanent node, updating the
packet count for all links between the temporary node and the base
station, and computing the maximum eigenvalue, r, and
corresponding power for the resulting network, PIN' The cost is
then a weighted function a· r + (1- a )PIN .
3. Update Labels. Update the labels and temporary routes for all
temporary nodes that have a lower cost associated with connecting
top.
4. Select Permanent Node. Select the temporary node with the smallest
label value and mark it as permanent and set p equal to this most
recent permanent node. If several nodes qualify then select the node
with the least total network power PTN •
5. Re-compute Labels. Update the labels of all temporary nodes based
on the most recent permanent node selection.
6. Check if Done. If there are no more temporary nodes then stop.
Otherwise, go to Step 2.
Clearly, the cost function is monotonically non-decreasing since the

addition of a temporary node cannot increase the SNIR or decrease the total
network power. Therefore, the algorithm will make choices in the direction
of maximizing SNIR and minimizing total power. Note that Step 5 is unique
to this algorithm and is not needed in least-cost routing algorithms since
typically selection of a permanent node has no impact on the labels for the
remaining temporary nodes (the label values for the permanent nodes change
as well, but since their value no longer impacts the route selection they are
not updated). However in the wireless environment, since the interference
experienced is dependent on the number of packets carried along a route, the
addition of a new permanent node impacts the interference experienced by
all other nodes, including the temporary nodes. Thus, to fairly compare the
previous route selections made for a temporary node, the labels must be
updated based on the new set of permanent nodes.
Also different from the usual least-cost routing algorithm is the
considerable computation associated with calculating label values. Usually
the costs associated with each node pair are known in advance (e.g. distance
or power) and label updates are trivial. Here, however, we cannot compute
the costs in advance. We address algorithm complexity in the next section.
4.2 Algorithm Complexity
We derive the BSP algorithm complexity as follows. For a Knode

network (including the BS) we will iterate K-l times. At iteration k we
calculate the eigenvalue, eigenvector, and total network power associated
with adding each temporary node. There are k-l links at iteration k and
therefore we are finding the eigenvalue and eigenvector, and inverting a (k-
1)x(k-l) matrix. The maximal eigenvalue and associated eigenvector can be
found in C(k-l)2 computations, C a relatively small constant (see [20] for a
number of fast, iterative algorithms for non-negative irreducible matrices).
The matrix inversion is also D((k-li). Once a permanent node is selected
we must recalculate the labels and therefore incur O(k?) computations since
the matrix is now kxk. Thus, each iteration is dominated by K- k, O(e)
computations. Summing over all iterations we have
Lf=l (K -k )k 2 = O(K4 - K2 ) = O(K4) computations as the computational
complexity of the BSP algorithm.
4.3 Optimality Characteristics
The optimality of the algorithm for SIR (a = 1) is somewhat trivially

achieved for a single cell system since a SH network (all nodes directly
connected to the BS) will achieve the optimal SIR solution. To see this,
consider that we require all traffic must flow to the BS on the last hop. With
power control we can achieve an SIR of (K - 2t (and SNIR arbitrarily near
(K - 2t). However, if there are any other active links not connected to the
BS then the SIR is necessarily degraded. Since the first iteration of the
algorithm considers all nodes attached to the BS, the algorithm will
necessarily explore the optimal SIR solution. Therefore the algorithm is
optimal for SIR. Obviously this is a lot of computation to come up with a
trivial solution. As stated above and shown in the next section, the advantage
of the algorithm is its ability to trade off between SIR and power. Results in
the next section show that near optimal SIR solutions may be obtained with
considerably less total network (and therefore average per node) power.
As for optimality in network power, the algorithm is not optimal. In
situations where links experience similar path attenuation characteristics, the
inclusion of a link early on by the algorithm may exclude a better solution
later. Routes that use less power than found via the BSP algorithm with
a = 0 have been found in simulation studies. However, as will be shown,
the algorithm finds routes with significantly lower average power
consumption, as compared to both SH and more typical routing algorithms.
Furthermore, we applied simulated annealing methods (see Section 10.9 of
[21]) to search for a near-optimal route, in terms of the cost function defined
above, for a given node location and shadowing realization. We have found
that for negligible to moderate levels of noise and a ~ 0.9 the BSP algorithm
finds routes that are near optimal.
5. SIMULATION RESULTS
In this section we show the performance of the algorithm on 20 nodes

(excluding the BS) uniformly distributed in a unit circle, averaged over 50
realizations of node position and path loss. Path loss follows a d 4 law, d
being distance. The noise power experienced at each node is arbitrarily set
to 11 . W =10-9 so that performance is interference limited, not noise limited,
and we set J3 = 1.1· r. For comparison, we also show results for a more
traditional least-cost routing based strictly on the path loss between nodes,
ignoring the impact of interference and traffic requirements due to link
selections. We also consider a channel model with log-normal shadowing,
and with fading in one case. Thus, the path loss is proportional to
&2 .1 O~ /10 d- 4 where E accounts for fading (when modeled) as a Rayleigh
distributed random variable with parameter C = 1/../2 (to ensure the average
received power is not biased), and ~ accounts for shadowing as a Gaussian
random variable with zero mean and standard deviation a .
Figure 2) shows the average SNIR and total network power characteristics
as a function of a without shadowing, and Figure 3) presents the results for
the same parameters with a = 8 dB shadowing. The SNIR is normalized to
the maximum SNIR of -1010g(K -2), which does not include the 3 dB
penalty for multihop. Note that the power level for a = 1 corresponds to the
SH power. The performance shown is typical in that for a near 1, the SNIR
is minimally impacted while still achieving substantial power savings over
the SH solution (a = 1). We see that the least-cost routing performs well
when no shadowing is present. It achieves good SNIR (within about 2 dB of
SH) while using about 10 dB less power than the single hop solution. The
BSP achieves higher SNIR for almost all values of a at the expense of about
3 dB more power than the least-cost route (but still -7 dB less power then
SH). However, when shadowing is present, as shown in Figure 3, on
average the BSP route is significantly better. For a wide range of a the BSP
route achieves greater SNIR with less power. With 8 dB shadowing and
a = 0.9 we achieve on average 9 dB better SNIR than least-cost routing and
within -0.5 dB of optimal SNIR while using 4 dB less power than the least-
cost route and 15 dB less power then SH.
Figure 4) shows the performance for 20 nodes and fixed a = 0.9 as a
function of shadowing standard deviation (in dB) without fading. We see
that the BSP algorithm's SIR performance is essentially independent of
shadowing, whereas, the greater the shadowing the better the algorithm is
able to take advantage of routing to decrease total network power

consumption. Figure 5) includes the effects of fading. We see that the BSP
algorithm SIR performance remains stable and near optimal. In addition, the
BSP power profile is nearly identical to the case without fading. Given the
difficulty of tracking fading rapidly enough to facilitate central routing, these
results indicate little advantage. We also see that the performance of the
least-cost route has degraded both in terms of a slight decrease in SIR and
increase power required.
SIR Characteristics
i: 8
v-
---------------------------------------------------------------
I _.... SSP r
z ·3 Route
least-cost Route
c·
.4 L--::L---::'=--_="=-_:-L-_-:"::-_-:'-:-~~2=1n::::od:::es N=o=S=ha;:::dOW\='=ng~
o ~ u ~ u ~ u U U M
alpha
Power Characteristics
.251-,---.--.,---.---y--.--;=~~;:r:;:====il
·30
I SSP Route
--- least-cost Route
21 nodes, No Shadowing
'"
E
~ ·35
~ T-----------------------------------------------------------
-40
-450L---::0~.1-~0~.2-~0~.3-~0~.4~-0~.5~~0.6~~0.7~~0.~8-~0~.9-~
alpha
Figure 2.. BSP performance as a function of a. for 20 nodes, no shadowing.

SIR Characteristics
0:
Z
C/)
-4
i!
1
-6
z -8
_________________________________________ jI __ .. 21~;:st~c~~e Route
nodes. 8 dB ShadO\\ing
-10
o ~ ~ ~ U M M ~ M M
a~ha
-20 1-.-----,---,-----,------,----rr==E;;::::;;:~====il
BSP Route
-25 1 ....... least-cost Route
21 nodes, 8 dB Shadowing
-30
i -35 ----------------------------------------------------------------
~
:::~ o ~ ~ ~ U M M ~ M M
alpha
Figure 3. BSP performance as a function of a for 20 nodes, 8 dB shadowing.
SIR Characteristics
L: -_I-----~:=----------------I------------~------.~-----._
~ --- Least-cost Route - ____ _
~ -15 20 nodes, Shadowing & No Fading
-200~----=------'-----:------:-----1:'::0------,J12
Sigma (dB)
_I
-25!r==-=;;;:;~====:::r:==~c::;==:.::.::.....-.------,--1
BSP Route
--- Least-cost Route
I
-30 20 nodes, Shadowing & No Fading
~ -35 -----------------------------------------
J -40 --------------------
-450L-----=------'----~---~----1~O---~12·
Sigma (dB)
Figure 4. Comparison ofBSP and Least-cost SNIR and Power characteristics versus
shadowing (no fading) for a fixed at 0.9.
SIR Characteristics
I:: --;----~:=--------------i----------------------
8 -15
'ij --- Least-cost Route
20 nodes, Shadowing & Fading ............. ...
z ,
·20'---------"-------"-----'----'------'-----'
o 10 12
Sigma (dB)
.251r==~~====='===:::;-,----,---r--1
1
___ ~::st~c~~eROut. I /"------------
20 nodes, Shadowing & Fading .., ....",'
S~ ~"
S -
~ -35 1......,.--:..::.--::-::.:--;..---------------------------------"-""-"-""__
-40 L __-'---__--"--__-'-__-'--__==::::::::::=:l
o 10 12
Sigma (dB)
Figure 5. Comparison ofBSP and Least-cost SNIR and Power characteristics versus
shadowing with fading for IX fixed at 0.9
6. CONCLUSIONS
This paper considered a CMHN architecture using CDMA and
investigated the impact of routing on the SNIR and power consumption
characteristics. We developed a route cost based on the SNIR and power
consumption characteristics and presented a new polynomial complexity
centralized routing algorithm to balance between SNIR and total network
power consumption. The algorithm shows that route selection has a
significant impact on the achievable SNIR and power consumption of the
network and that significant SNIR and power benefits are achievable,
especially when lognormal shadowing and fading are present. Hence,
considerable advantage can be achieved by combining network layer routing
decisions with physical layer characteristics. Furthermore, routing decisions
that do not account for the physical layer may result in significantly
degraded performance.
7. REFERENCES
1) L. Ying-Oar and H. Yu-Ching, "Multihop cellular: a new architecture for wireless

communications," presented at Proceedings IEEE INFOCOM 2000. Conference on
Computer Communications. Nineteenth Annual Joint Conference of the IEEE Computer
and Communications Societies (Cat. No.00CH37064). IEEE. Part voi.3, 2000, pp.1273-
82 voi.3. Piscataway, NJ, USA., 2000.
2) E. S. Sousa and J. A. Silvester, "Spreading code protocols for distributed spread-
spectrum packet radio networks," IEEE Transactions on Communications, vol. 36, pp.
272-81, 1988.
3) K. S. Gilhousen, I. M. Jacobs, R. Padovani, A. J. Viterbi, L. A. Weaver, Jr., and C. E.
Wheatley, III, "On the capacity of a cellular COMA system," IEEE Transactions on
Vehicular Technology, vol. 40, pp. 303-12, 1991.
4) R. L. Pickholtz, L. B. Milstein, and o. L. Schilling, "Spread spectrum for mobile
communications," IEEE Transactions on Vehicular Technology, vol. 40, pp. 313-22,
1991.
5) A. J. Viterbi, A. M. Viterbi, and E. Zehavi, "Other-cell interference in cellular power-
controlled COMA," IEEE Transactions on Communications, vol. 42, pp. 1501-4,1994.
6) A. M. Viterbi and A. J. Viterbi, "Erlang capacity ofa power controlled COMA system,"
IEEE Journal on Selected Areas in Communications, vol. II, pp. 892-900, 1993.
7) S. Ariyavisitakul, "SIR-based power control in a COMA system," presented at
GLOBECOM '92. Communication for Global Users. IEEE Global Telecommunications
Conference. Conference Record. (Cat. No.92CH3130-2). IEEE. 1992, pp.868-73 vol.2.
New York, NY, USA., 1992.
8) S. Ariyavisitakul, "Autonomous SIR-based power control for a TDMA radio system,"
International Journal of Wireless Information Networks, vol. I, pp. 187-98,1994.
9) S. Seo, T. Oohi, and F. Adachi, "SIR-based transmit power control of reverse link for
coherent OS-COMA mobile radio," IEICE Transactions on Communications, vol. E81-
B, pp. 1508-16,1998.
10) J. Zander, "Performance of optimum transmitter power control in cellular radio
systems," IEEE Transactions on Vehicular Technology, vol. 41, pp. 57-62, 1992.
11) S. V. Hanly, "An algorithm for combined cell-site selection and power control to
maximize cellular spread spectrum capacity," IEEE Journal on Selected Areas in
Communications, vol. 13, pp. 1332-40,1995.
12) S. V. Hanly, "Capacity and power control in spread spectrum macrodiversity radio
networks," IEEE Transactions on Communications, vol. 44, pp. 247-56,1996.
13) R. W. Nettleton and H. Alavi, "Power control for a spread spectrum cellular mobile
radio system," presented at 33rd IEEE Vehicular Technology Conference. IEEE. 1983,
pp.242-6. New York, NY, USA., 1983.
14) A. J. Viterbi, CDMA : principles of spread spectrum communication. Reading, Mass.:
Addison-Wesley Pub. Co., 1995.
15) E. Seneta, Non-negative matrices and Markov chains, 2nd ed. ed. New York: Springer-
Verlag, 1981.
16) H. Minc, Nonnegative matrices. New York: Wiley, 1988.
17) N. Bambos, "Toward power-sensitive network architectures in wireless
communications: concepts, issues, and design aspects," IEEE Personal
Communications, vol. 5, pp. 50-9, 1998.
18) R. H. Jones and N. C. Steele, Mathematics in communication theory. Chichester, West
Sussex, England New York: E. Horwood; Halsted Press, 1989.
19) 1. A. Bondy and U. S. R. Murty, Graph theory with applications. New York: American
Elsevier Pub. Co., 1976.
20) W. Bunse, "A class of diagonal transformation methods for the computation of the
spectral radius of a nonnegative irreducible matrix," SIAM Journal on Numerical
Analysis, vol. 18, pp. 693-704, 1981.
21) W. H. Press, Numerical recipes in C .' the art of scientific computing, 2nd ed. ed.
Cambridge [Cambridgeshire] ; New York: Cambridge University Press, 1992.
Terminal Migration Model in Which Cell Dwell Time
is Defined by Different Probability Distributions in
Different Cells
Hirotoshi Hidakat *, Kazuyoshi Saitoht , Noriteru Shinagawat , and Takehiko

Kobayashitt
tYRP Mobile Telecommunications Key Technology Research Laboratories Co., Ltd.
* Presently, Sumitomo Electric Industries Co., Ltd.
hidaka-hirotoshi@sei.cojp
ttTokyo Denki University
koba@c.dendai.acjp
Abstract: In evaluating the teletraffic of mobile communication networks, it is important

to model the motion of terminals. In the previous migration model, mobility
characteristics of terminals, such as cell dwell time, have been expressed by a
single probability distribution. In this paper, we discuss the modeling of the
cell dwell time of terminals in each cell. Using measured data we show that
cell dwell time differs from cell to cell and follows log-normal distributions
rather than conventional exponential distributions. We also use computer
simulations to compare the results predicted by a migration model in which
cell dwell time is defined by probability distributions that differ from cell to
cell with the results predicted by the conventional migration.
Key words: terminal migration model, cell dwell time, mobile terminal, probability
distribution, mobile communication networks
1. INTRODUCTION
The rapid increase in the use of cellular telephones and the great
improvements of mobile phone services have created a demand for networks
that can provide various kinds of fast and stable mobile multimedia
communications services to users moving at high speeds. Because a terminal
can move from one cell to another during a call, it is necessary for the
designers and operators of such networks to consider the mobility
characteristics of terminals when they evaluate the network teletraffic. There
133
Multiaccess, Mobility and Teletrajfic for Wireless Communications, Volume 6, 133-142.
134 Hirotoshi Hidaka, et al.
have therefore been many studies on the modeling of the behavior of

terminals within cellular systems, and "cell dwell time" - the time that a
terminal remains in each of the cells - has been modeled by using various
probability distributions: an exponential distribution [1], a sum of hyper-
exponential distribution [2], a generally distribution [3], a generalized
gamma distribution [4], and various others [5]-[7]. To model the mobility
characteristics of terminals more realistically, we have used GPS receivers to
measure the motion of various kinds of vehicles and have also evaluated cell
dwell time and channel occupancy time in virtual cellular systems [8]-[10].
We found as a result that the cell dwell time of taxis follows a log-normal
distribution, which is a long-tailed distribution, rather than an exponential
distribution, which has generally been used to model the cell dwell time of
terminals.
Until now, we and other researchers simulating migration models of
terminals have used distributions having the same mean and standard
deviation in all cells, assuming that a terminal's motion is independent of the
terminal's location. The actual mobility characteristics of a terminal,
however, will usually vary from cell to cell. Urban areas, for example, and
the areas surrounding train stations will contain cells in which the dwell time
will be long and in which the number of channels allocated should be greater
than that allocated in adjacent cells. Future mobile multimedia
communication systems are thus expected to use complicated cell
configurations arranged according to the teletraffic density in each cell. It is
therefore important that those who design the networks take into
consideration the teletraffic that would be generated in each cell. In the work
reported in this paper, we calculated the cell dwell time for every cell by
using the taxi data we had already obtained so far. In addition, we use
computer simulations to compare the results predicted by a migration model
in which cell dwell time differs from cell to cell with the results of
measurements and the results predicted by the conventional model.
2. MEAN AND STANDARD DEVIATION OF CELL

DWELL TIME FOR EACH CELL
We chose to evaluate the cell dwell time for taxis driving in the
Kurihama district of Yokosuka in Japan, because it is the largest body of
data we have (about 60 days). We analyzed the 5-km-square area the center
of which is Kurihama Station, dividing the area into cells 200 or 1,000 m
square and calculating the dwell time in each cell.
Terminal Migration Model 135
Figures l(a) and l(b) show the mean and standard deviation of cell dwell
time for each cell when the cell size is 200 m. We can see that both the mean
and standard deviation are large in the cells surrounding stations, presumably
because taxi drivers there are waiting for passengers or taking a break. On
the general roads between stations, on the other hand, the cell dwell time is
small because the taxis are only passing through the cells there. Figures 2(a)
and 2(b) show the mean and standard deviation of cell dwell time when the
cell size is 1,000 m. The difference from cell to cell obviously becomes
smaller when cells are larger. As described above, in the circumstance of
micro-cell configuration, the mean and standard deviation of cell dwell time
would differ greatly between cells.
Kita-kurihama Station Kuriharre Station
(a) Mean
Kita-kurihama Station
(b) Standard deviation

Figure 1. Mean and standard deviation of cell dwell time in a service area with 200-m-square
cells.
(a) Mean
(b) Standard deviation

Figure 2. Mean and standard deviation of cell dwell time in a service area
with IOOO-m-square cells.
3. DISTRIBUTION FOLLOWED BY CELL DWELL

TIME FOR EACH CELL
Up to now, the cell dwell time of a terminal has been modeled by one
probability distribution having the same mean and standard deviation in all
the cells. In particular, an exponential distribution, which is easy to
understand the phenomenon and calculate, has been used frequently. It was
shown in Section 2, however, the mean and standard deviation of the cell
dwell time differs greatly between cells in the micro-cell configuration. This
means that the probability distribution followed by the cell dwell time also
differs greatly between cells. This section discusses the probability

distribution approximating the cell dwell time calculated for each cell.
As shown in Fig. 1, a 5-km-square service area contains 625 cells (25 x
25) when the cell size is 200 m. Since the probability distributions followed
by the cell dwell time are not easily evaluated for all cells, we consider only
three typical cells. The first (called "cell A") is the cell containing Kurihama
Station, for which both mean and standard deviation are larger than in any
other cell. The second cell ("cell B") is one of the cells surrounding that
station, for which the mean is relatively small but standard deviation is large.
The third cell ("cell C") is one of the cells along one of the general roads, for
which both the mean and standard deviation are small. The mean and
standard deviation of cell dwell time for each of the three cells are listed in
Table 1.
The mean f.1, standard deviation 0; and regression coefficient R2 obtained
when we used the least-squares method to approximate the probability
distribution of cell dwell time by exponential, normal, and log-normal
distributions are listed in Table 2, where we see that for all cells the largest
regression coefficients were obtained when the actual distributions were
approximated by log-normal distributions. Especially, for cell A and B,
because the standard deviation is much greater than the mean value as shown
in Table 1, the cell dwell time of those cells followed long-tailed log-normal
distributions well.
Table 1. Mean and standard deviation of cell dwell time (sec.) for three typical cells
Mean Standard deviation

CellA 194 300
Cell B 47.8 110
Cell C 12.2 6.34
Table 2. Parameters of distributions approximating the cell dwell time for three typical cells
Exponential Normal Log-normal

z
PDF
f(x) =lex~- x f f(x) =Texpf - ~f f(x) = lr:- expf- (I!!~-,Il)'f
f..J= 170 f..J= 107, a= 248 f..J = 4.49, a= 1.34
Cell A
R2 = 0.990 R2 = 0.981 R2 = 0.998
f..J=24.3 f..J = 16.6, a= 30.0 f..J = 2.48, a= 1.67
Cell B
R2 = 0.966 R2 = 0.941 R2 = 0.995
f..J= 12.0 f..J = 11.3, a= 2.43 f..J= 2.41, a= 0.211
Cell C
R2 = 0.939 R2 = 0.984 R2 = 0.992
4. SIMULATION REFLECTING RESULTS OF

ANALYSIS
In the conventional model of tenninal migration, when a tenninal enters a
new cell the cell dwell time of the new cell has been detennined according to
one probability distribution (usually an exponential distribution) having the
same mean and standard deviation in all cells. The previous section, however,
showed that the mean and standard deviation of the probability distributions
followed by the dwell time in various cells differed greatly. In this section
we therefore define the cell dwell time by the probability distribution unique
to each cell and report the results obtained when we used computer
simulations to compare the cell dwell times predicted by this model with
those predicted by the conventional migration model.
4.1 Simulation Model

In the simulation we considered a configuration of 10 x 10 two-
dimensional torus cells as shown in Fig. 3, where the cells in the areas
labeled A, B, and C correspond to the cells A, B, and C described in Section
3. We used the exponential distribution, which has usually been used, and
the log-nonnal distribution, for which the regression coefficients reported in
Section 3 were the largest. The simulation in which the exponential
distribution was used is called "Type I" here, and that in which the log-
nonnal distribution was used is called "Type II." And to compare the results
of those simulations with one using the conventional migration model, we
also ran a simulation in which all I 00 cells were assumed like cell B in Table
2 and the cell dwell time of all cells followed an exponential distribution
with a mean of 24.3. This simulation is called "Type III." The direction
taken by a tenninal when moving into an adjacent cell ("straight," "right,"
"left," or "reverse") was detennined by probability values that were based on
vehicle measurements [9] and are listed in Table 3.
4.2 Simulation Results

The mean and standard deviation of a tenninal' s cell dwell time in each
of the three simulations are listed in Table 4, and cumulative probability
distribution of cell dwell time in each simulation is shown in Fig. 4.
Although the exponential distribution was used in both Type I and Type III
simulations, the characteristics of cell dwell time differ greatly between the
two simulations. In the Type I simulation the standard deviation is much
greater than the mean because the mean of dwell time given in each cell
differs between cells as shown in Fig. 3. In the Type III simulation, on the
other hand, the mean is of course nearly equal to the standard deviation
because the cell dwell time was determined by one exponential distribution
for which the mean was the same in all the cells.
We also ran simulations in which the cumulative probability of cell dwell
time was approximated by exponential and log-normal distributions
estimated using the least-squares method. The results are listed in Table 5.
The results listed for the Type II simulations clearly followed a log-normal
distribution more closely than they did an exponential distribution. In the
Type III simulations, although the regression coefficient for the log-normal
distribution is large, the mean approximated by the exponential distribution
is so close to that of original data that the data follows the exponential
distribution very well. With regard to the Type I simulation, it is worthy of
attention that the results follow the log-normal distribution well though the
dwell time in each cell was given by the exponential distribution. This seems
to result from the effect of long-tailed characteristic because, in the cell A,
random exponential numbers, of which the mean is great, may be generated.
Figure 3. Simulation model.
Table 3. Transition probability
Straight 0.3
Right 0.3
Left 0.3
Reverse 0.1
Table 4. Mean and standard deviation of the cell dwell time ofa terminal in the simulation
Used distribution Mean Standard deviation

Type I Exponential 18.6 44.8
Type II Log-normal 21.9 90.2
Type III Exponential 24.4 24.6
1.0 I-------:::;~~---I
0.8
0.6
lJ..
C>
o
0.4
-0- Type I
- . - Type II
-x- Type III
10 100 1.000
Celldwelltin e [5]
Figure. 4. Cumulative distribution ofthe cell dwell time of a terminal in the simulation.
Table 5. Parameters of distributions approximating the cell dwell time for three types.
Used distribution Exponential Log-nonnal

f.1 = 13.6 f.1 = 2.06, a= 1.24
Type I Exponential
R2 = 0.992 R2 = 0.997
f.1 = 12.5 f.1 = 2.34, a= 0.34
Type II Log-nonnal
R2 = 0.952 R2 = 0.991
f.1 = 23.9 f.1 = 2.74, a= 1.03
Type III Exponential
R2 = 0.999 R2 = 0.993
In Ref. 8 we showed that the actual cell dwell time measured for a taxi
followed a log-normal distribution rather than exponential distribution,
seemingly because the cell dwell time of taxis is large around a station and is
small in other places. In other words, the cell dwell time differs between
cells. Although the results reported in this paper may be applicable only for
taxis, cell dwell time will generally vary from cell to cell. It is therefore
necessary to analyze the mobility characteristics of other platforms for every
cell in the same way that it is necessary to analyze them the mobility
characteristics of taxis for every cell.
5. CONCLUSIONS
In the conventional model of terminal migration the cell dwell time has
been expressed by one probability distribution which is used commonly in
all the cells, assuming that a terminal moves at random independent of
locality. In actual communication systems, however, the dwell time in each
cell is expected to differ greatly. This paper analyzed the cell dwell time of
taxis we have measured so far and it showed that the mean and standard
deviation of cell dwell time are large in certain places, such as those within
or close to train stations, and are small in cells along general roads. It also
showed that the dwell time in each cell followed a log-normal distribution
rather than a conventionally used exponential distribution.
We also considered a migration model in which the cell dwell time is
defined by the different probability distribution from each cell, and we used
computer simulations to evaluate the cell dwell time of a terminal. Even if
the dwell time in all cells was given by an exponential distribution, the
results of those simulations showed that the cell dwell time followed a long-
tailed log-normal distribution. Although these results may be applicable only
to taxis, the methodology described in this paper-taking account of cell
dwell time following different probability distributions in different cells-
offers a realistic and effective way to evaluate teletraffic characteristics in
future mobile multimedia communication systems.
REFERENCES
[1] D. Hong and S. S. Rappaport, "Traffic model and performance analysis for cellular mobile
radio telephone systems with prioritized and non-prioritized handoffprocedures," IEEE
Trans. Veh. Tech., vol. VT-35, no. 3, pp. 77-92, Aug. 1986.
[2] P. Orlik and S. S. Rappaport, "Traffic performance and mobility modeling of cellular
communications with mixed platforms and highly variable mobilities," Proc. IEEE, vol. 86,
no. 7, pp. 1464-1479, July 1998.
[3] Y. Fang and I.Chlamtac, "Teletraffic analysis and mobility modeling of PCS networks,"
IEEE Trans. Commun., vol. 47, no. 7, pp. 1062-1071, July 1999.
[4] M. M. Zonoozi and P. Dassanayake, "User mobility modeling and characterization of
mobility patterns," IEEE 1. Select. Areas Commun., vol. IS, no. 7, Sept. 1997.
[5] F. Khan and D. Zeghlache, "Effect of cell residence time distribution on the performance
of cellular mobile networks," Proc. IEEE VTC'97, pp. 949-953, May 1997.
[6] K. K. Leung, W. A. Massey, and W. Whitt, "Traffic models for wireless communication
networks," IEEE J. Select. Areas Commun., vol. 12, no. 8, pp. 1353-1364,1994.
[7] Bar-Noy and I. Kessler, "Mobile users: to update or not to update?," Proc. INFOCOM'94,
pp. 570-576, 1994.
142 Hirotoshi Hidaka, et ai.
[8] T. Kobayashi, N. Shinagawa, and Y. Watanabe, "Vehicle mobility characterization based

on measurements and its application to cellular communication systems," IEICE Trans.
Commun., vol. E82-B, no. 12, pp. 2055-2060, Dec. 1999.
[9]K. Saitoh, H. Hidaka, N. Shinagawa, and T. Kobayashi, "Vehicle motion in large and
small cities and teletraffic characterization in cellular communication systems," IEICE
Trans. Commun., vol. E84-B, pp. 805-813, Apr. 2001.
[10] H. Hidaka, K. Saitoh, N. Shinagawa, and T. Kobayashi, "Teletraffic characteristics of
cellular communication for different types of vehicle motion," IEICE Trans. Commun.,
vol. E84-B, pp. 558-565, Mar. 2001.
Concatenated Location Management
Koji SASADA, Satoshi HIYAMA, and Masami YABUSAKI

Network Laboratories, NIT DaCoMa, Inc.
Abstract: Conventional mobile communication systems perform location management

for each mobile terminal (MT). But when many MTs share the same mobility
characteristics (for example when the MTs are on the same train), the number
of signals needed for location management can be reduced. We described an
effective concatenated location management that handles groups of MTs.
Using this scheme, the train (as one example) recognizes the MTs as
passengers, and updates the location information for all such MTs when it
updates its own location.
Key words: location management, concatenation, paging
1. INTRODUCTION
The recent increase in the number of mobile communication subscribers

is imposing huge loads on the system. This trend will continue with the
commercialization of the 3rd generation mobile communication system
(IMT-2000); IP based IMT Platform (IP2)[1] has been proposed as a target
for the next mobile communication system beyond IMT -2000.
In the conventional approach to location management, the mobile
terminals (MTs) are considered to move independently and so register their
location with the network individually[2][3][4]. In UMTS[4], an MT sends a
location area update request signal when it moves into a new location area,
and it sends a routing area update request signal when it moves into a new
routing area. We can reduce the overhead significantly by considering that in
many cases groups of MTs move as one for some distance. This paper
describes an effective Concatenated Location Management (CLM) that does
exactly this.
143
144 Koji SASADA, Satoshi HIYAMA, and Masami YABUSAKl
2. PROBLEMS
In the conventional mobile communication network, a mobile

communication operator deploys location registers (LRs) in the network and
provides mobile communication services by managing the user's location
information through hislher telephone number. The operator defines a
location registration area as covering multiple base stations (BSs), and
manages the area as location information. When an MT moves from one
location registration area to another, location information is updated as
follows.
(1) BS broadcasts location registration area information periodically.

(2) MT receives the location registration area information, and MT
registers its location to the network if the area information differs
from that received last.
(3) The network updates MT's location information in LR.
With this approach, a group of MTs that moves simultaneously into a new
location registration area tries to register their location simultaneously as
shown as Fig. 1. This is a serious problem because it is impractical to
provide sufficient resources to handle the flood of signals thus generated.
location registnuion area A location rcgiStrlllion area B
Figure 1. Problem with conventional location registration

3. CONCATENATED LOCATION MANAGEMENT
Since we must anticipate that many MTs will share the same moving
platform, such as a train, they will have the same movement characteristics,
and a more effective location registration approach appears possible.
Therefore, we propose Concatenated Location Management (CLM), which
enables location information to be updated without individual location
registration.
The logical solution is to equip the train with an Intermediate Radio
Station (IRS) and update the location information of all MTs with one action.
This relationship between the MT and the IRS is called "Concatenation" in
this paper. There are two ways of implementing this concatenation: IRS
(method 1) and the network (method 2).
In method 1, the IRS updates the location information of all MTs that are
traveling with the IRS. In method 2, the MTs update their location
information to indicate that they are traveling with the IRS when they join
the IRS. The network is responsible for updating the location information of
the MTs moving with the IRS.
Fig. 2 shows the basic procedures of method 1. The details of method 1
are as follows.
location rcgistr31ion area A

->Iocalion registration area B (S)
location registration area B
Figure 2. Concatenated location registration method I

(1) IRS broadcasts IRS information periodically.

(2) When MT enters the train, it receives IRS information. It recognizes
that it is moving with the IRS, and it sends a concatenation location
registration request to IRS.
(3) IRS sends a concatenation location registration response to the MT.
This informs the MT that initial concatenation phase is completed.
(4) When the IRS enters a new location registration area, it sends a
location registration request to the network that includes MT
information.
(5) The network updates the location information of the MTs on the
train.
Fig. 3 shows the procedures for method 2. The details of the method are as
follows.
MT#2;
location regim-alion ar~a A
·>concatenal;ng to IRS #1 (3)
Figure 3. Concatenated location registration method 2
(1) IRS broadcasts IRS information periodically.

(2) When an MT enters the train, it receives IRS information. It
recognizes that it is moving with IRS and sends a concatenation
location registration request to the network that contains the IRS

information.
(3) The network manages the concatenation between the IRS and MT,
and sends a concatenation location registration response to the MT.
This informs the MT that initial concatenation phase is completed.
(4) When IRS enters a new location registration area, it sends a location
registration request to the network.
In method 1, the IRS sends one location registration request to the

network on behalf of the MTs in the train. When the IRS changes its location
registration area, it sends only one signal for location registration including
the information of concatenated MTs. This greatly reduces the number of
signals compared to the conventional location registration method. Moreover,
when an MT joins the IRS, the system is not impacted because signals are
transmitted only between the MT and IRS. However, each location
registration request signal includes information on all MTs traveling with the
IRS so the signal may be quite large.
In method 2, the IRS sends one location registration request to the
network on behalf of the MTs as is true in method 1. When the IRS changes
its location registration area, it sends only one location registration signal.
Moreover, this signal is much smaller than the comparable signal in method
1 because it contains only information on the IRS. However, each MT must
send a signal to notify the network of concatenation registration or
concatenation breakage. Therefore, if the average time spent by the MT
traveling with the IRS is small, there is a commensurate increase in the
number of concatenation registration and breakage signals from the MTs.
This increase in traffic may overwhelm the benefits of the smaller location
registration signals.
4. PAGING CONTROL
In the conventional method, the network pages an MT from all BSs in the
registered location registration area of the MT. The network identifies the
BS where the MT visits when the MT responds to the paging signal. In this
section, we propose paging control for CLM.
The conventional method has the network page the MT directly, and this
idea can also be adopted for CLM (alternative 1: AI). An alternative is for
the network to notify the IRS of the paging request, and have the IRS page
the MT. There are two responses possible; one is that the MT responds to the
148 Koji SASADA, Satoshi HIYAMA, and Masami YABUSAKI
network directly (alternative 2: A2), and the other is that the MT responds to
the network via the IRS (alternative 3: A3). Fig. 4 shows each alternative.
Figure 4. Paging procedures
In method 1, the IRS notifies the network of the MT information together

with location registration, but the relation between the IRS and the MTs is
only temporary. Therefore, it is unknown whether the MT is traveling with
the IRS at the time of paging, so A2 and A3 are not applicable without
additional procedures. To ensure MT's concatenation to the IRS, a
mechanism like notification to the network of concatenation breakage is
necessary.
As mentioned above, paging is performed from the BSs in the location
registration area. When an incoming call to MT occurs, the network needs to
identify the location registration area that the MT is visiting. In AI, the
network identifies the location registration area registered for the MT, and
pages it within the area. In A2 and A3, on the other hand, the network
identifies the IRS to which the MT is concatenated and the location
registration area registered for the IRS. The network then notifies the IRS of
the incoming call to the MT.
Concerning the number of signals, the network and the IRS can aggregate
paging signals to MTs and paging response signals from MTs in A2 and A3
because they are transmitted via the IRS. If multiple calls are received by the
network for MTs under the same IRS, the network can transmit only one
paging signal to the IRS indicating the MTs' information (Fig. 5). Moreover,
in A3 the IRS transmits just one paging response signal to the network as the
responses from the MTs (Fig. 6).
Figure 5. Aggregation of paging signals Figure 6. Aggregation of paging response signals
5. LOCATION REGISTER CONFIGURATION
There are two ways to manage location information: Bl and B2. In Bl,
the network manages the conventional location registration area of the MT.
In B2, the network identifies the location registration area through the
concatenation relationship.
B 1 requires only one database that holds the location information of the
IRS and all MTs as shown in Fig. 7. This enables the network to identify the
MT's location registration area from the table and page the MT directly. The
network does not manage the concatenation between the MT and IRS so
there is no need to notify the location register about changes in the
concatenation relationships.
B2 requires two databases as shown in Fig. 7. This figure shows that MT
#1 is concatenated to IRS #1, and that IRS #1 is in location registration area
A. Clearly MT #1 is visiting location registration area A. The network uses
the concatenation relationship when paging the MT. Note that the location
register must be informed of any change in the concatenation relationships.
location table for Mrs
location infonnation
location table for Mrs and IRSs
concatenating to IR...o;; #1
10 location infonnation
location registration area B i
MT #1 location registration area A
i················································..
.,i
MT #2 location registration area B
IRS #1 location registration area A location infonnation
locationrc&istcr
location registration area A
(alternative B 1) (alternative B2)
Figure 7. Location register configuration

B 1 is basically the same as the conventional approach, so movement of

the IRS demands that the location information of all concatenated MTs must
be updated. B2 is simpler because the location information of all
concatenated MTs does not need to be updated when the IRS moves across a
location registration boundary. On the other hand, the location registration
area where the MT visits is not indicated directly. This incurs some overhead
in identifying the concatenated IRS and determining its location information.
6. RESULT
As an example of realizing a system using the methods and alternatives

mentioned above, Fig. 8 shows the procedures used when applying method 2
for concatenated location registration, A2 for paging control, and B2 for
location register configuration.
Figure 8. Overview of example procedure
We conducted simulations to evaluate the procedures shown in Figure 8.

The simulations assumed that 1,323 MTs were aboard the Japanese express
train called the Shinkansen, which runs for 1,174.9km at speeds of up to
285km/h. The conditions assumed are shown in Table 1 and Fig. 9. The
result is that eLM method 2 could reduce the number of signals by 91.2%
compared to the conventional location registration scheme (Fig. 10). This is
because the express runs long distances with few stops.
Table 1. Simulation conditions
passenger model shown in Fig. 9

max. passenger number 1,323
MTs per passenger 1
location registration interval 20km
total distance 1,174.9km
1.400
1.200
1.000
800
600
400
200
o
Shin- Okayama Hiroshima KoItUlO Hakala
ToIt)'O Yoltohama Nagoya Kyoco
0.0 28.8 366.0 513.6 552.6 732.9 894.2 1107.7 1174.9
Distan<. from slart (Icrn]
Figure 9. Passenger model
20.000
18.000
16.000
-il14.ooo
Iii.
;: 12.000
r::
o
6,000
4,000
2,000
o
s.....
KjOIO ow. Obyamo W..... ima K<>I<u.. llabla
Figure 10. Simulation result

7. REQUIRED RADIO FUNCTIONALITIES
This section describes the key radio functions.
(1) broadcast from IRS

IRS broadcasts information needed for MT to perform concatenated
location registration. This broadcast includes IRS identification.
(2) concatenated location registration and cancellation

When IRS manages MT registration (as in CLM method 1), the MT
performs concatenated location registration to notify the IRS of
concatenation. When the concatenation is broken, the MT performs
concatenated location registration cancellation to the IRS. If the network
manages the relation between MT and IRS (as in CLM method 2), the MT
performs concatenated location registration to notify the network of
concatenation to the IRS. When the concatenation is broken, the MT
performs concatenated location registration cancellation to the network.
(3) paging within IRS

It is can be considered that the network notifies MT of incoming call via
concatenated IRS. At that time, the IRS performs paging.
(4) location registration by IRS

IRS registers its location to the network. In CLM method 1 the IRS
manages the registration of MTs, so it needs to notify the network of the
registered MTs' information at the time it registers its location. In CLM
method 2, the network manages the relation between the MTs and IRS, so
the network can determine MT location information from IRS location
registration.
(5) radio channel used in IRS

Transmitted signals between MT and IRS mentioned above are useful
within the IRS. Therefore, a radio frequency that does not affect signals
transmitted between BS and MT/IRS should be used as the radio channel
linking MT to IRS.
8. CONCLUSIONS
Considering that multiple Mobile Terminals may share the same

movement characteristics, we proposed the Concatenated Location
Management scheme to reduce the number of signals required for location
management. CLM requires that an Intermediate Radio Station (IRS) be

placed on the common moving platform. We call relation between the IRS
and the MTs concatenation.
We consider two methods for location registration control and determine
their strengths and weaknesses. Paging control is another functionality that
must be supported efficiently and we introduce and analyze three
alternatives. Location registration configuration is the third important
function and we introduce two alternatives. These techniques can be
combined in many different ways and the optimum combination depends on
the network situation. As an example, we described simulations that
examined many MTs on a limited express train. The results showed that
CLM method 2 could reduce the number of location registration signals by
91.2%.
REFERENCES
[1] H. Yumiba et ai, "IP-Based IMT Network Platform," IEEE Personal Communications,
pp.18-23, Oct. 2001.
[2] "Location registration procedures (GSM 03.12 version 7.0.0 Release 1998)," ETSI, Aug.
1999.
[3] "Location Registration Control," Personal digital cellular telecommunication system
ARIB standard RCR STD-27H, pp.889-895, Feb. 1999.
[4] "General Packet Radio Service (GPRS) Service description (3GPP TS 23.060 version
3.10.0 Release 1999)," ETSI, Jan. 2001.
HANDOFF SCHEME IMPROVEMENT IN WIRELESS
NETWORKS
Anna Hac and Yongcan Zhang

University of Hawaii at Manoa, Honolulu, Hawaii 96822
Abstract
Chaining followed by make-break algorithm reduces signaling traffic in
wireless ATM network and improves the efficiency of the virtual channel. We
propose a new handoff call management scheme which complies with existing
ATM signaling standard and its implementation leaves commercially ATM com-
ponents unaffected. This method considers traffic condition when chaining and it
is easy to implement in the chaining followed by the make-break scheme.
Keywords:
wireless ATM networks, routing, handoff algorithms.
1. INTRODUCTION
Wireless data services use small-coverage high-bandwidth data networks
such as IEEE 802.11 whenever they are available, and switch to an overlay ser-
vice such as the General Packet Radio Service (GPRS) network with low
bandwidth when the coverage of a wireless local area network (WLAN) is not
available [3].
From the service point of view, ATM (Asynchronous Transfer Mode) com-
bines both the data and multimedia information into the wired networks while
scaling well from backbones to the customer premises networks. In wireless
ATM (W ATM) networks, end user devices are connected to switches via wired or
wireless channels. The switch is responsible for establishing connections with the
fixed infrastructure network component, either through wired or wireless channel.
A mobile end user establishes a virtual circuit (YC) to communicate with another
end user (either mobile or ATM end user) [I]. When the mobile end user moves
from one AP (access point) to another AP, a handoff is required. To minimize the
interruption of cell transport, an efficient switching of the active YCs from the old
data path to the new data path is needed. Also the switching should be fast
enough to make the new YCs available to the mobile users.
When a handoff occurs, the current QoS (Quality of Service) may not be
supported by the new data path. In this case, a negotiation is required to set up a
new QoS. Since a mobile user may be in the access range of several APs, it will
155
156
select the AP that provides the best QoS.

During the handoff, an old path is released and then a new path is esta-
blished. For the mobility feature of mobile ATM, routing of signaling is slightly
different from that of the wired ATM network. First, mapping of mobile terminal
routing-id's to paths in the network is necessary. Also rerouting is needcd to rc-
establish connection whcn thc mobilcs move around. It is one of the most impor-
tant challenges to rcroute ongoing conncctions to/from mobile users as those users
move among base stations. Connection rerouting schemes must exhibit low
handoff latency, maintain efficient routes, and limit disruption to continuous
media traffic while minimizing reroute updates to the network switches [4].
Limiting handoff latency is essential, particularly in microcellular networks
whcre handoffs may occur frequently and users may suddenly lose contact with
the previous wireless access point. To reduce the signaling traffic and to maintain
an cfficient route may Icad to disruptions in service to the user that are intolerable
for continuous media applications such as packetized audio and video. Thus, it is
important to achieve a suitable tradeoff between the goals of reducing signaling
traffic and maintaining an efficient route and limiting disruption to continuous
media traffic, while at the same time maintaining low handoff latency [4]. Con-
nection rerouting procedures for ATM -based wireless networks have been pro-
posed for performing connection rerouting during handoff. A few handoff schemes
have been proposed:
Break-make and make-break schemes are categorized as optimistic schemes
because their goals are to perform simple and fast handoff with the optimistic
view that disruption to user traffic will be minimal. The crossover switch simply
reroutes data traffic through a different path to the new base station, with the con-
nection from the source to the crossover switch remaining unmodified. In the
makc-brcak schcmc, a new translation tablc entry in the ATM switch (make) is
created and later the old translation entry (break) is removed. This results in cells
being multicast from the crossover switch to both the new and old base stations
for a short period of time during the handoff process.
The key idea of predictive approaches [2] is to predict the next base station
of the mobile endpoint and perform advance multicasting of data to the base sta-
tion. This approach requires the maintenance of multiple connection paths to
many or all of the neighbors of the current base station of the mobile endpoint.
Chaining approaches to connection rerouting have been proposed [I]. The
basic idea is to extend the connection from the old to the new base station in a
form of a chain. Chaining results in increased end-to-end delay, and less efficient
routing of the connection.
Chaining followed by make-break scheme, which involved a real-time
handoff using the chaining scheme and, if necessary, a non-real-time rerouting
using the make-break scheme shows good performance in connection rerouting,
because the separation of the real-time nature of handoffs and efficient route
identification in this scheme allows it to perform handoffs quickly, and, at the
same time, maintain efficient routes in the fixed part of the network.
In this paper, a Chain Routing algorithm is proposed and combined into an
existing handoff scheme. This improved scheme complies with an ATM signaling
157
standard and can be easily implemented into WATM network.
2. A HAND OFF SCHEME IN THE CHAIN ROUTING ALGORITHM

Handoff algorithms in terrestrial wireless networks focus on the connection
rerouting problem. Basically, there are three connection rerouting approaches: full
connection establishment, partial connection re-establishment, and multicast con-
nection re-establishment. Full connection establishment algorithms calculate a
new optimum route for the call as for a new call request. The resulting route is
always optimal; however, the call rerouting delay and the signaling overhead are
high. To alleviate these problems, partial connection re-establishment algorithm
re-establishes certain parts of the connection route while preserving the remaining
route. This way the route update process involves only local changes in the route
and can be performed faster. However, the resulting route may not be optimal.
In the multicast connection re-establishment algorithm, a virtual connection tree is
created during the initial call admission process. The root of the tree is a fixed
switching node, while the leaves are the switching centers to serve the user termi-
nal in the future. By using the multicast connection re-establishment method,
when a call moves to a cell with a new switching center, connection rerouting is
done immediately due to the already established routes. The disadvantage of this
algorithm is that network resources can be underutilized as a result of resources
allocated in the connection tree.
We define the Chain Routing algorithm and implement it as a partial con-
nection re-establishment in the new handoff scheme. This process is done during
chain elongation. The new scheme can be used in wireless ATM network.
2.1. The algorithm

Handoff procedures involve a set of protocols to notify all the related enti-
ties of a particular connection for which a handoff has been executed, and the
connection has been redefined. During the process, conventional signaling and
additional signaling for mobility requirements are needed. The mobile user is
usually registered with a particular point of attachment. In the voice networks, an
idle mobile user selects a base station that is serving the cell in which it is
located. This is for the purpose of routing incoming data packets or voice calls.
When the mobile user moves and executes a handoff from one point of attach-
ment to another, the old serving point of attachment has to be informed about the
change. This is called dissociation. The mobile user will also have to reassociate
itself with the new point of access to the fixed network. Other network entities
involve routing data packets to the mobile user, and switching voice calls which
have to be aware of the handoff in order to seamlessly continue the ongoing con-
nection or call. Depending on whether a new connection is created before break-
ing the old connection or not, handoffs are classified into hard and seamless
handoffs.
The Chaining scheme extends the connection route from the previous base
station to the new base station by provisioning some bandwidth using virtual
channel (VC) or virtual path (VP) reservations between neighboring base stations.
Chaining can simplify the protocols and reduce signaling traffic significantly, and
it can be accomplished quickly. However, chaining wilJ typically degrade the
158
end-to-end performance of the connection and the connection route is no longer

the most efficient. This could lead to dropped calls if resources in the W ATM are
not available for chaining. To improve the route efficiently and reduce the
number of dropped calls, we propose the Chain Routing algorithm.
We consider a broadband cellular network based on hierarchical ATM net-
work. In the planar environment, each cell is hexagonal, and the base station of
each cell has some PVCs (permanent virtual circuits) connected to the other base
stations in neighboring cells. Each base station has a number of PVCs connected
to the ATM switch only for the use of handoff calls. A parameter describing
occupancy rate of the PVC is proposed for each base station. Overall occupancy
rate (ORP) of PVCs is the larger number of occupancy rate of PVCs between the
base station and its neighboring base stations and occupancy rate of PVCs
between the base station and ATM switch.
When a mobile user makes a new call, its base station will establish a SVC
(switched virtual circuit) to carry the new call. When the terminal user moves to
an adjacent cell, the traffic path will be extended by a PVC from the current cell
to the adjacent cell. The chain length will be elongated by 1. Whenever the
chain is elongated, one bit will be send back to check the ORP of all the base sta-
tions on the chain route.
After elongation has been set up, and all the base stations on this route have
low occupancy rate, the network will follow the PVC-based scheme. In this
scheme, if a user roams from its current cell to a new cell, the traffic path is
elongated by the PVCs between these two cells. The traffic path will keep grow-
ing if the user keeps roaming. However, maintaining connections by continuously
elongating paths from original cells to the new cells will cause the path to be
inefficient.
When some parts of the route have high occupancy rate, we propose two
ways to reroute the chain parts of the route:
From the last station on the chain after each elongation, we propose sending
one bit back through the chain and check the ORP of each base station on the
chain.
The path will be rerouted according to one of the following two schemes.
I. Short path. Select a route in which the length of the path is the shortest. If
length of the route is shorter, it is more likely to be selected.
2. Select the path in which the PVCs have lower occupancy rate. That is, a
PVC between ATM switch and any base station in the elongation route can
be set up in order to obtain a low ORP. The number of options that are
available is N, where N is equal to the length of the chain.
The chain has to be rerouted whenever there is a better chain route, and the
speed of elongation will be slowed down. The network efficiency can be improved
significantly.
The path can be rerouted following the first scheme.
From the last station on the chain after each elongation, we send one bit
back through the chain and check the ORP of each base stations on the chain. If
the found ORP of a base station is close to jam, we stop, move back one base
159
station, and use this base station's PVC to connect to the ATM switch. If the base
station at the end of chain has very high ORP or it is jammed, we have to send a
signal to connection server to reroute the call.
If the speed of elongation is high, the signaling and calculation cost is
reduced, and the network efficiency is lower than in the Chain Routing algorithm.
2.2. The handoff scheme

The Chain Routing algorithm has to be implemented in the handoff scheme.
The Chain Routing algorithm is included into the handoff scheme (chaining fol-
lowed by make-break) in step (e) as follows.
(a) The mobile host sends a handoff request message to the new base station,
identifying the old base station and its connection server.
(b) The new base station adds local translation table entries for its internal rout-
ing.
(c) The new base station asks the old base station to forward packets pertaining
to the mobile host.
(d) The new base station sends back a handoff response message to the mobile
host, instructing the mobile host to transmit/receive through the new station.
(e) We include the Chain Routing algorithm. A single bit is transferred from
the mobile host back to the starting point of the chain route. It checks ORP
of each base station. After the new route is found and the new base station
chosen which is connected to the ATM switch, the new base station sends a
message to the ATM switch channel server (performing make, break, and
break-make). The new base station can change its translation table entries in
its base station channel server immediately, and the new connection
between the chain to the ATM switch is established.
This way, the chaining portion of the handoff is completed.
Note that these five steps (a), (b), (c), (d), and (e) are accomplished in real
time.
(f) The new base station passes the updated route information to the connection
server.
(g) The connection server performs necessary QoS computations on the new
route. Note that the connection server has a centralized knowledge of a
significant portion of the route and can perform this calculation easily. If
the connection server detects a possible QoS guarantee violation, or if the
fixed links are becoming congested and the route efficiency is desired, the
connection server undertakes the following steps (h), (i), and 0).
In all the other cases, the handoff flow terminates at this point.
(h) This is the first step of the make-break portion of the handoff. The connec-
tion server identifies the best route to the crossover switch, allocates
resources along the new route, and sets up a new routing entry in the cross-
over switch. The switch multicasts cells received from the source to both
base stations.
160
(i) The connection server informs the new base station of the completion of the
route change, which then starts using the new route.
(j) The connection server exchanges messages with the ATM switch, removing
the old routing entry. The connection server also requests the old and new
base stations and switches in the old route to release the old resources.
3. THE ANALYSIS OF THE CHAIN ROUTING ALGORITHM

Upon receiving a handoff request from the mobile host, the new base station
first executes the procedures for the Chain Routing algorithm scheme. The new
base station then transmits the handoff response message to the mobile host so
that the mobile host starts listening and transmitting via the new base station. The
new base station then initiates the make-break rerouting procedure. The scheme
combines the advantages of both make-break and Chaining schemes. It results in
fast handoffs so that the mobile host is quickly connected to the new base station
during hand off. Furthermore, an optimistic scheme can be later employed, as
needed, in order to make more effective use of bandwidth, and to minimize disr-
uption. This scheme is useful in cases when a user is handed over in the network
that is lightly loaded or when the mobile user does not travel far during a connec-
tion. In such cases, the handoff performed using chaining does not disrupt the
communication, and since the network is lightly loaded, there will be no notice-
able performance degradation due to the increased hop count. If the network
becomes congested or if the user moves far enough so that the effects of the
extended chain are undesirable, the make-break scheme can be applied to reroute
the connection.
3.1. Comparison with hop-limited scheme

The elongation pattern in Chain Routing algorithm is one adjustment of
Hop-limited handoff scheme and it is based on Chaining scheme. By using an
analysis of Chaining scheme and Hop-limited handoff scheme, we compare the
results with Chain Routing algorithm scheme.
We consider a mobile in a cell. A virtual circuit (VC) connecting the cell's
base station to the ATM switch or to an adjacent cell's base station is occupied by
the mobile. The VC can be released in three cases: I) the connection is naturally
terminated; 2) the connection is forced to be terminated due to handoff blocking;
3) the mobile has already successively made r-I handoffs since it came to the
current cell, and it is making the rth handoff attempt (here r is a system parame-
ter).
Comparing with the Hop-limited handoff scheme with large r values, Chain
Routing algorithm tends to use higher number of required PVCs connecting each
base station to the ATM switch for rerouting requests. When the occupancy rate
of the route path increases, the Chain Routing algorithm needs to revoke more
rerouting at the chain part of the route. At the same time, the Chain Routing
algorithm tends to use lower number of PVCs to connect the base station to the
neighboring base station compared with the Hop-limited handoff scheme with
relatively small r values. Because Chain Routing algorithm needs to revoke more
rerouting at the chain part of the route, generally the length of the route is smaller
161
than in the Hop-limited handoff scheme with relatively small r values.

Chain Routing algorithm produces less signaling traffic and network pro-
cessing load than the Hop-limited handoff scheme with a small number of r,
because it will not evoke the network to reroute the traffic path so often. At the
same time, it has lower bandwidth efficiency than the Hop-limited handoff scheme
with small number of r, because it will need more VCs to connect the base station
to a neighboring base station.
Chain Routing algorithm produces more signaling traffic and network pro-
cessing load than the Hop-limited handoff scheme with the large number of r,
because it needs to do rerouting process in the chaining parts and it evokes the
network to reroute the traffic path more often. At the same time, it has higher
bandwidth efficiency than the Hop-limited handoff scheme with large number of r,
because it will need less VCs to connect the base station to a neighboring base
station.
Chain Routing algorithm is another option which can be selected besides the
Hop-limited handoff scheme. It can give better performance than the Hop-limited
handoff scheme in certain cases. It's performance can be adjusted by tuning the
threshold at which it performs the chain routing calculation.
3.2. Signaling traffic cost

Signaling traffic is caused by reroute-related updates and modifications
occurring in the ATM switches. In the Chaining scheme we can provision
bandwidth between neighboring base stations and thereby avoid modifying the
switch routing entries. Thus, there is a clear tradeoff between the amount of
bandwidth provisioned and the number of reroute updates. The amount of pro-
visioned bandwidth can be used as a tunable parameter for engineering network
resources.
We analyze the signaling traffic cost in Chain Routing algorithm scheme.
When no reroute is found, the starting base station has a bit that remembers this
base station is the starting point of the chain. The signal needs to be transferred
in a single bit which is transferred from the mobile host back to the starting point
of the chain when the mobile host performs a handoff.
When rerouting is needed, one message is sent to the ATM switch channel
server and one message is sent to the base station server. The messages to the
ATM switch channel server contain the necessary 3-tuple (VPI, VCI, and port) for
modifYing the switch translation table entry. The messages to the base station
channel server (add entry, delete entry, delete forwarding entry, and forward) also
contain only the necessary 3-tuples for the base station to update its translation
table entries.
Because QoS computation is not involved, the Chain Routing algorithm
scheme can be performed in real time, it reduces the risk of lost connection
because of limitation of bandwidth availability and it improves the efficiency of
the PVC between neighboring base stations and PVC between base station and the
ATM switches. A possible QoS guarantee violation, or congested fixed links are
reduced because previous routes before handoff are optimized through connection
server. The most likely problem is the handoff part. If chain part is improved,
162
the entire route is improved. As the result, the chance of going through steps (h)
and (i) is reduced so is the signaling traffic involved in (h) and (i).
Signaling traffic depends on the network configuration and protocols
involved. When a mobile user roams within the A TM switch area, the signaling
traffic is low, and it performs according to the Chain Routing algorithm scheme.
When a rerouting process is required, the signaling messages are a few bytes long,
because only one ATM switch is involved. The longest message is the handoff
request message from the mobile user. This message is 44 bytes long and
includes the mobile identity, old base station channel server identifier, and the 3-
tuple (VPI, vcr, and port) of the translation table entry at the same ATM switch.
The route update message to the connection server contains the identity of the
mobile endpoint and two base stations involved in the Chaining scheme.
When the mobile user roams outside the original ATM switch and a reroute
is requested because of the overload of links or QoS problem, the new base sta-
tion needs to identify the best route to the crossover switch, allocate resources
along the route, and then exchange messages with the crossover switch, which
executes break-make or make-break operations. In this case, Chain Routing algo-
rithm performs better than the Chaining scheme.
1. In certain cases, because of the Chain Routing algorithm, the links connect-
ing base stations, and the links connecting A TM switch and base station are
used more efficiently, thus this re-route does not occur as often as the
Chaining scheme.
2. In certain cases, when the mobile user roams to the other A TM switch area,
the chain part inside the original A TM switch area will be rerouted accord-
ing to the Chain Routing algorithm, thus this portion of the routing path
will not likely have overload problem and the QoS problems as the Chain-
ing scheme does, and the overall routing path is not likely to be rerouted as
it is in the Chaining scheme.
The difference can be demonstrated by the different call drop rates in cer-
tain network configurations.
3.3. Handoff latency

The Chaining scheme, Chaining with break-make and make-break extend
the connection route from the previous base station to the new base station. By
provisioning some bandwidth by using virtual channel (VC) reservations between
neighboring base station, the chaining can be accomplished quickly (since the
crossover switch is not involved). However, chaining will typically degrade the
end-to-end performance (e.g., end-to-end delay) of the connection, and the con-
nection route is no longer the most efficient. This can lead to dropped calls if
resources in the wired network are not available for chaining.
Handoff latency is defined to be the time duration between the following
two events at the mobile host: the initiation of handoff request and the reception
of handoff response.
The handoff latency is slightly higher for the break-make scheme as com-
pared to the make-break scheme because the break-make scheme involves two
operations (break and make) at the switch before the handoff response can be
163
sent, whereas only one operation (make) is needed in the make-break scheme.
The Chaining scheme is fast because it preassigns VCs between neighboring base
stations and, thus, translation entries at the crossover switch need not be to
changed. If VC's were not preassigned, the handoff latency in the Chaining
scheme would be comparable to that of the make-break scheme.
Chaining with break-make and Chaining with make-break perform their
rerouting operations after handoff and, thus, those operations do not affect the
handoff latency of these schemes. Note that the handoff latency measurements
depend on the number of connections of the mobile endpoint that must be
rerouted. This is because each connection corresponds to a translation table entry
in the switch. Therefore, rerouting mUltiple connections implies that multiple
translation table entries have to be modified, resulting in higher latencies.
Regarding the impact of connection rerouting involving multiple ATM switches
on handoff latency, the handoff latency in Chaining with break-make and Chain-
ing with make-break will not be affected since latency is determined only by the
chaining of the neighboring base stations. On the other hand, in break-make
scheme and make-break scheme, handoff latency is directly proportional to the
number of ATM switches that need to be updated along the new route. Thus, the
separation of connection rerouting from the real-time phase in schemes Chaining
with break-make, and Chaining with make-break results in a low hand off latency
regardless of the number of switches involved in the rerouting operations. In the
Chain Routing algorithm scheme, the Chain Routing algorithm only applies to the
chain part of the route path, translation entries at the crossover switch need not to
be changed. The time cost of chain routing attributed to handoff latency will be
comparable to the Chaining algorithm.
In certain cases, some calls are blocked because of the overload of chain
part of the route path and those links between ATM switch and base stations. In
cases that those calls have to been rerouted, the handoff latency is comparable to
make-break scheme or break-make scheme. In the Hop-limited handoff and
Chain Routing algorithm schemes, the routes have to be rerouted at certain cir-
cumstances.
In the Hop-limited handoff scheme, when a mobile has successfully made
r-l handoffs, and its rth handoff request is also successful, its traffic path would
be rerouted from the new base station to the ATM switch to which it belongs.
Regardless whether the mobile user roams out of the current ATM switch area to
the new ATM switch or not, the connection server performs make-break or
break-make and necessary QoS computations on the new route. If the mobile user
roams out of the current ATM switch area, the handoff latency is directly propor-
tional to the number of ATM switches that need to be updated along the new
route.
The handoff latency is similar to make-break or break-make scheme. For
the Chain Routing algorithm scheme, the route will be rerouted when the occu-
pancy of the VCs between current base station and its neighboring base station or
of the VCs between ATM switch and base station reach a certain value. Because
two cases are considered, one is that the user roams inside an ATM switch area,
and the other is that the user roams outside the current ATM switch area, and the
handoff latency is different from the Hop-limited handoff scheme. Regardless
164
whether the user roams inside an ATM switch area or the user roams outside the
current ATM switch area, only the chain part of the route path inside the original
ATM switch area will be rerouted. The handoff latency is similar to the Chaining
scheme. Because handoff latency of the Chain Routing algorithm consists of
rerouting cost and chaining cost and handoff latency of the Chaining scheme con-
sists of chaining cost only, the latency of Chain Routing algorithm is higher than
of the Chaining scheme. Depending on the r value of the Hop-limited handoff
scheme, the latency of Chain Routing algorithm is higher than of the Hop-limited
handoff scheme with a large r value but shorter than of the Hop-limited handoff
scheme with a small r value. The routes that are blocked have to be rerouted as
those in break-make and make-break schemes.
Regarding the impact of connection rerouting involving multiple ATM
switches on handoff latency, the handoff latency in Chaining with make-break
will not be affected since latency is determined only by the chaining of the neigh-
boring base stations. Thus, the separation of connection rerouting from the real-
time phase in Chaining with make-break results in a low handoff latency regard-
less of the number of switches involved in the rerouting operation. While connec-
tion rerouting due to handoffs is similar to rerouting due to the failure of network
components, thcre are two important differences. First, handoffs are much more
frequent than network faults. With frequent reroutes, the disruption caused to
ongoing connections has to be minimized. On the other hand, in many cases
applications will be willing to tolerate some disruption due to rare network fault
rerouting scenarios. Second, handoffs result in connection reroutes that are lim-
ited to a small geographic locality (e.g., neighboring base stations). On the other
hand, reroutes due to failures may involve reestablishing the entire connection.
In ATM networks, all data is transmitted in small, fixed-size packets. Due
to the high-speed transfer rate (in the range of hundreds to thousands of Mb/s) and
short cell length (53 bytes), the ratio of propagation delay to cell transmission
time and the ratio of processing time to cell transmission time of ATM networks
will increase significantly more than that in the existing networks. This leads to a
shift in the network's performance bottleneck from channel transmission speed (in
most existing networks) to the propagation delay of the channel and the process-
ing speed at the network switching nodes. This chain routing method will
decrease the work load in the network switching nodes.
1. There is no need to identify the crossover switch when rerouting, because
chain routing method works only with one ATM switch.
2. There is no need to calculate the best route through the connection server
because it is done locally to reduce signaling traffic.
3. It is easy to implement. Only one parameter ORP is added to the new
scheme, and the calculation is very simple. It complies with existing ATM
signaling standard, and its implementation leaves commercially available
ATM components unaffected.
4. It is in real time.
5. It can significantly reduce signaling traffic.
6. In the new handoff scheme, the concern of traffic jam is included. This
scheme can handle different kinds of situations efficiently. By doing that,
165
the entire PVC in this ATM switch will have the highest utility efficient, so
that system adopting this scheme can handle much more handoffs.
4. CONCLUSION
The signaling traffic is significantly reduced by using the Chain Routing
algorithm scheme or the Hop-limited handoff scheme. The Hop-limited handoff
scheme or the Chain Routing algorithm scheme should be added to the Chaining
and make-break schemes depending on different situations. These schemes
significantly reduce the signaling traffic in the network, which causes lower
number of call drops, and there is no need to check QoS often, and to re-route
often, in the chaining part of the route path. Without the two methods, more
checking or rerouting needs to be done, with more signaling traffic in the W ATM
network. These methods significantly reduce the number of call drops in the
chaining parts. Chain Routing algorithm scheme is more suitable when the PYCs
between the base stations and the ATM switches, and the PYCs between the base
stations are very limited or there is a rush of handoff requests in the W ATM net-
work.
The Hop-limited handoff scheme is more suitable when the PYCs between
the base stations and the ATM switches, and the PYCs between the base stations,
are not limited, and the traffic in the W ATM network is not heavy. Depending on
different situation, different r values should be selected.
Configuration of the network does not make much difference, but the
number of cells inside one ATM area. When selecting a scheme, one factor must
be considered that overhead of Chain Routing algorithm scheme is higher than the
Hop-limited handoff scheme because it needs to check the ORA when it decides
if rerouting is needed.
Chaining with make-break is ideally suited for performing connection
rerouting, Chain Routing algorithm scheme or the Hop-limited handoff scheme
make the best use of it. The limitation of Chain Routing algorithm scheme is that
it has no effect on the mobile users coming from the other ATM switches.
5. REFERENCES
1. Chan, K.S. Hop-limited handoff scheme for ATM-based broadband cellular
networks. Electronics Letters 1998; 34:26-27
2. Ghai, R., Singh, S. An architecture and communication protocol for picocel-
lular networks. IEEE Personal Communications 1994; 1:36-46
3. Pahlavan, K. Handoff in hybrid mobile data networks. IEEE Personal Com-
munications 2000; 7:34-47
4. Ramjee, R. Performance evaluation of connection rerouting schemes for
ATM-based wireless networks. IEEE/ACM Transactions on Networking
1998; 6:249-61
Hierarchical Mobility Controlled by the Network
Youssef Khouaja, Karine Guillouard, Philippe Bertin

France Telecom R&D, 35512 Cesson-sevigne France
Jean-Marie Bonnin
ENST Bretagne, 35512 Cesson-Sevigne France
Abstract: The majority of IP based existing mobility solutions have significant handover
execution time and movement detection delay. In this article, we present
NCHMIPv6 (Network Controlled Hierarchical Mobile Internet Protocol
version 6), a management protocol for hierarchical mobility controlled by the
network.
NCHMIPv6 includes the decision phase of handover execution in the

management mechanism and allows the control of handover by the network.
An NCHMIPv6 mobile node monitors the link quality with its current access
point. If the quality is degraded under a given threshold configurable by the
network, it collects quality measurements with neighbouring access points and
reports them to a manager in the network. In response, the manager selects a
target access point to which the mobile node shall handoff. In parallel, it
duplicates packets destined for the mobile node and redirects them towards the
future mobile node localization. Thus, NCHMIPv6 reduces handover
execution time and limits considerably packets loss.
Key words: Mobile IP, Micro-mobility, HMIPv6, NCHMIPv6, Network Control
1. INTRODUCTION
These last years, Internet became very popular and made IP protocol
essential for the development of telecommunications networks. At the same
time, the miniaturization of data-processing equipment like portable
computers and the development of wireless networks and services increased
the need for mobility.
An IP mobile node, connected to the Internet, is localized compared to its
attachment point with its IP address; if it moves to a new attachment point, it
must imperatively:
• change its IP address: it is for example the solution adopted by the
dynamic address allocation scheme allowing users nomadism between
several sites; this solution however requires to stop all IP transfers in
progress; it thus allows the nomadism but does not provide servIce
continuity in the course of mobility
167
168 Y. Khouaja, K. Guillouard, JM. Bonnin, P. Bertin
• or inform all the routers of the Internet that its IP address locates a new
attachment point, it is an unrealistic solution.
Mobile IP [Perk97] is the standard protocol for the support of macro-
mobility i.e. mobility between networks. It allows a transparent routing of
IP packets destined for the mobile nodes in Internet and ensures
consequently service continuity for communications in progress. However,
if Mobile IP is used to manage micro-mobility, i.e. mobility localized in the
same network, it results in introducing delays in the diffusion of the new
localization and generates significant control traffic in the Internet core.
These last years, several micro-mobility protocols were developed [EMSOO]
to solve these problems. Nevertheless, the handover execution time
managed by these protocols remains always considerable and the movement
detection is delayed. We then propose in this paper NCHMIPv6, a new IP
micro-mobility protocol extended from HMIPv6 [COO], which includes the
decision-making phase of the handover execution in the mobility
management protocol and uses a handover scheme controlled by the
network.
This article consists of three principal sections. In the first section, we
briefly describe the protocol Mobile IPv4 and we develop a short
comparison with Mobile IPv6. The second section presents a classification
in two families of micro-mobility protocols: Proxy Agents Architecture
protocols and Localized Enhanced-Routing protocols. Then, the limits of
these protocols are detailed. In the third section, we present the architecture,
procedures and evaluation ofNCHMIPv6.
2. IP MACRO-MOBILITY
Mobile IP [PerkO 1] protocol is the current standard for supporting macro-

mobility in IP networks i.e. host mobility across IP domains while
maintaining transport level connections. It is transparent for applications and
transport protocols, which work equally with fixed or mobile hosts. It can be
scaled to provide mobility across the mobile Internet. It allows nodes using
Mobile IP to interoperate with nodes using the standard IP protocol. There
are two versions of Mobile IP: Mobile IPv4 and Mobile IPv6. Each one
addresses a particular version ofIP.
2.1 Mobile IPv4
Mobile IPv4 protocol defines three functional entities: the mobile node,
the home agent and the foreign agent. The mobile node is configured with a
permanent IP address belonging to its home network. It is called the home

address. All packets sent to the mobile node are addressed to its home
address. The home agent is a router in the home network of the mobile
node. It is continuously aware of the mobile node current location. The
foreign agent is a router in the visited network. The mobile node uses it to
obtain a new temporary address and generally to register it with the home
agent. The home and the foreign agent are called mobility agents.
Mobile IP performs as follows: when the mobile node is connected to a
network, it listens to mobility agent advertisements broadcasted by mobility
agents. If the network prefix changes, the mobile node detects that a
movement appeared. It is then located in a visited network and tries to
acquire a new temporary address. This new address can either be obtained by
an auto-configuration mechanism like DHCP or be the actual foreign agent
address [Drom97]. The former is called Co-located Care-of Address
(CCOA) and the latter is called Care-of address (COA). If CCOA is
acquired, the mobile node registers this new temporary address with the
home agent by exchanging registration requests and responses using CCOA
as source address. If COA is acquired, the mobile node cannot register itself
using this address. In this case, the foreign agent will perform the
registration with the home agent. Once registration is completed, the home
agent intercepts packets sent to the mobile node and uses IP-in-IP
encapsulation [Perk96] to tunnel them to the new temporary address. The
home agent must also answer to the ARP (Address Resolution Protocol)
requests destined for the mobile node hardware address with its own
hardware address in order to intercept packets generated inside the home
network CARP proxy scheme) [Plum92]. If the mobile node uses a COA
address, the corresponding foreign agent decapsulates the packets and
delivers them to the mobile node. Otherwise, the mobile node decapsulates
its packets itself, as it is directly reachable using the CCOA address.
To maintain the registration, the mobile node has to periodically renew it.
When the mobile node returns to its home network, it has to remove its
current registration. This will result in home agent stopping to intercept
traffic destined for the mobile node.
In Mobile IP, the mobile node has to use its home address as a source IP
address in order to maintain the current communications, i.e. it sends its
packets through a router on the visited network, and assumes that routing is
independent of the source address. Nevertheless, due to security concerns,
the use of routers that perform ingress filtering [FS98] breaks this
assumption and imposes on the mobile node the use of a topologically
correct source IP address in its packets. Consequently, an extension of
Mobile IP, known as Reverse Tunnelling [MontOO], has been proposed to
170 Y Khouaja, K. Guillouard, JM. Bonnin, P. Bertin
establish a topologically correct reverse tunnel from the care-of address, i.e.
either the mobile node or the foreign agent depending on the temporary
address of the mobile node, to the home agent. Sent packets are then
decapsulated by the home agent and delivered to correspondent nodes with
the home address as IP source address.
As the home agent intercepts all packets addressed to the mobile node
and tunnels them to the visited network, a "triangle routing" effect is
produced: all packets must first pass through the home agent even if the
current access router is in the same network as the correspondent node. An
extension of Mobile IP, known as Route Optimisation [PJOI], has been
proposed to overcome this problem: it allows data packets to be routed
directly from the correspondent node to the mobile node using a binding
cache in the correspondent node that keeps track of the current temporary
address. These binding caches are created and updated by Binding Update
messages sent by the home agent or the mobile node in response to mobile
node warnings or correspondent node requests.
2.2 Comparison of Mobile IPv4 with Mobile IPv6
The design of Mobile IP support in IPv6 [JPO I] is based on the

experiences gained from the development of Mobile IP support in IPv4, and
the opportunities provided by the new features of the IP version 6 itself such
as an increased number of available IP addresses and additional automatic IP
auto-configuration features.
Firstly, in Mobile IPv6, the route optimisation process is integrated into
the protocol. In fact, the route optimisation and the registration procedure
with the home agent are both done by new defined Binding Updates.
Furthermore, Mobile IPv6 and IPv6 itself, allow mobile nodes and
Mobile IP to coexist efficiently with routers performing ingress filtering, as
the mobile node uses its temporary address as the source address. The home
address of the mobile node is indicated in a Home Address destination
option of the IP packet.
Also the use of IPv6 destination options, that carry optional information
only addressed to the destination, allows all Mobile IPv6 control traffic to be
piggybacked on any existing IPv6 packet, whereas in Mobile IPv4 and its
extensions, separate UDP packets are required for each control message.
Finally, in Mobile IPv6, there is no longer any need to deploy foreign
agents. Mobile nodes make use of the enhanced features of IPv6 to operate
in any location away from the home network without any special support
required from its local router.
3. IP MICRO-MOBILITY
Micro-mobility, i.e. mobility restricted to the same domain or limited to

the same site, is characterized by local, frequent and fast displacements. If
Mobile IP is used for the management of this movement type, the mobile
node must emit registration request and binding updates at each
displacement, even local, towards its home agent and its correspondent
nodes, which can be very far from its visited network. The IP network core
is then loaded by update messages, the diffusion of the new localization is
late and the communication suffers from the long interruptions with
significant packets losses. A micro-mobility protocol is necessary to mitigate
these disadvantages. These last years, several solutions were developed,
which can be divided into two categories, those based on a hierarchy of
mobility agents and those using localized enhanced-routing.
3.1 Proxy agents architecture protocols
These protocols extend the idea of Mobile IP by using a hierarchy of

mobility. A mobile node is registered with its local agent of the lowest level
of the hierarchy; then the local agent is registered with the higher agent in
the hierarchy and the registrations go up the hierarchy until reaching the
home agent. When the mobile node moves, it emits only one registration
request located in a domain. The packets, which are destined for this mobile
node, follow then the hierarchy of the mobility agents and are delivered from
an agent to another via a tunnel. The most known protocols are Mobile IPv4
Regional Registration [GJPOl] and HMIPv6 [SCEBOl].
3.2 Localized enhanced-routing protocols
These protocols define the network architecture in two levels: the highest
level is standard Internet network implementing Mobile IP and the lowest
level is composed of domains able to manage micro-mobility. The router
connecting the two levels masks local movements compared to the rest of
Internet network. Inside a domain, a specific routing protocol is introduced.
The packets emitted by the mobile node update the routing entries in the
intermediate nodes. Then a routing table entry maps the mobile node address
with the address of the neighbouring node having transmitted the last packet
responsible for the entry update. The chain of these correspondences
represents the path traversed by the packets destined for the mobile node.
The most known protocols are Cellular IP [CGKTWVOO] and HAWAll
[RLTVOO].
172 Y. Khouaja, K. Guillouard, J.M Bonnin, P. Bertin
3.3 Limits
These protocols of micro-mobility offer powerful solutions for the

support of local mobility. They present the advantage to solve the principal
limits of Mobile IP without complicating the mechanism of mobility
management. However, improvements are still possible especially with
regard to the handover execution decision phase, the movement detection
procedure and the handover execution time.
Indeed, in these solutions, the decision-making phase of handover
execution is excluded from the protocol. A mobile node undergoes simply
the handover at the radio level: it changes access point each time the radio
connection quality with its current access point is degraded. When the
mobile node sticks to its new access point, it must wait the reception of
router advertisement to detect its movement and to start the localization
update. Thus there is a considerable delay between the physical movement
of the mobile node and its detection at the IP layer by the node itself, which
increases evidently the handover execution time.
In the following part, we present a new protocol, NCHMIPv6, which is
an extension of the HMIPv6 protocol [COO]. NCHMIPv6 proposes to solve
the problems described above by including the decision-making phase of the
handover execution in the mobility management protocol and by introducing
the concept of the handover management controlled by the network.
4. NCHMIPV6
4.1 Architecture
NCHMIPv6 domain comprises five functional entities: AR the access

router, AP the access point, MM the mobility manager, DB the database and
MN the mobile node. AR is the access router of mobile node to IPv6
network. Each access router is associated with a mobility manager. The
access point AP ensures the radio connection with the mobile node. Each
access point is attached to a single access router. The mobility manager MM
intercepts and delivers packets destined for mobile nodes of the domain like
the mobility server in HMIPv6 [COO]. More, it takes part in the decision-
making of the handover execution; it is involved in the choice of the target
access point elected for the handover. The database DB stores information
on the access points of the domain and parameters useful for the
management of the handovers.
MM Mobility Manager
MN Mobile Node
AP Access Point
AR Access Router
8.-.-.-.-.-.-..-.-.-.-.-.-.-..-.-.-.-.-.-.-~ The ~::~: node

Figure J: Architecture model ofNCHMIPv6 domain
4.2 Initialisation
In NCHMIPv6 domain, each access point is associated with an interface

of an access router, it belongs to a mobility manager and has a set of
neighbouring access points. The list of neighbouring access points of a given
access point is maintained in a database hosted by its mobility manager. This
database maintains also useful information for the decision-making phase of
the handover execution. A single identifier, the network prefix of the
associated access router interface and the length of this prefix, identifies an
access point.
The mobility manager announces its support of NCHMIPv6 by
periodically emitting to its access routers an information packet containing
its NCHMIPv6 operating mode, its address and the length of its prefix
network. When a mobile node enters a new NCHMIPv6 domain (stage 0 in
figure 2), it receives a router advertisement of its access router with options
announcing the support of NCHMIPv6 and information from the mobility
manager. Then the mobile node acquires two new temporary IP addresses:
the first address, virtual COA address @V, is valid at the level of the
mobility manager. It is given by the concatenation of the physical interface
identifier of the mobile node with the prefix network of the mobility
174 Y Khouaja, K. Guillouard, JM. Bonnin, P. Bertin
manager collected in router advertisements. The second address, local eOA

address @L, identifies the local connection. It's obtained by the
concatenation of the physical interface identifier of the mobile node with the
prefix network of the current access router (IPv6 stateless auto-
configuration).
Stages of Initia lization
I. Mobile JPv6 Binding pdates

1'. HMIPv6 registration
2. MM consu lts DB
\
l'
/
I /
/
/
3
. .. - . - . Signalling
1. @L -)@home. ~
MN
C Corrcspon dant ode
O. The mobile node accedes to
NCHMIPv6 domain
Figure 2: The mobile node initialisation in NCHMIPv6 domain
The virtual eOA address plays the role of Mobile IPv6 primary eOA
address. The mobile node informs its home agent and its correspondent
nodes, eN, of this eOA address by using Mobile IPv6 Binding Updates
(stage 1). Moreover it must be informed of the correspondence between the
virtual address and the local address, so that the mobility manager is able to
deliver these packets. For that, as soon as it acquires these two eOA
addresses, the mobile node informs the mobility manager of the
correspondence between the virtual address and the local address (@L, @V)
by transmitting an HMIPv6 registration request (stage 1'). Thus, all the
packets addressed to the mobile node are routed towards its virtual address
@V at the level of the mobility manager. Then, the mobility manager
intercepts these packets and delivers them to the mobile node local address
@L. The mobility manager, after having consulted its database (stage 2),
acknowledges the registration request while returning to the mobile node a
Mobile IPv6 Binding Acknowledgement which must include, moreover, the
list of the neighbouring access points of the mobile node current access point
in a new NCHMIPv6 option (stage 3).
4.3 Handover management
In an NCHMIPv6 domain (c.f. figure 3), a mobile node monitors the

quality of the radio link with its current access point. If it is degraded under a
given threshold Sl (stage 1), the mobile node collects new radio quality
measurements with the neighbouring access points of its current one (stage2)
and reports them to its mobility manager in a NCHMIPv6 handover request
(stage 3). The mobile node emits in this request only link qualities of the
access points included in its list of neighbouring access points sent by the
mobility manager. This reduces the size of the request thanks to a first
selection made by the mobility manager.
The mobility manager examines these measurements, consults its
database and selects the optimal target access point for the mobile node
(stage 4). The database maintains various information that can help the
decision-making of the handover execution. For example, if information on
the current load of the access points is stored, the mobility manager can
make a load repartition. The choice of the algorithm is specific to the
operator and is out of the scope of this paper.
The mobility manager then returns NCHMIPv6 handover response to the
mobile node by specifying the selected target access point, the address of the
associated access router interface and the length of the corresponding
network prefix (stage 5). In parallel, it creates a new binding between the
current virtual address and the future local address of the mobile node
(@V1, @L2). The new address can be predicted since it is built through the
concatenation of the prefix network of the new access router interface and
the physical identifier of the mobile node. The mobile node receives the
handover response, builds its future local temporary address @L2 and
changes its attachment. Once its attachment with the new access point is
completed; the mobile node emits neighbour advertisements in order to
inform its neighbours of its new local address and a router solicitation to
receive the router advertisement instantaneously from its access router (stage
7). This decreases the frequency of diffusion of the router advertisements in
NCHMIPv6 domain and minimizes consequently the load on radio links.
The access router receives the neighbour advertisement of the mobile
node with the new local COA address. It updates its routing table while
176 Y. Khouaja, K. Guillouard, JM. Bonnin, P. Bertin
mapping the new local mobile node IP address to the hardware address
diffused in the neighbour advertisement. From this moment, the
communications towards the mobile node can start again since the packets
are already redirected at the level of the mobility manager. With the
reception of the router advertisement from its new access router, the mobile
node completes and updates the parameters of its new local COA address
(lifetime, configuration flags ... ).
Signalling
5
..
.... :::::....,..~..
~
'. ~
6. @L2 -+@home . ••••••
'. '. " ':::.2 MN
.... O. @L1 -+@ home.
6 S. @L l -+@bo me.
...." ,,: I
@L2 @home .
....... ........ 1
Stages of bandover management:
1. Quality < S, 4. MM consults DB 7. Neighbour Advertisement

2. Measurements collection 5. "andover Response 8. Registration Request
3. "andover Request 6. MN moves 9. Acknowledgement +
list of APs
Figure 3: The handover management in NCHMIPv6 domain
The mobile node must also check the nature of the movement. If it is a
movement inside the same domain, it emits an HMIPv6 registration request
towards its mobility manager in order to confirm the new binding (@Vl,
@L2) and to eliminate the old one (@Vl, @L1) (stage 8). If it's a
movement between domains, the mobile node acquires, in addition to its
local address @L2, a new virtual address @V2 and announces it to its new
mobility manager and to its corresponding nodes by emitting respectively a
new HMIPv6 registration request and Mobile IPv6 Binding Updates. The
new mobility manager acknowledges the request by a Mobile IPv6 Binding
Acknowledgement containing the list of the neighbouring access points of
the new access point in NCHMIPv6 option.
During all the handover management phase, the mobile node continues to
monitor the radio link quality with its current access point. If quality is
degraded under a critical threshold S2, the mobile node keeps the possibility
of executing a handover towards the access point having the best radio
quality, without awaiting the handover response from the mobility manager.
This ensures a minimal quality for established communication.
4.4 Evaluation of the handover execution time
To compare the handover execution times, we use the FreeBSD 3.4

INRIA HMIPv6 implementation [Dupont99],[BOO], on which we have
developed extensions to support NCHMIPv6. The mobile nodes use Lucent
Orinoco IEEE802.11 [IEEE97] cards to communicate with access points
attached to Ethernet segments.
In HMIPv6 (c.f. figure 4), the mobile node controls continuously the
radio connection quality with its current access point. If it is degraded under
a threshold imposed by the radio card (stage 1), the mobile node must move
to a new access point (stage 2). After while, the mobile node receives a not
solicited router advertisement coming from the access router to which is
attached its new access point (stage 3), and discovers a new network prefix.
Then it builds its new local address @L2, diffuses it in a neighbour
advertisement and emits a router solicitation in order to make sure that it
changed sub-network (stage 4). If the old access router does not return a
router advertisement before the expiry of a timer configurable by the
network administrator, the mobile node concludes that it has moved,
depreciates its old prefix and starts the localization update procedure.
• If it is a movement in the same domain, the mobile node emits HMIPv6
registration request to its server of mobility in order to inform it of @L2
(stage 5). The mobility server acknowledges the request and replaces
the old binding (@VI, @L1) by the new one (@V1, @L2) (stage 6).
Then the packets destined for the mobile node are redirected by the
mobility server towards @L2, and the communications in progress can
continue.
178 Y. Khouaja, K Guillouard, JM. Bonnin, P. Bertin
Signalling
/
/ 5
2. @Ll
3. @L1
~@home.
~@home,
2 ,.... ,.. ,"""",[ ~N I o.
I
@L1 ~@home.
@L2~@home. 1
5. @L2 ~@home.
1. Quality < S 4. Neighbour Advertisement
+ Router Sollicitation
Stages of handover
management: 2. MN moves 5. Registration Request
3. Not solicited Router 6. Acknowledgement

Advertisement
Figure 4: The handover management in HMIPv6 domain
• If it is a movement between domains, the mobile node acquires, in

addition to its new local address @L2, a new virtual address @V2.
Then it transmits Mobile IPV6 Binding Updates to its correspondent
nodes in order to inform them of @V2, and HMIPv6 registration
request to its new mobility server to create the binding (@V2, @L2).
With the reception of the HMIPv6 registration request, the new
mobility server transmits HMIPv6 forward request to the old
mobility server in order to inform it of @V2. Then the packets
destined for @Vl are redirected towards @V2. The new mobility
server intercepts the packets and delivers them to @L2. After the
reception of Mobile IPv6 Binding Updates the correspondent nodes

send the packets destined for the mobile node towards @V2. Then,
they are intercepted by the new mobility server and are delivered to
@L2. In HMIPv6 domain, the communications begin again as soon
as the packets are redirected by the old mobility server or by the
correspondent nodes towards the new mobility server. We assume
that the transmission time between the mobility server and two
neighbouring access points is the same. The HMIPv6 handover
execution time during communication is thus equal to:
THMIPv6Handover = TRadio + TAdvertisement + TConjirmation + 2TTransmission + TExecution
Where:
• TRadio is the time of re-association with the new access point. It is
estimated at 0.l5 seconds in our testbed.
• TAdvertisement is the latency before the reception of router advertisement
of the new access router and the detection of the new network prefix.
This latency depends on the diffusion interval of the not solicited
router advertisements by the access routers. The specifications of
ND6 (Neighbour Discovery for IPv6) [NNS98] suggest a 3 second
minimal value whereas Mobile IPv6 suggests a value alternative
between 0.05 seconds and 1.5 seconds depending on the network.
We thus retain for this estimation an average of 1 second.
• TConjirmation is a sufficient latency to depreciate the old prefixes and to
conclude that the mobile node has moved. The minimal value used
by INRIA HMIPv6 implementation is 1 second.
• T Execution is the configuration and execution time of the control
commands on the mobile node user level (new prefix acquisition,
new COA address configuration, neighbour advertisements
emission, movement type identification ... ). It is estimated at 0.1
second.
• 2TTransmission represents the transmission time of the new binding
registration request and the transmission time of the first packet
redirected towards the new localization.
THMIPv6Handover = 2.25 + 2TTransmission
In NCHMIPv6, according to the preceding description of the protocol

(see section 3.3), the movement detection is done instantaneously and the
transmission of the first redirected packet and the handover response are
done in parallel. The NCHMIPv6 handover execution time during
communication is thus equal to:
TNCHMIPv6Handover = TRadio + TExecution
Where:
180 Y Khouaja, K. Guillouard, J.M. Bonnin, P. Bertin
• TRadio is the time of re-association with the new access point It is

estimated at 0.15 seconds.
• T Execution is the configuration and execution time of the control
commands at the mobile node user level. It's estimated at 0.05
second since a whole set of commands is done before and during the
radio handover execution.
TNCHMIPv6 Handover= 0.2
3,5
3 ..... .-.
..--.-
<II
2,5
~... 2
<II
> 1,5
0
'C
c;;
",,- ;
(II
:c:
1 ,
• • • • • • • • •
0,5
0
C)~ C)~
C), C),
,,~ C)'r "v~
C),
c)'? C"J~) , "
,,~ "bi,~
v
Transmission time
Figure 5: Evaluation of hand over execution time
Figure 5 presents the evaluation of HMIPv6 and NCHMIPv6 handover

execution times depending the transmission time of packets between the
mobility manager (mobility server for HMIPv6) and the mobile node.
We notice that HMIPv6 handover execution time is more important than
the NCHMIPv6 one. This is due to the movement detection phase. The
HMIPv6 handover time is proportional to twice the transmission time of
packets between the mobility server and the mobile node in the network
2 T Transmission'
In NCHMIPv6, the handover time is constant and equal to the radio
handover time.
In figure 6, we study the loss of data during a communication that

undergoes HMIPv6 or NCHMIPv6 handover. In the study scenario, a
correspondent node plays the role of MPEG video server, which transmits
128 UDP bytes, every 50 ms (20.48 kb/s) to a client mobile node.
In HMIPv6, the packets are lost during all the handover execution
procedure. The number of lost UDP packets, NHMIPv6, is thus:
NHMIPv6 = THMIPv6Handover X Emission bit rate
NHMIPv6 = (2.25 + 2TTransmission) X Emission bit rate
0,05 0,1 0,15 0,2 0,25 0,3 0,35 0,4 0,45

Transmission time
Figure 6: Evaluation of the packets loss
In NCHMIPv6, the redirection of the packets is done in the first phase of

the handover execution. If we assume that the packets, which arrive at the
new access router of the mobile node during the radio handover execution
are lost, the number oflost UDP packets NNCHMIPv6 is:
N NCHMIPv6 = TNCHMIPv6Handover X Emission bit rate
NNCHMIPv6 = 0.2 x Emission bit rate
Figure 6 shows that the loss ofUDP packets in HMIPv6 is proportional to
handover execution time and that it grows according to the transmission time
between the mobility server and the mobile node. For NCHMIPv6, since we
suppose that the redirected packets, arrived before the radio handover
execution, are lost, the loss ofUDP packets remains constant.
5. CONCLUSION
In this article, we have presented NCHMIPv6 a protocol for hierarchical

mobility management controlled by the network. This protocol proposes to
include the phase of handover execution decision-making in the management
mechanism. Thus it reduces the handover execution time and decreases the
loss of packets considerably. Moreover, the concept of the handover
controlled by the network makes possible to do repartition of load and to
elect intelligently the target access point of the handover (prediction of the
movement, assignment of priorities ... ). Lastly, NCHMIPv6 has several other
advantages, whose description comes out of the scope this article. For
example, in NCHMIPv6, a mobile node starts the handover execution phase
when its quality with its current access point falls under a given threshold
182 Y. Khouaja, K. Guillouard, J.M. Bonnin, P. Bertin
S 1. This threshold can be suggested by the mobility manager to the mobile

node depending on flows during transmission and the current environment of
the radio cell. Moreover, when the radio cell is loaded, the mobility manager
can filter the redirected packets according to their class of service and the
priority of the mobile node destination.
REFERENCES
[COO] Castelluccia, "An Hierarchical Mobile IPv6 Proposal", INRIA, November 1998.
[BOO] Bellier, http://www.inrialpes.fr/planete/people/bellier/hmip.html.
[CGKTWVOO] Campbell, A., Gomez, J., Kim, S., Turanyi, Z., Wan, C-Y., Valko, A.,
"Design, Implementation and Evaluation of Cellular IP", Communication IEEE, July 2000.
[DH98] Deering, S., Hinden, R., "Internet Protocol, Version 6 (IPv6) Specification", RFC
2460 IETF, December 1998.[Drom97] Droms, R. ,"Dynamic Host Configuration Protocol",
RFC 2131 IETF, March 1997.
[Dupont99] Dupont, F., ftp:/Iftp.inria.fr/networklipv6/.
[EMSOO] Eardley, P., Mihailovic, A., Suihko, T., "A Framework for The Evaluation of IP
Mobility Protocols", In Proceedings ofPIMRC 2000, London, UK, September 2000.
[FS98] Ferguson, P., Senie, D. ,"Network Ingress Filtering: Defeating Denial of Service
Attacks which employ IP Source Address Spoofing", RFC 2267 IETF, January 1998.
[GJPOI] Gustafsson, E., Jonsson, A., Perkins, C., "Mobile IP Regional Registration", Draft
IETF, September 200 I.
[IEEE97] IEEE Std 802.11-1997, "Wireless LAN Medium Access control (MAC) and
Physical Layer (PH) Specification".
[JPOI] Johnson, D., Perkins, C.,"Mobility Support in IPv6", Draft IETF, July 2001.
[MontOI] Montenegro, G. ,"Reverse Tunneling for Mobile IP, revised", RFC 3024 IETF,
January 2001.
[NNS98] Narten, T., Nordmark, E. Simpson, W., "Neighbor Discovery for IP Version 6
(IPv6)", RFC 2461 IETF, December 1998.
[Perk97] Perkins, C., "Mobile IP", IEEE Communications Magazine, May 1997.
[PerkO I] Perkins, C., "IP Mobility Support for IPv4, revised", RFC 2002, September 200 I.
[PJOI] Perkins, C., Johnson, D., "Route Optimization in Mobile IP", Draft IETF, September
2001.
[RLTVOO] Ramjee, R., La Porta, T., Thuei, S., Varadhan, K., "IP micro-mobility support
using HAWAII", Draft IETF, July 2000.
[SCEBOI] Soliman, H., Castelluccia, C., El-Malki, K., Bellier, L.,"Hierarchical MIPv6
mobility management", Draft IETF, July 200 I.
Approximate and exact ML detectors
for CDMA and MIMO systems:
a tree detection approach
Sandrine Vaton 1 , Thierry Chonavel2 and Samir Saoudi 2

Ecole Nationale Superieure des Telecommunications de Bretagne
1 Departement Electronique, 2 departement Signal et Communications
Technopole Brest Iroise, BP 832, 29285 BREST Cedex, FRANCE

Email: sandrine.vaton.thierry.chonavel.samir.saoudi@enst-bretagne.fr
abstract This paper deals with Direct-Sequence Code Division Multiple Access
(DS-CDMA) transmissions over mobile radio channels. Different detection tech-
niques have been proposed in the past years, among which approximate ML detec-
tors. In this paper we propose an exact ML detector with a low complexity. We show
that a QR factorization of the matrix of users' signatures is appropriate: the upper
triangular form of the R matrix makes it possible to state the ML detection in terms
of a shortest path detection in a tree diagram. Different algorithms based on the tree
diagram are compared: the stack algorithm (exact ML), and the feedback decoding
algorithm (approximate ML). The numerical complexity of the proposed techniques
is studied in detail; in particular, the low complexity of the stack algorithm at high
SNR is pointed out. Simulations show that the performance of the stack algorithm in
terms of Bit Error Rate (BER) are very close to the single user bound. Furthermore,
we point out the fact that the same kind of approach can be used to perform ML
detection in Multiple Input Multiple Output (MIMO) systems.
1 Introd uction
Until the mid 80's, the techniques that were used to detect symbols trans-
mitted by a multiuser DS-CDMA system were based on single user detection
strategies. Unfortunately, this kind of approach yields very poor performance
due to the lack of orthonormality of the signals arising from distinct users.
Interference among users is known as Multiple Access Interference (MAl) and
it is all the more sensitive as the interfering users have high power, yielding
what is called the near-far-effect. In order to suppress MAl, multiuser detec-
tion strategies have been developped since 1986 [11].
In particular, MAl can be suppressed using the popular decorrelator or
linear Minimum Mean Square Error (MMSE) techniques. They are not so
efficient in terms of Binary Error Rate (BER) performance as the Maximum
Likelihood (ML) detector, but they require smaller computational effort [13].
In fact, the ML detector consists in selecting the most probable combination of
users bits among 2K , where J{ is the number of users. If an exhaustive search
is performed, ML detection is impracticable in many real systems [12]. It is
183
184
thus highly desirable to propose ML detection techniques with a reasonable

numerical complexity.
In [15], an approximate ML criterion has been proposed, based on the use
of a modified Fano metric. The optimum path is searched by means of a
stack algorithm [16, 5]. More recently, Rasmussen et al. have remarked that
applying an orthogonal transform to the received data vector enables to derive
an exact ML metric [9]. In fact, assuming a Gaussian noise, this is simply
because the noise independance property is preserved among the transformed
data vector entries. However, in [9], instead of looking for a fast exact ML
detection procedure, a complexity constrained structure is proposed yielding
an approximate ML detection. Another approximate ML detection technique
can be found in [7] where a group detection strategy has been proposed.
In this paper we propose a new exact ML detection technique with a nu-
merical complexity roughly equivalent to that of the decorrelator at usual
SNRs for CDMA systems operation. The detection can be decomposed into
two steps: (i) first, a QR factorization of the matrix of users' signatures is
performed (ii) then, the detection is performed as an optimal path selection
in a tree diagram. More precisely, this can be done by considering methods
developped in the framework of decoding algorithms for convolutional codes
[8]. Different exact or approximate ML detection algorithms are based on the
tree diagram, among which the stack algorithm (exact ML). Note in particular
that step (i) yields the exact ML metric in the tree diagram that we build and
that using the stack algorithm in step (ii) enables finding the exact optimal
path in this tree.
We also propose a new, complexity-constrained, approximate ML algo-
rithm, which is simply obtained by replacing the stack algorithm by the feed-
back decoding algorithm (see for instance [8]) that performs approximate ML
detection. The comparison of both algorithms shows in particular that the
computational complexity of the stack algorithm rapidly becomes linear with
the number of users for increasing signal to noise ratios.
Furthermore, transmission communication systems working between multi
sensors arrays at both sides have been considered with much interest for a few
years [14, 1, 2]. We will show that it is straightforward to extend the use of
the ML decoding techniques that we consider in this paper for non convolutive
Multiple Input-Multiple Output (MIMO) systems.
The paper is organized as follows: the problem of the detection for DS-
CDMA and MIMO systems is stated in Section 2. For the sake of conciseness,
further sections will only adress the case of DS-CDMA systems. In Section 3
we show that ML detection can be addressed as a shortest path detection in
a tree diagram. The stack algorithm and the feedback decoding algorithm are
then described in Section 4. In Section 5 we discuss the numerical complexity
of the proposed methods. We present simulation results in Section 6 and the
conclusion in Section 7.
185
2 Problem statment
For the sake of conciseness we will only consider real valued signals, extension
to complex signals being straightforward. We consider in this section a syn-
chronous DS-CDMA communication system where K users transmit data sym-
bols (bkh=l,K with bk = ±l. (Sk)k=l,K represent the corresponding spreading
sequences with length N (N 2: K). Then noting b = [b 1 , ... ,b K F the vector
of users bits, and S = [S1, ... ,SK] the matrix of signatures, the outputs of the
BPSK or QPSK demodulator sampled at the chip period is the vector
r = Sb+n, (1 )
where n is a zero mean uncorrelated gaussian noise vector with uncorrelated

entries.
In classical detection techniques (RAKE receiver, linear MMSE and decor-
relator detectors) the received data vector is considered at the output of the
signatures matched filters bank:
(2)
In our approach the matched filtering is not necessary and we consider expres-
sion (1).
Similarly, in MIMO communications, the data are transmitted from an ar-
ray of K sensors to an array of N sensors. Denoting by r the vector observed
on the receiver array yields data model (1), where b represents the transmit-
ted data vector and S is the channel mixing matrix. This model is valid when
instantaneous, that is to say non convolutive, transmission channels are con-
sidered. Working from model (1), it is clear that the detection of b can be
adressed in the same way for CDMA and MIMO systems.
3 QR factorization and the tree diagram

The ML detection of b from the model r = Sb + n can be restated as a
constrained minimization problem:
{ minb II Sb - r 112 (3)

bk = ±1, k = 1, ... ,K,
where II x 112= x T x.
Consider now the QR factorization of S,
S=QR, (4)
where Q is a N X K matrix such that QTQ = IK(IK is the K x K identity

matrix) and R is a K X K upper triangular matrix. Then, straightforward
186
User 3 User 2 User 1
Figure 1: Tree diagram for K = 3 users. The bold path is b = (b 1 , b2 , b3 )T =

(+1, +1, -If.
calculations lead to:
(5)
The second term on the right-hand side of this equality does not depend on
b and can therefore be omitted for the optimization. Problem (3) is then
equivalent to:
{ minb II Rb - z W (6)
bk = ±1, k = 1, ... ,K,
where z = QTr . Since the matrix R is upper triangular, it is possible to state

the detection as a shortest path detection in a tree diagram.
To see this more clearly, note that
(7)
i=K,K -1, ··,1
(Rb)i is the component number i of the vector Rb; that is to say (Rb)i =
"L,j=i,K Rijbj , and (Rb)i only depends on bits bj , for j = i, i + 1, ... , I<.
Then, the vector b can be identified to one path in a tree diagram, starting
from the root with user K, then coming users number K - l,I{ - 2, ... ,1. The
criterion II Rb - z Win problem (6) can be decomposed into a sum of K
187
Userbit= "+1"1
Userbit= "-1"1
User 4 User3 User 2 User 1
Figure 2: Illustration of the stack algorithm for [{ = 4 users. The detected

path is in bold line.
branch metrics of this diagram: the metric of the branch bk - 1 extending the
path bK , bK- 1, ... ,bk is I(Rbh-l - zk_11 2 = I~j=k-l,K Rk-l,jbj - zk_11 2 .
Figure 1 corresponds to the case [{ = 3 users. The bold path stands
for the vector of users bits b = (b 1, b2 , b3 f = (+1, +1, _1)T. Finally, the
criterion II Rb - z Wis the cumulated branch metrics along the path b =
(b K , bK -1, ... ,b 1 ) in the tree. The solution of the problem (6) is therefore the
path b with the shortest cumulated metric.
4 Tree detection algorithms

We now present two algorithms that can be used to detect the path with
the shortest metric in the tree: the stack algorithm and the feedback decoding
algorithm. They implement exact and approximate ML detection respectively.
4.1 The stack algorithm

The stack algorithm (Zigangirov, 1966 [16]; Jelinek, 1969 [5]) keeps track of a
few paths and their corresponding metrics in a stack; the head of the stack is
always the path with the shortest metric among all the paths in the stack. At
each step of the algorithm, the path at the head of the stack is extended by
one branch, thus yielding two successors. The two successors along with the
other paths in the stack are reordered, so that the head of the stack is always
the path with the shortest metric. The process of extending the path is then
188
step 1
User K User K-l User K-2
Figure 3: Illustration of the feedback decoding algorithm with K :::: 4 users

and L = 3.
repeated until a complete path with K branchs is found. The stack algorithm
with L:::: 3 is illustrated on Figure 2.
4.2 Feedback decoding

The feedback decoding algorithm (Heller, 1975 [4]) is a sliding window algo-
rithm. The first step of the algorithm is to make a decision upon the bit bK,
based on the metrics 1(Rb)i - Zj 12 for i :::: K, K -1, ... , K - L + 1; the decision
is bK :::: +1 or bK :::: -1, depending whether the path (b K , bK - 1 , ... , bK - L +1 )
with minimum metric stems from a branch with bK = +1 or bK :::: -1. Once
the decision upon bK has been made, the part of the tree that stems from the
branch that has not been selected is discarded (half the tree). The other half is
extended by one stage, yielding 2L successors. The decision on the bit bK - I is
then made, based on the metrics I(Rb)i _z;l2 with i :::: K -1, K -2, ... , K - L
and on the decided bit bK. The procedure is repeated until the tree is extended
to bit bl . The feedback decoding algorithm with a window of size L :::: 3 is
illustrated on Figure 3.
5 Computational issue
The computational costs involved by the detection techniques that we have
just presented consist in two contributions: (i) a preprocessing that requires
the calculation of the QR decomposition of the matrix S (ii) the search of a
shortest path in a tree diagram. We are going to see that in many situations
189
(i) represents the main contribution to the computational effort.
5.1 Preprocessing
The computational effort required to change problem (3) into problem (6)
mainly amounts to the QR decomposition of the matrix S. It involves O(N [{2)
operations, and it can be implemented in several ways [3].
At this point, we can note that this is not more than performing the
decorrelator detection. Indeed, a direct implementation of the decorrelator
detector involves O(N [{2) operations for calculating the correlation matrix
STS of users signatures and O([{3) operations for its inversion. In fact, the
decorrelator detector is generally implemented iteratively by means of an iter-
ative Successive Interference Cancellation (SIC) detector [6, 10]. This involves
O(N [{) operations per iteration. But the number of iterations that must be
used to ensure convergence of the SIC detector to the decorrelator increases
significantly with the number of users. Note that the linear MMSE detector
can be implemented with similar computational complexity. This shows that
performing the QR decomposition of the matrix S is not significantly more
complex than considering the decorrelator or the MMSE detector.
5.2 Detection of the shortest path in the tree diagram

In order to compare the numerical complexity of the detection algorithms,
we evaluate only the number of branch metrics 1(Rb)i - Zi 12 that must be
computed. In a general case, computing I(Rb)i - zil 2 involves [{ - i + 1
additions and one multiplication (because bk = ±1).
The numerical complexity of the stack algorithm becomes lower as the SNR
increases. At a medium to high SNR the stack algorithm explores with a high
probability only that part of the tree that corresponds to the shortest path,
yielding the computation of only 2[{ metrics at high SNR. Note that the factor
2 is because at each step one metric is calculated for the survivor and one for
the concurrent.
At extremely low SNRs, the stack algorithm may require the computation
of, at most, 2K metrics. Anyhow, for memory storage and computational re-
quirements, the number of positions in the stack can be limited to a predefined
value, thus possibly discarding the paths in the stack with the shortest metric.
In that case the stack algorithm performs an approximate ML detection.
The numerical complexity of the feedback decoding algorithm does not
depend on the SNR; it requires the computation of ([{ - L + 2)2£ - 2 branch
metrics. At medium to high SNRs the stack algorithm is preferable to the
feedback decoding algorithm because (1) it performs exact ML detection and
(2) the number of metrics required by the stack algorithm is lower than that
required by the feedback decoding algorithm. At extremely low SNRs the
190
feedback decoding algorithm requires less computation than the stack, but it
performs only approximate ML detection.
6 Simulation Results
We consider the case of K = 9 users with signatures of length N = 20 chips.
The signatures are Gold sequences of length 63, truncated at N = 20. Then,
the matrix S is of the form
S = (l/VNJx T
+-++-+--+---+--++--+
+-++-++---+++-+----+
--+--+++------+--+-+
+--+--+++--++++-----
+++++-+-+-+--++-+-++ (8)
--+-+---++-+-+++++--
+---++----++-+-+--++
++---+-+++++----++-+
+--+-+++++--++------
-+-+-++--++++-++---+
where sign + stands for +1 and sign - stands for -1. Typically, this kind
of matrix can be met in the case of a DS-CDMA communication. Note that
the absolute value of the correlation among the columns ranges from 0 to 0.5
and that the mean and the standard deviation of the off-diagonal terms of the
covariance matrix STS are given by 0.044 and 0.224.
Figure 4 shows a comparison of the numerical complexity of the stack
algorithm and the feedback decoding algorithm. For the stack algorithm, the
number of branch metrics is averaged on a high number of blocks of K users. As
stated in Section 5 the numerical complexity of the stack algorithm decreases
rapidly as the SNR increases. For Eb/No 2': 5dB (BER:S 8.0 10- 3 ) the
numerical complexity of the stack algorithm is very close to its lower bound
(2K branch metrics computed).
The numerical complexity of the stack algorithm is then compared to that
of the feedback decoding algorithm, for window sizes L = 3, L = 2 and
L =1. Note that in the case L = 1 the number of branch metrics computed
is (K - L + 2)2£ - 2 = 2I<. In the case of the feedback decoding algorithm,
when the window is shortened by one unit (L -t L - 1), the number of branch
metrics computed is divided by about 2.
Figure 5 shows the BER obtained with the stack algorithm, the feedback
algorithm (L = 1,2,3), the decorrelator, as well as the single user bound.
As one can see from Figure 5 the stack algorithm (exact ML) performs always
better than the feedback decoding algorithm (approximate ML). What is more,
the feedback decoding algorithm performs always better than the decorrelator,
even when L = 1.
191
70~------------'------r~~~~~==========~
"0
Ql
:J
a.60
E
8
~
",S 50
Ql
E
..c
o
c
t1l40
15
'0
lii
~30
::J
c
Ql
Ol
~ 20
Ql ~~--~~--~~--~~--~~--~~--~~~~~
~
1~5L--------------L--------------L-------------~10
Eb/NO (dB)
Figure 4: Average number of branch metrics computed per block (K 9

users) .
The gain obtained with the stack algorithm, by comparison to the feedback
decoding algorithm, is all the more important as the SNR is high. At high
SNR the stack algorithm achieves the single user bound (Figure 5) with the
same computational complexity as that of the feedback algorithm for L = 1
(Figure 4). Compared to the decorrelator a gain of about 3.6 dB is achieved
by the stack algorithm at BER= 10- 5 .
7 Conclusions
In this paper, we have proposed a new approach to perform ML detection
for symbol vectors in DS-CDMA and MIMO systems. It is based on a QR
decomposition of the signatures matrix, or of the channel mixing matrix, fol-
lowed by an exact or approximate shortest path detection in a tree diagram.
In particular, using the stack and the feedback tree search algorithms yield
very attractive ML detectors both in terms of computational complexity and
performance.
Acknowledgment The authors are grateful to the anonymous reviewer who

supplied them with references [9] and [15].
192
lO-'L----'-----'-----'-----'----'----'--'-----'
o 4 6 8 10 12 14
Eb/NO (dB)
Figure 5: Comparison of the performances of different detectors (K = 9 users).
References
[1] G. Foschini. Layered space-time architecture for wireless communication in a
fading environment when using multi-element antennas. Bell Labs Technical
Journal, pages 41-59, 1996.
[2] R. V. G.D. Golden, G.J. Foschlni and P. Wolniansky. Detection algorithm and
initial laboratory results using v-blast space-time communication architecture.
Electronic Letters, 35:14-16, January 1999.
[3] G. Golub and C. Loan. Matrix Computations. The Johns Hopkins University
Press, Baltimore, Maryland, 1984.
[4] J. Heller. Advances in Communication Systems, volume 4, chapter Feedback
decoding of convolutional codes. A.J. Viterbi, New York, academic edition,
1975.
[5] F. Jelinek. Fast sequential decoding algorithm using a stack. IBM Jour. Res.
Dev., 13:675-685, November 1969.
[6] M. Junti. Multiuser demodulation for DS-CDMA systems in fading channels.
PhD thesis, Gulu University, Department of Electrical Engineering, Gulu, Fin-
land, 1998.
[7] I. Medvedev and V. Tarokh. A channel shortening multiuser detector for ds-
cdma systems. In VTC'Ol, Rhodes, April 2001.
[8] J. G. Proakis. Digital Communication. New York McGraw-Hill, second edition,
1989.
193
[9] L. Rasmussen, T. Lim, and T. Aulin. Breadth-first maximum likelihood de-

tection in multiuser cdma. IEEE Trans. on Comm., 45(10):1176-1178, oct.
1997.
[10] L. Rasmussen, T. Lim, and A. Johanson. A matrix-algebraic approach to suc-
cessive interference cancellation in CDMA. IEEE Transactions on Communi-
cations, 48(1):145-151, January 2000.
[11] S. Verdu. Minimum probability of error for asynchronous gaussian multiple-
access channels. IEEE Trans. Information Theory, IT-32:85-96, Januray 1986.
[12] S. Verdu. Optimum multiuser asymptotic efficiency. IEEE Trans. Communica-
tions, COM-34:890-897, sept 1986.
[13] S. Verdu. Multiuser Detection. Cambridge University Press, 1998.
[14] A. Wittneben. A new bandwidth efficient transmit antenna modulation diver-
sity scheme for linear digital modulation. In ICC'93, pages 1630-1634, 1993.
[15] Z. Xie, C. Rushforth, and R. Short. Multiuser signal detection using sequential
decoding. IEEE Trans. on Comm., 38(5):578-583, may 1990.
[16] K. Zigangirov. Some sequential decoding procedures. Probl. Peredach. Inform.,
2:13-25, 1966.
BLOCK TURBO CODE WITH BINARY INPUT FOR
IMPROVING QUALITY OF SERVICE
Patrick ADDE, Ramesh PYNDIAH and Sylvie KEROUEDAN

ENST de Bretagne, BP 832, 29285 BREST CEDEX FRANCE.
Mail: Patrick.Adde Ramesh.Pyndiah Sylvie.Kerouedan@enst-bretagne.fT
INTRODUCTION
An iterative decoding algorithm ("Block Turbo Code (BTC) algorithm") for product
codes based on soft decoding and soft decision output of the component codes was
introduced by R. Pyndiah in 1994 [1][2]. It uses the concepts developed by C.
Berrou who proposed a technique to encode and decode a class of error correcting
codes, called "Turbo codes" CTC [3]. The BTC is based on the product code which
is a series concatenated coding scheme, introduced by Elias [4]. The information bits
are placed in a matrix. The rows of the matrix are encoded by a linear block code
and the columns by a second block code.
Numerous papers, [5] to [12], describe this algorithm, give its performance, attempt
to simplify it and propose architectures for implementing it in integrated circuits or
DSPs.
For a better understanding of this paper we recall the basic principle of product
codes and their properties. We give a brief description of the soft decoding
algorithm (Section 1) and the soft decision algorithm (Section 2). In Sections 3 and
4, we analyze the performance of BTCs when the input data are binary, which is the
case in several systems (optical transmission, data storage, networking, ... etc).
1. ITERATIVE DECODING OF BTCs
The concatenated code used for BTCs is the product code proposed by Elias in
1954. Let us consider the concatenation of two systematic linear block codes C\ and
CZ with parameters (nt.kt.bt) and (nz,kz,b z) where nj, kj, and bi stand for code length,
number of information bits and minimum Hamming distance respectively. The
product code is obtained by :
1) placing the k\.kz information bits in an array of k\ rows and kz columns;
2) coding the k t rows using code CZ;
3) coding the nz columns using ct.
195
196
------ n2 --------1.~
I •
k 2
Lhecks
kl Information symbols on
n1 rows
1
j Checks on columns
LI.eCI<$
on
checks
Fig. 1 : Product code p=c IX C 2
The parameters of the product code are given by: n=nl.n2, k=kl.k2 and <5=<5 1.<52 • The
code rate is given by R=R 1.R2 where Ri is the code rate of code C i. The fact that the
minimum Hamming distance of the product code is the product of the minimum
Hamming distance of the elementary codes gives product codes a significant
advantage over parallel concatenation [6]. This is a direct result of a property of
product codes, which states that all the rows are code words of C 2 and all the
columns are code words of C I.
The elementary code used is the Bose-Chauduri-Hocquenghem (BCH) code which
is a systematic cyclic code, able to correct at least t errors in a block of n symbols.
The minimum distance dmin is at least equal to 2t+ 1. Thus, it is possible to add a
parity bit in order to increase the minimum Hamming distance and improve the
coding gain. These particular codes are named extended BCH codes. The concept of
product codes is a simple and relatively efficient method to construct powerful codes
(that is, having a large minimum Hamming distance dmin ) using simpler linear block
codes.
The iterative turbo decoding process can be achieved by cascading several
elementary decoders illustrated in Fig.2, where k represents the current half-
iteration, and
1. [R] is the received vector,
2. [W(k)] is a vector that contains the extrinsic information (which is the
difference between the output information and the input information) given by
the previous decoder concerning the reliability of the decoded bit,
3. [R'(k)]=[R]+a(k).[W(k-l)],
4. a(k) and ~(k) are constants determined by simulations.
197
[R] [R]
D::layline
Fig. 2 : Block diagram of an elementary decoder
The basic element of a turbo decoder is the SISO decoder used for decoding the rows
and columns of the product code. It is a modified Chase algorithm [13] which starts
by computing the maximum-likelihood (ML) code word using the log-likelihood
ratio (LLR) of the bits at the input of the SISO decoder. For each bit of the ML code
word, it then computes the log-likelihood ratio which is the soft output of the
decoder [2].
2. RELIABILITY COMPUTING
Let us consider the transmission of binary elements {O, 1} coded by a linear block
code C with parameters (n,k,J) on a Gaussian channel using binary symbols {-1,+1}.
We shall consider the following mapping of the symbols: 0-->-1 and 1-->+1. The
output of the Gaussian channel R=(r[, ... ,rj, ... ,rn) for a given transmitted codeword
E=(e[, ... ,ej, ... ,en) is given by:
R=E+G (1)
where components gl of G=(g[, ... ,gj, ... ,gn) are AWGN (Additive White Gaussian
Noise) samples of standard deviation cr.
By using ML decoding, it can be shown that the optimum decision

D=(d[, ... A, ... ,dn ) is given by:
D=C i if IR_C i I2 IR-C'12; 'if Ie [1,2 k] I (2)
.. ' . th
where C'=(c\, ... ,cj, ... ,dn) is the i codeword ofC and:
'2 ~ '2
IR-C'I = L (rrd,) (3)
'~l
is the squared Euclidean distance between Rand C i . When using an exhaustive
search for the optimum codeword D, the computation complexity increases
exponentially with k and becomes prohibitive for long block codes. The Chase
algorithm offers a good trade-off between complexity and performance. Instead of
considering the whole set of code words it generates a sub-set of ML code words
by decoding test patterns using an algebraic decoder. The optimum decision D is the
code word in at minimum Euclidean distance from R. Once we have determined
D, we have to compute the reliability of each of the components of vector D in order
to generate soft decisions at the output of the decoder. The reliability of decision dj is
the LLR of transmitted symbol ej which is given by :
198
A(dj) = (pre j=+lIR) (4)

Pr ej=-lIR
The LLR computation complexity is of order 2k and is prohibitive for long codes.A
simplified expression was been proposed in [2]:
A'(dj~IR--C-l{J)121R--C+l{J)n (5)
where C+1(1) and C-1(/J are code words at minimum distance from input R with a + 1
and a -1 in position} respectively. By expanding (5) using (3) we obtain the
following relation:
N~}~l~+ 1=I/oti
f~~~) (J
w
where:
r0 if c7 1Ul = ci 1(j) (7)
PI = ~ll if c+ 1(j)
I
oF- c- 1(j)
I
If we assume that (J is constant, we can normalize N(dj ) with respect to the constant
2/(J2 and we obtain the following equation:
(8)
with:
n 1(j) (9)
wi = "[.rIC: PI
1=lhi
The normalized LLR r j is taken as the soft output of the decoder. The term Wj is a
correction term applied to the input data and is called extrinsic information. The
extrinsic information is a random variable with a Gaussian distribution since it is a
linear combination of identically distributed random variables. Furthermore, it is
uncorrelated with the input data rj. Like for CTC, the extrinsic information plays a
very important role in the iterative decoding of product codes.
Computing the reliability of decision 0 at the output of the soft-input decoder
requires two code words C+l(/J and C1(J), see (5). Obviously D is one of these two
code words and we must fmd the second one which we shall call C. C can be viewed
as a competing code word of D at minimum Euclidean distance from R with Cj 0.
Given code words C and D, it can be shown that the soft output is given by the
following equation:
r' .=(IR-Cl 2 -IR-Dl 2
) 4
)d.
J
(10)
In the event where code word C is not found, we use the following equation:
rj=~ ..dj with ~ o. (11)
This sub-optimal solution uses reliability ~ (a constant for the current half-iteration)
at many positions}. As a refmement, we propose to estimate N(0) when there is no
competitor by:
N(0) = Irjl + i: Irk I

k=x
(12)
x and y being chosen among the least reliable binary symbols of [ R'm] [14] [11].
199
If we compare the results for fixed and variable ~ (12), the experimental result is
much closer to the theoretical curve (gap of 0.1 dB to 0.2dB).
In order to further improve the estimation of ~ we can adopt the following strategy:
A'(~) = I rjl + (2, Irk I)/A (13)

k=O
where A is an integer. The improvement with respect to (12) is not significant but
for practical implementation, (13) is of a much complexity when A is a power of 2.
3. IMPLEMENTATION OF BTC
The "optimal" turbo decoding algorithm of linear block codes follows the following
steps:
1. Search for the least reliable binary symbols of [R']; their positions are called I],
I 2 .... I m ,
2. Generate test sequences [TQ] that are a combination of elementary test vectors
[TQy having "1" in position Ij and "0" elsewhere,
3. For each test word [TQ], compute [ZQ]:
[ZQ]=[TQ]ffi sign of [R'],
4. Decode [ZQ] by the algebraic algorithm (result [CQ]),
5. For each vector [C Q], compute the square Euclidean distance between [R] and
[C Q],
6. Select code word [Cd] at minimum distance from [R']; then [D]=[C d] is the
result of binary decoding,
7. Compute reliability Fj for each element dj of [D]; this involves searching for a
code word which has a minimal square Euclidean distance from [R'] (called a
competitor) with C/"*C/,
8. Compute extrinsic information:
d ' d
Wr[FrCj Rj ]Cj .
The implementation of this algorithm has shown that the complexity depends mainly
on the computation of the reliability at the output of the decoder. It is possible,
without any significant degradation, to decrease this complexity by reducing the
number of competitors. When using only one competitor, the loss is 0.12dB but the
complexity for this function is divided by 10. We show that the gap with theoretical
reference [14] is small, lower than 0.2dB at 10-6 • This very small value shows that
the algorithm allows the theoretical limit to be reached. The complexity of the
elementary decoder is very low: fewer than 10,000 gates are necessary to implement
the BCH decoder in the BTC solution.
200
ffi
III
I,OO&OS t--,--=+=-=-c*",=t~rlr=-~-=-_-=-_-="_=""'
__="",_,""_,""",_,...,,-1_
1,00&07
2 2,5 3 3,5
BllNOindB
Fig 3: Performance for the product code BCH(32,26, 4/ at iteration 6, as afunction

of the number ofcompetitors.
Fig. 3 gives the Bit Error Rate (BER) as a function of the Signal Noise Ratio (SNR)
when the modulation used is QPSK modulation [9][11]. For this product code
BCH(32,26,4)2, the number of competitors varies. Sequential decoding with 6
iterations is used with the Chase algorithm using 16 test patterns. We measure the
performance of the BTC when using 5 bits for the linear quantization of the data
[R] and [W]. The code rate is about 0.66. These results show that BTCs are very
attractive for a wide range of applications because they offer the best trade-off
between performance and complexity.
4. BTC WITH BINARY INPUT
Initially, turbo codes were designed for systems where soft data inputs to the
decoder were available. However there are many applications where the inputs are
binary. This is true for optical transmission systems and high speed data storage
applications. In classical forward error correction there is a loss of 2-3 dB in the
coding gain when using binary input instead of soft inputs. This is due to the fact
201
that using binary inputs, optimum decoding consists in finding a code word at
minimum Hamming distance which yields a coding gain of
G a 10.log (R(t+ 1))
where R is the code rate and tthe error correction capability given by t =r (<5-1 )/ 21
With soft inputs, optimum decoding consists in finding a code word at minimum
Euclidean distance which yields a gain of
G a 1O.log (R <5).
For product codes with <5=36 or <5=16 the difference in coding gain is ~G 3 dB.
In this paper we consider the degradation in coding gain as we decrease the number
of quantization bits down to one with BTC.
Figure 4 shows the behaviour of BTCs, as we change the number of quantization
levels. The code used in this diagram is the product code BCH(32,26,4)2. The
simulations are given for 2, 3, 4, 7, 8 and 31 levels. The first noteworthy result is
that, between binary quantization (2 levels) and 31 levels, the difference is around
2dB at BER = 10-6 , which is less than 3 dB predicted by theory. If we introduce
erasures in the input data (3 levels), an improvement of O.5dB is obtained at BER=
10-6 . The impact of quantization levels depends on factors such as the error
correcting capability or length of the codes. When the error correcting capability
increases, the difference between 31 levels and 2 levels decreases. For example, for
the product code BCH(32,21 ,6)2 the loss between 31 levels and 2 levels is 1.5 dB at
BER=1O-6 instead of3 dB.
1,OOE-01 -+-8 levels

_ 7 levels
--'-4 levels
1,OOE-02 -+-3 levels
-+-2 levels
--+-
1,OOE-03
0::
w 1,OOE-04
III
1,OOE-05
1,OOE-06
1,OOE-07
Eb/NO in dB
Fig 4: Performance for the product code BCH(32,26,4/ at iteration 8, as afunction

of the number of quantization levels.
202
Figure 5 shows that for the product code BCH(128,113,6)2 the difference between
31 levels and 2 levels is around IdB at BER=1O·6 , and is only O.8dB when the
product code is BCH(256,239,6)2, which is a very long code.
-= - -
- --
- ~ --
-- . ~
-- -- --
-- --. ---
-- -- ~
1,00E-07 -+-
_
!256,239) 31 levels
256,239) 2 levels ~ ~ ~
--
- . - 128,113) 31 levels :::
___ 128,113) 2 levels - - -
1,00&08 .Ll::==~:::I::i::ci:=ii:r..u...u..Ll
2,8 3,1 3,4 3,7 4 4,3 4,6
Eb'NOindB
Fig 5: Performance for the product code BCH(J28,113,6/ and BCH(256,239,6/ at
iteration 8 as a jUnction of the number ofquantization levels.
Another way to evaluate the performance of the block turbo code is to compare the
bit error ratio BER(OUT) at the output of the decoder with the BER(lN) at the input
(see Fig.6).
203
BER(OUT)
10° ,------------------,
BER(IN)
Fig 6: Bit error ratio BER(OUT) as a function of BER(IN) for different BTCs.
We observe the very rapid variation of the output BER with input BER as we
increase minimum distance of component codes and their code length. For product
code BCH(256,239,6)2, the extrapolated BER(OUT) is very small «10- 15 ) for an
input BER < 1%.
5. CONCLUSION
BTC have been extensively investigated in the last eight years and their main
properties are:
-large minimum distance (16,24,36".),
-very high code rate (up to 0.96),
-low complexity « 10.000 gates),
-very high data rate (>50 Mbps).
In this paper we have investigated the effect of input data quantization on the
performance of BTC. Simulation results given here show that the loss in
performance when feeding binary data input instead of soft data input to a BTC
decreases as the code gain increases. In the case of BCH(256,239,6)2 the loss is of
0.8dB instead of the theoretical 3.0 dB loss. Although the theoretical justification is
not yet established, this result is extremely important for applications where only
binary data is available at the decoder input and should open the way to considerable
improvement of quality of service on heterogeneous networks.
REFERENCES
[1] R. Pyndiah, A. Glavieux, A. Picart and S. Jacq, "Near optimum decoding of
products codes," in proc. of IEEE GLOBECOM '94 Conference, vol. 113, Nov.-Dec.
1994, pp. 339-343.
204
[2] R. Pyndiah, "Near optimum decoding of product codes : Block Turbo Codes,"
IEEE Trans. on Comm., vol 46, nO 8, August 1998, pp. 1003-1010.
[3] C. Berrou, A. Glavieux and P. Thitirnajshima, "Near Shannon limit error-
correcting coding and decoding: Turbo-codes," IEEE Int. Con! on Comm.ICC'93,
vol 2/3, May 1993, pp. 1064-1071.
[4] P. Elias, "Error-free coding," IRE Trans. on Inf. Theory, vol. IT-4, Sept. 1954,
pp.29-37.
[5] P. Adde, R. Pyndiah, O. Raoul, "Performance and complexity of block turbo
decoder circuits," Third International Conference on Electronics, Circuits and
System ICECS'96, Rodos, Greece, 13-16 October 1996 -, pp 172-175.
[6] R. Pyndiah, "Iterative decoding of product codes: Block Turbo Codes," in proc.
of IEEE International Symposium on Turbo Codes & Related topics, Sep. 1997,
pp.71-79.
[7] P. Adde, R. Pyndiah, O. Raoul and J.R. Inisan, "Block turbo decoder design,"
International Symposium on turbo codes and related topics, Brest, Sept. 1997,
pp.166-169.
[8] A. Goalic and R. Pyndiah, "Real time turbo decoding of product codes on a
digital signal processor," in proc. of IEEE International Symposium on turbo codes
and related topics, Brest, Sept. 1997,pp. 267-270.
[9] S. Kerouedan, P. Adde et P. Ferry, "Comparaison performances/complexite de
decodeurs de codes BCH utilises en turbo decodage", GRETSI'99, Vannes, 13-17
Sept. 1999, pp 67-70.
[10] S. Kerouedan and P. Adde, "Implementation of a block turbo decoder in a
single chip", 2nd International Symposium on turbo codes and related topics, Brest,
Sept. 2000, pp. 243-246.
[11] S. Kerouedan, P. Adde and R. Pyndiah, "How we implemented block turbo
codes?", Annales des Telecommunications, tome 56, juilletlaout 2001, pp. 447-454.
[12] S. Kerouedan and P. Adde, "Block turbo codes: Towards implementation", 8th
International Conference on Electronics, Circuits and System ICECS'200J, Malta,
September 5-7, pp. 1219-1222.
[13] D. Chase, "A class of algorithms for decoding block codes with channel
measurement information," IEEE Trans. Inform. Theory, vol IT-18, Jan. 1972, pp.
170-182.
[14] P. Adde and R. Pyndiah, "Recent simplifications and improvements of Block
Turbo Codes", 2nd International Symposium on turbo codes and related topics,
Brest, Sept. 2000, pp. 133-136.
Gallager Codes for Asynchronous Multiple Access.
A. de Baynast t and D. Declercq

Laboratoire ETIS - ENSEA
6, avenue du Ponceau F95014 Cergy-Pontoise Cedex
France
email: debaynas.declercq@ensea.fr
Abstract. In the next wireless communication systems (UMTS or IMT-2000), the

symbol asynchronism will occure in the uplink transmission . The classical methods
(matched filter or RAKE receiver) are unsuitable in that case, and a joint decoding is
necessary. In this paper, we propose a robust method to the symbol asynchronism for
a multiple access system based on the Gallager codes. As in the synchronous case (de
Baynast and Declercq, 2002), we show that we can bypass the spreading sequences
in an asynchronous multiple access system. The" algebraic" orthogonality is replaced
by a "statistical" orthogonality which is provided by the randomness of the Gallager
codes. The performance of our system are close to the optimum achievable bound
(2dB or less at BER= Ie - 3 for a block length N = 2000) for many configurations.
Keywords: Joint multiple access detection, symbol asynchronism, Gallager codes,

capacity for an asynchronous multiple access system.
Introd uction
In the scope of multiple access techniques, spread spectrum has been

proposed for code division multiple access (DS-CDMA) and is success-
fully used in IS-95 system. Spread spectrum multiple access employs a
set of spreading sequences which provide an "algebraic" orthogonality
between all users in a cell. This orthogonality helps to mitigate the
multiple access interference (MAl) at the receiver.
For third generation systems, IMT -2000 or UMTS, the increase in
rates implies a smaller spreading factor: 1 ~ SF ~ 16 instead of 1024 in
IS-95. In the uplink transmission, asynchronism - chip rate or symbol
rate - between the users dramatically destroys the orthogonality of the
spreading sequences; interference penalizing too much the performance
of the RAKE receiver, e.g. (Johansson and Svensson, 1995).
For the asynchronous DS-CDMA system, (Scaglione et al., 2000)
allow each user to activate - at each symbol interval - just one of
Nf frequencies (Nf > N u , the number of users). This family of codes
A. de Baynast is supported by a DGA (French Defense) PhD grant.
205
Multiaccess, Mobility and Teletraffic for Wireless Communications. Volume 6, 205-220.
206
for CDMA systems eliminate MAl completely. However, it requires the

introduction of trailing zeros also called guard chips at the end of the
spreading sequences. The number of these guard chips is at least equal
to the maximum offset between the users in terms of chips. Because this
number should not be too large, the users have to be quasi-synchronous.
This method is more adequate to IS-95 than to IMT-2000.
More generally, none set of spreading sequences can achieve the
Welsh's bound in a completely asynchronous transmission even with
strong assumptions like quasi-synchronous transmission, Gaussian in-
terference, etc.
Since the asynchronism penalizes too much the performance of a DS-
CDMA system, it seems to be judicious to consider an asynchronous
multiple access scheme which favors coding at the expense of spreading:
the" algebraic" orthogonality between the spreading sequences is then
replaced by a kind of orthogonality in the code structures which we will
denote "statistical" orthogonality. Such a method has been successfully
proposed in the synchronous case (de Baynast and Declercq, 2002) and
would intuitively not be affected by the asynchronism (symbol or frame
asynchronism). That will be confirmed by simulations means (see section
4).
We describe now the proposed approach: we assign at each user a Gal-
lager code - also described in the literature as low density parity check
code (LDPC), (Gallager, 1962; MacKay, 1999) - and at the receiver we
adopt a joint decoding algorithm. This random coding scheme with long
enough block can reach the capacity of the AWGN synchronous MAC,
(Cover and Thomas, 1991) and (Chung et al., 2001). Whereas this result
has not been extended to the case of the AWGN asynchronous MAC,
we will show with simulations that the performance achieved by such
a system are very close to the capacity. We suppose in the sequel that
the offsets between all users are known from the receiver. Since the type
of synchronism that is difficult to achieve in many practical situations
is symbol synchronism, we limit our study on the symbol asynchronous
multiple access system.
The paper will proceed as follows: after describing the model in
the second part, we present the theoretical performance bounds of the
AWGN asynchronous multiple access channel from information theory.
In the fourth part, we describe the proposed joint multiple user decoding
algorithm. Finally, we present some simulation results.
207
1. Channel model
We supposed without loss of generality that we are frame-synchronous

and symbol asynchronous. The simultaneous Nu users are sorted in
ascending order of their respective offsets T1 :S T2 :S ... :S TNu with
T1 = O. Furthermore, we supposed that the sampling clock at the re-
ceiver is able to synchronize at any offset Tn and this for any value of
.6.Tn ,n+1 = Tn +1 - Tn. This synchronization can be realized by the use of
scrambling sequences and the use of a powerful digital downconverter.
Since the receiver knows the assigned rectangular pulse waveforms as
well the symbol period T and the relative offsets .6.Tn ,n+1, it can com-
N-1
pute y[l] = { y(l)[I], y(2)[I], ... ,y(Nu)[l] } /=0 by sampling the continuous
observation data at each instant IT + Tn (see figure 1). Because of the
sdl-l]
..
i
symbol period
..
s~dl-l]
...
. ....
1- TNu
Figure 1. Asynchronous model channel
N-1
factorization theorem (Lehman, 1959), { y(1) [1], y(2)[I], ... ,y(Nu ) [1] } /=0
are sufficient statistics to estimate the transmitted messages. This im-
plies that the channel output {y(t)} enters in the computation of the
posterior probability of each message only through y(n) [1]. See (Verd ti,
1989a) for more details. The basic discrete time AWGN asynchronous
multiple access channel with total input power constraint, rectangular
pulse and noise variance N o/2 can also be modeled by:
Nu Nu
Y= L anMT,nsn + n = L MT,nXn + n = MTx + n = z + n (1)
n=l n=l
where y = [yT[o]yT[l] .. . yT[N - l]t are the observed data with

y[1] = [y(1) [1] y(2) [l] ... y(Nu ) [l]], an the real fading coefficients, MT,n
208
the matrix which contains the asynchronism coefficients ~Tn,n+1' Sn =

{sn [I]} I=O, ... ,N -1 the transmitted codeword of length N and power
JE[s;[I]] = 1 from the nth user and n = {n[l]}1=0, ... ,N-1 the additive
white noise following the normal distribution.
Let en be a binary linear (N, K) code, i.e., a code of block length
N, dimension K. Let M = N - K (the code rate R is equal to R =
KIN); then en is defined by a M X N parity-check matrix Hn and every
codeword Sn E en satisfies the parity-check equation Hnsn = o. Together
with the parity-check matrix H n , we associate a generator matrix G n
(of full column rank) of size N X K which satisfies:
(2)
where b n is a vector of K information bits. We suppose in the sequel
that the frame length N, the code rate R are equal for all users, we are
frame synchronous and one trailing zero is added at the beginning of
each frame to ensure the causality of the transmission.
2. Performance limits of a asynchronous multiple user system
The aim of this section is to recall the performance limits of the AWGN
asynchronous MAC for BPSK sources (E {-I, I} ). This analysis will be
useful to evaluate the performance of the proposed method. Under the
assumptions of the previous section, we adopt the following definition
according to (Gilhousen et al., 1991; Verdu and Shamai, 1999):
DEFINITION 1. The spectral efficiency D [bits.s- 1.Hz-1} of the global

system (sometimes referred also as the system load) is equal to the sum
of achievable rates:
(3)
2.1. CAPACITY OF THE ASYNCHRONOUS AWGN-MAC WITH

GAUSSIAN SOURCES
Before we derive the capacity, i.e. the maximal spectral efficiency, of the
asynchronous AWGN-MAC with BPSK sources, we recall the formula of
the capacity of the asynchronous AWGN-MAC with Gaussian sources.
Under the global power constraint, i.e. the average power per symbol
209
Ex = 1 defined as Ex IlL L;;~l IE [x;[I]], the capacity is given by

(Verdu, 1989b):
C = N~~OO 2~ log2 [II + 2C~~ MrM;ll (4)
where I denotes the identity matrix, 1.1 the determinant operator and
N the code block length.
With two users, it has been shown in (Verdu, 1989a) p.737 that the
worst case offset between the signals is zero, i.e. in which case the channel
is symbol synchronous. The most favorable case occurs when the symbol
offset is equal to half the symbol period. We observe the same results
with two BPSK sources.
2.2. CAPACITY OF THE ASYNCHRONOUS AWGN-MAC WITH BINARY

SOURCES
Under the following assumptions,

AI) global power constraint: the average power per symbol Ex 1
defined as Ex = l\t
L;;~l IE [s;[l]] ,
A2) all users are equal power: IE [s;[I]] = 1, \In = 1, ... ,Nu
A3) distinct offsets: LlTn ,n+l : :J. LlTn ',n'+l, \In':::J. n,

the capacity of the system is given by (5).
with a~, the noise variance defined as: a~ = 2CE~/No'

PROPOSITION 1. Under the conditions Al and A2, the capacity for
BPSK sources is:
- maximized if all offsets are distributed uniformly over a period

symbol (LlTn ,n+l = J),
210
minimized if the users are synchronous (Tn = 0, Vn = 1, ... ,Nu ).

As shown on figure 2, limEb/N0-400 C = 1.5 for 2 synchronous users
(0 dashdotted line) and limEb/N0-400 C = 2 when the offset ~T i= 0 (x
solid line). Indeed, in the noiseless case, if the users are synchronous and
have equal power, an ambiguity occurs when 0 is observed: it may come
from {-I, +I} or {+ 1, -I} and a coding step is necessary even in the
noiseless case. The asynchronism suppresses this ambiguity by taking
account of the values of the neighbouring samples.
8rr=~~~~~~~~----~r------'
+ BPSK Nu=2 - a~=OT (sync.)
7 - - BPSK Nu=2 - a~=O.25T
o
-1
-~L-------~------~------~~------~2
Figure 2. Ebj No vs. capacity: Gaussian sources, BPSK 2 users with equal power
(asynchronous case: AT = O.25T and O.50T, synchronous case: AT = OT).
2.3. ACHIEVABLE BER FOR GAUSSIAN ASYNCHRONOUS MAC WITH

BINARY SOURCES
In practice, the system works at a certain non-vanishing BER depending

of the required Quality of Service (QoS). The aim of this section is to
derive the minimum Eb/NO required with respect to the expected BER.
The rate R' achieving a certain non-zero BER will obviously be greater
that the rate R achieving a transmission without error.
In a multiple user context (Nu BPSK modulations), in order to derive
the new rate R', we assume that each user has the same BER for a given
Eb/NO. This assumption is verified if all users have the same coding rate
211
R and the same power. In that case, the new code rate R'(R) is given
by (6).
R' R[Nu + ; (;;u ) BER u- n(1- BER)n.

N
log2 (BER N u- n (1- BERt) 1 (6)
Figure 3 shows the achievable BER as a function of Eb/NO for rates

R = ~ and ~, 2 users with different offset (asynchronous case: bo r =
O.25T and O.50T, synchronous case: bor = OT).
Figure 3. BER vs Eb/ No for different rates (R = t, t) - Equal power BPSK async.
2 user (b.T = OT, O.25T, O.50T).
These abacuses will be useful to compare the performance of our simu-

lation results with the theoretical bounds with respect to the parameters
of the system.
After having given the performance bounds for AWGN MAC with
respect to the signal to noise ratio Eb/NO, the rate R and the BER, we
describe the proposed algorithm based on Gallager random-like codes.
212
3. Joint multiple user detection based on Gallager codes
Since the introduction of turbo-codes in (Berrou et al., 1993), many new

coding and decoding techniques have been proposed (MacKay, 1999).
It turns out that all good codes are random-like codes and that they
share a common decoding algorithm: the belief propagation on graphical
representations (Kschischang et al., 2001). In this paper, we use the
factor graphs that are powerful tools to develop decoding algorithms.
Factor graphs have been proposed by (Wiberg, 1996) as a generalization
of Tanner graphs in coding theory (Tanner, 1981). They are bipartite
representations of systems composed of data nodes and functional nodes.
The data nodes represent observations and input symbols while the func-
tion nodes describe how their adjacent data nodes interact. The branches
of the graph carry probability weights that comes in and out the data
nodes. Belief propagation in a graph depicts how the weights are updated
until a fixed point has been reached (Kschischang et aI., 2001). It can
be shown that exact a posteriori weights can be computed if the factor
graph is indeed a tree, that is there is no cycles in the graph. Besides,
if the cycles in the graph are "sufficiently" long, iterative decoding with
probability propagation yields excellent (though approximate) results,
close to optimum performance.
First, we briefly describe the decoding problem for Gallager codes in
the single user case and in the second paragraph, we derive the joint
asynchronous multiple user decoding problem for Gallager codes.
3.1. GALLAGER CODES IN THE SINGLE USER CASE
These block codes have been proposed by Gallager in 1963, together with
a stochastic decoding algorithm which is very close to belief propagation.
Mc Kay & al. have rediscovered and extended LDPC Gallager codes
recently (MacKay, 1999) and have shown that Gallager codes can be
easily decoded with iterations of belief propagation on their factor graph
(d. figure 4).
First, we describe the functional nodes (black square). Since it is the
single user case, we have: x[/] = s[/]. Each channel node calculates the
conditional probability densities:
p(x[/]Jy[l]) ex ~exp
27r(J2
(-(y[l] - x[/])2/2(J2) (7)
213
Each parity-check node indicates that the set Qk of the codeword bits
{s[l]} E Qk to which the parity-check is connected have even parity:
Ls[l] = 0 mod 2 (8)

Qk
When a channel output vector y is observed, the iterative probability

propagation decoder begins by sending messages P(x[I]ly[l]) according to
(7) from y to x. Then messages Q~1) (x[l]) are sent from the codeword x
to the parity-check constraints and messages R~l) (x [I]) from the parity-
check constraints back to the codeword x according to (8). Each time an
iteration i is completed, new estimates of APP(x[I]ly) for I = 1, ... ,N
are obtained. After a prespecified stopping rule such as the maximum
number of iteration or no change in the estimated codeword has been
reached, the iterative decoding stops. For more details on the Gallager
codes decoding, refer to (MacKay, 1999; Kschischang et al., 2001).
channel
output
channel
node
codeword
parity-check
node
Figure 4. A factor Graph for a Gallager code C(N, M)
3.2. GALLAGER CODES IN THE ASYNCHRONOUS MULTIPLE USER CASE
!
In the multiple user case, we rewrite the model described by (1):
y(1)[IJ [0 ~Tl,2 ~Tl,2 0 ~Tl,2 OJ x[IJ + n(1)[I]

y(2) [IJ [0 ~Tl,2 0 ~Tl,2 ~Tl,2 OJ x[IJ + n(l)[l]
y(Nu ) [I] ~
214
or equivalently in a vectorial form
y[l] = M~x[l] + n[l] = z[l] + n[l) (9)

with
Then, each codeword xn[l] is connected to both variables z[l-l] and z[I].
Figure 5 shows the factor graph for a joint asynchronous multiple user
system using Gallager codes. The fading coefficients an are supposed to
be perfectly known at the receiver.
As in the single user case, each channel node calculates the conditional
probability densities for the channel:
p(z[I]Jy[l]) ex 211"1a 2 exp (-JJy[l] - z[I]JJV2az) (11)
In the multiple user case, z[l] is described by (10). The variable z[l] has
2Nu components and then 22Nu possible states.
We define, as the "spine-check" node, the functional node to which
z[I], z[1 - 1] and Sn[l] , iiI :S n :S Nu are connected. Using (10), this
functional node is described by:
Nu
L anMT,nsn[l] = z[l] (12)
n=l
with Sn[l] = [Sn[l- l]sn[l]]T

Such as in the single user case, once a channel output vector y is ob-
served, the iterative probability propagation decoder begins by sending
messages P(z[I]Jy[l]) from y[l] to z.
Messages 01/[1- 1] are sent from z[l-l] to the spinal-check and T'[l]
are sent from the spinal-check to z[l]: the forward step. Messages O'[l]
are sent from z[l] to the spinal-check and TI/[l - 1] are sent from the
spinal-check to z[l- 1]: the backward step.
Then messages P?)(st[l]) (resp. S?)(st[l])) are sent from the spine-
check node to Sl [I] (resp. from sdl] to the spinal-check node). The user
index is arbitrary in the case of all users with equal power. As in the
single user case, messages Q~l) (Sl [I)) are sent from the codeword Sl to the
parity-check constraints and messages R~l) (sdl]) from the parity-check
constraints back to the codeword 81 according to (8). For the others
users (2 :S n :S N u ), the procedure is exactly the same. Each time an
215
Gallager code for user 2
Gallager code for user 1
Figure 5. A factor Graph for a joint asynchronous multiple user decoding algorithm
using Gallager codes C(N, M)
iteration i for all users is completed, new estimates of APP(mn[I]ly) for

1= 1, ... ,N are obtained.
3.3. ENCODING AND DECODING COMPUTATIONAL COMPLEXITY
3.3.1. Fast-encoding Gallager codes

One of the drawbacks of Gallager codes is that their encoding time gen-
erally scales as 0 (N 2 ) because the generator matrix G is not generally
sparse. In fact, some methods exist to ensure the generator matrix is
sparse (see for instance (MacKayet al., 1998)). In our case, we compute
the Gaussian elimination to calculate the generator matrix G n from
the parity-check matrix Hn using the Markowitz criterion, (Duff et al.,
1986). It ensures a good sparsity for G n , roughly say O(3N), despite it
is a local criterion.
3.3.2. Decoding
The major limitation of such joint multiple user detection algorithm is its
exponential complexity in the number of users. Fortunately, there exist
various means in order to reduce this complexity. A well-known result
in the graph theory, see for instance (Frey, 2000) is that the exponential
complexity at the spinal node can be reduced to a polynomial complex-
ity o(N~) with no loss in performance. In most of cases, this remains
too complex. Several suboptimal strategies are possible. For example,
216
a hybrid structure" joint decoding/successive interference cancellation

(SIC)" can be applied. The users are gathered in subsets (4 to 8 users
in each subset). A joint decoding algorithm for each subset is used while
the SIC procedure (Patel and Holtzman, 1994) is used to go of a subset
to another. In a future work, we plan to quantify the loss induced by
this suboptimal structure.
4. Simulation results
In this set of simulations, we present the performance of the proposed

asynchronous multiple user joint decoding algorithm for several code
rates R and several system loads over the AWGN multiple access chan-
nel. Each user has the same code rate R. The system load is defined as
the ratio between the number of user Nu and the processing gain R (see
section 2 for more details). We run the decoder on 103 frames for a block
length N = 2000. Although the Gallager codes are much better for a
larger block length (~ 20000), we limit our simulations to this length
for realistic implementation issues. To limit the computation duration,
we set the maximum number of iteration to 100 whereas the algorithm
did completely not finish converging. To be in agreement with a UMTS
system, the system should work at an high spectral efficiency: we set the
system load to values greater than 0.5 in all the cases. The transmission
is BPSK modulated. Perfect knowledge of the channel fading coefficients
(equal to one in these simulations) and of the offsets is assumed. We
compare in presence of asynchronism our algorithm with a DS-CDMA
system. To simulate the DS-CDMA system, the spreading sequences are
OVSF sequences of length 2 ([11] and [1 - 1]). At the receiver, we use
a matched filter followed by a single user Gallager decoding described
in section 3.1. The simulations are reported in figure 6 and 7. Since the
performance of all users are very close (less than O.ldB), the average
BER between all users is pictured.
In the first simulation, we set the system load to 1 with R = 1/2 and
Nu = 2 for several asynchronism offsets: ~T = OT (sync.) , 0.2ST and
0.5T. The number of ones per row tr in the parity-check matrices Hn is
set to 5. Note that in this paper, we do not optimize the parity-check
matrices Hn. This optimization will be reported in future work. Addi-
tionally, we have also drawn the achievable BER(Eb/No) given in section
2 for R = ~ and ~T = OT, 0.25T and 0.5T. These curves correspond to
the lower achievable bounds of the transmission. Note that we do not
compare the obtained performance to the single user performance since
217
in a multiple access system the single user performance is not generally

the optimum achievable bound (even it is true at high signal to noise
ratio and unequal power or ~T =I 0).
In the second simulation, the system load is equal to 0.5 i) with a
rate R = 1/4, no spreading and Nu = 2 (MAC based on Gallager codes,
o solid line) and ii) R = 1/2, SF = 2 which gives a "total" redundancy
factor equal to 4 and Nu = 2 (DS-CDMA, 0 dashed line). The number
of ones per row tr in the parity-check matrices Hn is set to 3. The results
are pictured on figure 7. Additionally, we have drawn the performance of
the synchronous multiple user joint decoding algorithm for a rate R = ~
(+ dashdotted line).
1O~'0'---0"-.5----'--'------'--'-2_----:'-c_--'-_----,-'------'
EblNO [dB]
Figure 6. BER vs Eb/No - R ~ - Equal power BPSK async. 2 user

(.6.7 = OT, O.25T, O.50T).
We can notice that our asynchronous multiple user joint decoding

algorithm provides good performance since the iterated process is close
to the optimum curve (2dB or less at BER= Ie - 3) for the rates
R = 1/2,1/4 (for R = 1/4, the theorical bounds are pictured on fig-
ure 3). These results are in agreement with the results on the coding
theory of section 2 and confirm both following facts: i) a multiple access
system do not necessarily need an "algebraic" orthogonality between
the users to be powerful, ii) the proposed joint algorithm is robust to
the asynchronous (moreover, the best performances are obtained for an
asynchronism offset equal to an half symbol period whereas the worst
performances are obtained in the synchronous case).
218
10-'oL--:'-:----'---:":---:---:'-::----:'-----::'-:-----'
EblN~[dBl
Figure 7. BER vs Eb/No - BPSK 2 async. user (boT = O.5T) - 0: DS-CDMA (R = t,

SF = 2), 0: the proposed method based on Gallager codes (R = :D
As expected, the performance of the DS-CDMA system using a

matched filter and a Gallager single user decoder is badly affected by
the asynchronism. Note that the offset ~r = 0.5T is is not the worst.
Indeed the average multiple access interference is equal to kIE[s2] in this
case and for an offset ~r = IT (i.e. a chip period), the average MAl
is equal to lE[s2]. However, the complexity of the DS-CDMA system is
linear with the number of users whereas the complexity of the proposed
algorithm is exponential in the number of users.
Conclusion
In a symbol asynchronous DS-CDMA system at high spectral efficiency

(uplink transmission in IMT-2000 with a small spreading factor), the
performance of the decoder are too much penalized by the MAl af-
ter the despreading step. In this paper, we have also proposed a joint
asynchronous multiple user decoding algorithm using Gallager codes.
We have shown that i) the asynchronism can improve the performance
if a joint decoding algorithm is used, ii) we can bypass the spreading
sequences in an asynchronous multiple access system as in a synchronous
system, see (de Baynast and Declercq, 2002). To make a comparison with
the theorical AWGN MAC capacity, we have derived the performance
219
bounds with respect to the code rate R, the number of users N u , the sig-
nal to noise ratio Eb/NO and the BER. Note that these bounds generally
are not equal to the single user bounds. The performance of our system
are close to the optimum achievable bound (2dB or less at BER= Ie - 3
for a block length N = 2000) for a system load RNu equal to 1 and 0.5
(R = 1/2 and 1/4). The major problem is still the complexity of the
algorithm exponential in the number of users. As described in section
3.3.2, a suboptimal method combining multistage and iterative detection
is possible.
References
Berrou, C., A. Glavieux, and P. Thitimajshima: 1993, 'Near Shannon Limit Error-
Correcting Coding and Decoding: Turbo Codes'. In: ICC. pp. 1064-1070.
Chung, S.-Y., T. Richardson, and R Urbanke: 2001, 'Analysis of sum-product de-
coding of low-density parity-check codes using a Gaussian approximation'. IEEE
Trans. Inform. Theory 47, 657-670.
Cover, T. and J. Thomas: 1991, Elements of information theory. New york: Wiley
edition.
de Baynast, A. and D. Declercq: 2002, 'Gallager codes for multiple access'. accepted
to IEEE Symposium on Information Theory.
Duff, I. S., A. M. Erisman, and J. K. Reid: 1986, Direct methods for sparse matrices.
Clarendon Press; Oxford University Press, oxford: New-York edition.
Frey, B.: 2000, Graphical Models for Machine Learning and Digital Communication.
Cambridge, Massachussets: The MIT Press.
Gallager, R: 1962, 'Low-Density Parity-Check codes'. IRE Transactions on Infor-
mation Theory.
Gilhousen, K. S., I. Jacobs, R Padovani, A. Viterbi, 1. Weaver, and C. Wheatley:
1991, 'On the Capacity of a Cellular CDMA System'. IEEE Trans. on Vehicular
Technology 40(2).
Johansson, A. and A. Svensson: 1995, 'Multi-stage interference cancellation in multi-
rate DS/CDMA systems'. In: PIMRC'95. Toronto, Canada.
Kschischang, F., B. Frey, and 1. H.-A.: 2001, 'Factor graphs and the sum-product
algorithm'. IEEE Trans. Inform. Theory 47(2), 498-519.
Lehman, E.: 1959, Testing Statistical Hypotheses. New york: Wiley edition.
MacKay, D.: 1999, 'Good Error-Correcting Codes Based on Very Sparse Matrices'.
IEEE Transactions on Information Theory 45.
MacKay, D., S. Wilson, and M. Davey: 1998, 'Comparison of constructions of irregular
Gallager codes'.
Patel, P. and J. Holtzman: 1994, 'Analysis of a simple successive interference can-
cellation scheme in a DS/CDMA system'. IEEE Journal on Selected Areas in
Communications 12(5).
Scaglione, A., G. Giannakis, and S. Barbarossa: 2000, 'Lagrange/Vandermonde MUI
Eliminating User Codes for Quasi-Synchronous CDMA in Unknown Multipath'.
IEEE Transactions on Signal Processing 48(7), 2057-2073.
220
Tanner, R.: 1981, 'A recursive approach to low complexity codes'.

Verdli, S.: 1989a, 'The Capacity Region of the Symbol-Asynchronous Gaussian
Multiple-Access Channel'. IEEE Trans. Information Theory 35( 4).
Verdli, S.: 1989b, 'Multiple-Access Channels with Memory with and without Frame-
Synchronism'. IEEE Trans. Information Theory 35(3).
Verdli, S. and S. Shamai: 1999, 'Spectral Efficiency of CDMA with Random
Spreading'. IEEE Trans. Information Theory 45(4).
Wiberg, N.: 1996, 'Codes and decoding on general graphs'. Ph.D. thesis, Linkiipings
Universitet, Sweden.
Bounding Techniques for the Design of Channel Coding
and Modulation Systems
Yufei W. Blankenship and Brian K. Classon

Motorola Labs - Communication Systems Research Laboratory
1301 E. Algonquin Road, Schaumburg, IL 60196 USA
{yufei, classon}@labs.mot.com
Keywords Channel coding, channel capacity, sphere packing bound
Abstract The composite channel capacity of various modulations in the AWGN

channel is presented. Optimal constellation shaping is found to be only
slightly better than the conventional rectangular constellation and thus not
justified for practical implementation for the purpose of capacity increase.
Shannon's sphere packing is extended to lower bound the performance of
channel coding with realistic modulations. An adjusted sphere packing
bound is proposed to more closely bound the channel coding performance
with realistic modulations. The adjusted sphere packing bound takes finite
block size and realistic modulation into consideration, and is closely
approached when the channel coding is well-designed. The bounds are
useful to expedite system design.
1. INTRODUCTION
Communication systems use increasingly complicated channel coding and
modulation techniques to gain a higher throughput with lower power
consumption. Designs such as adaptive modulation and coding (AMC) and
hybrid automatic repeat request (HARQ) are likely to be used in 3G and 4G
systems. A set of operating modes must be determined, each with a different
combination of parameters including the frame size, the code rate, and the
modulation order. It is desirable to have a quick and systematic method to
find the correct combination.
One method is to use bounding techniques. With the capacity-approaching
channel coding methods, such as turbo codes [1] and low-density parity-
check codes, it is possible to find a theoretic bound which is closely
approached by a well-designed code [2]. The bound includes the effect of
finite block size and realistic modulation, where the modulation is
considered a part of a composite channel. The bounds can quickly provide
valuable information on the potential performance of the system without
requiring extensive simulations. Methods making use of the bounds have
been applied to improve the design of a HARQ system for 1xEV-DV [3].
221
222 Yufei W Blankensip and Brian K. Classon
In this paper, the theoretical channel coding performance limits are

studied using the system model in Figure 1. The bounds defined by the
channel capacity and composite channel capacity are presented for additive
white Gaussian noise (AWGN) channels with commonly-used modulation
schemes. The channel noise is assumed to have a zero mean and a variance
of N o/2. Bounds of AWGN channels are used for determining the system
capacity of a fast AMC system since the instantaneous channel of certain
fading channel can be characterized as an AWGN channel [4].
The composite channel capacity of a modulation with different
constellation shaping is studied. It is found that the mathematically optimal
constellations may not be worth the implementation complexity if the only
purpose is to gain a higher capacity, although constellation shaping can be
useful for other purposes, such as reducing peak-to-average power ratio.
Shannon's sphere packing bounds are extended to higher order modulations
to show the lower performance bounds as functions of frame size k, the
signaling rate r, and the word error probability Pw' In order to find a tighter
bound, the sphere packing bound is combined with the capacity limit to
provide an adjusted sphere packing bound. Simulations of turbo codes with
combinations of modulation scheme, code rate, and frame size show that the
simulation is about 1 dB away from the adjusted sphere packing bound
uniformly.
rate R M = 2 q constellation
k info bits
estimated
info bits
Figure 1. The communication system model.
2. COMPOSITE CHANNEL CAPACITY LIMITS
2.1. Composite Channel Capacity of Conventional Modulations

Let a discrete memoryless channel have an input alphabet X= {xo, XI,""
xK-d and an output alphabet Y= {Yo,Y), ... , YJ-I}' Its channel capacity is
defined as [5]
C = max J(X; Y) (info bits/symbol), (1)
q(x,)
where q(Xi) is the probability distribution function of the channel input Xi. Let
Cw and C2D represent the capacity limits for anyone dimensional (1-D) and
Bounding Techniquesfor the Design of Channel Coding and Modulation Systems 223
two-dimensional (2-D) signaling technique, respectively, on an AWGN

channel. These limits can be reached when the channel input is allowed to
have continuous values, or when the number of discrete channel symbols
approaches infinity.
Using (1), Cw ofthe static channel is found to be
Cw = .!..10 g (1 + 2Es)
2 (info bits/symbol). (2)
2 No
When the capacity limit is achieved, the signaling rate is r = Cw . Thus, the
minimum bit SNR requirement is
(~E)w 22ClD -1
2Cw
(3)
For 2-D signaling with equal noise power on each dimension, the capacity of
the 2-D AWGN channel is
(4)
where Es is the average energy of the vector symbol and equal to twice the
symbol energy of I-D signaling. Considering Es = rEb = C2nEb, the minimum
Eb/NO is
(5)
(6)
where
224 Yulei W Blankensip and Brian K. Classon
(7)
and the average power per symbol Es = _1_ I (Ai 2 Bi 2) is used to normalize
M i~O
+
the symbol energy in (7). The function D(u,v) accounts for the location of the
constellation points. For instance, an M-PSK constellation has Es = 1, and
can be defined as Aj = cos 27rj/M and B j = sin 27rj/M .
Figure 2 shows the composite channel capacity found by (6) for QPSK,
8-PSK, 16-PSK, 16-QAM, 32-QAM, and 64-QAM. The 2-D capacity limit
C2D is also plotted for comparison. Each modulation uses the conventional
constellation structure, i.e., PSK symbols are evenly spaced on a circle, and
the QAM symbols are evenly spaced on a rectangular grid.
Some design guidelines can be derived from Figure 2. The maximum
number of information bits that can be transmitted per channel symbol is
determined by the lowest amount of redundancy introduced by channel
coding (i.e., highest code rate). Since when the capacity is reached,
r = C = Rq, R = C/q, thus the minimum Eb/NO is actually a function of the
channel coding rate R. As expected, each modulation with M (= 2q)
constellation points achieves the maximum of q information bits/symbol, or
code rate R = 1, as Eb/NO becomes large (Eb/NO> 16 dB for all under
discussion). As M increases, the number of information bits per pulse-
amplitude modulation (PAM) symbol approaches Cw , while that of PSK and
QAM approaches C2D • The gap between the C2D and the modulations in
Figure 2 is due to the nonideality of the modulation techniques.
Bounding Techniques for the Design of Channel Coding and Modulation Systems 225
13PSK •
QPSK :
BPSK :
4 6 8 10 12 16
Minimum E"INo (dB)
Figure 2. The composite channel capacity (info bits/symbol) vs. minimum E,)No (dB) for
modulations BPSK, QPSK, 8-PSK, 16-PSK, 16-QAM, 32-QAM, and 64-QAM. Unquantized
detection and a static channel are assumed. The 2-D capacity limit C2D is plotted for reference.
Within the same modulation category (PSK or QAM), Figure 2 shows that
the bandwidth efficiency, measured by signaling rate r (= C), is always larger
when M is higher. However, in general, high M implies high signal
processing complexity. Thus, for the bandwidth efficiency region where
there is little difference between higher- and lower-order modulation, the
lower-order modulation should be used. For example, both R = liz 64-QAM
and R = % 16-QAM provide r = 3 info bits/symbol. Since there is only
0.35 dB difference in performance and 16-QAM requires less signal
processing, it may be more appropriate to use R = % 16-QAM. Figure 2
suggests that for the same modulation categoryl, the system can use a 2Q-ary
modulation for a code rate R < 0.8 at the expense of less than 0.5 dB instead
of a 2Q+ l_ary modulation with a code rate Rq/(q+ 1). In other words, higher
code rate with lower order modulation may be more appropriate than lower
code rate with higher order modulation. However, as R gets closer to 1
(C~q), the curve of a given modulation flattens, and the increase of the
minimum E,JNo brings quickly diminishing returns in bandwidth efficiency
for the increased code rate. It is better then to use a higher level modulation.
2.2. Constellation Shaping

In the above composite channel capacity of conventional I-D and 2-D
modulation techniques is discussed. In this section the possibility of
maximizing the composite channel capacity of 2-D modulations by
1 The rule does not apply to BPSK and QPSK since the dimensionality changes between
BPSK and QPSK.
226 Yufei W. Blankensip and Brian K. Classon
optimlZlng the constellation is investigated. Possible gain of capacity is

measured with reference to conventional PSK or QAM design. In general, a
ND-dimensional cube, formed by Cartesian product of PAM constellation
with itself ND times, can be used as a reference since it has the same power
efficiency as PAM [7]. When ND approaches infinity, a constellation with
points in an ND-dimensional sphere has 1.53 dB shaping gain over a
constellation with points in an ND-dimensional cube [8].
A figure of merit of a constellation is the minimum squared distance
between signal points normalized by the average energy [7], i.e., d~n/Es . As
can be seen from expressions such as the union bound, the error probability
performance depends on d~n/Es as well as the number of nearest neighbors,
which could be fairly large in high-dimensional space. Neglecting the
influence from the number of nearest neighbors, the power efficiency of a
constellation is inversely proportional to d~n / Es .
1 ...... ,. ..... .. ........ 0.5 .. ; ....... :....... , ....

0.5 .
o ..... : .. """".' .. : ....... o .........

-0.5 .
-0.5
-1 ....•. . . . . . ...: ..
-1 '--4-~---I~~--<"""'"
-1 0 1 -1 -0.5 0 0.5 1
(a) d;iJaverage energy ::O.58S (b)d~in/average energy =0.667
1.... .. ....................... .
....... :
0.5 ..
...... ,.......
-0.5 ...... , ..........,.................. .
-0.5
-1
-1 0 1 -1 0 1
(e) d ~in/average energy =0.845 (d) d~in/average energy =0.928
Figure 3. Four constellations of 8-QAM and their figures of merit. (a) 8-PSK, (b) rectangular
8-QAM, (c) optimal rectangular 8-QAM, (d) hexagonaI8-QAM.
Figure 4. The composite channel capacity (info bits/symbol) vs. minimum EJNo (dB) for the
four 8-QAM constellations in Figure 3 with AWGN channel.
In Figure 3, four typical constellations of 8-QAM are shown with unity
distance between adjacent points. Constellation (a) is 8-PSK and (d) is
constructed on the hexagonal lattice2 . Constellations (b) and (c) are
constructed on a rectangular lattice, with (c) being the best rectangular
design [9].
From the d~n/Es shown underneath their graphs in Figure 3, it can be
predicted tentatively that 8-PSK has the worst power efficiency, and the
hexagonal 8---QAM has the best. To test their performance, their composite
channel capacities vs. minimum EblNo are calculated and plotted in Figure 4.
The curves do show that the hexagonal design outperforms the other three.
The difference between the four designs is mainly in the high code-rate
region 0.6 < R < 1 (i.e., 1.8 < C < 3). The hexagonal 8-QAM shows as much
as 1.2 dB advantage over 8-PSK. However, there is very little difference
between the hexagonal 8-QAM and the optimal rectangular design.
The same study was also carried out on 16-QAM constellations. The
optimal 16-QAM constellation is about 0.2 dB better than the conventional
16-QAM. Thus there is little advantage to use the optimal constellation in
terms of the composite channel capacity when a good rectangular
constellation can be used. Considering the extra complexity required to
implement a mathematically optimal constellation, the small gain may not
justify abandoning the conventional rectangular QAM constellations.
2 The shape refers to the lattice, or the regular array of points, from which the signal points are
selected [6]. The densest two-dimensional lattice is the hexagonal lattice [8].
228 Yulei W. Blankensip and Brian K. Classon
3. SHANNON'S SPHERE PACKING BOUND

The sphere packing bound describes the performance of the optimal
spherical code, which is a mathematical model with the codewords being
distributed evenly on the sphere with maximum possible distance between
each other. The random coding bound describes the average performance of
all spherical codes. For a given bit SNR EblNo, the word error probability of
the optimal code, Pw.opt, is lower bounded by the sphere packing bound Pw.spb
and upper bounded by the random coding bound Pw.rcb,
i.e., Pw,spb ::;; Pw,oPt ::;; Pw,rcb' Since the sphere packing bound and the random
coding bound allow a code symbol to be any real number, i.e., not restricted
to any digital modulation constellation, they are lower (i.e., better) than the
optimal and average performance, respectively, achievable by an actual
channel code with any digital modulation technique. Nevertheless, because
the lower bound has become uniformly tight since the advent of turbo codes,
the sphere packing analysis is an effective tool to use in the design of
complex communication systems.
Consider a block code. Let the information block size be k, the signaling
rate be r, n = klr, the information frame be (UI, ... , Uk), the codeword be (x},
... , x n), XjE fll, and the energy of each codeword be nEs • In a spherical code,
all codewords have exactly the same power, and a codeword can be seen as a
point on the surface of a sphere in an n-dimensional space. The disturbance
of the noise moves this point to a nearby location according to a spherical
Gaussian distribution. When maximum likelihood decoding is assumed, the
probability of making a word error is equivalent to the probability of moving
a codeword point outside its region to make the perturbed point closer to
another codeword. Shannon [10] found that the average probability of
moving a codeword point outside its region, or the word error probability P w>
is bounded by
Pw ~ pw.spb = Q(Os) , (8)
where Pw,spb is the sphere packing bound3 of the word error probability, Q( 8.)
is the probability of a codeword being placed outside a cone of half-angle 8..
Variable 8. is a measurement of each codeword's region, whose value is such
3 The set of sequences at distance d or less from a codeword can be interpreted as a sphere of
radius d around the codeword. Sphere packed codes are the set of codewords such that the set
of spheres of radius d around the different codewords exhaust the space of binary n-tuples and
intersect each other only on the outer shells of radius d. Shannon's perfect code for the
continuous-input channel requires that the entire continuum of n-dimensional Euclidean space,
not just the discrete points represented by binary n-vectors, be filled up by the nonintersecting
cones. This is much stricter than the requirements of a normal sphere-packed code.
that an n-dimensional cone of half-angle Os encompasses a fraction 2-k of the

total solid angle of the n-dimensional space. The value of Os is the solution to
n(es ) = rk , (9)
where the function n( B) is the ratio of the solid angle in the cone of half-
angle 0 to the total solid angle. Together (8) and (9) define the sphere
packing bound for the continuous-input AWGN channel.
Conceptually the bound Pw.spb is exactly the word error probability that
would occur if it were possible to subdivide the Euclidean space into 2k
congruent cones, one for each codeword, and place the codewords on the
axes of these cones4 . Intuitively, any actual code would have a higher word
error probability than the code with such a conical partitioning. Thus Pw,spb is
the lower bound on the word error probability of the best code, PW,opt.
Shannon [10] derived the exact expressions of Q( B) and n( B) as
0(0)= n-l r(%+l) f(,m<l>)"'d<l>, (10)

n r( n;I)~ 1
r
and
Q(e) = r(sin~)"-2 Sn-l exp[- (S2 + nA 2- 2s,J;A COs~ )) dsd~ ,
n; 1)
n -1
2nI2~r( 2
(11)
where A is defined as the square root of the ratio of the signal power to the
noise power, i.e., A = ~2EslNo = ~2rEb/No . The computation in (10) and
(11) becomes prohibitive for large n. Fortunately, when n ~ 00, nCB) and
Q(B) can be asymptotically approximated as follows [10]:
r(~+ 1) (sin e)n-l
[1+0('!')] ""
n I l
n(e) < 2 (sine) - = (sine)"- (12)

- nr( n; 1)~ cose ·b1m cose n ~27rn cose, '
and
. 11
G smuexp -A + AGcose
2 )]"
[ (
Q(e)"" 2 , when eo <e<!'!!..., (13)

&~I+G2 sine(AGsin 2 e-cosB) 2
4Shannon pointed out that the perfect cone partitioning, and thus the perfect spherical code, is
only possible for n = 1 or 2, if k > 1 [10].
where G = ( AcosB +~ A2 cos 2 B + 4 )/2 and Bo = coC 1 A. For a given r, the

function Q( 0) decreases when e, A, or n increases, which agrees with the
intuition that the word error probability is smaller when codewords are
farther apart, have higher power, or have bigger block size. Half-angle es
decreases sharply as k increases when k < 10 and flattens in the k > 10
region. This makes the required Ei/No decrease slowly as k increases in the
k < 10 region for Pw 10-2 . As shown later, the required EblNo even
increases to a peak before decreasing for P w = 10- 1.
Let (EbINo)spb be the Ei/No required by the sphere packing bound to
achieve a certain P w performance. For a given set of P w, k, and r, (Ei/No)spb
can be found using the following two steps:
a) Find Os as a solution to (9). Use the exact n(O) in (10) for small n
(n < 100), and the asymptotic expression in (12) for large n
(n 100);
b) Substitute Os into the expression of Q( 0), and find the solution to
equation Q(es) = P W ' Use the exact Q(O) in (11) for small n (n < 100),
and the asymptotic expression in (13) for large n (n> 100).
In the above, the threshold n = 100 between the exact and asymptotic
expressions is an empirical value. Calculations show that the threshold is
related to the given P w.
In Figure 5, the curves of (Ei/No)spb vs. block size k are shown for
P w,spb = 10- 1 and 10-3 for various code rates. The curves derived from the
exact expressions are not connected to the curves derived from the
asymptotic expressions in order to show the transition region. In Figure 6,
the curves of sphere packing bounds are shown (labeled "Sphere") where
curve fitting is used to link the exact and the asymptotic parts, In Figure 6,
two other bounds are also shown for comparison: the bound obtained from
channel capacity Cw (labeled "Limit") and the random coding bound
(labeled "Random") [11]. The capacity limit curves are obtained by solving
C= r{l- Hb )= r{1 + 111og211 + (1-11)log2{1-11)) (14)
while assuming independent and equiprobable bit errors so that P b is related
to the block size k and the word error probability Pw by
P =1-{1-PS =tG)P {I-PS-

w b
i i
• (15)
While all three bounds converge as k ~ 00, both the random coding bound
and the channel capacity bound have little meaning for low k.
Figure 6 shows that the word error probability Pw, not the signaling rate r,
affects the shape of the curve. Regardless of rand P w, the sphere packing
bound approaches (EbINo)w as k ~ 00. For a given Pw, curves of all rates
approach their respective capacity limit similarly. The difference between the
asymptotic bound and the exact bound in the transition region is more
pronounced for higher Pw than for lower Pw. For the code rates examined, the
asymptotic curves merge to the exact curves around k = 1 ~ 100, with
smaller k for lower P w, and larger k for higher P w. Thus, for the block sizes
°
k> 100, the asymptotic expressions of the sphere packing bound can be used
instead of the computationally prohibitive exact expressions.
The discussion above of the sphere packing bound is limited to I-D
signaling. However, the parameters of ND-dimensional signaling can be
normalized so that the sphere packing bound can be calculated for a system
with ND-dimensional modulation as well. As stated, the derivation of the
sphere packing bound is based on each codeword of n symbols being seen as
a point in an n-dimensional Euclidean space. For any ND-dimensional
modulation, a frame of n symbols is a point in an ND x n-dimensional
Euclidean space. Thus finding the optimal performance of a code with k info
bits/frame, r info bits/symbol, and ND-dimensional modulation is equivalent
to finding the sphere packing bound with the normalized parameters: info
frame size k, codeword size n' = (k/r) x N D , and per-dimensional signaling
rate r' = r/ND = Rq/ND where r can be greater than 1. When applied to the
system in Figure 1 with a 2-D modulation (such as QPSK and 16-QAM), the
lower bound Pw,spb is found by using the set of parameters (k, n' = 2k/R/q,
r' = Rq/2) in the sphere packing equations in place of (k, n = k/R/q, Rq). This
lower bound applies to any channel coding with realistic modulation.
10 2 10' 10' 102 10'

Information Block Size (~sl Information Block Size (bilS)
(a)Pw= 10,1 (b)Pw = 10,3

Figure 5. The required EJJNo defined by Shannon's sphere packing bound for a range of
information block sizes k and signaling rates r E {1/3, 1/2,2/3,3/4, 7/8} to achieve word error
probabilities of Pw = 10,1 and 10,3 in a continuous-input AWON channel. Exact expressions
are used for small k (solid lines) while asymptotic expressions are used for large k (dashed
lines).
10 2 'O~ 102 10'

Information Block. Size (bits) Information Block Size (bits)
(a)Pw =IO- 1
102 10' 1O!

Information Block Size (bits)
Figure 6. The required Et/No derived from (i) channel capacity limit C w , (ii) Shannon's sphere
packing bound, and (iii) Shannon's random coding bound for a range of information block
sizes k and signaling rates r E {1I3, 112,2/3,3/4, 7/S} to achieve word error probabilities of
P w E {lO-I, 10-2, 10-3 , 1O-4 } in a continuous-input AWGN channel. For the sphere packing
bound and the random coding bound, exact expressions are used for small k while asymptotic
expressions are used for large k. The transition region is drawn using polynomial curve fitting.
4. APPLICATION EXAMPLE OF BOUNDS

The sphere packing bound assumes that the perfect sphere packing code
uses ideal modulation where a transmitted symbol can take any real value.
Thus, the sphere packing bound does not account for any realistic digital
modulation design, although the signaling rate r parameter includes the
modulation order q (bits/symbol)_
When comparing the sphere packing bound with simulation results, the
gap between the bound and results can be viewed as the summation of two
terms; one term for the imperfect design of the channel code and the other
term for the inherent imperfection of the modulation constellation. To
account for the modulation imperfection, an adjustment factor is proposed
for the sphere packing bound.
Bounding Techniques/or the Design o/Channel Coding and Modulation Systems 233
In the following, the discussion is limited to the commonly-used I-D and

2-D modulations that were studied in previous sections. As presented in
Figure 2, the composite channel capacity of any realistic modulation is
inferior to the capacity limit of a given dimension. For instance, I-D
modulations, such as BPSK, have a composite capacity CBPSK that is lower
than the I-D capacity limit C ID , and 2-D modulations, such as QPSK and
I6-QAM, have a composite capacity that is lower than the 2-D capacity limit
C2D • It can be shown that as the frame size approaches infinity, the sphere
packing bound approaches C ID , the sphere packing bound modified for 2-D
modulations approaches C2D , and the bounds of BPSK approaches CBPSK .
Thus if the gap (called modulation imperfection in the following) between
the composite channel capacity CMOD of a given realistic modulation (MOD)
and CID (or C2D ) is included in the sphere packing bound, then the adjusted
sphere packing bound approximates the performance when perfect channel
coding is used with the given modulation. This is true especially for large
frame sizes where the required EblNo is almost flat with respect to the frame
size. As an example, the sphere packing bound shifted by the imperfection of
BPSK shows little difference from Gallager's binary random coding bound
for large frame sizes (> 104), as illustrated in Figure 7. The small difference
indicates that the adjusted sphere packing bound is a good approximation
since Gallager's binary random coding bound is independently derived from
the probability density functions of the input signals and the channel noise.
For a given code rate R and a given modulation of order q, the adjustment
LlMOD(R) (in dB) is calculated using the following procedure:
1. Let c = Rq. Find (EJNo)cap(dB) that yields a capacity c. Use CID = c
and (3) if the modulation is I-D. Otherwise use C2D = c and (5).
2. Find the (EbINo)MoD(dB) that yields a composite channel capacity of
c for the given modulation.
3. LlMoD(R) = (EbINo)MoD(dB) (EbINo)cap(dB).
Since (EJNo)MOD (EbINo)cap = (EiNo)MoD (EiNo)cap in dB units, LlMOD(R)
could be used for either EblNo or EiNo. The curves of LlQPSK(R), Ll8PSK(R),
Ll I6QAM(R), and Ll64QAM(R) are found and plotted in Figure 8. For all
modulations, LlMOD(R) is an increasing function of R. This figure also
indicates the performance drawbacks of using 8-PSK for high code rates. In
general, LlMOD(R) is smaller if the modulation constellation is well-designed.
234 Yufei W. Blankensip and Brian K. Classon
4.5 Pw ·D.Ol.. • ...::.} .

Solid: binary nindPmcoding ~olJ1d •
4 ....... Das~ed'.d]jj;(ed:Siihete'Piifkifili·btiUiiil
: :. .:.
-O'~O~,-=~.E=Et::±[jl0~'~~::!::h:b:::kWl0'
Intannatian Block Size (bits)
Figure 7. The adjusted sphere packing bound ofBPSK and Gallager's upper bound of binary-
input AWGN channel.
Code rate
Figure 8. Modulation imperfection LlMOD(R) in dB for QPSK, 8-PSK, 16-QAM, and 64-QAM.
For a given frame size, code rate R, and 2Q-modulation, the adjusted
sphere packing bound is obtained by adding ~MOD(R) to the sphere packing
bound of the given frame size and signaling rate Rq. Figure 9 illustrates the
procedure of calculating the adjusted sphere packing bound using QPSK as
an example.
In Figure 10, the adjusted sphere packing bound is plotted against the
simulated performance of a turbo-coded orthogonal frequency division
multiplexing (OFDM) system for an AWGN channel. In this system, the
number of OFDM subcarriers is fixed, and each subcarrier employs the same
modulation scheme (e.g., QPSK, 16-QAM, 64-QAM). The channel coding
comprises a standard 3GPP2 turbo encoder [12], a puncture module to vary

the code rate R, and a block channel interleaver. The turbo decoder, which
uses log-MAP decoding, operates on log-likelihood ratios (LLRs) generated
by the optimal LLR generation technique in [13].
Figure 10 shows that the curve of the adjusted sphere packing bound
almost parallels the simulation results curve, i.e., the gap between them is
approximately constant for a given modulation. This gap varies with the
modulation, from 0.8 dB for QPSK to 1.2 dB for 64-QAM. This gap can be
interpreted as the imperfection of the channel coding design since the
imperfection for the modulation technique was accounted.
For the exceptions, R = 7/8 16-QAM and R = 113 64-QAM, the large gap
between the simulation results and the lower bound indicates potential flaws
in the design of these two channel coding/modulation schemes.
The tightness of the adjusted sphere packing bound for a wide range of
code rates and modulation techniques shows that channel coding schemes
utilizing turbo codes, when well-designed, can uniformly approach the lower
bound within 1 dB for all block sizes and modulations.
(a) Measure LlQPSK(R) (b) Shift the sphere packing bound up by ~psK(R)
Figure 9. A QPSK example is used to illustrate the procedure of obtaining the adjusted sphere
packing bound. LlQPSK(R) is obtained by measuring the gap between the composite capacity of
QPSK and the 2-D capacity limit. Then the adjusted sphere packing bound is obtained by
shifting the sphere packing bound by LlQPSK(R).
10 ...
750 subc~rl~rs :
F:;= 1O-"'i8PSK,-tlJ~o code
av~rage gap = 0.99 dB
Code Rate
(a) QPSK (b) 8-PSK
16
14
12
in
~10
~:
ill 8
Code Rate
(c) 16-QAM (b) 64-QAM

Figure 10. Comparison between simulation results and the adjusted sphere packing bound for
Pw of 10-2 . The simulation curves are generated by using 750 OFDM subcarriers and log-MAP
turbo decoding. Note that the R = 7/8 point of 16-QAM and the R = 113 point of 64-QAM are
excluded from the average gap calculation.
5. CONCLUSION
Performance bounds based on channel capacity and Shannon's sphere
packing bounds are discussed with application to. coding/modulation system
design. Unlike other bounding techniques, these bounds strive to take the
specific modulation into account. The performance bounds discussed here
can be used in many system design applications. They have been used to
theoretically calculate the maximum throughput of a HARQ system with a
given set of coding/modulation parameters for lxEV-DV [3]. They have
also been used to validate the coding/modulation design of 3G and 4G
systems [13]. Although this paper only discusses AWGN channels, similar
analysis can be applied to fading channels.
References
[1] C. Berrou, A. Glavieux, and P. Thitimajshima, "Near Shannon limit error-correcting

coding and decoding," in Proc. Int. Communications Con! (ICC'93), Geneva,
Switzerland, May 1993, pp. 1064-1070.
[2] S. Dolinar, D. Divsalar, and F. Pollara, "Code performance as a function of block size,"
The Telecommunications and Mission Operations Progress Report 42-133, Jan.-Mar.
1998, Jet Propulsion Laboratory, Pasadena, California, May 1998. Available online:
http://tmo.jpl.nasa.gov/tmo/progress_reportl42-133/133K.pdf
[3] Y. W. Blankenship and B. K. Classon, "Bounding techniques applied to modulation and
coding design," in 2002 Int. Con! on Third Generation Wireless and Beyond, May 28-
May 31, 2002. To appear.
[4] "lxEV-DV Evaluation Methodology - Addendum (V6)," 3GPP2 WG5 Evaluation Ad
Hoc, July 25, 2001.
[5] T. M. Cover and J. A. Thomas, Elements ofInformation Theory. New York: Wiley, 1991.
[6] S. G. Wilson, Digital Modulation and Coding. Upper Saddle River, New Jersey:
Prentice-Hall, 1996.
[7] G. D. Forney and L. Wei, "Multidimensional constellations-Part I: introduction, figures
of merit, and generalized cross constellations," IEEE J. Select. Areas Commun., vol. 7,
No.6, pp. 877-892, Aug. 1989.
[8] G. D. Forney, R. G. Gallager, G. R. Lang, F. M. Longstaff, and S. U. Qureshi, "Efficient
modulation for band-limited channels," IEEE J. Select. Areas Commun., vol. SAC-2, No.
5, pp. 632-647, Sept. 1984.
[9] J. G. Proakis, Digital Communications. New York: McGraw-Hili, 1995.
[10] C. E. Shannon, "Probability of error for optimal codes in a Gaussian channel," Bell
System Technical Journal, vol. 38, pp. 611-656, 1959.
[11] R. G. Gallager, Information Theory and Reliable Communication. New York: Wiley,
1968.
[12] Third Generation Partnership Project, Technical Specification TS 25.212,version 3.1.0,
December 1999.
[13] B. Classon, K. Blankenship, and V. Desai, "Channel coding for 4G systems with adaptive
modulation and coding," in 2001 Int. Con! on Third Generation Wireless and Beyond,
May 30-June I, 2001.
Quality of Service of Internet Applications over
the UMTS Radio Interface
Silke Heier, Andreas Kemper, Sebastian Grabner, Jan-Oliver Rock

Communication Networks
RWTH Aachen, Germany
{sheikemisgrijor}@comnets.rwth-aachen.de
http://www.comnets.rwth-aachen.de;- {sheikem}
Abstract The continuing evolution of exponential growth of the Internet market and grow-
ing demand of data services in fixed networks is also expected for mobile users.
Universal Mobile Telecommunications System (UMTS) as the 3rd generation
mobile system is designed to provide high bit data rates to the user. The mo-
bile Internet access will playa key role to ensure success of UMTS introduc-
tion. Typical Internet applications are running on a Transmission Control Proto-
col/Internet Protocol (TCPIIP) stack which should ensure a reliable end-to-end
communication in systems with limited quality of service guarantee. In contrast,
most protocols of radio access systems will grant traffic contracts. This raises
the question if the benefits ofTCP might be a drawback for the overall perfor-
mance of the radio interface with its own quality of service management mecha-
nisms. The aim of this paper is to examine the interaction between TCP and the
UMTS Radio Link Control (RLC) protocol. Of special concern is the interaction
of TCP retransmissions with the RLC Automatic Repeat Request (ARQ) mech-
anisms. Simulations have been performed to evaluate the performance of the
implemented protocol stack with and without TCP. Simulation results of quality
of service parameters like delay and packet throughput depict the performance
of mobile Internet access over UMTS.
Keywords: UMTS, UMTS Radio Access, Radio Interface Protocols, WWW, Quality ofSer-
vice, TCPIIP, RLC
1. Introduction
The aim of Universal Mobile Telecommunications System (UMTS) is the
provisioning of high bit rate data services to the mobile user. Hence, typi-
cal Internet applications like World Wide Web (WWW) browsing will migrate
from fixed access network systems to the mobile environment. Normally these
applications are running on a Transmission Control Protocol/Internet Protocol
(TCP/IP) protocol stack which guarantees a reliable end-to-end communica-
tion. Nevertheless, these protocols will not ensure quality of service in terms
239
X. Lagrance and B. fabbari (eds.),
240
of delay and throughput. In most cases the customer faces best-effort resource
management strategies in fixed networks. This might be applicable since over-
provisioning of bandwidth is feasible in fixed networks.
Due to the nature of radio links, the radio interface protocols have to cope
with limited bandwidth, higher delays and error rates. Therefore, the UMTS
radio interface protocols provide various and highly complex functions and
mechanisms to realize a reliable communication. Traffic contracts and quality
of service requirements are supported if desired.
This paper demonstrates the opportunities of the UMTS Radio Link Control
(RLC) protocol to support quality of service. RLC provides a reliable link with
ARQ mechanisms. Both, TCP and RLC protocol are using retransmissions to
recover lost data packets. Running an Internet application over UMTS rises
the question if it is necessary to use TCP. On the one hand side, protocol
overhead burdens the radio interface with its limited bandwidth. On the other
hand side, TCP retransmissions and time outs may interfere with the Automatic
Repeat Request (ARQ) of the RLC protocol. Another opportunity is running
a TCP application over an unacknowledged RLC connection. Since TCP is
not designed for radio access, it has to be shown how it will perform in an
environment with high error rates and delays.
The aim of this paper is to present performance results with the help of a
UMTS Radio Interface Simulator. The performance evaluation considers the
detailed UMTS protocol implementation as well as the traffic load generation
for WWW browsing sessions including TCP/IP modeling. Thus, it is feasible
to present the performance of a TCP application in an unreliable mobile envi-
ronment. The simulation results will show the performance in terms of quality
of service parameters a mobile user will experience while surfing the Internet.
Following this introduction, Sec. 2 gives a description of the traffic model.
Furthermore it presents the used simulation methodology and the simulation
environment. Sec. 3 and Sec. 4 illustrate the quality of service mechanisms of
the TCP and the RLC protocol. Performance evaluation, simulation scenarios
and discussion of simulation results is depicted in Sec. 5 and Sec. 6. The paper
concludes in Sec. 7 with a summary.
2. Simulation Environment
A WWW browsing session is a typical application which is running on the
TCP/IP protocol stack. A WWW traffic model is necessary for simulative
examinations of the performance of data services of mobile radio networks. A
typical WWW browsing session consists of a sequence of page requests. These
pages contain a number of objects with a dedicated object size each. During
a page request, several packets for each object may be generated which means
that the page request constitutes of a bursty sequence of packets. The bursty-
Ness during the page request is a characteristic feature of packet transmission.
After the page has entirely arrived at the terminal, the user is consuming certain
amount of time for studying the information. This time interval is called rea-
241
ding time. Tab. 1 gives an overview of the used WWW traffic model described
by parameters and their distributions.
Related documentation can be found in Frost and Melamed, 1994; Arlitt and
Williamson, 1995; Paxson, 1994. The main part of the later implementation is
based on the work of Arlitt and Williamson, 1995.
For this study the simulation tool UMTS Radio Inteiface Simulator (URIS)
was developed at the Chair of Communication Networks. This simulation en-
vironment is used to investigate, optimize and develop features of the radio
interface protocol stack. In addition, it offers the opportunity of capacity and
quality of service evaluation by simulations of various scenarios. The simu-
lator is a pure software solution in the programming language C++. The si-
mulation model is implemented with the help of a powerful C++ class library
which was developed by the Chair of Communication Networks and is called
the SDL Peiformance Evaluation Tool Class Library (SPEETCL) Steppler and
Lott, 1997. This generic, object oriented library is well suited for telecom-
munication network simulation purposes and can be used in event driven, bit
accurate simulation environments. The UMTS protocols at the radio interface
enhanced by a TCPIIP protocol stack were specified with the Specification and
Description Language (SDL). To generate an executable out of the SDL phrase
notation and the C++ library, a SDL2SPEETCL code generator is used.
The software architecture of the URIS simulator is shown in Fig. 1. Up
to now the simulator consists of various traffic generators and a traffic load
mixture unit which is used to adjust scenarios with desired load mixtures. The
physical channel module models the transmission of bursts in radio frames on
the radio interface. This includes discarding of erroneous bursts depending on
the error model.
The core of the simulator are the modules User Equipment (UE) and UMTS
Terrestrial Radio Access (UTRA) which are built formally similar. Each UE
and UTRA is implemented as an SDL system which contains the protocol im-
plementation of the layers. Fig. 2 gives an overview of the protocol structure
in a generic SDL system. The complex protocols like Medium Access Control
Table 1. Model Parameters ofWWWBrowsing Session
WWW Parameter Distribution Mean Variance

Session Arrival Rate [h -1 ] neg. expo 2.0
Pages per Session geometric 20.0
Reading Time between Pages [s] neg. expo 33.0
Objects per Page geometric 2.5
Inter Arrival Time between Objects [s] neg. expo 0.5
Page Request Size [byte] normal 1136.0 80.0
Object Size [byte] log2-Erlang-k ~ 11.3 5.4
242
www Speech www Speech
Physical Channel
Figure 1. Structure of the URIS Simulator
(MAC), RLC, Packet Data Convergence Protocol (PDCP) and Radio Resource
Control (RRC) based on UMTS Release 4 and the TCP/IP and User Data-
gram Protocol (UDP)/IP are specified in SDL with object oriented methods.
For simulation purposes, the TCP/IP protocol stack can be easily bypassed by
switching the transport and the network layer to transparent transmission.
Usual simulator approaches model protocols and functions on the basis of
abstractions and simplifications. The aim of URIS is a detailed, bit accurate
implementation of the standardized protocols. This offers the opportunity to
determine the performance ofUMTS in a realistic manner.
3. The TCP lIP Protocol Suite

The modem Internet is mainly based on TCP and IP. In order to provide
Internet services to the mobile user, the TCP/IP protocol suite has to be adapted
to the UMTS protocols. The UMTS radio interface protocol stack offers this
functionality by providing the PDCP.
TCP provides the overlying protocols with a reliable connection. The fol-
lowing services are offered:
• Retransmission of lost packets,
• Recovery from out-of-order delivery,
• Flow control.
243
Traffic Generators
SOL System
!
ENV
I---- LLME Transport Layer
DC TCP UDP CS
I I
TCP UDP CS
I- RRC Network Layer
LLME IP AM UMTI
t !
LLME
Ipocpl
!
-1 Radio Link Control
Ilogical channels
-1 Medium Access Control

I transport channels
I
--- Physical Layer
Iphysical channels
SOL Environment
Figure 2. URIS SDL System Structure
The TCP implementation realized in URIS is based on the so called "Reno"

TCP stack and uses the following flow control mechanisms:
• Slow start and congestion avoidance,
• Fast retransmit and fast recovery,
• Delayed acknowledgments,
• Selective acknowledgments Mathis et aI., 1996; Fall and Floyd, 1996.
IP is a connection-less and unreliable protocol that operates on a best-effort
basis. This means that IP packets may be lost, out-of-order or even duplicated
without IP handling these situations. This has to be done by higher layers. In
URIS, the IP protocol implementation currently performs data encapsulation.
4. Radio Link Control Protocol

The RLC realizes segmentation and retransmission services for both user
and control data. The RLC protocol provides three different data transfer ser-
vices:
244
• Transparent data transfer service mode (TR),
• Unacknowledged data transfer service Mode (UM),
• Acknowledged data transfer service Mode (AM).
The TR mode is an unidirectional service typically used for broadcast or

paging services, where it is not necessary to guarantee an error-free transmis-
sion. A use for transmission of streaming data (e.g. audio or video) is also
feasible, especially if real-time transmission is more important than error reli-
ability. A dropping mechanism prevents delivery of already expired Protocol
Data Units (PDUs).
The UM is an unidirectional service typically used for streaming applica-
tions (streaming class) where it is not necessary to guarantee an error-free
transmission. Voice over IP (conversational class) is also a feasible service
conveyed by the UM. The UM transmits higher layer data packets without
guaranteeing delivery to the peer entity. By using a sequence-number check
function in the receiving entity, the UM is capable of detecting missing RLC
PDUs, but error recovery is not performed. A dropping mechanism prevents
transmission of already expired PDUs.
The AM transmits higher layer data packets with guaranteed delivery to the
peer entity. Therefore it is mainly used for traffic of the interactive class or
background class type. ARQ mechanisms are applied for correction of trans-
mission errors. It is possible for the higher layers to request a transmission
confirmation from the AM. A Go-Back-N ARQ has been implemented as a
first approach for the AM. Segmentation and reassembling as well as concate-
nation are fully implemented as specified in 25.322,2001.
5. Simulation Scenarios
The main parameters concerning Quality of Service (QoS), which are af-
fected by RLC and TCP/IP, are delay and throughput. Two measurements are
made during the simulation in both the RLC layer and the transport layer:
Packet Delay: time from sending a traffic load packet to the respective
layer until correct reception of the packet by the traffic load receiver,
2 Packet Throughput: size of the traffic load packet divided by the packet
delay. The throughput is given in bytes per second [byte/s).
Three types of simulation scenarios have been executed:
1 RLC AM without TCP/IP,
2 TCP/IP with RLC AM,
3 TCP/IP with RLC UM.

245
-.~----==acket E~r Rate = 0%

acket Error Rate = 1%
acket Error Rate =3%
~~!;!ror Rate =5%
i~
0.8 O.B
i
c
0.6 0.6
!
" 0.4 i" 0.4
~
0.2
aekel Eno( Rale = 0%
acket EITOf' Rate = 1%
acket EnOl" Rate'" 3%
JJ
----- ....
I a:-"
A
0.2
acket~~
0 0
0 0.5 1 1.5 2 0 20000 40000 60000 80000 100000
Downlink Delay [s] Downlink Throughput [byteJs]
(a) Traflic Load Packet Delay with RLC AM and (b) Traflic Load Packet Throughput with RLC
without TCP/IP AM and without TCP/IP
1 --
acket Error Rate 5%=
i~
0.8 0.8
i
~
0.8 0.6
! 1"
""
v
0.4 0.4
a:-
J
0.2
0%
1%
~ 0.2
3%
5%
10%
0 0
0 0.5 1.5 0 20000 40000 60000 80000 100000
Downlink Delay [5] Downlink Throughput (byte/5]
(c) Traflic Load Packet Delay with RLC AM and (d) Traffic Load Packet Throughput with RLC
TCP/IP AM and TCP/IP
I . "\
1 \ r\\· .
1
o 8' . . \...
acket Error
~acket
,ekel
Rate =
Error Rale
Erro< Rate
acket Error
--0%
1%
=
Rale - 5%
10%
1 0. 6 \
1
~ . ~,
0.4
R
it" 0.2
10 20000 40000 60000 80000 100000

Downlink Delay [sl Downlink Throughput (byte/51
(e) Traflic Load Packet Delay with RLC UM and (I) Traffic Load Packet Throughput with RLC
TCP/IP UM and TCP/IP
Figure 3. Simulation Results for Traffic Load Packet Delay and Throughput
246
The cumulative distribution function of the measured packet delay and the
complementary cumulative distribution function of the measured packet through-
put have been calculated. Measurements have been made for increasing packet
error rates from 0% up to 10%. Tab. 2 shows the parameters of RLC and
TCP/IP used for these simulations.
6. Simulation Results
The results of the simulations are given in Fig. 3. As can be seen a mini-
mum delay is always present, caused by the constant transmission delay of
the physical layer. Packet error rates of 0% represent the ideally achievable
throughput. The throughput is not limited by the underlying dedicated channel
capacity since the offered load is always below 2 Mbitls.
Set 1: RLC Acknowledged Mode without TCP lIP

The first scenario evaluates the capabilities of the implemented Go-Back-N
ARQ of the RLC AM without interference of the TCP/IP protocols. The data
packets generated by the load generator are sent directly to the RLC protocol.
The results of this scenario are shown in Fig. 3(a) and Fig. 3(b). The distinct
steps of the delay curves in size of one Transmission Time Interval (TTl) length
are caused by the retransmission requests.
As expected, the packet delay and packet throughput degrade while the
packet error rates increase. At packet error rates up to 3%, 90% of the pack-
Table 2. Simulation Parameters
Parameter Value
TTl Length 0.04 s
Dedicated Channel Capacity 260 kbyte/s
RLC TxWinSize 1024 PDUs
Max. No. ofPDU Retransmissions 40
Status Prohibit Timer 0.08 s
Poll Timer 1.0 s
Maximum TCP Segment Size 512 byte
Maximum TCP Send Window 16 kbyte
Min. TCP Retransmission Timeout 3s
Max. TCP Retransmission Timeout 64 s
Delayed Acknowledgment Not used
Selective Acknowledgment Not used
Header Compression Not used
Packet Error Rate 0%,1%,3%,5%,10%
247
ets have a delay of less than I s. The throughput is approximately 2/3 of the
throughput at 0% error rate. For a 5% packet errorrate, the Go-Back-N mecha-
nism is no longer able to provide sufficient data throughput. Due to the un-
necessary Go-Back-N retransmissions of correctly received packets even more
transmission errors are caused and congestion occurs within the RLC protocol.
To avoid unnecessary retransmissions and congestion a Selective-Reject-
ARQ (SR-ARQ) mechanism should be used. The SR-ARQ is currently under
study. It is expected that the performance will be better but a Go-Back-N ARQ
as a first approach is sufficient to study the general behavior.
Set 2: TCP lIP with RLC Acknowledged Mode

The RLC AM has been used to investigate the interaction ofthe TCP proto-
col with the RLC ARQ. Fig. 3(c) shows the results of the measured TCP delay
for several packet error rates. The distinct steps of multiple TTIs are caused by
retransmissions on RLC protocol level. As expected, the delay imposed by the
TCP protocol and the underlying layers increases for higher packet error rates.
At packet error rates up to 5%, 90% of the packets have a delay of less than
1.5 s. Fig. 3(d) shows the respective throughput. As a result of higher delays,
lower data throughput is achieved.
Compared to the first scenario, the delays of traffic load packets are higher
and the peak throughput is lower. This is caused by two effects. First, TCP
adds protocol overhead which results in a higher gross data rate that has to
be transmitted over the radio interface. Second, the window mechanism of
TCP realizes a flow control of bursty WWW traffic where packets have to wait
since the send window is closed. Overhead and window waiting times produce
additional delays and lower throughput. Even if the radio link layer guarantees
a higher peak throughput (see Fig. 3(b» the TCP flow control mechanism is too
slow to use the guaranteed bandwidth efficiently. RLC radio bearer capacity is
left unused since the TCP send window is closed due to bursty traffic.
An interesting effect has been observed at high loss rates (10%). Due to
the heavy losses on the radio link, the Go-Back-N ARQ mechanism of the
RLC must perform many retransmissions. In the mean time the Retransmis-
sion Timeout (RTO) for the transmitted TCP segments expires which leads to
a retransmission by the TCP protocol. In this case the delays get even worse
since data packets will be send twice on different protocol layers which in-
creases the traffic unnecessarily. As a result, duplicate TCP packets will arrive
at the TCP receiver since the original packets are delivered by the RLC ARQ
first and second by TCP retransmission. Resulting duplicate acknowledgments
are the reason why TCP's fast retransmit algorithm is triggered. By mistake,
TCP assumes congestion and reduces the throughput which results in unused
but guaranteed RLC radio bearer capacity.
248
Set 3: TCP lIP with RLC Unacknowledged Mode

The RLC UM has been used to evaluate the performance ofTCP's retrans-
mission mechanisms. Fig. 3(e) depict the TCP delay which shows higher de-
lays with growing packet error rates. It can be noted that the steps of the curves
mark the expiration of a TCP RTO. In all simulations a minimum RTO of three
seconds has been used. It is obvious that the retransmission mechanism ofTCP
is far too slow to handle the heavy packet losses of a radio link. Fig. 3(f) gives
the respective throughput which clearly shows the heavily reduced throughput
for higher error rates. Even if the radio link layer guarantees high radio bearer
capacity the TCP retransmission mechanism is too slow to use the guaranteed
bandwidth efficiently. RLC radio bearer capacity is left unused since the TCP
send window is closed due to bursty traffic, missing acknowledgments and re-
transmission timeouts.
7. Conclusion
This paper comprises simulative examinations of the performance ofWWW
surfing over the UMTS radio interface. The simulation results show that a
WWW session using TCP and the RLC UM is not feasible to satisfy a mobile
user. If packets are lost the TCP protocol infers that there must be congestion
in the network between the two peers. In consequence, a TCP retransmission
leads to a decrease of the send window size which results in less throughput.
This assumption is correct in wired environments but TCP is unable to handle
the unreliable radio link of mobile users due to the high error rates and TCP's
slow error handling. Very high delays and small throughput lead to unsatisfied
mobile users. Improvements may be archived by using the TCP SACK option
which allows the receiver to convey more information about its state to the
sender, thus increasing the efficiency of the retransmission strategy.
Running a WWW session over a standard TCP and the RLC AM will be
an ordinary scenario during the introduction ofUMTS. First UEs will rely on
standard application solutions which will be adopted from the fixed world.
Hence, TCP will run end-to-end including the radio interface. This paper
shows that this solution is applicable to satisfy the mobile user but the perfor-
mance suffers since TCP mechanisms will not efficiently use the guaranteed
QoS of the UMTS radio bearers in terms of delay and throughput. It has to
be mentioned that effects of the core network are not included in our simu-
lations. This is not only due to currently missing simulation capabilities but
also partly unknown structure of the core network. While it is hard to model
the behaviour of packets traveling along the internet, it can't even be assumed
that the entire core network is IP-based. Thus it can be concluded that the
overall performance will get worse since TCP as an end-to-end protocol will
experience additional delays and congestion in the core network. Using heav-
ily disturbed radio links the mobile user will face a reduced throughput and
longer waiting times for WWW page downloads than in fixed networks. This
249
relies on the weak TCP performance in a mobile environment and can not be
influenced by assignment of radio bearers with higher capacity.
The simulation results of a WWW session running without TCP/IP in RLC
AM show the best performance. As a result of all simulations, future UEs are
suggested where typical packet data applications should run without a standard
TCP at the radio interface. The plain UMTS radio interface protocol stack
offers reliable radio bearers with sufficient QoS for WWW browsing sessions.
To face the requirements of the Internet world, TCP might terminate or start
running end-to-end in the UMTS core network.
Biographies
Silke Heier and Andreas Kemper are Ph.D students at the Chair of Commu-
nication Networks. Together with her students, including Sebastian Grabner
and Jan-Oliver Rock, Mrs. Heier provided significant improvements to the
UMTS protocol simulator URIS. While she is almost finished with her degree,
mainly her students implemented layers two and three under her supervision.
Andreas Kemper is her successor with respect to further developments on the
simulator. Due to the current status of implementation, his focus is mainly
on the development of physical layer and radio resource control (RRC) rou-
tines. The horizon in development appears to be the coupling of this protocol
simulator with a system-level simulator to provide more detailed information
on channel quality, resulting for instance from propagation and interference
situation.
Acknowledgment
The authors would like to thank Prof. B. Walke of the Chair of Communi-
cation Networks for his support and friendly advice to this work.
250
References
25.322, G. T. (2001). Radio Link Control (RLC) Protocol Specification. Technical Specification
V 4.1.0, Release 4, 3rd Generation Partnerschip Project, Technical Specification Group Radio
Access Network.
Arlitt, M. F. and Williamson, C. L. (1995). A Synthetic Workload Model for Internet Mo-
saic Traffic. Proceedings of the 1995 Summer Computer Simulation Conference, Ottawa,
Canada, July, pp. 24-26.
Fall, K. and Floyd, S. (1996). Simulation-based Comparisons of Tahoe, Reno, and SACK TCP.
Lawrence Berkeley National Laboratory.
Frost, V. S. and Melamed, B. (1994). Traffic Modeling For Telecommunications Networks.
IEEE Communications Magazine. pp. 70-81.
Grabner, S. (2001). Simulative Performance Evaluation of the UMTS Radio Link Control Pro-
tocol. Master's thesis, Chair of Communication Networks, RWTH Aachen University of
Technology.
Holma, H. and Toskala, A., editors (200 I). WCDMA for UMTS: Radio Access for Third Gener-
ation Mobile Communications. John Wiley & Sons Ltd.
Mathis, M., Mahdavi, 1., Floyd, S., and Romanow, A. (1996). Selective Acknowledgement Op-
tions. RFC, Internet Engeneering Task Force.
Paxson, V. (1994). Empirically-Derived Analytic Models of Wide-Area TCP Connections. IEEEI
ACM Transactions on Networking, 2 (4), pp. 316-336, August 1994;
ftp:l/ftp.ee.lbl.govlpapersIWAN-TCP-models.ps.Z.
Rock, J.-O. (2001). Simulative Performance Evaluation ofIP based Services over UMTS. Mas-
ter's thesis, Chair of Communication Networks, RWTH Aachen University of Technology.
Steppler, M. and Lott, M. (1997). SPEET - SOL Performance Evaluation Tool. 8th SDL Forum
'97.
Walke, B. (2001). Mobile Radio Networks, 2nd Edition. John Wiley & Sons Ltd., Chichester,
England.
Interactions between the TCP and RLC
Protocols in UMTS
Robert Bestak, Philippe Godlewski and Philippe Martins
Ecole Nationale Superieure des Telecommunications, Department Infres

46, rue Barrault, 75634 Paris Cedex 13, FRANCE
E-mail: {bestak.godlewski.martins}@enst.fr
Abstract - Congestion control mechanisms, used in TCP, suffer from some

performance issues when using them over wireless links. In this paper we
investigate a TCP connection over the UMTS radio interface when using the
RLC protocol in acknowledged mode. We compare protocol mechanisms used
in TCP and RLC. The influence of the TCP congestion control mechanisms on
the RLC buffer occupancy is analyzed and interactions between these two
protocols are investigated.
I. Introduction
Due to the rapid advances in the area of wireless communication and
Internet, provision of data services for applications such as e-mail, web
browsing, telnet, etc., over a wireless network is gaining importance. Because of
wireless channel constraints (mainly time varying conditions and bandwidth
limitation), the efficiency of the transmission protocols defined and optimized
for wired networks suffer when applied directly to wireless networks.
One important objective of UMTS (Universal Mobile Telecommunication
System) is to offer an Internet connection to a mobile subscriber by using an
efficient and reliable packet transmission protocol stack. The 3G- system will
provide for users higher data bit rates (around 384 kbps for mobile users and up
to 2 Mbps for indoor mobile users) than their 2G counterparts.
The reliable data transfer on the radio interface of UMTS, i.e. between an
user equipment (UE) and a Radio Network Control entity (RNC), is ensured by
the Radio Link Control (RLC) protocol. The RLC layer and the MAC layer
(Medium Access Control) manage QoS parameters required by the services.
In Internet, the dominant protocol used for end-to-end reliable data transfer
is the Transmission Control Protocol (TCP). The TCP protocol has been tuned
to perform well in wired networks where bit error rates are very low and
congestion is supposed to be the primary cause of packet losses. However, there
are some design issues in TCP, which make it difficult to use efficiently over the
251
X. Lagrance and B. fabbari (eds.).
Multiaccess, Mobility and Teletrcif.ficfor Wireless Communications, Volume 6,251-262.
252
wireless links. In recent years, there have been large research activities in order
to improve the TCP performance in these networks (e.g., [1], [6]).
The suggested approaches can be classified in two categories ([4]). The
approaches in the first category try to hide non-congestion related losses from
the TCP sender (e.g., [3], [5]). They do not imply any changes in the existing
sender implementation. The intuition behind these approaches is that since the
problem is local, it should be solved locally. TCP does not have to be aware of
the characteristics of the individual links. In the corresponding solutions, most
of the losses, detected by the TCP sender, are caused by congestions. The
approaches in the second category attempt to make the sender aware of the
existence of wireless links and realize that some packet losses are not due to
congestion in the wired network (e.g., [6], [7]). The TCP sender can then avoid
invoking the congestion control algorithms when detecting non-congestion
related loss. A comprehensive comparison between different solutions for
improving TCP performance over wireless channels can be found for instance in
[4].
This paper intends to focus on the use of TCP over the UMTS radio
interface. The functionalities of TCP and RLC protocols are compared and
possible interactions between these two protocols are studied. A metric
indication concerning the RLC buffer occupancy is analyzed and we discuss the
impact of retransmissions at the RLC level on the TCP performance.
The paper is organized as follows. The next section contains a rapid
description of the radio protocol stack of UMTS. Section III develops a
communication scenario using TCP in the higher layer. The comparison
between the RLC and TCP protocols is the subject of section IV. Results of
simulation are discussed in section V. In the last section, our conclusions are
presented.
II. The UMTS radio interface

The radio interface of UMTS (interface Uu, [8]) is decomposed into three
layers. The Physical layer (PRY) of UMTS offers transport services to Medium
°
Access Control layer (MAC) via transport channels. The information of
transport channels are transmitted on radio frames of 1 ms. A small multiple of
the radio frame duration can be used to transmit data from radio transport
channel (i x 10 ms, where i E {O, 1,2, 3}).
The second layer (link layer) is split into two sub layers, RLC and MAC.
The MAC layer offers services to the RLC layer via logical channels. MAC
layer handles different types of logical channels with different QoS parameters
and supports fast adaptation mechanisms provided by the Radio Resource
Control layer (RRC). RLC entities handle Protocol Data Units (RLC PDUs or
blocks). RLC blocks can transport user data (e.g. a TCP segments) or it can
transport signaling messages (for instance RRC messages).
253
The RLC protocol can operate in three modes:

(a) Transparent mode (Tr). It is the simplest mode. No protocol overhead is
added to a Service Data Unit (SDU). This mode is used principally for a
service that requires low delay - such as voice transmission.
(b) Unacknowledged mode (UM). For this mode the RLC blocks are
transmitted without guaranteeing delivery to the peer entity. The blocks are
numbered to allow the segmentation/reassembly and the concatenation of
SDUs. Unacknowledged mode is used for cell broadcast services or for
certain RRC signaling procedures.
(c) Acknowledged mode (AM). This mode is the most complex and guarantees
delivery to the peer entity: an ARQ (Automatic Repeat Request)
mechanism is used for the error correction. It offers more functions than
UM. It provides in sequence/out of sequence delivery of SDUs to the
higher layers. RLC entity delivers only SDUs, that are error free and
ensures that a SDU is delivered only once. This mode can be used for
example for Internet browsing or file transfer.
The acknowledged mode is bi-directional, while the transparent and
unacknowledged modes are unidirectional.
III. TCP over the radio interface

We consider in this paragraph a typical TCP connection from an UE to a
server via the radio interface of UMTS (figure 1). The TCP congestion
mechanisms have been originally developed for wired network with low Bit
Error Rate (BER) less than 10-7• The TCP sender detects a packet loss either by
the arrival of three duplicated acknowledgments or of the absence of an
acknowledgment for the packet within timeout interval. TCP reacts to a packet
loss by dropping its transmission (or congestion) window size. The congestion
control mechanisms are then initialized and the retransmission timers are backed
off. These steps result in a reduction of the transmitted data into the network.
'I
FIl' FIl'
IITIP IfITP
Figure 1. A TCP connection over RLC.

254
In wireless networks the packets losses are produced for other reason than
congestion. The losses are mainly due to high bit error rates (as high as 10·3) and
handoff procedures. These lost packets are interpreted by TCP as network
congestion event and therefore TCP invokes for each event the congestion
mechanisms. This misinterpretation results in an unnecessary reduction in end-
to-end throughput and the TCP performance degradation.
In cellular systems, there are defmed protocols that implement mechanisms
that allow rapid detection of erroneous blocks transmitted over the radio
interface. These protocols use mechanisms such as ARQ, FEC (Forward Error
Correction) or hybrid ARQIFEC.
IV. Comparison TCP/RLC

The architecture of UMTS radio interface makes it possible to adapt the IP
world to the optimized radio world. Two types of data units can be found in
these two worlds: TCP segment (packet) and RLC PDU (or block). The
numbering differs in the TCP and RLC protocols. TCP transmits segments and
numbers bytes (modulo 232). Unlike TCP, RLC handles blocks and number
blocks (modulo 27 for UM, i2 for AM). The size of a block, for each RLC
entity, is fixed at the beginning of a communication (it can be changed through
bearer reconfiguration by the RRC layer). The maximal TCP segment size can
be negotiated while establishing a TCP connection (TCP option). The Maximum
TCP Segment Size (MSS) is per default 536 bytes.
The RLC (AM) acknowledgement policy is more flexible than that ofTCP.
There are two types of policies that RLC can use: (i) the transmitter explicitly
demands the receiver to send an Ack (by setting the Polling bit P in the header
of a RLC block) or (ii) the receiver sends an Ack without being asked. In case
(i), the transmitter may set the P bit by several ways: the transmitter sets the P
bit every N RLC block or every M SDU or each time a part of the transmission
window (x%) is fIlled etc. In case (ii), the receiver can send an Ack periodically
or when detecting missing RLC blocks. The policy and the mechanism used
depend on the configuration of the RLC entity.
The TCP protocol (version Reno) uses fast retransmit, fast recovery or
retransmission timer for detecting packet losses. TCP supposes that a loss is the
result of congestion in the wired network. Standard TCP uses a cumulative
acknowledgment scheme. It often does not provide the sender with sufficient
information to recover quickly from multiple loses within a single transmission
window. To improve the TCP performance in this case, Selective
Acknowledgment (SACK) option has been defmed.
Acknowledgments are sent in a very similar way in both protocols. TCP
uses either Piggybacking or specific Ack (Ack may be delayed a certain time -
200 ms). In RLC, an acknowledgment is sent in a stand-alone control block,
255
called STATUS PDD. The piggybacking mechanism can be used. In this case, a
special control block named Piggybacked STATUS PDU is defmed.
Both TCP and RLC can change dynamically the transmission window
during a communication. For TCP, the current transmitter's window size txwnd
is the minimum between the congestion window size cwnd and the advertised
reception window size rxwnd. In case of RLC, the receiver can change the
transmission window by sending STATUS PDU with a new size of the window
(by using the "Super Field Window" described in [13]).
A RLC entity can transfer data by using either one logical channel or two
logical channels. If it uses one logical channel, the size of a RLC control block
(STATUS PDU) must have the same size as the RLC block for data (PDU).
The unused space in the control block is filled by padding. In order to optimize
the size of the blocks for data and control, RLC makes it possible to use two
logical channels. One channel is used for transport of data and the second
channel for transport of control.
Due to data services that TCP support (Web browsing, file transfer etc.),
the downstream and upstream traffics are asymmetric. The size of a TCP
segment containing data is much higher than the size of a TCP segment with
Ack. This feature has an impact on the radio interface. In order to use the radio
resource efficiently, it is desirable to allocate channels with different capacity
for the downlink and the uplink connection. This data rate ratio has do be
determined carefully. Otherwise, the adverse effects on the TCP performance
can take place when this ratio is not appropriate (e.g., [2], [10], [12]). Table 1.
sunnnarizes some of the features ofTCP and RLC.
RLC TCP
Concerned link MS-RNC End-to-end
Go-Back-N
fast retransmission,
ARQ mechanisms Selective Repeat
fast recovery
SACK (optional)
SNmodulo PDUs (blocks) Bytes
Numbering AM: i _1
2
232 - I
(Sequence Number) UM: 27-1
Tr: unidirectional
Mode UM: unidirectional Bi-directional
AM: bi-directional
Piggybacking Piggybacking
How to send an Ack
STATUSPDU Frame with an Ack
Tr: '" 0 to 80
Aver. size of unit
UM:",20 '" 500 to 1500
[Bytes]
AM:",16t080
Table 1. Some of characteristics of RLC and TCP.

256
V. Simulation
The simulation is carried out on a RLC/TCP test-bed developed at ENST.
The architecture of the test-bed is described in figure 2. The test-bed comprises
2 PC that are connected via Ethernet. The first PC is used as UE and the second
PC simulates RNC entity, wired network and Server. The data transmission
between these PCs is based on the socket interface via Ethernet using a
tunneling protocol. The time of simulation is controlled by Supervisor in the
UE.
In our analysis we consider an uplink data transmission (file transfer). The
main parameters used during the simulations are given in table 2. The average
round trip time in the wired network is equal to 200 ms and the standard
deviation is equal to 20 ms ([9]). The round trip time in the wireless network,
i.e. between UE and RNC entity, is set to 20 ms when no retransmission occurs.
--------------------------------------., ,,------------------------------------------,,
: : Superviso '
VE
: : : j
,______________________________________ 2 ,.------------------------ ----------------,
medium
Figure 2. Architecture of the TCPIRLC test-bed.
We use Reno TCP, which incorporates slow start, congestion avoidance,

fast retransmit and fast recovery algorithms. The receiver acknowledges without
delay every TCP segment received. The slow start threshold ssthresh is set to
32*MSS. The maximum value of the receiver window size is set to 40*MSS.
size of the file rkbytes1 500

TCP MSS [bytes] 600
RLC block [bytes] 80
radio frame duration rms1 10
average wired round trip time [ms] 200
Table 2. Parameters during simulations
257
The RLC entity works in acknowledged mode. The transmitting RLC
entity segments delivered SDUs from higher layers. If possible, the entity
employs concatenation (i.e. a block may contain fragments of 2 SDUs). The
receiving entity reassembles received blocks into SDUs and delivers them to the
higher layer in sequence. Every radio frame transmits at most six RLC blocks.
The duration of a radio frame is 10 ms. The radio channel is dedicated to the
TCP connection until the whole file is transferred. Handoffs are not considered.
The error process on the radio channel is supposed to be memory less.
In our simulation scenarios, we introduce a congestion event in the wired
network (denoted Er in the following). This is realized by introducing a single
TCP segment loss with the same sequence number in all our simulation
scenarios (the transmission always starts with the same sequence number).
The RLC buffer occupancy

In the RLC layer, the input transmission buffer stores the SDUs (TCP
segments) delivered from TCP. The size of this transmission buffer may have an
impact on the TCP performance. If TCP segments fill up the RLC buffer, they
begin to be dropped. These lost segments are interpreted by TCP as a congestion
event and congestions mechanisms are invoked. However, in our simulations,
the RLC buffer size is assumed to be infinite, i.e. there is no TCP packet loss at
RLC level.
The impact of the TCP mechanisms slow start (SISt) and congestion
avoidance (CgAv) on the RLC buffer occupancy is illustrated in figure 3. The
graph (3a) is for the case where the radio channel is without errors (no RLC
block is retransmitted); graph (3b) represents the radio channel with errors that
are processed by the RLC protocol. Block error rate (BIER) equals 10%.
The buffer occupancy starts to growth exponentially (phase SlSt in figure
3a). This is due to slow start phase in the TCP level. As soon as, TCP reaches
ssthresh it changes to the congestion avoidance phase and the buffer occupancy
starts to growth linearly (phase CgAvl in figure 3a).
When txwnd reaches a certain value, txwndRLc, the RLC buffer occupancy
is proportional to txwnd. Each time the txwnd is increased by one MSS, the RLC
buffer occupancy is increased by one SDU. The approximate value of txwndRLc
can be found according to the following expression:
txwndRLC "" C (RTTwireless + RTTwired)
where C denotes the data rate of the radio channel and RTTwireless (respectively
RTTwired) is the mean value of round trip time in the wireless network
(respectively in the wired network).
At the end of the phase CgAvl, the TCP parameter txwnd reaches the
maximum value (in our example, it is set to 40 MSS); then it stays constant.
Likewise, the RLC buffer occupancy stops increasing linearly and oscillates
258
around a mean value (phase CgAv2 in figure 3a). In the graph, its value is
around 23 SDUs.
A congestion event, introduced by a segment loss in the wired network (Er
in figure 3a), can be observed at time 7s. TCP detects the congestion event by
receiving three duplicated Acks. When receiving the third duplicated Ack, TCP
retransmits the segment loss and reduce its congestion window by one half.
Subsequently, the TCP sender enters into the fast recovery phase. During this
phase, the TCP sender "inflates" its window by the number of DupAcks
(duplicated Acks) it receives. Each received DupAck indicates that a segment
has been removed from the network and is now cached at the receiver side. The
sender waits until a certain number of DupAck are received without delivering
to the RLC entity any segments. During this period, the buffer occupancy
dramatically decreases because (i) there are no new delivered SDUs from TCP
and (ii) the RLC entity continues to serve queued SDUs in the buffer.
Once the TCP sender has received a fresh Ack, it exits the fast recovery
phase and enters into the congestion avoidance phase. The buffer occupancy
starts again to growth linearly (phase CgAvl' in figure 3a).
Figure (3b) shows the buffer occupancy for a radio channel with
retransmission of RLC blocks. Due to the retransmissions, the buffer occupancy
can reach during certain periods a much higher value than in figure (3a). The
RLC buffer size should be chosen with respect to rxwnd (TCP receiver window
size) and to the BIER on the radio interface.
Transmission delay of a SDU over the radio interface

The radio transmission delay for each TCP segment over the radio interface
is shown in figure 4 (as in figure 3, we consider a single TCP segment loss and
the same parameters). Figure (4a) shows the error free radio channel case and
figure (4b) the noisy case (radio channel with errors). The radio transmission
delay considered in this paper is calculated from the moment where a SDU is
queued in the RLC buffer till the moment where it is erased from the buffer (i.e.
it has been correctly transferred and acknowledged).
The transmission delay also reflects the different TCP phases that have
been discussed in the previous section. Retransmission of RLC blocks over the
radio interface increases the sojourn time of each queued TCP segment in the
buffer. Error recovery employed at the RLC level may interfere with the timeout
mechanism at the TCP level. Due to a timeout event, TCP could retransmit a
segment that the RLC can be still trying to transmit or has already succeeded to
transmit over the radio interface. Thus, the TCP sender unnecessarily
retransmits a segment and then enters the slow start phase.
This unnecessarily retransmission may lead to the so called "retransmission
ambiguity" that is described in [11]. The redundant retransmissions at the TCP
level cause degradation of the TCP performance and lead to useless waste of
radio resources.
259
The RLC blJffer occupency.
'SISt CgAvl CgAv2 CgAvl' End

30 I I I I I I
Er
10 15
Timels1
Figure 3a. Number of SDUs in the RLC buffer as function of time: error free
radio channel. There is one TCP segment loss introduced in the wired
network (at time around 7s).
The RLC buffer occupency.
1Q 15
Timels)
Figure 3b. Number of SDUs in the RLC buffer as function of time: radio channel
with errors (BIER=10%). As in figure (3a), a TCP segment loss is
introduced.
260
Figure 4a. Transmission delay over the radio interface for each SDU: error free
radio channel. As in figure (3a), a TCP segment loss is introduced.
Figure 4b. Transmission delay over the radio interface for each SDU: radio
channel with errors (BIER=10%). As in figure (3a), a TCP segment
loss is introduced.
261
VI. Conclusion
This paper has investigated some issues that emerge when using TCP over
the RLC protocol in UMTS. We have compared and analyzed the protocol
mechanisms used by TCP and RLC in acknowledged mode. Some mechanisms
can be found in both protocols: segmentation/reassembly, flow control and
ARQ. We have noticed that the RLC mechanisms are more adaptable to the
radio constraints than those used in TCP.
In our simulation, we have considered a single uplink TCP connection. We
have shown that the RLC buffer occupancy reflects the different TCP phases:
slow start, congestion avoidance, fast retransmit and fast recovery. The
dimensioning of the RLC buffer should be done with regard to the TCP receiver
window size and to the radio interface BIER.
This study has shown reciprocal interactions between the protocols. We
presently investigate some possibilities of coupling TCP and RLC mechanisms.
References
[1] E. Ayanoglu, S. Paul, T. F. LaPorta, K. K. Sabnami, and R. D Gitlin,
..AIRMAIL: A link-layer protocol for wireless networks," ACM
ACMIBaltazer Wireless Networks Journal, February 1995.
[2] Y. Bai, A. T. Ogielski, "TCP over asymmetric CDMA radio links," Proc.
IEEE Con! on Vehicular Technology (VTC 2000 Fall), September 2000.
[3] A. Bakre, B.R. Badrinath, "I-TCP:Indirect TCP for mobile hosts," Int.
Con! Distributed Computing Syst. (ICDCS), May 1995.
[4] H. Balakrishnan, V. Padmanabhan, S. Seshan, and R. H. Katz, "A
comparison of mechanisms for improving TCP performance over
wireless links," ACM SIGCOM'96, August 1996.
[5] H. Balakrishnan, S. Seshan, and R. H. Katz, "Improving reliable
transport and handoff performance in cellular wireless networks," ACM
Wireless Networks, December 1995.
[6] S. Biaz, N. H. Vaidya, "Distinguishing congestion losses from wireless
transmission losses: a negative result," Computer Communications and
Networks, 1998.
[7] W. Ding and A. Jamalipour, "A new explicit loss notification with
acknowledgment for wireless TCP," Personal, Indoor and Mobile Radio
Communications, September 2001.
[8] H. Holma, A. Toskala, "WCDMA for UMTS, Radio Access for Third
Generation Mobile Communications," John Wiley & Sons, Ltd, England
2000.
262
[9] T. J. Kostats, M. S. Borella, I. Sidhu, G. M. Schuster, J. Grabiec, J.
Mahler, "Real-time voice over packet-switched networks," IEEE
Network, Jan.-Feb. 1998.
[10] T. V. Lakshrnan, U. Madlow, B. Suter, "Window based error recovery
and flow control with a slow acknowledgment channel: a study of
TCPIIP performance," Proc. INFOCOM, April 1997.
[11] R. Ludwig, R. H. Katz, "The Eifel algorithm: Making TCP robust against
spurious retransmission," ACM Computer Communication Review, Vol.
30, No.1, January 2000.
[12] S. Varna, "Performance and Buffering Requirements ofTCP Application
in Asymmetric Networks," Proc. INFOCOM, March 1999.
[13] 3G TS 25.322 (2001-06), RLC protocol specification (Release 4).
IMPACT OF SR-ARQ WITH FINITE
BUFFER ON TDDjTDMA WIRELESS LAN
Emilio Strinati,
Jeremy Gosteau,
Sebastien Simoens
and Pietro Pellati
Motorola Labs Paris, Saint Aubin 91193 Gi/-sur- Yvette France
tel. :+33-{0)1-69-35-25-64
fax. :+33-(0) 1-69-35-25-01
jeremy.gosteau@crm.mot.com
Abstract In this paper, the influence of some implementation parameters of Selective Re-
peat Automatic Repeat Request (SR-ARQ) on system performance is investi-
gated. In the framework of the specific SR-ARQ algorithm specified by the
ETSI BRAN HIPERLAN12 (H/2) Wireless LAN standard, the need for optimiz-
ing the ARQ signalling bandwidth is illustrated and several signalling strategies
are presented. Even with optimum management of the signalling bandwidth, the
finite transmit and receive buffers can seriously limit the throughput. This effect
is modeled by using a simple probabilistic approach, relying on the TDDITDMA
access scheme, and is evaluated by simulation. The interaction of SR-ARQ with
scheduling and Link Adaptation is also discussed and finally, an ARQ aware
scheduling strategy is proposed.
Keywords: Wireless LAN, HIPERLAN/2, ARQ, SR-ARQ, scheduling, Link Adaptation
1. Introduction
Any communication system implements mechanisms for limiting the trans-
mission of erroneous messages. These techniques can be sorted in two main
categories depending on their using error-correcting or error-detecting codes.
In the latter, if a receiver detects an erroneous packet, it sends a message back
to the transmitter to signal the error.
The scheme used by such a system is called Automatic Repeat Request
(ARQ). Many instanciations have already been thoroughly studied (Lin and
Costello, 1983), (Gibson, 1997). They range from the stop-and-wait to the se-
263
X. Lagrance and B. fabbari (eds.),
264
lective repeat (SR) algorithms. This paper just focuses on the latter. Indeed,
the SR scheme provides efficient retransmission with a limited overhead of ac-
knowledgment and a reasonable delay: only the erroneous packets which are
negatively acknowledged or for which the time out has expired are repeated.
This implies that buffers must be provided at both the transmitter and receiver
side in order to store the not yet positively acknowledged packets.
In this paper, three issues related to the SR-ARQ are addressed: the sig-
nalling strategy, the limitations due to the finite transmitting or receiving buffer
and the impact of the scheme on the scheduling and on the Link Adaptation
(LA).
In (Li et al., 2000), the authors study the signalling mechanism of an SR-
ARQ scheme in the framework of Hl2 Wireless LAN. They show that, in the
case of a downlink (DL) connection, an incremental allocation of the ARQ ac-
knowledgment messages optimizes the throughput on the Data Link Control
(DLC) layer. In the case of an uplink (UL) connection, we show in section
2 that the algorithms can be further improved if the exact required number of
feedback messages is dynamically granted. We then compare various algo-
rithms for DL or UL connections, and select the most efficient one. With this
first step, a strategy is proposed to optimize the throughput with respect to the
signalling constraints in the case of an Hl2 based network.
Nevertheless, a limitation in the throughput can still be observed even for
large buffer sizes when the Packet Error Rate (PER) grows. In section 3, we
model the phenomenon with simple discrete probabilities calculus, relying on
a TDDffDMA access scheme, which differs by the approach and the assump-
tions from what can be found in (Miller and Lin, 1981), (Saeki and Rubin,
1982) or (Jianhua et al., 1999). The analytical approach is compared with re-
sults simulated with a H12 network simulator.
To complete the study of the SR-ARQ scheme with a finite buffer size, sec-
tion 4 evaluates its impact on the scheduling and the LA. Actually, the schedul-
ing of resources has already been studied in papers like (Kadelka et al., 1999)
or (Ranasinghe et al., 2001) for TDDffDMA based systems. These studies are
yet limited to alternatives of round robin algorithms without involving ARQ
parameters. We propose here to explain to what extent the choice of some
ARQ parameters can greatly influence the choice of a scheduling algorithm and
thereof the resulting overall system performance. We will show that a metic-
ulous setting of the parameters is key to avoid a drop in throughput. Simple
guidelines can be drawn out of this study. Based on these results, a scheduling
strategy is proposed. In the same way, the Link Adaptation (LA) (Goldsmith
and Chua, 1998), (Simoens and Bartolome, 2001) needs some slight tuning in
order to take the ARQ into account. This issue is also discussed in section 4.
Impact of SR-ARQ with finite buffer on TDD/TDMA Wireless LAN 265
2. Signalling strategies of SR-ARQ in H/2

Signalling in H/2. In H/2, the Medium Access Control (MAC) frame
structure relies on a TDD{[DMA subdivision (see in (Kadelka et aI., 1999),
(ETSIIBRANIDLC, 2000)). The frame lasts 2 ms, splits into Broadcast, DL,
Direct Link and UL and is made of Packet Data Unit (PDU) trains.
Table 1. Notations used in the description of a MAC frame
Acronym Description Comments
BCH Broadcast CHannel carried in broadcast PDU train

FCH Frame CHannel carried in broadcast PDU train
ACH Access feedback CHannel carried in broadcast PDU train
SCH Short transport CHannel carried in DLiUL PDU trains
LCH Long transport CHannel carried in DLiUL PDU trains
RCH Random access CHannel carried in UL PDU train
GT Guard Time duration determined by scheduler
Prxs very short Preamble 8 ps
Prs short Preamble 12 ps
PrL Long Preamble 16 ps
RTT Radio Turn around Time 6 ps
.. ..
Broadcast
POUTrain
- OL
POUTrain
..
DL
POUTrain
..
UL
POUTrain
- --
UL UL
POU Train POU Train
Figure 1. HIPERLAN/2 MAC frame layout
The PDU trains which are of concern here consist of SCHs (see table I)
to carry control information and LCHs mainly used to carry payload. In the
direct link phase (that will not be considered any more in what follows), Mobile
Terminals (MTs) send PDU trains to each others in a peer-to-peer manner.
During the DL phase, the access point sends PDU trains either in multicast or
266
to a specific MT. Lastly, MTs send PDU trains to their access point during the
UL phase. Figure 1 illustrates this.
A connection is identified with a DLC User Connection IDentifier (DUC
ID) and a given MT can manage several connections either DL or UL:
- if a DL connection is considered, the access point sends LCHs contain-
ing payload to the MT and the latter acknowledges the receipt of the
packets in SCHs containing ARQ feedback messages. In these mes-
sages, the MT can request more bandwidth for acknowledgement using
the ABIR bit (ARQ Bandwidth Increase Request);
- if an UL connection is considered, the MT sends the payload in LCHs
and the acknowledgement is done in SCHs sent by the access point.
In both cases, as the resource allocation is centralized, the scheduler in the
access point grants the number of LCHs and SCHs for each connection. To
get the resources, the MT makes its request via an RCH or an SCH. Thus, for
limiting the overhead and for the buffer management, an adequate choice of
the signalling strategy is key.
This section deals with the choice of the strategy. This will be Hl2 ori-
ented but the algorithms can be extended to any other system relying on a
TDDrrDMA access scheme.
Signalling strategy algorithms. As illustrated in table 2, several

algorithms can be envisaged either for a DL or for an UL connection. For all
these algorithms, the scheduler grants dynamically the number of SCHs (they
can vary from frame to frame) and the maximum number of SCHs can be fixed
or not. Let outline their principles.
Table 2. Signalling strategies
Strategy name Connection type Comments
I ABIR based DL dynamic allocation according to the ABIR

2 upper boundary UL dynamic allocation but limited to a max. value
3 no boundary UL dynamic allocation without limit
The ABIR based algorithm is close to the one proposed in (Li et aI., 2000).
Indeed, when the access point receives an ARQ feedback message with the
ABIR bit set, the number of SCHs is increased by one the frame after with no
upper limit. Otherwise, this number is decreased by one. Yet, this value is kept
greater than one.
In the case of the second algorithm, the scheduler grants all the needed ARQ
feedback messages to the connection up to a given limit denoted Max. We
tested this scheme for Max ranging from 1 to 3. This limits the overhead with
a highly dynamic bandwidth.
The third algorithm is an extension of the second with no upper limit. This
may grow the overhead but the scheme responds to any variation of the traffic
with no delay.
The performance of these three algorithms is compared in terms of through-
put calculated on top of the DLC layer (i.e. provided to the IP layer). Before
dealing with the simulation results, let first derive an expression for the ideal
throughput.
Analytical ideal approach of the throughput. Based on the Hl2

MAC overhead calculation of (Kadelka et al., 1999) and the ideal ARQ study
of (Lin and Costello, 1983), the H/2 throughput on top of DLC layer is given
by:
'toverhead)
P[Mbps] = r mode[Mbps]' ( I - A
. 1-"
(
I - PER
)
(1)
'tframe
where
- rmode is the nominal bit rate ranging from 6 to 54 (Mbps) depending
on the physical mode selected for the transmission of the LCHs (see in
(ETSIIBRANIDLC, 2000))
- 'toverhead is the part of the MAC frame containing no payload. Referring
to (Kadelka et al" 1999) and (ETSIIBRANIDLC, 2000), this is worth
'toverhead[ ,us] = 146[ ,uS] + 'tSCHs[ ,uS] (2)
The duration of the SCHs also depends on the physical mode used by the
LCHs. Also note that the MAC overhead includes propagation delays
(guard times).
- 't frame equals 2 ms
- ~ represents the overhead ratio introduced by the Cyclic Redundancy
Check (CRC) bits (ETSIIBRANIDLC, 2000) and the convergence layer
header (ETSIIBRAN/CL, 2000). This value is worth
~= 48 (3)
54
- PER is the Packet Error Rate ranging from 0 to 20% in our simulations,
which corresponds to typical operating conditions.
Equation (1) reflects ideal ARQ assumptions that is an infinite buffer size,
an unlimited number of retransmissions, error-free signalling and no signalling
bandwidth limitation. Ideal ARQ assumptions are not realistic but will provide
268
an upper bound to the throughput. Finally, packet errors are assumed inde-
pendent. In the H/2 context, such an assumption is valid in noise-limited en-
vironments, where thermal noise produces bit error bursts at the output of the
Viterbi decoder which are much shorter than the packet length.
Simulation results. The throughput obtained during the transmission

of data between one access point and one MT, with one connection activated
(either UL or DL) and full system load (large file transfer) is depicted on figure
2. In this simulation, the transmitter or receiver buffer has a window size of
512 and the payload is transmitted with the fastest mode (64 - QAM), which
gives a nominal bit rate of 54 Mbps.
Figure 2. Throughput vs. PER for different signalling algorithms
From figure 2, several observations can be driven:

- the second algorithm gives better results for higher values of Max; this
is natural since the retransmission of the packets is faster so that the new
arriving packets are transmitted faster as well;
- as no limit is set on the number of SCHs, the third algorithm performs
even better; even though the non limited overhead can be fatal to the
throughput, it is observed that under these circumstances, the number of
SCHs is regulated and does not exceed 8 or 9;
- the first algorithm (using the ABIR bit in a DL connection) reaches a

comparable performance provided that the maximum number of SCHs
is not limited; as this scheme does not respond as fast as the third one,
we obtain a throughput which is slightly inferior.
Note that for other sources of traffic (like VBR - Variable Bit Rate), the
algorithm ranking is expected to be similar but potentially with larger gaps
between the throughput curves. For instance, if a VBR traffic source is used,
the third algorithm will be better suited to the dynamic of the traffic than the
first one.
Nevertheless, even with the best fitted scheduling algorithm, the discrepancy
between the theoretical curve obtained with ideal hypothesis on the ARQ and
that obtained with the simulation is rather large. This phenomenon is explained
in the next section.
3. Influence of a finite buffer on ARQ

This section analyzes the effect of finite buffer space on the throughput
performance of SR-ARQ. As stressed on figure 2, when the PER increases,
the throughput obtained by simulations becomes significantly lower than that
computed using expression (1). The issue of SR-ARQ under limited buffer
space has already been investigated in several papers. Generally (Miller and
Lin, 1981), the transmission of packets and the reception of acknowledgements
can occur simultaneously, which cannot be assumed in a TDDffDMA access
scheme. In (Saeki and Rubin, 1982), the analysis of SR-ARQ with TDMA is
restricted to messages of one packet. Several approaches are possible for the
study of SR-ARQ performance. In (Rosberg and Sidi, 1990), given the statis-
tics of the traffic source, the queuing and resequencing delays are obtained by
modeling buffers as Markov chains. (Jianhua et aI., 1999) provide a method
called "sequential method" to compute SR-ARQ throughput. Here we derive
an approximation of the throughput by simple discrete probabilities calcula-
tion, assuming a TDDffDMA scheme with multiple packets transmitted per
frame. Then we compare it to simulation results in the Hl2 context.
Derivation of analytical throughput expression. The same as-

sumptions as in section 2 hold, except that now the buffers are of limited size
W at transmitter and receiver side. Furthermore, the following behavior of the
SR-ARQ algorithm is supposed:
- The new packets are sent with ascending sequence number. Including
the retransmissions, M packets are allocated to the connection per frame.
We assume M ::; W.
- Let i be the index of the oldest packet not yet positively acknowledged.
It is allowed to free only the packets of index smaller than i.
270
- Only packets of sequence number smaller than i + W can be sent. There-

fore, it can happen that less than M packets are transmitted in a frame, or
even no packet if the buffer is full. This results in a stalled connection.
- There is no ARQ signalling bandwidth limitation.
In order to isolate the buffer saturation effect, we compute the efficiency coef-
ficient X defined as the ratio between the observed throughput and the "ideal
ARQ" throughput: X~ Pet!eCtive. An analytical expression of X is derived in
Ideal
the appendix. A block of packets is considered, which is defined as the se-
quence of packets sent in a given frame. We assume that the transmission was
not stalled at the first transmission attempt of the block, therefore the block
size is M. The age of a packet is defined by the number of attempts required to
successfully transmit it (the age equals 1 if the first attempt is successful). The
age N of a block is defined as the number of attempts needed to successfully
transmit all its packets (N = 1 if every packet of the block is correctly received
at the first attempt). If N :2 2, denote by n the position (1 ~ n ~ M) of the
first erroneous packet in the block at the (N - l)th attempt. The block will be
responsible for a transmission stall if all the following conditions are true:
1) N:2 2
2) There is no block in the buffer older than N (otherwise, there would be
a transmission stall, but the considered block would not be responsible
for it).
3) NM - (n - 1) > W
Condition 3 deserves some comments: when the considered block reaches age
N, the next N - 1 blocks of size M have already been transmitted. These blocks
contain new packets and retransmissions. Condition 3 is obtained by assuming
that these blocks contain only new packets that need new storage space in the
buffer. Therefore at the Nth attempt, only n - 1 packets out of NM have been
freed, which leads to the above condition. Rigorously, this simplification is
not valid, since each of the (N - 1) blocks contains at least one retransmission
(that of the erroneous packet of the considered block). The simplification leads
to an over-estimation of the buffer space required, and as a result to under-
estimating X. A more complex calculation (not derived in the paper) would
take into account the proportion of retransmitted packets in each of the N - 1
blocks, which requires conditional probabilities calculations. The considered
block contributes to the average efficiency X. If the block does not produce any
transmission stall, then the associated efficiency equals 1. Otherwise, it equals
M
M+NM-(n-l)-W'
The expression of X derived in the appendix is plotted on figure 3 versus sim-
ulation results obtained with the Hl2 network simulator. The model matches
the simulation results with good accuracy. Yet, when a~t\? grows, the buffer
saturation occurs more frequently and the assumption that blocks are of size
M is no longer valid. Also, some implementations specific to the H/2 standard

can be accounted for minor differences. Still, the validity of the assumptions
is credited by the similarity of the curve shapes: the efficiency at a given PER
decreases when the ratio a~ {\? approaches 1. At common PER values (below
15 %), the efficiency remains high (above 95 %) when IX is below 30 %. This
can have an impact on resource allocation as explained in next section.
5 10 IS ro ~ ~ ~ ~ ~ M ~ ~ ~ ~ n ~ ~ ~ ~ 100
a..MIW{%)
Figure 3. Illustration of the stall phenomenon for PER=5,10 and 15 %
4. ARQ impact on some layer-2 algorithms

Description of some scheduling algorithms. In this section, the
impact of SR-ARQ on two specific scheduling algorithms is investigated. This
analysis is mainly based on two simple scheduling techniques described in
(Kadelka et aI., 1999) in the Hl2 context and illustrated in figures 4a and 4b.
More sophisticated techniques exist in the literature. For instance, (Ranasinghe
et aI., 2001) optimize the resource allocation by classifying the terminals and
by using the dual queue method. In general, it is possible to trade off fairness
between terminals and connections against maximum throughput in the cell.
For instance, the following algorithms are classified by fairness and through-
put efficiency in figure 5:
272
~ :.'!
I""
l
k=J
C2 .
Cl
Figure 4a. Non Exhaustive Round Figure 4b. Exhaustive Round Robin
Robin Algorithm (NERR) Algorithm (ERR)
- best-SNIR ERR: the connection having the best Signal to Noise plus
Interference Ratio is served first - this implies a high throughput but the
slowest connections may never be served;
- time-based NERR: the scheduler allocates the same duration for each
connection no matter what their modulation is - even the slow connec-
tions will be served;
- data-based NERR: the scheduler allocates the same amount of data for
each connection - this will provide the fairest algorithm at the expense of
the cell throughput. In the following NERR will stand for this algorithm.
Fairness Cell throughput
..
t t t
data-based time-based best-SNIR
NERR NERR ERR
Figure 5. Comparison between various scheduling algorithms
Let now see how we can choose a scheduling algorithm based on the ARQ
configuration used.
Impact of ARQ on the resource allocation. In order to illustrate

the influence of ARQ on resource allocation, figure 6 plots the throughput of
two Hl2 connections in 64 - QAM mode (54 Mbps nominal bit rate) at full load
Impact of SR-ARQ with finite buffer on TDDjTDMA Wireless LAN 273
versus PER with a fixed ARQ window size set to 512 and served by NERR. The
total throughput almost reaches the ideal ARQ upper-bound. As a reference,
the throughput obtained in the same conditions but with a single connection is
also plotted. The latter can be viewed as a "worst case" of what can be reached
in the multi-connection case when only one connection is served per frame,
and all connections are stalled simultaneously. Basically, NERR performs very
well because the ratio a (cf. section 3) was divided by two.
Figure 6. Throughput of 2 active twined connections served by NERR
Now this phenomenon has been clearly highlighted, let see how this trans-
lates into recommendations for tuning scheduling algorithms. For that purpose,
let consider simulation results plotted on figure 3. If a throughput efficiency
greater than 98% is imposed, with a PER of 10%, a needs to be less than
28%. For simplification sake, the scenario is restricted to n identical connec-
tions, each set to the same physical mode <p (from 0 for BPSK rate ~ to 6 for
64 - QAM rate i). If each frame is completely filled, the number of LCHs per
frame is thus related to the physical mode. A relation between <p, n and the
window size W can therefore be derived as represented on figure 7.
This graph can be read in the following manners:
- if we have one user in the cell, we cannot have a W smaller than or equal
to 64;
274
512
X maximum PHY mode (O=BPSKII2 -> 6=64QAM3/4)
.~
nb users
Figure 7. Maximum physical mode to use for a given number of users and a given window
size
- if we have several users in the cell, all in physical mode 5 (36 Mbps) and
with a W of 128, then this number of users must be greater than or equal
to 5;
- if all the users have a W of 32 and a physical mode 6 (54 Mbps), there
needs to be at least 26 users in the cell;
- if 4 users share the cell with the 6th physical mode (54 Mbps), W has to
be greater than or equal to 256;
- if 7 users share a cell with a W of 64, then their physical mode should
not exceed 4 (27 Mbps);
- if all connections have a physical mode 5 and a W of 64, the optimum
number of users in the cell using a NERR algorithm is 9.
An ARQ aware scheduling strategy proposal. The above re-

sults and figure 5 suggest the following algorithm to allocate resources in a
TDDffDMA access based network. Connections are gathered in groups shar-
ing similar QoS defined by a priority (figure 8). These groups are served by
ERR (priority order) and among each group, the connections are served by
NERR. If the conditions shown on figure 7 are met, the throughput efficiency
in the cell can reach 98 %. We can go further by referring to (Kadelka et aI.,
1999) where the authors show that the number of users served by NERR must
be minimized in order to limit the MAC overhead. Since figure 7 provides a
lower limit Afor the number of users per group, we can thus impose an upper
limit A to reduce the overhead. For instance, the QoS based priority groups
can be further divided into sub-groups (of A< nb users < A) served by NERR.
These sub-groups being served by a fair ERR, in which the first served group
changes cyclically. For clarification sake, let consider the simplistic hypothesis
which led to figure 7, if all users have a window size of 128 and are transmitted
in physical mode 5, Aequals 5 and the optimum number of users in the cell can
be taken equal to 5. Note that such an algorithm is not simulated in this paper
and is currently under evaluation.
Transmission of new PDUs:

NERR-ERR strategy to insert PDUs in Tx queue
1) NERR inside QoS priority groups
2) ERR to serve the different groups
Group 1 Group2 Group3

QoS1 QoS2 QoS3
NERR NERR NERR
~J~
ERR
Figure 8. Scheduling algorithm proposal for a TDDffDMA access based WLAN
Impact of SR-ARQ on Link Adaptation. Link Adaptation is a

technique that has been extensively studied (Goldsmith and Chua, 1998): it
consists in adapting the constellation size and the coding rate to the fluctuating
link quality. For instance, when the estimated PER exceeds a pre-computed
threshold, a more robust physical mode (i.e. the association of a coding rate
and a constellation) is selected with a lower nominal data rate but a higher
throughput in the current transmission conditions. A measure of the link qual-
ity can be the PER. In the H/2 context (Lin et aI., 2000), (Simoens and Bar-
tolome, 2001), the physical mode switching points correspond typically to a
PER of 30 % and are computed assuming ideal ARQ. However, as illustrated
on figure 2, the throughput with non-ideal ARQ can be much lower than that
of ideal ARQ at such PER values. Therefore, the thresholds computed assum-
ing ideal ARQ can lead to a wrong behavior of LA algorithm and a significant
throughput degradation. This problem can be partially solved by carefully de-
signing the resource allocation algorithm, as explained before. A simple solu-
tion, which is often proposed in the literature, consists in taking some margin
in the switching points (at the expense of a slightly sub-optimum throughput
performance) so that for instance the PER never reaches 30 % but rather 5 %.
276
5. Conclusion
In this paper, two SR-ARQ signalling strategies well adapted to TDD/-
TDMA access based systems either for an UL or for a DL connection are
proposed. Nevertheless, a gap is observed between the theoretical throughput
and the one obtained by simulation in a Hl2 network. We then develop an an-
alytical approach to derive a new formula for the throughput which takes the
finite buffer into account. By doing so, we verify that finite buffer space is a
major factor limiting the throughput of the SR-ARQ scheme. We also show
that the throughput loss can be recovered by carefully setting ARQ buffer size
of each connection, modifying the link adaptation switching points and care-
fully designing resource allocation algorithm.
Appendix: Derivation of analytical throughput effi-

ciency expression
With the definitions of section 3, the first thing to compute is PNo,no the probability of the
event (N = No ~ 2 and n = no). An intermediate result is the probability of event A « the mth
Nb
packet of the block is correctly received at the latest at the h attempt », which of course does
not depend on m since packet errors are independent.
No
p(A) = L p(Aj} (A.l)
j=1
with A j the event: « The mth packet of the block is not received correctly until the /h attempt
(which is successful) ». We have p(Aj} = Ej-l(1- E) where E denotes the PER. Thus (A.I) can
be rewritten as:
No. I_ENo
p(A) = (I-E) L EJ - 1 = (I-E)-- = I_ENo (A.2)
j=l I-E
The event (N = No and n = no) can be expressed as BnCnD with:
- B: «the first no - 1 packets of the block are correctly received at the latest at the (No -
l)fh transmission»
p(B) = (1- ~o-l )no-l
- C: «the n~ packet is correctly received at No and not before »
p(C) = E 0-1(I- E)
- D: «the M - no remaining packets are correctly received at the latest at No »
p(D) = (I_ENO)M-nO
Since packet errors are independent, events B,C and D are also independent. Therefore, PNo,no =
p(B)p( C) p(D). Having PNo,no' the expression of X is direct:
X= p(N= 1)+ L PNo,no+ L

PNo,no (M+NoM -~no-I) _ W y(No) + l-y(No))
No?2 No?2
l:5,no:5,M l:5,no:5,M
NoM - (no -I) :5, W NoM - (no - 1) > W
(A.3)
Withp(N= 1) = (I-E)M and
+~
y(No) = TI(I- ENo+i)M (A.4)

i=O
The equation A.3 reflects the three conditions necessary to produce a transmission stall as de-
scribed in section 3. y(No) is the probability of the second event: «There is no block in the buffer
older than No ». This means that the block sent just before the considered block was received
correctly at age No and that the block sent before the previous one was received correctly at age
No + I and so on. Since blocks are independent just like packet errors arc, y(No) is the product
of these probabilities as written in A.4.
278
References
ETSIIBRAN/CL (2000). Broadband Radio Access Networks (BRAN); HIPERLAN Type 2
Functional Specification; Packet based Convergence Layer; Part 1: Common Part. Norme
ETSI, document TS101493-1 vl.l.l, European Telecommunications Standards Institute,
Sophia-Antipolis, Valbonne, France.
ETSIIBRAN/DLC (2000). Broadband Radio Access Networks (BRAN); HlPERLAN Type 2
Functional Specification; Data Link Control (DLC) layer; Part I: Basic Data Transport Func-
tion. Norme ETSI, document TSI01761-1 v1.2.1, European Telecommunications Standards
Institute, Sophia-Antipolis, Valbonne, France.
Gibson, J. D. (1997). The Communications Handbook. CRC Press, IEEE Press.
Goldsmith, A. J. and Chua, S. (1998). Adaptive coded modulation for fading channels. IEEE
Trans. on Communications, 46:595-602.
Jianhua, H., Zongkai, Y., and Donghua, D. (1999). Performance analysis on selective repeat
ARQ with finite buffer. IEEE Trans. on Communications, 2:1101-1104.
Kadelka, A., Hettich, A., and Dick, S. (1999). n Performance Evaluation of the MAC protocol
of the ETSI BRAN HIPERLAN/2 standard n. Proceedings of the European Wireless '99.
Li, H., Lindskog, J., G.Malmgren, Miklos, G., Nilson, F., and Rydnell, G. (2000). n Automatic
Repeat Request (ARQ) Mechanism in HIPERLAN/2 n. Proceedings of the IEEE VTC Spring
conference.
Lin, S. and Costello, D. J. (1983). Error Control Coding: Fundamentals and Applications. En-
glewood Cliffs, NJ: Prentice-Hall.
Lin, Z., Malmgren, G., and Torsner, J. (2000). System Performance Analysis of Link Adaptation
in HIPERLAN Type 2. In Proceedings of the IEEE VTC Fall conference, volume 4, pages
1719-1725.
Miller, M. J. and Lin, S. (1981). The analysis of some selective-repeat ARQ schemes with finite
receiver buffer. IEEE Trans. on Communications, 29: 1307-1315.
Ranasinghe, R. S., Andrew, L. L. H., Hayes, D. A., and Everitt, D. (2001). Scheduling disci-
plines for multimedia WLANs: Embedded round robin and wireless dual queue. In Proceed-
ings of the IEEE ICC Conference, volume 4, pages 1243-1248.
Rosberg, Z. and Sidi, M. (1990). Selective Repeat ARQ: the joint distribution of the trans-
mitter and the receiver resequencing buffer occupancies. IEEE Trans. on Communications,
38: 1430-1438.
Saeki, B. H. and Rubin, I. (1982). An analysis of a TDMA channel using stop-and-wait, block,
and select-and-repeat ARQ error control. IEEE Trans. on Communications, 30: 1162-1173.
Simoens, S. and Bartolome, D. (2001). Optimum performance of link adaptation in HIPER-
LAN/2 networks. In Proceedings of the IEEE Vehicular Technology Conference, Rhodes,
Greece.
Traffic Performance Analysis of Multimedia
Applications in Evolved GSM Networks
Peter Stuckmann, Christian Hoymann

Communication Networks
Aachen University of Technology (RWTH), Germany
pst@comnets.rwth-aachen.de
http://www.comnets.rwth-aachen.de;-pst
Abstract This paper presents traffic performance results for different traffic mixes in cellu-
lar packet radio networks. These results are useful to estimate the radio capacity
that is needed during the evolution of GSM networks towards third-generation
(3G) mobile communication systems like GSMIEDGE Radio Access Networks
(GERAN). First a traffic mix of applications based on the Wireless Application
Protocol (WAP) and conventional Internet applications like WWW browsing and
e-mail over the General Packet Radio Service (GPRS) are regarded. In the next
step traffic performance results for Streaming applications over GPRS and EG-
PRS are presented and the feasibility of Streaming with coexisting interactive
and background applications like WWW and e-mail is examined. Simulation
results for quality of service measures for the different applications and GPRS
system measures are based on the simulation tool GPRSim that models the appli-
cation and user behavior, the TCPIIP and WAP protocol architecture, the GPRS
protocol architecture and the radio channel.
Keywords: GPRS, EDGE, Multimedia, GPRSim, Stochastic Simulation
1. Introduction
The driving force for the evolution of second generation mobile communi-
cation systems such as the Global System for Mobile Communication (GSM)
is the predicted user demand for mobile data services that will offer mobile
Multimedia applications and mobile Internet access.
After High Speed Circuit Switched Data (HSCSD) has been introduced in
some countries in 1999, the first GPRS-based services have been available
since 2001 in Europe. Many countries worldwide will introduce GPRS in the
next years. With these new services mobile data applications with net bit rates
of up to 117 kbitls will be offered and established on the market. To real-
ize higher data rates the European Standardization Institute (ETSI) and the
3rd Generation Partnership Project (3GPP) have developed the Enhanced Data
279
x. Lagrance and B. Jabbari (eds.),
280
Rates for GSM Evolution (EDGE) standard, which offers a net bit rate of up
to 384 kbitls by means of modified modulation, coding and medium access
schemes (see Furuskar et aI., 1999; Stuckmann and Franke, 2001).
In parallel to the GSM evolution, the data applications performed by mo-
bile users will evolve. In the first phase of the GSM evolution, where the data
services Circuit Switched Data (CSD) and GPRS are available, WAP-based
applications as defined in Wireless Application Protocol Forum, 1999 running
on smart phones and PDAs besides conventional Internet applications running
on laptop computers or enhanced PDAs will dominate. Then Video Stream-
ing applications (see Elsen et aI., 2001) and Large Data Transfer (LDT) ap-
plications including the Multimedia Message Service (MMS) based on WAP
version 2.0 as defined in Wireless Application Protocol Forum, 2001 are felt to
become more popular with the optimization of GPRS and with the introduction
of EDGE and the related packet data service Enhanced GPRS (EGPRS).
While for the time period right after the service introduction minimal con-
figurations were chosen supporting only a basic availability of GPRS, with
increasing data traffic load in the next years GSMlGPRS cell capacity will
have to be extended. For this evolution of GSMlGPRS networks and for
the introduction of EGPRS, dimensioning guidelines are needed for operators,
equipment manufactures and system integrators. They should describe the re-
lationship between the offered traffic and the radio resources to be allocated
to reach a desired quality of service for the different applications (see Walke,
2001; Stuckmann and Paul, 2001; Stuckmann, 2002).
This paper aims at presenting simulation results for two predicted traffic
mixes, one for a GPRS evolution scenario and one for an EDGE introduction
scenario. The first one is composed of WAP, WWW and e-mail, the second is
defined by Streaming, WWW and e-mail.
In Section 2 the potential applications and the related traffic models are
introduced. After the description of the simulation tool GPRSim in Section 3
the traffic performance results are presented and interpreted in Section 4.
2. Applications and Traffic Models

This section describes the traffic characteristics that are expected in 2.5 and
3G mobile radio networks. After the presentation of traffic models for the
conventional Internet applications WWW and e-mail, WAP applications are
depicted. Finally an introduction into Streaming applications and the related
traffic models are given.
www
All applications summarized by World Wide Web (WWW) are based on the
Hypertext Transfer Protocol (HTTP), which uses the TCPIIP protocol stack.
HTTP organizes the transfer of Hypertext Markup Language (HTML) docu-
ments (web pages).
281
WWW sessions consist of requests for a number of pages. These pages

consist of a number of objects with a certain object size. Another characteristic
parameter is the delay between two pages depending on the user's behavior to
surf around the Web (see Arlitt and Williamson, 1995; ETSI 3GPP, 1998).
Table 1 gives an overview of the WWW traffic parameters. The small number
of objects per page (2.5 objects), and the small object size (3700 byte) were
chosen, since Web pages with a large number of objects or large objects are
not suitable for thin clients such as PDAs or smart phones served by (E)GPRS.
The traffic characteristics of the WWW model can be seen in the distribution
functions of the object size.
E-mail
E-mails are transmitted by using the Simple Mail Transfer Protocol (SMTP)
or the Post Office Protocol version 3 (POP3) for e-mail download. Since the
size of an e-mail download on a mobile device is the crucial parameter for this
research, a traffic model defining e-mail sizes is suitable. The introduced e-
mail model based on Paxson, 1994 describes the load arising with the transfer
of messages performed by an SMTP user. The only parameter is the e-mail size
that is characterized by two log2-normal distributions plus an additional fixed
quota of 300 byte (see Table 1). The base quota was assumed to be a fixed
overhead. Subtracting the overhead, a bimodal distribution remained. The
lower 80 % were said to be text-based mails, while the upper 20 % represent
mails with attached files, which can be rather large. The transition between
these two distributions is 2 kbyte. The maximum e-mail size is set to 100 kbyte.
WAP
The WAP specifications, which are the basis for the implementation in to-
day's mobile terminals, including the June 2000 Conformance Release, also
known as WAP 1.2.1, aim at optimizing the operation in 2G networks. There-
fore WAP 1.2.1 defines a distinct technology comprising protocols and content
representation. WAP is a suite of specifications that defines an architecture
framework containing optimized protocols (e.g., WDP, WTP, WSP), a com-
pact XML-based content representation (WML, WBXML) and other mobile-
specific features like Wireless Telephony Applications (WTA) as defined in
Wireless Application Protocol Forum, 1999.
W AP Release l.x. In addition to the goal of the optimized operation

in 2G networks, WAP has been developed because today's graphics-enhanced
web services cannot be brought to and displayed on thin clients, e.g., GSM
mobile phones, and IP as the network layer may not be applicable in some
environments, e.g., WAP over Short Message Service (SMS) or Unstructured
Supplementary Service Data (USSD).
282
Table 1. Traffic model parameters
WWW Parameter Distribution Mean Variance

Pages per session geometric 5.0 20.0
Intervals between pages [s] negative exponential 12.0 144.0
Objects per page geometric 2.5 3.75
Object size [byte] log2-Erlang-k (k = 17) 3700 1.36· 10 6
e-mail Parameter Distribution Mean Variance
e-mail size (lower 80 %) [byte] log2-normal 1700 5.5.106
e-mail size (upper 20 %) [byte] log2-normal 15700 62.9.10 9
Base quota [byte] constant 300 0
WAP Parameter Distribution Mean Variance
Decks per session geometric 20.0 3800
Intervals between decks [s] negative exponential 14.1 198.8
Size of 'Get Request' packet [byte] log2-normal 108.2 4.1 . 10 3
Size of 'Content' packet [byte] log2-normal 511.0 3.63.10 5
Because of the optimizations and different protocols it is not possible to

run WAP end-to-end to a regular Internet site. Instead, a WAP Gateway must
be used. The main services a WAP Gateway provides is protocol conversion
between WAP stack and Internet stack. In addition to this standardized func-
tionality, many gateway vendors provide a variety of value-added services that
allow for personalization, for example.
WAP Release 2.0. In the specification WAP 2.0 as defined in Wireless

Application Protocol Forum, 2001 some existing WAP protocols have been
extended by new capabilities. WAP 2.0 converges with widely used Inter-
net protocols like the Transmission Control Protocol (TCP) and the Hypertext
Transfer Protocol (HTTP). Internet Engineering Task Force (IETF) work in the
Performance Implications of Link Characteristics (PILC) Working Group has
been leveraged to develop a mobile profile of TCP for wireless links. This pro-
file is fully interoperable with the common TCP that operates over the Internet
today. Further, WAP 2.0 does not require a WAP proxy, since the communi-
cation between the client and the server can be conducted using HTTP 1.1.
However, deploying a WAP proxy can still optimize the communication pro-
cess and may offer mobile service enhancements, such as location, privacy,
and presence based services. In addition, a WAP proxy remains necessary to
offer Push functionality.
In addition to protocol work, the WAP Forum has continued its work on
service-enabling features for the mobile environment, like the Push service or
synchronization issues. Although WAP 2.0 has been finished in 2001, WAP 1.x
283
protocol stacks will still be used in the mobile terminals in the next years. In
this paper, only WAP l.x is regarded.
WAP Traffic Model. A WAP traffic model has been developed and
applied in Stuckmann et aI., 2001; Stuckmann and Hoymann, 2002.
A WAP session consists of several requests for a deck performed by the
user. The maximum amount of data that can be transferred by one request
defaults to 1400 bytes. The parameters are summarized in Table 1. The main
characteristic is a very small mean packet size (511 byte) modelled by a log2-
normal distribution with a limited maximum packet size of 1400 byte (see
Table 1).
Video Streaming
Many Internet portal sites are offering video services for accessing news and
entertainment content from a Personal Computer (PC). Beside Motion Picture
Expert Group (MPEG), H.263 is the currently most accepted video coding
standard for Video Streaming applications. In the near future, mobile commu-
nication systems are expected to extend the scope oftoday's Internet Streaming
solutions by introducing standardized Streaming services as described in Elsen
et aI., 2001.
In the scope of modelling video sources, a lot of attention has been paid to
long range dependent or self-similar models of traffic streams in telecommuni-
cation networks (see Willinger et aI., 1997). Many of such models have been
used to investigate Variable Bit Rate (VBR) video sources with a statistical
analysis of empirical sequences and estimation of the grade of self-similarity
(see Rose and Frater, 1994). Since MPEG and H.263 video traffic consists of
a highly correlated sequence of images due to its encoding, the correct mod-
elling of the correlation structure of the video streams is essential (see Zaddach
and Heidtmann, 2001).
In this work no stochastical models of video streams with self-similar or
high-correlated traffic characteristics are applied. Real video sequences coded
by an H.263 coder are used to generate the Streaming traffic.
The Video Streaming traffic model used within the scope of this work is
based on three video sequences in the format QCIF (Quarter Common Inter-
mediate Format) with the resolution of 176 x 144 pixels. The sequences are
proposed by the Video Quality Expert Group (VQEG) and are for this reason
commonly used. Each sequence is representing a particular group of videos
with different intensities of motion.
• Claire stands for a very low motion intensity and can be seen as a
characteristic video conferencing sequence or inactive visual telephony.
• Carphone includes both, periods with rather high motion and peri-
ods of low motion intensity. It represents many kinds of vivid or active
video-conferences or even visual telephony.
284
• The third video, Foreman, is a sequence with permanently high motion

intensity of both, the actor and the background. This permanent motion
is characteristic for sport events or movie trailers.
The H.263 coder was used with a skip factor of 2, which means that every
second frame of the original sequence was skipped so that the frame rate of the
coded sequences was reduced from 25 to only 12.5 frames/so
The quantization level 20 (Q20) was adjusted for Intra (1)- and Predictive
(P)-frames. The resulting video quality is marginal. But it is acceptable for
mobile devices with its limited visual output capacities.
A conservative mix of sequences including 80 % Claire, 10 % Car-
phone and 10 % Foreman has been selected for the simulations performed.
The mix shall represent video streams with low motion and only a few streams
with higher motion intensity.
Due to the negligible size of Real-Time Streaming Protocol (RTSP) and
Real-Time Protocol (RTP) control messages in comparison to the size of real-
time data, they have been neglected. The resulting average IP traffic offered by
this particular mix is 14.39 kbitls (see Table 2).
Sequences offered IP traffic

Q20 80-10-10 Mix
Claire 10.9 kbitls
Carphone 26.7 kbitls 14.39kbiUs
Foreman 31.7 kbitls
Table 2. Offered IP traffic of video sequences
Beside visual telephony all of the new emerging applications are relatively
short in duration. So called heavy users, generating long streams with huge
amounts of data, have not been taken into account. The duration of video
sessions is modelled by a negative-exponential distribution with an average
value of 60 S. This is an assumption with regards to the prognosis for 3G
networks in ETSI 3GPP, 1998 where the duration of real-time calls is proposed
to be modelled by a negative-exponential distribution.
3. Simulation Environment
The full details of the GPRS protocol stacks of the radio and the fixed net-
work and of the Internet protocols including the characteristics of TCP cur-
rently cannot be described by formulas usable in practice. Since GPRS net-
works are presently introduced in the field, traffic engineering and related per-
formance results are needed soon, so that capacity and performance estima-
tions become possible for GPRSIEDGE introduction and evolution scenarios.
Measuring the traffic performance in an existing GPRS network is not pos-
sible, since a scenario with a well-defined traffic load is hard to set-up, the
285
Circuit Packet Internet load Generator (SOL)

Generator
Switched
Generator I Funet I HTTP FTP "MrP WAPI "'it*'-
IMobIIe,1 TCPIUOP
IRallwayl IP
Generator
GIST
Figure 1. The (E)GPRS Simulator GPRSim
evaluation of the perfonnance by measurement is very difficult, and the analy-

sis of different protocol options is not possible in an existing radio network.
Therefore computer simulation based on the prototypical implementation
(called emulation) of the standardized GPRS protocols and the Internet proto-
cols in combination with traffic generators for the regarded applications and
models for the radio channel are chosen as the methodology to get the needed
results rapidly.
The (E)GPRS Simulator GPRSim is a pure software solution based on the
programming language C++. Up to now models of Mobile Station (MS), Base
Station (BS), Serving GPRS Support Node (SGSN), and Gateway GPRS Sup-
port Node (GGSN) have been implemented. The simulator offers interfaces to
be upgraded by additional modules (see Figure 1).
For the implementation of the simulation model in C++ the Communication
Networks Class Library (CNCL) (see M. Junius et aI., 1993) is used, a pre-
decessor to the SDL Perfonnance Evaluation Tool Class Library (SPEETCL)
presented in Steppler, 1998. This enforces an object oriented structure of pro-
grams and is especially suited for event driven simulation.
Different from usual approaches to establish a simulator, where abstrac-
tions of functions and protocols are being implemented, the approach of the
GPRSim is based on the detailed implementation of the standardized GSM
and (E)GPRS protocols. This enables a realistic study of the behavior of EG-
PRS and GPRS. The real protocol stacks of (E)GPRS are used during system
286
simulation and are statistically analyzed under a well-defined and reproducible

traffic load.
The complex layers of the protocol stacks like SNDCP, LLC, RLCIMAC
based on (E)GPRS Release 99, the Internet traffic load generators and TCPIIP
itself are specified formally with the Specification and Description Language
(SDL), translated to C++ code by means of the Code Generator SDL2CNCL
(see Steppler, 1998) and finally integrated into the simulator.
4. Traffic Performance Evaluation

Simulation Scenario Parameter Settings
The cell configuration is given by the number of Packet Data Channels (PD-
CHs) permanently available for GPRS. In this paper 1,4,6 and 8 fixed PDCHs
have been regarded. For the GPRS simulation series a CII of 12 dB (13.5%
BLEP) has been regarded and Coding Scheme 2 (CS-2) has been used. For
EGPRS the channel conditions are determined by the cell and cluster size that
are the basis for the CII calculation as described in Stuckmann and Franke,
2001. Cluster size 7, a cell size with a radius of 3000 meters and a velocity
of 6 kmIh has been regarded. Both Link Adaptation (LA) and Incremental
Redundancy (IR) are applied.
LLC and RLCIMAC are operating in acknowledged mode for WWW, e-
mail and WAP and in unacknowledged mode for Streaming. The multislot
capability is 1 uplink and 4 downlink slots. The MAC protocol instances
in the simulations are operating with 3 random access subchannels per 52-
multiframe. All conventional MAC requests have the radio priority levelland
are scheduled with a FIFO strategy. LLC has a window size of 16 frames.
TCPIIP header compression in SNDCP is performed. The maximum IP data-
gram size is set to 1500 byte for WAP and 552 byte for the TCP-based appli-
cations. In the Internet stack for WWW and e-mail TCP is operating with a
maximum congestion window size of 8 Kbyte. The transmission delay in the
core network and external networks, i.e., the public Internet is neglected, since
it is assumed that the servers are located in the operator's domain and the core
network is well dimensioned. Since the high round-trip time in GPRS networks
is mainly caused by Temporary Block Flow (TBF) establishment procedures at
the air interface, the delay in well dimensioned IP subnetworks does not have
a great effect on the end-to-end performance.
Performance and System Measures

To characterize the traffic performance of GPRS several performance and
system measures are defined in the following. For the different applications
different critical performance measures will be regarded, since different per-
formance characteristics are required for transaction-oriented applications and
real-time applications, respectively.
287
Mean IP throughput per user is the downlink IP throughput measured dur-
ing transmission periods, e. g., the download period of a single object
of a web page. This is an important QoS parameter from a user's point
of view. The statistical evaluation of this measure is done by counting
the amount of IP bytes transmitted in each TDMA frame period for each
user, if a packet train is running. Thus, the throughput is not averaged
over inactive periods. The number of IP bytes transmitted divided by
the TDMA frame duration represents a simulation sample value in the
evaluation sequence. At the end of the simulation the mean throughput
is calculated from this evaluation sequence.
Mean application response time is the difference between the time when a
user is requesting a web page, a WAP deck or an e-mail and the time
when it is completely received.
Mean IP datagram delay is the end-to-end delay of IP datagrams evaluated

by means of time stamps given to the datagrams, when the IP layer per-
forms an SNDCP data request for transmission. When the datagram ar-
rives at the receiver, the difference of the actual time and the time stamp
value is calculated as a sample of the respective evaluation sequence.
Mean throughput per cell is also called system throughput and is calculated
from the total IP data transmitted on all PDCHs of the regarded radio
cell and for all users during the whole simulation duration, divided by the
simulation duration. Since a loss of IP datagrams over fixed subnetworks
is not modelled, this parameter equals the offered IP traffic in the radio
cell.
PDCH utilization: is the number of MAC blocks utilized for MAC data and
control blocks normalized to the sum of data, control and idle blocks.
Thus existing capacity reserves in the scenario under consideration can
be seen from this measure.
WAP in Comparison to Internet Applications over

GPRS
To be able to compare the user-perceived performance of WAP in com-
parison to conventional Internet applications, the application response time is
shown in Figure 2(a) for pure WWW, e-mail and WAP traffic.
In situations with low traffic load the response time for a WAP deck is below
2 s, while the response time for a web page is around 4 s. The reason is that a
web page has a larger content size and is transmitted over TCP.
In load situations with higher traffic load the response time for a WAP deck
remains nearly constant for up to 20 MS. If only 1 PDCH is available, the WAP
response time increases to more than lOs for 20 MS in the radio cell. Because
of the larger content size the response time for web pages passes 20 s already
288
Application response time

~r------r------'----=~='-47.POC=H~-'
e-maJ 4 PDCH -----
Wpp, 4 PDCH -----
e·
-----------------
•• .
:=::::==:::::~.::::::::::::::::=:=;======~==~=~~----------------
,
NumbefofMS
15 20
(a) Pure WWW/e-mail and WAP traffic (b) Traffic Mix
Figure 2. Mean application response time
DL IP throughput per .,... DL IP ltYou(tlput per uaar
WWW4PDCH-
... e-mail 4 PDCH -------
............., WAP, 4 POOH --.... -.
2. 20 ..... ~":- ...,
I,.
t
;
0;. I,.
t -----~-~----
e:
..
~ 10 10
~ ~
.....................c....................... ,............. --.. --.~ ............... J

.
. L -____- L____ ~~
NumberofMS
____ ~ ____ ~
~
. L -____- L____
•
~~
NurnbarofMS
____ ~ ____ ~
20
(a) Pure WWW/e-mail and WAP traffic (b) Traffic Mix
Figure 3. Mean downlink IP throughput per user
with 10 active MS in the radio cell even if 4 PDCHs are available. The reason
for the strong increase in response time for WWW and e-mail can be seen in
other evaluated measures like the downlink PDCH utilization in Figure 4(a).
100 % PDCH utilization is reached for WWW/e-mail traffic with 15 MS, while
15 WAP users are only utilizing the PDCHs with 30 % for the same PDCH
configuration.
289
Figure 3(a) shows the mean downlink IP throughput per user during trans-
mission periods. While the throughput performance for pure WAP traffic re-
mains relatively constant with an increasing number of mobile stations and
4 PDCHs, it decreases dramatically for pure WWW/e-mail traffic because of
the higher offered traffic and the higher utilization. The poor throughput per-
formance for WAP traffic can be explained by the low WAP deck size. Such
transaction-oriented applications are more influenced by the high round-trip-
time, which is mainly caused by the high delay over the air interface, than by
the available bit rate. Since the response time for a WAP deck is less than 1.5 s,
which should be acceptable for a wireless application, the user is not aware of
this low throughput performance.
Since WWW and e-mail applications comprise larger file sizes to down-
load than WAP-based applications do, the throughput performance perceived
by a user in situations with low traffic load ranges from 14 to 24 kbitls. These
performance values are mainly influenced by the characteristics of the offered
traffic. Since the e-mail traffic model has larger file sizes than WWW, the
throughput performance is better. With an increasing number of mobile sta-
tions up to 15 the saturation is reached and the performance for WWW and
e-mail users gets unacceptable and even gets worse than the low throughput
for pure WAP traffic. In this situation with high traffic load the WWW and
e-mail traffic performance is less influenced by the characteristics of the traffic
model like the file size, but by the load on the air interface.
Downlink PDCH utll1zatlon Downlink PDCH utilization

100 r---.,----,--------c:o=--r-----,
Traffic-Mix -
pure WAP -------
pureWWW···
80
I6 60
]
s
Ii
~ 40
I 20
°0L---~--~--~--~20
Number of MS Number of MS
(a) Pure WWW/e-mail and WAP traffic (b) Traffic Mix compared to pure WWW and
pure WAP traffic
Figure 4. Mean downlink PDCH utilization

290
Traffic Mix with WAP and WWW / e-mail over

GPRS
Since the predicted traffic mix for GPRS networks will be composed of
WAP traffic and conventional Internet applications like WWW and e-mail, the
GPRS traffic performance for a traffic mix of 60 % WAP, 28 % e-mail and
12 % WWW sessions will be regarded, here.
Figure 2(b) shows the application response time for WAP decks, e-mails and
WWW pages, respectively. Compared to the graphs in the previous section, the
WWW and e-mail performance is not strongly affected by WAP traffic, since
small WAP packets can be multiplexed seamlessly with the TCP-based WWW
and e-mail traffic. The throughput (see Figure 3(b» decreases slower with an
increasing number of mobile stations than in Figure 3(a) with pure WWW, e-
mail and WAP traffic regarded separately, since here WAP represents the main
part of a traffic mix and the total offered traffic per radio cell is increasing much
slower.
The same applies for the response time. In the scenario with traffic mix
WWW pages have a response time of 5 s with 10 active stations generating a
traffic mix, while 10 stations generating pure WWW traffic have to wait for
more than 20 s.
The WAP response time increases slightly from 1.2 s for pure WAP traffic to
2.1 s for the traffic mix scenario. The reason is that WWW and e-mail sessions
are composed of larger application packets that leave less resources open for
WAP users (see Figure 4(b». Nevertheless a response time for WAP decks of
2.1 s still should be acceptable.
Traffic Mix with Video Streaming Applications over

EGPRS
As a typical EGPRS introduction scenario, Video Streaming applications
over EGPRS are examined in coexistence with WWW and e-mail applications.
Due to the conservative predictions concerning the future usage of Streaming
applications the mix only contains 10 % Streaming sessions. The remaining
part is assumed to 63 % e-mail and 27 % WWW session. As the critical perfor-
mance measure the downlink IP datagram delay is regarded (see Figure 5(a».
Additionally the downlink IP throughput per user in Figure 5(b) indicates, if
the Streaming data rate of 14.39 kbitls can be maintained under a regarded
number of mobile stations offering the Multimedia traffic mix. In addition the
distribution of the downlink IP datagram delay for 1 and 10 mobile stations is
shown in Figure 6(a) and 6(b).
With 4 available PDCHs in the regarded radio cell the IP datagram delay
for Streaming increases dramatically to more than 10 s with more than 6 active
stations that generate a traffic mix. With more than 6 stations the throughput
of 14.39 kbitls can not be maintained, which shows that the performance for
Streaming users becomes unacceptable. With 6 and 8 PDCHs available in the
291
OL IP datagram delay (video streaming) DL IPthroughput par user (video streaming)

10000
8000
1
~ 6O<lO
-8
{•
E
4000
"-
5
10
2000
NumberofMS NumberofMS
(a) Mean downlink IP datagram delay (b) Mean downlink IP troughput per user
Figure 5. Performance of Video Streaming applications (traffic mix)
Probability distribution function (1 MS par call, video streaming) Probability disbbution function (10 MS pat eell, video streaming)
D.•
0.'
~
it:
0.'
0.2
0
0
DL IP datagram dalay [msJ
(a) 1 mobile station (b) 10 mobile stations
Figure 6. Distribution of the downlink IP datagram delay for Video Streaming applications
(traffic mix)
cell 15-20 users generating the traffic with 10 % Streaming can be satisfied.
The delay starts increasing dramatically with 15 and 20 users, respectively.
Regarding the downlink IP throughput per user in Figure 5(b) there is no
significant difference in the performance between 6 and 8 available PDCHs.
The throughput for 4, 6 and 8 PDCHs start at the same level of 14.39 kbitJs.
This is exactly the data rate needed for the chosen video sequence. The down-
292
link IP throughput per user is remaining constant as long as the necessary data
rate for Streaming is provided. Depending on the number of fixed PDCHs the
real time data rate is decreasing below the required rate of 14.39 kbitls. At
this point the IP datagram delay is increasing dramatically. With 15 users the
required data rate can not be maintained any more.
The distribution functions in Figure 6(a) and 6(b) confirm these interpreta-
tion. For one mobile station 90 percent of the IP packets for the Streaming
applications are delivered within 150 ms. The performance is not depending
on the number of PDCHs available, since the regarded mobile stations can use
only maximum 4 slots on the downlink. For 10 mobile stations and 4 PD-
CHs available more than 50 percent of the IP packets are delayed more than
300 ms and the slow increase of the distribution function indicates a high de-
lay variance, which makes the delay performance for Streaming applications
unacceptable. With 6 and 8 PDCHs 85 % of the IP packets or Streaming ap-
plications are delivered within less than 300 ms, which makes the Streaming
performance just acceptable for 10 users generating the Multimedia traffic mix.
The steps in the distribution functions are affected by the segmentation of IP
packets into radio blocks and the number of radio blocks transmitted within a
GPRS radio block period of 20 ms.
The different requirements of the applications can be supported by Quality
of Service (QoS) management functions in the RLCIMAC layer (see Stuck-
mann, 2002). The transmission of Streaming data may be privileged on the ex-
pense of background traffic. While the application response times for WWW
and e-mail would increase, the Streaming application would be able to proceed,
although high background traffic load occurred in the cell.
5. Conclusions
In this paper the performance of different Multimedia applications in packet-
switched cellular radio networks based on GPRS and EGPRS is presented. For
GPRS introduction and evolution scenarios WAP applications and a traffic mix
of WAP and conventional Internet applications over GPRS are examined. Af-
ter the performance characteristics of WAP and Internet applications have been
regarded separately, the effects of coexisting Internet traffic on WAP traffic and
vice versa are outlined. It has been shown that WAP traffic can be multiplexed
seamlessly with the Internet traffic because of the small and limited WAP deck
size, while Internet traffic slightly slows down WAP traffic in situations with
high traffic load. Regarding Video Streaming applications in coexistence with
TCP-based applications over EGPRS it has been shown that only a small num-
ber of Streaming users can be served by EGPRS, even if the percentage of
Streaming in the traffic mix is low. At least more than 4 fixed PDCHs should
be available to support Streaming applications together with background TCP
traffic. Privileged transmission of real-time data, realized by QoS manage-
ment, is one approach to provide the required bitrate for video streaming in
situations with high traffic load.
293
References
Arlitt, M. and Williamson, C. (1995). A Synthetic Workload Model for Internet Mosaic Traf-
fic. In Proc. of the 1995 Summer Computer Simulation Conference, pages 24-26, Ottawa,
Canada.
Elsen, I., Hartung, F., Hom, U., Kampmann, M., and Peters, L. (2001). Streaming Technology
in 3G Mobile Communication Systems. IEEE Computer, 34(9):46-52.
ETSI 3GPP (1998). Selection Procedures for the Choice of Radio Transmission Technologies
of the Universal Mobile Telecommunication System UMTS (UMTS 30.03, 3G TR 101
112). Technical report, European Telecommunications Standards Institute, Sophia Antipolis,
France.
Furuskar, A., Mazur, S., Milller, F., and Olofsson, H. (1999). EDGE: Enhanced Data Rates for
GSM and TDMNl36 Evolution. IEEE Personal Communications, pages 56-65.
M. Junius et al. (1993). CNCL: a C++ library for event driven simulation, statistical evalu-
ation and random number generators and distributions. Technical Report, Communication
Networks, Aachen University of Technology.
Paxson, V. (1994). Empirically-derived analytic models of wide-area tcp connections. IEEEI
ACM Transactions on Networking, 2(4):316-336.
Rose, O. and Frater, M. (1994). A comparison of models for VBR video traffic sources in B-
ISDN. IFIP Transactions C-24:Broadband communications II, pages 275-287.
Steppler, M. (1998). Performance Analysis of Communication Systems Formally Specified in
SDL. In Proc. of The First International Workshop on Simulation and Performance '98
(WaSp '98), pages 49-62.
Stuckrnann, P. (2002). Quality of Service Management in GPRS-based Radio Access Networks.
Telecommunication Systems 19:3, Kluwer Academic Publishers, pages 515-546.
Stuckrnann, P., Finck, H., and Bahls, T. (2001). A WAP Traffic Model and its Appliance for the
Performance Analysis of WAP over GPRS. In Proc. of the IEEE International Conference
on Third Generation Wireless and Beyond (3Gwireless '01), San Francisco, USA.
Stuckrnann, P. and Franke, J. (2001). The Capacity and Performance Gain Reachable with Link
Quality Control in EGPRS Networks. In Proc. of the IEEE International Conference on
Third Generation Wireless and Beyond (3Gwireless '01).
Stuckrnann, P. and Hoymann, C. (2002). Performance Evaluation of WAP-based Applications
over GPRS. In Proceedings of the IEEE International Conference on Communications (ICC
'02), New York, USA.
Stuckrnann, P. and Paul, O. (2001). Dimensioning GSM/GPRS Networks for Circuit- and Packet-
Switched Services. In Proceedings of the 10th Symposium on Wireless Personal Multimedia
Communications, ISBN 87-988568-0-4, pages 597-602, Aalborg, Denmark.
Walke, B. (2001). Mobile Radio Networks - Networking, Protocols and Traffic Performance,
2nd Edition. John Wiley & Sons, Chichester.
Willinger, W., Taqqu, M., and Sherman, R. (1997). Self-similarity through High-Variability:
Statistical Analysis of Ethernet LAN Traffic at the Source Level. 1EEEIACM Transactions
on Networking, 5:71-86.
Wireless Application Protocol Forum (1999). Wireless Application Protocol - Wireless Appli-
cation Environment Overview. Technical report.
Wireless Application Protocol Forum (2001). Wireless Application Protocol 2.0 - Wireless Ap-
plication Protocol Architecture Specification. Technical report.
Zaddach, M. and Heidtmann, K. (2001). Measurement and Traffic Characterization of H.26x-
coded Video Streams. In Proc. of the 11th GIIITG Conference on Measuring, Modelling and
Evaluation of Computer and Communication Systems. VDE-Verlag, Berlin.
VoiceNideo over IP with Multi-Class QoS in 3G
Mobile Networks
Liren Zhang, ROll Fang
Network Technology Research Centre, School ofEEE, S2, Nanyang Technological University,
SINGAPORE 639798, elzhang@ntu.edu.sg
Abstract: A novel encapsulation scheme in support of voice over IP and video

streaming over IP in the 3G mobile networks is presented. An end-to-
end connection across the 3G mobile networks is modeled as a tandem
queuing system associated with multiple QoS priorities. The
perfonnance analysis is done using fluid flow techniques. The numerical
results in tenns of network throughput, end-to-end delay and packet loss
probability demonstrate that the proposed scheme is able to guarantee
the multi-class of QoS for VoIP and video over IP services in the 3G
mobile networks. On the other hand, the network utilization is also
significantly improved.
Key words: Video over IP, Voice over IP, QoS, queuing systems.
295
296
1. Introduction
The services of voice over IP and video streaming over IP are going to
be introduced into the 3G mobile networks [2], since the capacity of such
network is powerful enough to handle those services. On the other hand, the
migration of IP into the 3G mobile networks is able to significantly increase
the network utilization. It also makes the connectivity of the mobile network
to be easily extended to the Internet anywhere irrespective of the user's
location. [5] [6]
The packet encapsulation is one of the most important issues to support

voice over IP and video streaming over IP in the 3G mobile networks [1]. It
requires that the encapsulation format must match the bit rate requirement for
both voice and video in order to reduce the effects of delay and delay jitter.
The overhead in the encapsulation must be significantly reduced to increase
the utilization of network resources. Another significant problem is the multi-
class QoS guarantee, since both voice and video are more sensitive to delay
and delay jitters comparing to data transfer applications. In fact, packet loss,
delay and delay jitter are the typical nature of IP networks. On the other hand,
the variable packet size may also cause a significant problem of delay and
delay jitter. Especially for IP over the 3G mobile networks, the problem may
be even significant.
The first focus of this paper is the packet encapsulation in support of

voice over IP and video streaming over IP in the 3G mobile networks. The
encapsulation format must match the bit rate requirement for both the voice
and video in order to reduce the effects of delay and delay jitter. The second
focus of this paper is multi-class QoS guarantee. The multi-class QoS is
implemented on differentiated service basis using priority scheme of 4 bits
defined in the proposed mobile IP packet header. MPLS is considered to
simplify the packet switching process over mobile IP network. A mobile IP
router is modeled as a tandem queuing system, in which each output link
consists of two space-priority output queues. The high-priority queue is used
to carry the delay sensitive traffic while the low-priority queue for delay
insensitive traffic. Multiple thresholds are deployed on each queue,
respectively, for loss priority control [3]. The performance analysis is based
297
on the fluid flow model. The performance evaluation focuses on the trade-off
between the delay sensitive traffic and delay insensitive traffic in terms of
traffic throughput and packet loss probability.
2. Packet Encapsulation Format

The disadvantages of current Mobile IP on IPv4 basis include (1) large
overhead, (2) variable packet size (3) variable delay and (4) long processing
delay for packet switching, which are not suitable for transferring the voice
packets and video packets across the networks. The proposed packet
encapsulation in support of voice over IP and video streaming over IP in the
3G mobile networks include the following features: (1) using fix and relative
small packet size to reduce the packetizationldeparcketization delays and to
make the end-to-end delay to be approximately constant, (2) matching the bit
rate for both voice traffic and video traffic and (3) label switching instead of
packet switching to reduce the processing delay.
FCC (Federal Communication Commission) has defined a bandwidth of
120MHz located at 2GHz for the 3G mobile networks to provide multimedia
services including video, voice and data. In this paper, a broadband CDMA
system with bandwidth of 4.096 MHz and the spreading factor of 64 is
considered as the fundamental network platform. The proposed encapsulation
is based on bit-rate of 4.096Mbps/64= 64 kbps. As shown in Figure 1, video-
conferencing service is supported at a typical bit-rate of 6 x 64 kbps, which is
equivalent to 10ms of video data based on the H.261 standard. The video data
stream is divided into segments. Each segment of 480 bytes video data is
divided into 6 packets with fixed packet size consisting of 80 bytes data. By
contrast, as shown in Figure 2, voice traffic is encapsulated into fix-size
packet consisting of 80 bytes data, which is equivalent to 10ms voice based on
G.711 standard at a typical bit rate of 64 kbps. Hence, each video packet is
equivalent to 6 voice packets.
Packet Format For H.261 (10 ms) Clear Channel Video P'64kb/s (p=6)
_I
=
480 bytes
4BO Byte Video Bu ndles
Video payload
Iso b~e'llso b~e'll b~e'llso

00 byte. I I SO b~e'l
Flag Label I Sequence I Priority I Re.erve Data
4bits
I Number Field
4bits
Sub SN
5 bytes 4bits 4 bits BObytes

U.er -3
Header Data
Figure 1, Video-conferencing Encapsulation

298
Packet Format for G.711 (10 ms) Clear Channel Voice 64kb/s
8o Byte Voice Bundles • Packet 80 bytes
Voice payload
User
Header --l
Data
Sequence Priority
Flag Label Reserve Data
Number Field
4b1ts 5 bytes 4bits 4blts 4 bits 80bytes
Figure 2, Voice Encapsulation
As shown in Figure 3, the packet header consists of Flag of 4 bits for

specifying the beginning of packet, Label of 5 bytes for label routing switch,
Sequence Number of 4 bits to provide the information needed when
reassembling the packet from the same voice/video source, Priority Field of 4
bits for handling multi-class QoS and Reserve Field of 4 bits for future use.
User
Header ----~~Data
Sequence Payload
Flag Label Reserve Data
Number type
80bytes
Figure 3, Proposed Packet Format

The priority field of 4 bits represents 16 QoS classes that are divided into
two categories: Values 0 through 7 specify the priority of traffic for which the
source is providing congestion control. Values 8 through 15 specify the
priority of traffic that does not back off in response to congestion.
For non-congestion-controlled traffic, the lowest priority value (8) is
used for those packets that the sender is most willing to be discarded under
conditions of congestion, and the highest priority value (15) is used for those
packets that the sender is least willing to be discarded.
3. Modeling and Performance Analysis

In the 3G mobile networks, mobile IP routers may be interconnected by
several input and output links. Each link may consist of a number of IP packet
streams corresponding to different priority levels, which are multiplexed
during the transmission. Packets carried in the same IP stream are assumed to
have the same priority level. It is expected that different priority levels are
associated with different IP streams. A typical 3G mobile IP router is modeled
as a non-blocking tandem switching node associated with output queue at the
299
output ports. The non-blocking switching function includes that the IP packet
streams carried on the same input link are demultiplexed at the input port and
then routed to the corresponding output port according to the label assigned to
the packet stream. At the output port, IP traffic streams with different priority
levels are multiplexed before they are transmitted onto the output link. The
multiplexer at the output port consists of two parallel output queues
corresponding to two different delay priorities, respectively. Each output
queue operates on a first-in-first-out (FIFO) non-preemptive basis while the
queue is being served.
The multiple loss priority [3] is implemented using threshold control
mechanism over the partitioned buffer of the two output queues, respectively.
As shown in Figure 4, the high priority queue is fed with non-congestion
controlled traffic (i.e., delay sensitive traffic) which consists of KI classes of
packet loss priorities corresponding to the threshold Ql, Q2, .. " Qk.... QKl ,
respectively, where k is the priority index and k= 1 (i.e., Ql ) corresponds to
the highest packet loss priority. When the buffer occupancy of the high delay
priority queue exceeds the threshold Qk" ,(J < k <KI), in this case, only the
non-congestion controlled packets with loss priority ranging from 1 to k-l are
permitted to input the queue while packets with the other priority classes are
discarded. Likewise, the low delay priority output queue consisting of QKl+l,
QKl+2, ... QK2 different thresholds for packet loss control is fed with congestion
controlled traffic, where k= KI +1 corresponds to the highest packet loss
priority. Since the non-congestion controlled traffic has higher priority than
those of the congestion controlled traffic, the packets contained in the high
priority queue are always served first and the packets contained in the low
priority queue are only served when the high priority queue is empty.
QJ Q 2 High Priority Queue
l--'Ii~----------------Tltl----~
:
QD~+~JQrK~J~+2_____ __Q_u_eu_e__-r~~
Lo_w_p_r_io_rity
K1+1---+O- . -
I r---~
K2~ L.......J'---_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _--'--'----'
K2-K1 priority classes of congestion controlled traffic
Figure4: Queuing modelfor multi-class QoS

IP packet stream generated by each user source is modeled as an ON-
OFF process, in which the transition rate from the ON state to the OFF state is
a and the rate of transition from the OFF state to the ON state is ~. Then the
probability that the process is in the ON state is aJ(a+~). When N of such
ON-OFF packet streams are multiplexed, the resultant stream can be
represented by an (N+ I)-state Markov modulated process, where the state i
represents that i ( i = 0, 1, 2, ... , N) packet streams are in the ON state. The
transition rate from the state i to the state (i-I) is i~, The transition rate from
300
the state i to the state (i+ 1) is (N-i)a.

The performance of statistically multiplexed ON-OFF traffic is evaluated
using the fluid flow model, in which each individual source is modeled as a
Markov Modulated fluid [4] source consisting of the ON state and the OFF
state. In the ON-OFF process, the ON state is uniformly distributed and the
transition between the ON state and the OFF states is controlled by a
continuous-time Markov chain, which determines the rate of fluid generating.
The transport of packets over a transmission link is operated in the same
manner. The fluid flow technique has been applied successfully to a variety of
problems in ATM networks. The conditions of using fluid flow technique are
that (1) the population of IP sources is large (N)>l) and (2) the output
transmission link is divided into equal time slots and each slot is equivalent to
the maximum transmission time of IP packet such that the transmission time
of IP packets is assumed to be uniformly distributed. These assumptions are
reasonable, when the population of mobile users, including voice, video and
data is large. in this case, each user only contributes a small fraction of the
link capacity. On the other hand, the packets including voice, video and data
in the system all have the fix and relatively small size. In this case, the bit rate
allocated to each mobile user comparing to the huge capacity of the
broadband CDMA link is negligible. Therefore, the variation of the
transmission time of IP packet is also negligible.
Queuing analysis for the high delay priority queue

Considering that the input streams of the high delay priority queue consist
of i packet steams, which are in the ON state at time t. Now we define
thatF;k(t,x) ( 0 ~ i ~ n, 1 ~ k < Kl) is the cumulative probability
distribution for the packet with k-th loss priority in the queue at time t, where i
packet streams are in the ON state. In fact, F/(t,x) ( 0 ~ i ~ n, 1 ~ k < Kl )
represents the probability that the queuing buffer occupancy is less than or
equal to x (Qk+I :::; X :::; Qk) while i packet steams in the ON state at time t.
F/ (t,x) can be calculated by setting up a generating equation of F/(t+,1 t,x)
which is the probability at an incremental time of t+,1 t. That is
Fk(t + ~t,x)= [N - (i -1)]cm1F;:1 (I,X)+ (i + l)pMFi~l (t,x)
I
+ {l- [(N - i)a + ip]M }r: k[I, x - (iAk - C)~t] + 0 (~I), (1)
=I
k
where Ak Aj and Aj is the arrival rate of the traffic with j priority class
j=l
in the ON state and the term x-(iA\C)Llt is the buffer occupancy. On the right
side of equation (1), the first term is the probability of transition from the state
(i-I) to the state i at time t+ LIt, the second term is the probability of transition
301
from the state (i+ J) to the state i, the third term is the probability that the
system state i is not changing at time t, and the term o(M) represents the all
higher order terms which go to zero much rapidly than /).t when /).t intends to
zero. Hence, the effects of o(/).t) is negligible when /).t is small enough. In
equation (1), it also assumes that F.j(t,x) and FN+J(t,x) are set equal to zero.
Now, F/(t+/).t,x) and F/(t,x-t1x) are expanded for t1x =( i).k_C ) /).t in
their respective Taylor series with the assumption that the appropriate
continuity conditions are satisfied. Let /).t go to zero, the equation (1)
represents the following differential function:
(i% -C)d(r;k(x)} tU =[N - (i -l)]ar;:l (x)+ (i + 1)PF;:1 (x)

-[(N -i)a+ i,B]Fk (x) (2)
I
By defining Fk (x) == [FOk (x), Flk (x ), ..... .F~ (x) ]r ,then the boundary
conditions in equation (2) can be obtained as below
F;Kl(O)=O if i E E uKJ ,
F;l( Ql)= 1; if i E E~, (3)

F;k ( Qk) =F;k-l ( Qk) 1'f'l E u E k.J U EkD' 2'5.k'5.Kl,
Hence, the steady state distributions under the boundary conditions of
equation (3) can be used to calculate the throughput for the traffic with
different priority classes, that is
T' = t, F,.-I (Q,) 0... - 'E~E:' [( o! - c ) (F;.-l (Q.-/) - F;' (Q. ) l]

k = 2,3, ....... , KO .
For k = J, the solution of r is given by
TI = t,F/(QI)iA I+c[ 1- t,F/(QI)]- (4)
The arrival rate for the traffic with the k-th loss priority IS given
302
N
by A k = I)'A k ~ , (J ~ k ~ Kl). Then the packet loss probability due to
;=0
buffer overflow is given by
PLk =1- Tk / Ak, 1< k ~ Kl
(5)
Queuing analysis for the low delay priority queue

The low priority queue carries the delay insensitive packet streams.
Since the low priority queue is only served when the high priority queue is
empty. The following analysis is considered into two cases as below: (1) the
high priority queue has input traffic and (2) the high priority queue does not
have input traffic. Let Pe be the probability that the high priority queue is
empty, then Pe can be obtained from equation (4) and (5) i.e.,
N Kl
Pe = IF; (0). In the first case, let C= O. Recall equation (2), the
;=0
differential function for the low priority queue is given by
CiA: - C) dF;k (x)/ dx = [N - (i -l)]aF;:l (x)

+ (i + 1) fJ F;!l (x) - [(N - i)a+ ifJ]F;k (x) (6)
k = Kl + 1, ...... K2 ,
where E/ (x) is cumulative probability distribution for the traffic with the k-th
loss priority in the low priority queue.
Likewise, recall equation (3), the boundary conditions for the low priority
queue is given by
F;K2(0)=0, l~i~N,
F;k (Qk) = F;k-l (QJ, 1 ~ i ~ N, Kl + 2 ~ k ~ K2.
Hence, the steady state distributions under the above boundary conditions is
given by
Ti k = I {AkF/ (Qk), Kl + 1 ~ k ~ K2. (7)

;=1
The throughput and packet loss probability for the delay insensitive traffic
with different loss priority can be calculated using equation (7).
In the second case, the high priority queue does not have any input traffic
stream. Assuming that the input streams of the high priority queue consist of j
streams which are in the ON state. We define C~ = C - P..J, where
303
Kl
Al = L A j ' Replacing F/ (x) with F/ (x), in equation (6), we have the
j=l
following differential function:
(i}.k - C~) dFj~(x)/ dx = [N - (i -I)]aFj~i-l) (x)+ (i + I)PFj~;+l)(x)

- [ (N - i) a + iP ] Fj~ (x) (8)
where F/
(x) is the cumulative probability distribution for the traffic with the
k-th loss priority when the system in the state i, The boundary conditions for
such a case is given by
~\O)=O l'f'
i E nriG
uj
~+l(Q)=P; l'fi E
' nrKltl
Dj
(9)
~(g)=~-l(QJ if i E l{/ U E;;j
Kl+2$k$K2
where E;j ={ il iAk <C~} E~ ={ il iAk >C~}.

From equation (8) with the boundary condition given by equation (9), the
throughput and packet loss probability for the delay insensitive traffic with
different loss priority is given by
T2kj = t
;=0
Fj;-l (Qk) i}.k - L [( i}.k - C~) (Fj~-l (Qk-J) - Fj~ (Qk ))l
iEE~nEtjl J
KI+2 $k $ K2
4. Numerical Results
The following numerical results focus on the effect of priority on the
steady-state performance including the delay sensitive traffic and the delay
insensitive traffic with different packet loss priority classes, For the
illustrative purpose only, the high delay priority queue oflength Q4 consists of
2 classes of packet loss priority, named Class 1 and Class 2, where the Class 1
traffic has the higher packet loss priority than the Class 2 traffic, The
304
threshold Q2 is used for the packet loss priority control in the high delay
priority queue. Likewise, the low delay priority queue of length Q3 also
consists of 2 classes of packet loss priority, named Class 3 and Class 4, where
the Class 3 traffic has the higher packet loss priority than the Class 4 traffic.
The threshold Q4 is used for the packet loss priority control in the low delay
priority queue. The traffic of Class h (h=1,2, ... , 4) consists of Nh
homogeneous independent ON-OFF sources in which the ON state and the
OFF state are exponentially distributed, respectively, with different mean
values.
From the QoS requirement of view, video generally requires the
network to provide QoS guarantees with respect of both packet delay
and loss. Voice is sensitive to delay and delay jitter rather than packet
loss and data transfer applications are more sensitive to packet loss
rather than packet delay and delay jitter. In the following evaluation,
video and voice are classified as the Class 1 traffic and the Class 2
traffic, respectively. The Data transfer applications are classified as the
Class 3 and the Class 4 traffic in the low delay priority queue
depending on their packet loss priorities, respectively. The threshold of
Q2 and Q4 control the impact of the trade-off between delay and loss in
the higher delay priority queue and the lower delay priority queue,
respectively. However, the buffer length of Q4 and Q3 determine the
trade-off between the packet loss probability and the maximum delay.
The choice of the buffer length should be reasonably large so that the
loss probability of higher priority traffic is very small.
The effects of priority on the perfonnance of packet loss probability and
throughput for the delay sensitive traffic and the delay insensitive traffic are
illustrated in Figure 5 and Figure 6, respectively. It can be seen that the traffic
load offered by the Class 1 has the significant effects on the perfonnance of
packet loss probability for the all the other lower priority traffic classes.
Therefore, in order to achieve the desired packet loss probability for different
traffic classes, both the overall traffic load in the network and the traffic load
offered by the Class 1 need to be properly controlled. In addition, when the
traffic load of Class 1 decreases, the throughput of Class 1 traffic decreases
obviously. By contrast, the throughput of the other lower priority traffic
classes increases significantly.
Figure 7 and Figure 8 illustrate the throughput and the packet loss
probability for different threshold values, respectively, where the traffic
offered load of the Class 1 is fixed at 15% of the link capacity. Figure7
demonstrates that the effect of the different threshold values on the
perfonnance of packet loss probability is significant for all traffic classes. For
example, when the traffic loading is 80% and the threshold value of Q2 and Q4
increase from 2 to 6 respectively, the packet loss probability for the Class 2
traffic is reduced and the packet loss probability for the Class 4 is also
305
improved. Likewise, the different threshold values make the same significant
effect on performance of the throughput for the Class 3 and the Class 4 traffic.
5. Conclusion:
A novel encapsulation format for adapting voice, video-conferencing and
data traffic with multi-class QoS in the 3G mobile networks is presented. The
3G mobile IP router is modeled as a non-blocking tandem switching system
associated with output queues. The introduction of multi-class priority defined
makes the QoS control in the 3G mobile networks flexible. The multi-class of
QoS is implemented using multiple space-queues for delay priority control
and threshold controlled partial queues for loss priority control. The
performance is evaluated using fluid flow model. The illustrated numerical
results have demonstrated that the proposed incorporating priority is able to
guarantee the QoS for the transmission of VoIP and video-stream over IP over
the 3G mobile networks. However, the priority schemes do not reduce the
total packet loss but do protect the high priority traffic from packet loss while
allowing the performance of the low priority traffic to degrade as little as
possible, especially when the traffic loading and the threshold value are
properly controlled. The behavior of multi-class priority scheme is studied
with a variety of traffic conditions. The obtained results show that the high
priority traffic improve vastly with the use of multi-class priority scheme
under the condition that the proportion of high priority traffic including the
offered load and the user population must be kept to a small percentage. On
the other hand, the traffic burstiness must be also carefully controlled.
Reference:
[1) Das. S.; Misra. A.; Agrawal. P .. TeleMIP: telecommunications-enhanced mobile IP
architecture for fast intradomain mobility, IEEE Personal Communications, Volume: 7 Issue: 4,
Page(s): 50 -58, Aug. 2000
[2) Lee. W.CY.; Lee. D.J.Y., Mobile IP, Personal, Indoor and Mobile Radio Communications,
2001 12th IEEE International Symposium on , Volume: I , Page(s): 88 -92, Sept. 2001
[3) S.o.Bradner. "IPng, Internet Protocol next generation", Addison Wesley, 1996
[4) A.Elwalid. D.Mitra. "Fluid Models for the Analysis and Design of Statistical Multiplexing
with Loss Priorities on Multiple Classes of Bursty Traffic", IEEE INFOCOM' 92, 0415-0425,
May, 1992.
[5) Le Grand, G.; Horlail. E., A predictive end-to-end QoS scheme in a mobile environment,
Computers and Communications, 2001. Proceedings. Page(s): 534 -539Sixth IEEE Symposium
on ,2001
[6) Goodman, D.J., Packet reservation multiple access for local wireless communications,
IEEE Transactions on Communications, Aug. 1989
306
-0- ( 4%,No Priority)

0.9 -'- ( 10%,No Priority)
-+- ( 15%,No Priority)
0.8 .0. ( 4%,WlIh Priority)
.'. ( 1O%,WlIh Priority)
.+. ( 15%,WlIh Priority)
~~L----·--0~~~----~0~.7----~OB~----~0~.9-----
OIIered load

-'- ( 10%,No Priority)
0.9 -+- ( 15%.No Priority)
.0. ( 4%,Wilh Priority)
.'. ( 1O%,Wlth Priority)
.+. ( 1S%,Wlth Priority)
0.5
0.4 "---------
0.5 0.6 0.7 0.8 0.9
Offered load
Figure 5: Effect of class} traffic load on throughputfor four classes
10"

-*- ( 10%,No Priority)
.0. ( 4%,Wlth Priority)
.*. ( 10%,Wlth Priority)
.+. ( 15%,Wlth Priority)
10-14c'>--__--+____f!l-__-+__~®_--_----<__--. .:...~-'---f'r-~"'lJ
0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
Offered load
307
10
-2
. -~:~~=;~=~~~~~~0~
*~~~*~ ~~6~~
.p .• ~:>//
/ /// ~/ -<l-( 4%,No Priority)
-'- ( 10%,No Priority)
10-8
.0. ( 4%.With Priority)
.'. ( 10%,with Priority)
10-12'----_~_~_ _ "_~~<::~~~~
0_55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
Offered load
Figure 6: Effect of class I traffic load on packet loss prob. for four classes
-'- (No Threshold.No Priority)

0.9 (Threshold=2,with Priority)
.0.
.'. (Threshold=4.With Priority)
.+. (Threshold=6.With Priority)
O.B
gj
~'" 0.7
~
,g.
~O.6
e
~
0.5
0/
0.4
0.3 ~--~---~---~--~----.
OA O. O. OJ O. O.
Offered load
-0- (No Threshold,No Priority) ~-e---

.0. (Threshold=2,with Priority) ~
0.9 .'. (Threshold=4,With Priority) //Y
.+. (Threshold=6,With Pnority) h /:~- -+- ~ + - -" _
/ -:.:-/- --""'- - * ~ -*-, ~ ~- :"
/ ,,¥ _ 6 - - 6 --<J.__
/ __ ....0--- '0,,-
$r /0' e
//
?,///0' .-'
~/
/
/
0.5
0.4 '------.~--~-:------~--~.--~
0.4 0.5 0.6 0.7 O.B 0.9
Offered load
Figure 7: Effect ofthreshold on throughput for four classes

308
r ~---<;------.<!>---~l
-------
~~---<fT--------
-{)- (No 11Yeshold ,No Priority)

.0. (Threshold=2,Wlth Priority )
.'. (Threshold=4,With Priority )
.+. (Threshold=6,with Priority ) +-
- +- ..
_ _r - -
+ __ -
A-
__
t
/*/ . .-
... ¥ .----
;+ c/
*- o
1
"
/
B
10." <IJ ~ e-
0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
Offered load
1/ -0- (No Threshold,No Priority)
/
.0. (Threshold=2,With Priority)
.'. (Threshold=4,Wilh Priority)
.+. (Threshold=6,wllh Priority)
10-2 "----~~~~~~~~~~_~_._...L_._j__L __ ~
0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
Offered load
Figure 8: Effect of threshold on packet loss Prob, for four classes

Establishment of Mobile Extranets through Mobile
IPv6 and GPRS: Enabling Universal Access to
Corporate Intranetsl
K.Koutsopoulos, N.Alexiou, C.Konstantinopoulou, P.Demestichas, and

M.Theologou
National Technical University ofAthens, Electrical and Computer Engineering Department,
Computer Science Division, Telecommunications Laboratory, 9 Heroon Polytechneiou Street,
Zographou 15773, Athens, GREECE. Tel: + 30 17721493, Fax: + 30 17722534, E-mail:
pdemest@telecom.ntua.gr
Abstract: Mobile Extranets (MEs) are seen as a natural evolution of the Internet and
Intranet concepts. Their role is to provide universal (ubiquitous) access to
corporate Intranet services, along with security, location transparency, cost-
efficiency and QoS. The aim of this paper is to present and comment on an
approach for the realisation of an ME platform. GPRS and the IPv6 protocol
suite, with its integral mobility and security functionality, are key technologies
for the platform. Aspects covered in the platform presentation are the
requirements posed by various application types, the platform architecture, the
role and limitations of the technologies encompassed, and the prototype
terminal architectures. Finally, issues for further work are presented and
concluding remarks are made.
Key words: Wireless Internet, Macro-Mobility
1. INTRODUCTION
It is often stated that wireless communications and the Internet are the
fastest growing businesses in the telecommunications market. A driving
1 This work is partially funded by the Commission of the European Communities, under the
Fifth Framework Program, within the 1ST (Information Societies Technology) project
MOEBIUS (IST-1999-11591: Mobile Extranet Based Integrated User Services).
309
310 K.Koutsopoulos, NAlexiou, C.Konstantinopoulou, P.Demestichas,
MTheologou
force for wireless operators is the foreseen increased revenue that can come
from the (new) IP-based services. Internet Service Providers (ISPs) wish to
make their services more attractive by incorporating wireless access and
mobility capabilities. Moreover, recent times witness the explosion of the
corporate Intranet and, as a natural evolution, the Mobile Extranet concepts.
An Intranet has two fundamental functions: (a) to provide secure and
customised access to information found in transaction systems; (b) to allow
users to act on that information. Customisation refers to the user preferences,
requirements and privileges.
The aim of Mobile Extranets (MEs) is to enable ubiquitous (universal)
access to corporate Intranet services (which are otherwise geographically
limited). This access to the Intranet from the outside world should be
transparent to the applications (i.e., it should leave them unaffected), as well
as secure, cost-efficient and at the appropriate Quality of Service (QoS)
levels. In the light of these aspects, it is envisaged that the realisation of MEs
can rely on the cooperation of radio access technologies, public Internet
segments, Mobile IP (MIP) and security mechanisms. Following this trend,
the aim of this paper is to present a specific approach for the realisation of an
ME platform. The presented approach has been developed and validated in
[1]. In addition to the user equipment and the Intranet, our approach
discusses on the General Packet Radio Service (GPRS) [2], IPv6 [3], Mobile
IP (MIP) [4],[14] and IPsec [5] technologies. The paper discusses on the
role, and the potential limitations, of these technologies in the ME platform.
In this respect, the paper is on the interworkinglintegration of technologies.
This is not a trivial task. The usefulness and success of the resulting ME
platform will be a driving force for the adoption of the technologies. The
literature does not have many such studies and attempts.
The starting point for the ME platform presentation is a discussion on
high-level application types that can benefit from the ME concept (section
2). The aim of the discussion is to describe a general set of requirements that
should be met by the ME platform and to obtain insight on the added value
(additional capabilities) introduced in applications by the ME concept. In
this respect, the Integrated Mobile Health Care Solutions (IMHCS) and
Intranet Business Applications (rnA) concepts (which are the focus in [1])
will be briefly discussed.
The next, main phase of the ME presentation is the description of the
overall architecture, the presentation of the network technologies in the
platform, and the discussion on the corresponding prototype terminal
architectures and functionality (sections 3 - 5). The last phase of the ME
platform presentation focuses on sample interactions among the elements of
Establishment ofMobile Extranets through Mobile IPv6 and GPRS: 311
the architecture (section 5). Finally, concluding remarks are made and
directions for future work are identified (section 7).
2. APPLICATION REQUIREMENTS
This section starts for the brief description of general application types
(IMHCS and IBA) that can benefit from the ME concept. The aim is to
identify general requirements, in terms of data transfer, mobility and
security, that the ME platform should meet.
The IMHCS concept derives when legacy health care solutions are
offered through an ME platform. The concept should enable the remote
monitoring and provision of day-to-day support to roaming patients (which
should maintain their normal lifestyle as much as possible), as well as the
detection of irregularities in the patients' condition, as far as these can be
automatically ascertained from the measured parameters. The most common
usage scenario envisages patients uploading measured data (in a transparent
manner to them), and (in return) receiving regular advice (or notifications of
critical events), either by the system itself, or by their doctors. Patient
records are stored in a database system in the Intranet, which must be
secured against intrusion of any kind, whilst simultaneously being highly
available to authorised users (e.g., doctors).
The combination of the IBA and the ME concepts enables users (e.g.,
salesmen, managers, customers, etc.) to realise remotely the transactions
offered in the Intranet limits. Sample aspects are the acquisition of
supporting material, the placement of orders, and the inspection of customer
profiles, product catalogues, price lists, order and account status, etc.
In the light of the aspects above the rest of this section summarises some
general requirements that should be met by the ME platform for properly
supporting the IMHCS and IBA application types.
Data transfer characteristics. Typically, the data types correspond to
markup text, with more demanding information flows (e.g., multimedia
presentations in the IBA case) being less likely. The information flows are
bursty, and should be provided at certain quality levels (rate, delay and
reliability requirements). These general characteristics push for the use of
flexible resource allocation schemes for achieving cost-efficiency.
Mobility and availability. Users should be able to initiate sessions from
"anywhere" within the covered geographical area (especially crucial in the
IMHCS case). Moreover, users should be reachable, in principle at all times,
MTheologou
for enabling automated reception of information (e.g., advice and alarms in

the IMHCS case, directions in the IBA case). These two general
requirements necessitate the support of the, so-called, micro-mobility and
macro-mobility features. The former feature refers to movements inside a
domain, which typically corresponds to a segment of a radio access (cellular
or other) network. Macro-mobility refers to a situation where the terminal is
moving between different radio access networks (or distinct segments of the
same radio access network). Offering these features in a transparent to the
applications manner (i.e., hiding the fact that the Intranet is accessed from
outside the domain) is an important aspect, be associated with the following
indicative advantages. First, the applications and their development are
simplified. Second, the overheads, associated with the necessary location
information exchange, are restricted to the underlying (network,
middleware) layers. Finally, the applications that reside on the MS are
enabled to conduct transactions as if they were in the Intranet, and therefore,
they can obtain customised treatment.
Security. Access control, authentication and confidentiality are key
requirements posed on the ME platform. More specifically, only hosts with
proper credentials should be granted access to the Intranet that is to the
application servers for which they are authorised. Likewise, non-forgeable
authentication credentials are needed on both the user and server side.
Finally, the possibility of eavesdropping at any point between user and
server is unacceptable. Encryption of frames is therefore required.
3. OVERALL ARCHITECTURE OF THE MOBILE

EXTRANETPLATFORM
This section presents the overall architecture of the ME platform that is

developed in response to the general requirements identified in the previous
section. The role and limitations of the technologies in the architecture will
be discussed in the next section.
Figure lea) depicts the architecture of the ME platform. As already
introduced the ME platform architecture should include (in addition to users,
their equipment, and the Intranet) a radio access technology, public Internet
segments, Mobile IP (MIP) and security mechanisms.
With respect to terminals, it is assumed that the roaming Intranet users
are equipped with Mobile Stations (MSs). The MS architecture will be
presented in more detail in section 5. In general, the MSs should be capable
of co-operating with the selected network technologies, which are presented

in the rest of this section.
The ME platform is realised over three network segments, namely, the
GPRS access network, the (public) Internet and the (private) Intranet.
(a)
Appliclllw" A.p~
TCP TO.
II'><
''''' I--- '''''
..,.,
-
'''''' '''''' - '''''' - II'W
I
-
. _-----
I ';,.':~~ TqfDP
MT,.SGSN GTP. TCPlVDP, L_LnJ
,,--" '''y IP L_J..uI L_UwI ",.,.,..
H~I"~~~~ "':;~~lt:;;:;;N I---

GPRS Air SGSN.GGSN
J_fo-~ -,."
I"remel
MS BSS SGSN GGSN Firewall AS
Routers
(b)
Figure 1. a) Architecture of the Mobile Extranet platform. (b) High level description of
protocol stacks in ME platform
The GPRS access networks comprises three types of elements. First, the
Base Station System (BSS), which is the access element of the GPRS
network terminating the radio protocols and connecting to the core elements.
Second, the Serving GPRS Support Node (SGSN), which is the node serving
the MS; when the MS attaches to the network, the SGSN creates a record
containing information regarding mobility and security. Third, the Gateway
GPRS Support Node (GGSN), which is the node terminating all GPRS
protocols. The GGSN support edge-router specific functions and interfaces
to several ISPs or corporate networks.
The Internet Public Domain, which in principle belongs to different ISPs,
interconnects the GPRS access segment with the Intranet.
The Intranet comprises three types of entities, which are related to the
applications, the Mobile IP functionality and the security functions. The
Application Server (AS) provides the information and services offered to an
Intranet user. The firewall (FW) provides security functions, such as access
314 K.Koutsopouios, NAiexiou, C.Konstantinopouiou, P.Demestichas,
MTheoiogou
control, authentication, content security, etc. The Home Agent (HAG) is the
(MIP-related) entity of the Intranet that hosts the Home Address (HAddr)
and the Care ofAddress (CAddr) for MSs that roam outside the Intranet (and
access the services through the ME platform).
Figure 1(b) depicts the protocol stacks supporting the ME platform. It
should be observed that the platform architecture does not pose any special
requirements to the network segments. On the contrary, non-widespread
technologies, like IPv6, and associated functionality are restricted to the end-
segments, namely, the MSs and the Intranet.
4. NETWORK AND MIDDLEWARE

TECHNOLOGIES
This section presents the rational for selecting the network technologies
in the platform, as well as the role and limitations of the technologies.
4.1 GPRS
GPRS enables operators to offer radio access to IP-based services, with

micro-mobility (e.g., through the cell reselection, routing area update
procedures [2]), on a wide area basis. GPRS follows flexible radio-resource
allocation schemes, which means that radio resources are reserved only for
the duration of a packet transfer, and are shared by all "active" terminals
within a cell. GPRS uses the same FDMAlTDMA structure as GSM.
Spectrum is divided into carrier frequencies. Roughly, on each carrier there
are channels that can be assigned to circuit switched calls or to GPRS. In the
latter case the channel is called Packet Data Channel (PDCH). GPRS users
share, and obtain for the duration of packet transfers, portions of the PDCHs
(radio resources).
Prior to any data exchange, an IP address is assigned to the GPRS MS,
and a QoS profile is agreed, via the so-called PDP (Protocol Data Packet)
Context Activation procedure. The IP address can be assigned by the GPRS
operator or an ISP. The current commercial versions of GPRS support IPv4
addresses, which in the general case will be dynamic (temporary) ones
allocated only for the duration of the data session.
In this respect, the advantages of GPRS can be summarised. Specifically,
the features described above yield that GPRS is a (already deployed)
technology offering wireless access to IP services, micro-mobility, as well as
Establishment a/Mobile Extranets through Mobile IPv6 and GPRS: 315
flexible resource reservation schemes (and therefore, cost-efficiency), and

QoS support (through the PDP context activation procedure). On the
contrary, an alternative to GPRS could be the High-Speed Circuit Switched
Data (HSCSD) concept. HSCSD is not as widespread as GPRS. It can offer
higher QoS, compromising the resource utilisation efficiency (and therefore,
the cost-efficiency).
Of course, GPRS does not cover all the mobility and security features
necessary from the ME viewpoint, but it can cooperate with other
technologies that do so. As already explained an important ME requirement
is providing micro-mobility and macro-mobility transparency to the
applications. This cannot be achieved through the use of a dynamic, not
globally routable IP address provided to the MSs. On the contrary, a light
and flexible approach, affecting only the two ends, is to apply MIP
functionality in a manner that is explained in the next sub-section.
4.2 IPv6 and Mobile IP
This subsection starts from the reasons for introducing IPv6, and its
integral MIP functionality, at some parts of the ME platform (specifically,
the MSs and the Intranet). This is explained by discussing on some
disadvantages of the approach of introducing the MIP functionality based on
IPv4. Then, some important elements of the protocol operation, especially,
in common Internet infrastructures, supporting IPv4, are provided.
The MIP functionality can be implemented on the basis of IPv4 (MIPv4)
or IPv6 (MIPv6). Typically, MIPv4 can require the presence of a Foreign
Agent for representing the MS when the available addresses in the visited
network are scarce. In that case, the Foreign Agent's address plays the role
of a globally valid CAddr for the MS. Therefore, the Foreign Agent's role is
to receive information on behalf of the MS and to direct it to the MS.
Nevertheless, a Foreign Agent is not foreseen in the current versions of the
GPRS standard, and also not associated to the Intranet. These features
suggest the investigation of the potentials of using MIPv6 in conjunction
with GPRS.
Inherent advantages of IPv6 are the large address space, and the support
for Mobile IP and security, as an integral part of the protocol suite.
Moreover, through the "6 to 4" mechanisms ([9], [10]), IPv6 enables an MS
to operate in networks that are based on IPv4. An IPv6-capable MS has a
IPv6 compatible HAddr (HAddrv6). In roaming mode the MS can discover
its network neighborhood, as well as capture the frequent changes of the
M. Theologou
point of attachment to the network ([6], [7], [8]). In each ofthese changes an
IPv4 compatible CAddr (CAddrv4) is obtained. This address is converted
into a corresponding IPv6 compatible CAddr (CAddrv6) and registered at
the HAG, through a Binding Update Request (BUReq) message (which is
also discussed in the next subsection). This registration enables the roaming
MS to be reachable and its applications to remain unaware of the network
layer changes, which are due to mobility.
Figure 2 describes in a general manner the communication between the
MS and an AS in the Intranet.
The application at the MS sends packets with source address the
HAddrv6 and destination address the AS address. At the IPv6IIPv4 level the
following occur. At a first step, the source address is replaced by the
CAddrv6 and the HAddrv6 is indicated as an extension header. At a second
step the IPv6 message is encapsulated in an IPv4 message. The MS is
configured to use as a "6 to 4" gateway the FW of its Intranet. Therefore, the
source address of the IPv4 message is the CAddrv4, while the destination
address is that of the FW. This means that the MS specifies that the end of
the IPv4 path is the firewall of its Intranet.
.A.~icQlimr
/h611P.4 JPo611J'.4
Figure 2. High-level internal structure of the RMS component
At the other end of the communication the reverse procedure occurs. All
the messages transferred inside the Intranet are pure IPv6 messages. That is,
the Intranet is seen as an IPv6 native network. If the endpoints (MS and AS)
have been granted the permission to communicate (this aspect is discussed in
the next subsection) the firewall circulates the original IPv6 content inside
the Intranet.
Likewise, when information has to be forwarded outside of the Intranet,
the firewall, which is also a "6 to 4" gateway for the Intranet hosts,
encapsulates these inside IPv4 packets that reach their destination through
the CAddrv4 address.
A potential problem of the selected approach is when the GPRS operators

do not provide globally routable dynamic IP addresses, which means that
they apply address translation filters. A mid-term solution for overcoming
this problem is the request for the assignment of globally routable addresses
to selected classes of users, like the ME users.
A set of encapsulations has been presented at this section. These
encapsulations have been imposed by the following main needs. The first is
the provision of micro-mobility and macro-mobility transparency. The
second is to avoid the requirement for the avoidance of having to pose
special requirements of non-widespread technologies in the network
segments.
4.3 Security
Security is perhaps the most important issue in the establishment ofMEs.

In our current implementation security relies on the IPsec mechanisms that
are an integral part of the IPv6 suite. As already stated, the security features
required are access control, authentication and confidentiality.
The mechanism for achieving access control and authentication is
specified in [11],[12]. In the ME platform it involves the MS, the firewall
and the HAG. Initially, the MS and the HAG attempt to exchange BUReq
and corresponding Binding Update Acknowledge (BUAck) messages. The
firewall allows a BUReq message originated by an MS to enter the Intranet
and reach the HAG. The HAG authenticates the MS, based on information
on the BUReq message, and replies with a BUAck message. The successful
exchange of the BUReq and BUAck messages indicates to the firewall that
the MS has been authenticated by the HAG, and therefore, should be
allowed to attempt to establish data sessions with ASs in the Intranet. The
approach adopted for the establishment of MEs calls for dynamically
configurable firewalls.
At a next stage, the privacy of sessions and the integrity of the
information exchanged therein can be ensured by the encryption functions of
IPsec that rely on security associations established between the MS and the
AS.
Our approach is open for working with future upgrades of the IPsec
protocol suite, as well as concepts, still under consideration in the area of
security functions [13].
MTheologou
5. TERMINALS
This section presents in more detail the architecture of the MS. Figure 3
depicts the architecture of the MS.
In brief, in its prototype version an MS can be split into the Measuring
Device (MD), which is necessary for IMHCS applications, the Terminal
Equipment (TE) and the Mobile Terminal (MT).
The TE is independent from the particular type of radio access network,
and can be a laptop PC or a Personal Digital Assistant (PDA). The MT is a
GPRS capable phone. It is envisaged that future, commercial versions of the
MS will have the TE, MT (and MD if required) as functional components of
a single hardware entity.
Figure 3(a) depicts information on the communication between the
components of the MS. The following can be noted regarding the TE - MT
and the TE - MD communication.
The TE communicates with the MT and the MD via serial connections.
The majority oflaptops and PDAs have only one built-in RS-232 port. In our
approach this port is used for the communication with the MD. In the laptop
case a second serial port card be added via the standard PC-Card expansion
slot, present in all laptops. This second port is dedicated to the
communication with the MT. Of course, an alternative is to use an Infrared
connection for the TE and MT communication. In the PDA case the second
port can be based on the, so-called, compact flash cards, present in most
devices.
s"rlQI Llllk
s"rlalLlllk I"ernlillal ['llllIJIII£'1I1 (Ilifrared or (,I'IlS
\h'llfjllrili~ Del 'lce tx ruion
(.\1/1)
In) siD/from TE Ilabi/,' 1"'11111101
porta'TE) I.llplfll'. pc. "f) I los~riaJ (.111)
porlalM1)
(a)
Application
TCP
lPv6
IPv4
Relay
PPP r-- PPP Air

Infer/ace
Serial Serial Protocols
Communication Commua
TE MT
(b)
Figure 3. (a) Prototype MS architecture. The MD is required for the IMHCS concept. (b)
Protocol stacks in the TE - MT arrangement.
The detailed description of the protocol stack of the MS is depicted in

(Figure 3(b)). The Point-to-Point Protocol (PPP) is used in the TE - MT
communication. The TE sees the MT as a modem.
6. INTERACTIONS BETWEEN THE ELEMENTS

OF THE ME PLATFORM
This section presents in more detail the interactions between the elements
of the ME platform that were introduced in the previous section. Figure 4
depicts these interactions. The process consists of four main phases.
MTheologou
Figure 4. Interactions of the elements of the ME platform
The first phase is targeted to the establishment of a communication

between the TE and the MT. This phase is similar to that followed between a
PC and a modem. It relies on the exchange of AT commands and the
configuration ofPPP (Point-to-Point Protocol) components.
Through the second phase the MS attaches to the GPRS network and
conducts the PDP context activation. An outcome of this phase is the
consolidation on the QoS profile and the assignment of a CAddrv4
(temporary IPv4, care of address) to the MS.
Through the third phase the MS aims at accessing the Intranet. Therefore,
the MS issues a BUReq message for informing the HAG that its HAddrv6
should be associated with the CAddrv6, which has been derived through the
CAddrv4. The firewall enables the BUReq message to reach the HAG. The
HAG authenticates the MS, registers the CAddrv6, and returns a BUAck
message. The BUAck signifies to the firewall that the MS should be allowed
to access the Intranet.
During the fourth phase the MS establishes data sessions with ASs. The
MS and the AS agree on security issues so as to guarantee confidentiality
and integrity.
It should be noted that parts of the second and third phase can be repeated
in case the IP address assigned to the MS changes.
7. CONCLUSIONS
MEs are seen as a natural evolution of the Internet and Intranet concepts.
The aim of this paper was to present an approach for the realisation of an
ME platform. Our approach was the following. Starting point was the
collection of general requirements posed on the platform by various
application types. Next, the overall platform architecture was presented and
details on the role of the comprised network (middleware) and terminal
technologies were introduced. In the network side the discussion focused on
GPRS and the IPv6 protocol suite, with its encompassed mobility and
security mechanisms. At the next stage a sample of the cooperation of the
platform elements was given.
The following can be identified as indicative issues for further work. The
first is the investigation of alternate technologies (e.g., in the area of
security) on which the ME platform can be based. The second is the detailed
comparison of the selected ME approach with respect to alternate concepts
such as virtual private networks (VPNs). The third is the expansion of the set
of ME software components for the purpose of better managing aspects like
the QoS levels. In this direction the possibility to negotiate on these features
with operators and ISPs will be supported. The provision of QoS levels on
this heterogeneous platform is a work area enabling the success, and
providing evidence on the viability, of the ME platform.
REFERENCES
1ST project MOEBIUS (Mobile Extranet-Based Integrated User Services) Web site,
www.ist-moebius.org, Jan. 2001
2 European Telecommunications Standardisation Institute (ETSI), "Digital cellular
telecommunication system (Phase 2+). General Packet Radio Service (GPRS). Service
description. Stage 2", GSM 03.60, v. 6.3.0, Apr. 1999
3 S.Deering, R.Hinden, "Internet Protocol version 6 (IPv6) specification", RFC 2460, Dec.
1998
4 C.Perkins, "IP mobility support", RFC 2002, Oct. 1996
5 S.Kent, R.Atkinson, "Security architecture for the Internet Protocol", RFC 2401, Nov.
1998
6 T.Narten, E.Nordmark, W.Simpson, "Neighbour discovery for IP version 6 (IPv6)", RFC
2461, Dec. 1998
7 S.Thomson, T.Narten, "IPv6 stateless address auto-configuration", RFC 2462, Dec. 1998
8 A.Conta, S.Deering, "Internet Control Message Protocol (ICMP) for the Internet Protocol
version 6 (IPv6) Specification", RFC 2463, December 1998
322 K.Koutsopoulos, N.Alexiou, C.Konstantinopoulou, P.Demestichas,
M. Theologou
9 M.Crawford, "Transmission ofIPv6 Packets over Ethernet Networks", RFC 2464,

December 1998
10 B.Carpenter, K.Moore, "Connection ofIPv6 domains via IPv4 clouds", RFC 3056, Feb.
2001
11 S.Kent, R.Atkinson, "IP authentication header", RFC 2402, Nov. 1998
12 S.Kent, R.Atkinson, "IP encapsulating security payload", RFC 2406, Nov. 1998
13 L.Mathy, C.Edwards, D.Hutchison, "The Internet: A global telecommunications
solution?", IEEE Network, Vol. 14, No.4, July/Aug. 2000
14 D.Johnson, C.Perkins, "Mobility support in IPv6", Work in progress(draft-ietf-mobileip-
ipv6-15), July 2001
IP Traffic Control on UMTS Terminal Equipment
Manuel Ricardo, Rui Soares, Jaime Dias and Jose Ruela

FEUP - Fac. Eng. Univ. Porto. Rua Dr. Roberto Frias, 4200-465 Porto, Portugal
INESC Porto. Pra9a da Republica 93 ric, 4050-497 Porto, Portugal
Abstract: The paper presents the architecture of an UMTS terminal equipment optimised
for IP based communications and describes the traffic control mechanisms
required for supporting emerging 3G services.
Key words: IP, UMTS, Quality of Service, Traffic Control, Scheduling, PDP Context
1. INTRODUCTION
This paper presents some results of the work carried out in the European
1ST project ARROWS (Advanced Radio Resource Management for
Wireless Services) [1]. This project aims at providing advanced Radio
Resource Management (RRM) and Quality of Service (QoS) management
solutions for the support of integrated services within the context of
Universal Terrestrial Radio Access (UTRA). The project addresses packet
access, asymmetrical traffic and multimedia services, all based on IP. The
main objectives of ARROWS are: 1) to define and simulate RRM algorithms
for an efficient use of the radio resources; 2) to provide QoS bearer services
for packet switched flows at the UTRA; 3) to demonstrate the benefits of the
proposed algorithms and procedures by means of an IP based multimedia
testbed [2].
This paper is related to the third objective. The ARROWS multimedia
testbed consists of the following functional blocks: 1) an all-IP based UMTS
terminal; 2) an UTRAN (Universal Terrestrial Radio Access Network)
emulator, implementing the UMTS (Universal Mobile Telecommunications
System) radio interface and the relevant RNC (Radio Network Control)
functions; 3) a gateway implementing functions traditionally assigned to
323
324 Manuel Ricardo, Rui Soares, Jaime Dias and Jose Ruela
SSGN (Serving GPRS Support Node) and GGSN (Gateway GPRS Support
Node); 4) a backbone IP network and its associated routers; 5) a server.
Only the first functional block is addressed - the all-IP UMTS terminal,
which supports multimedia applications and is implemented in a LINUX
based PC. Its architecture, the selected applications, the classification,
scheduling and shaping of IP flows as well as the interface with the UMTS
network interface, assumed to implement the UMTS Non-Access Stratum
(NAS) functions, are described.
The paper is organised in seven parts. Section 2 introduces the selected
multimedia applications that will run on the terminal. Section 3 discusses the
UMTS terminal architecture, which is biased towards the compatibility
between the IP and UMTS worlds from flow and QoS points of view.
Section 4 reviews the IP QoS facilities currently available in Linux and
presents a strategy for using them in an all-IP UMTS terminal. Section 5
describes the mechanism proposed for controlling the traffic. Section 6 gives
experimental results that validate this strategy. Finally, Section 7 presents
the main conclusions.
2. MULTIMEDIA APPLICATIONS
Four traffic classes were identified in UMTS: conversational, streaming,

interactive and background [3]. One main aspect that distinguishes these
classes is how delay sensitive the traffic is. The conversational class is meant
for very delay-sensitive traffic, while the background class is delay tolerant.
Moreover, when comparing the conversational and streaming classes, the former
requires a tight bound on delay and stringent control of the delay jitter. This is
mostly due to the fact that conversational traffic is symmetric, while streaming is
highly asymmetric and therefore it is possible to use buffers for smoothing out jitter.
In conversational services, this would increase the delay acceptable for a natural
human conversation, turning the communication awkward.
One application representative of each UMTS traffic class was chosen for
the ARROWS testbed: Videoconference (conversational), Video streaming
(streaming), Web browsing (interactive) and Email (background). All
applications were required to satisfy three characteristics: 1) be widely used;
2) be open source, so that extensions to IPv6 or that incorporate new QoS
features, for instance, could be easy; 3) have port for LINUX.
For Videoconference, VIC and RAT, the well-known video and audio
conferencing tools, were selected as departing applications. Although
designed for multicast environments, they are configured in ARROWS as
point-to-point (unicast). Both applications rely on the Real Time Transport

Protocol (RTP).
VIC supports the H.261 and H.263 video codecs. RAT supports various
codecs, such as G.711 PCM (64 kbitls), G.726 ADPCM (16-40 kbitls), LPC
(5.6 kbitls) and GSM (13.2 kbitls). When used for an audio-video telephony
call, these applications generate two real-time and bi-directional IP flows
(audio and video) that have to be adequately transported through the UMTS
transport services, that is, Radio Access Bearers (RAB) and Packet Data
Protocol (PDP) contexts, as discussed in Section 3.
For Video streaming it was decided to use a coding scheme that generates two
streams with base and enhancement information, respectively. The application
that will be used is the one developed by the MPEG4IP group, which also
deploys video over the RTPIUDP/IP protocol stack. When playing a stream,
two unidirectional and real-time IP flows have to be transported over the
UMTS network.
Web browsing at the terminal requires a browser, that is, an HTTP client.
Email, at the mobile terminal, requires both POP3 and IMAP clients. The
three application protocols (HTTP, POP3 and IMAP) use the TCP/IP stack.
No real-time requirements are envisaged for the IP flows they generate.
Mozilla is being used as departing point for these applications.
3. TERMINAL ARCHITECTURE
The proposed UMTS terminal architecture that supports IP based

services is shown in "Figure 1".
3.1 Functional blocks
Forward includes the IP look-up routing tables and the encapsulation of

transport level segments in IP datagrams.
Classifier filters the packets and places them in different queues
according to their characteristics. TCP/UDP and IP header fields, such as source
port or source IP address, can be used as criteria to classify the packets.
Shaper, simply said, consists of a queue for an IP flow. A queue has
properties associated with it, such as bandwidth. A flow must be shaped so that
it does not violate the QoS previously negotiated for the associated PDP Context (in
the NAS Module).
Mapper is responsible for negotiating the activation, modification and
deactivation of PDP contexts, passing to the NAS Module the desired QoS
parameters. The mapping between RSVP QoS parameters and PDP context
QoS parameters is performed by this block.
RSVP implements RSVP (Reservation Protocol) [4] that, in ARROWS, is
used to guarantee end-to-end QoS to IP flows that traverse both the UMTS
and the IP backbone networks [5]. It can, in some circumstances, be avoided.
NAS module implements Non-Access Stratum functions, such as session
management and mobility management. It consists of two planes. On the
user plane, the module is offered as a standard LINUX network interface
(umtO, in "Figure I") and is able to exchange datagrams with the IP layer.
On the control plane, the NAS module is offered as a character device driver
(ldev/nasO, in "Figure 1"), through which messages for establishing and
terminating RABs are exchanged.
Applications
TCPIUDP
I RSVP
I
Forward I
I
I
Classifier I
I
Scheduler
~ I Mapper J
umtOI IdevlnasO
I NASModule
I
Figure 1 - UMTS Terminal Architecture
3.2 PDP context
A GPRS (Generic Packet Radio Service) subscription consists of one (or

more) PDP addresses that, in the case of "Figure I", will be the IP address
associated to the umtO interface. Each PDP address, in GPRS, is described
by one or more PDP contexts [6]. Each PDP context is associated to aRAB.
When more than one PDP context exists, the other PDP contexts must have a
TFT (Traffic Flow Template) associated. A TFT consists of up to eight
packet filters. Each filter contains a valid combination of the following
attributes: Source Address and Subnet Mask, Protocol Number (IPv4) /

NextHeader (IPv6) , Destination Port Range, Source Port Range, IPSec
Security Parameter Index (SPI), Type of Service (IPv4) / Traffic Class
(IPv6) and Mask, Flow Label (IPv6). A PDP context and a RAB, due to their
one to one relationship, are used interchangeably in the paper.
The mobile station should be able to support more than one PDP context
simultaneously and to forward IP packets into the appropriate RAB. This
justifies the use of the Classifier and the Shaper in "Figure 1". Therefore,
more than one PDP context may exist, each with different QoS parameters,
to which packets will be forwarded depending on the class of the traffic they
belong to. In videoconference, for instance, image and voice are carried as
two IP flows. These flows have different QoS requirements and thus may be
mapped to separate PDP contexts; synchronisation between the flows is
handled at the application level.
RSVP messages are themselves carried on IP datagrams. A RAB for best
effort traffic can be used to transport these messages. This RAB can also be
used to transport the IP flows for the Web browsing and Email applications.
4. IP TRAFFIC CONTROL IN LINUX
LINUX kernels have been growing to include a number of advanced

networking features such as firewalls, QoS and tunnelling. The QoS support,
available since kernel 2.1.90, provides a number of features. The working
principle adopted in recent kernels (e.g., 2.4.4) is shown in "Figure 2".
packet in
Output traffic control

Queue
packet out
Figure 2 - LINUX traffic control

The Input De-multiplexing module examines an incoming packet and

determines if it is addressed to the local node. If so, it is sent to higher layers
(TCPIUDP block) for processing. Otherwise, it is passed to the Forwarding
block. This block looks up the routing table and determines the next hop for
the packet. The packet is then placed in a queue kept for the device (e.g.,
ethO or umtO). Traffic control is performed on this Output Queue, just before
the packet is sent to the network interface. The traffic control in LINUX
consists of three building blocks: Queuing Discipline, Class and Filter.
A Queuing Discipline is a framework used to describe a policy for
scheduling output packets. It is usually associated to a network interface
such as ethO or, in the case of "Figure 1", umtO. LINUX provides several
disciplines for scheduling packets such as FIFO (First In First Out), TBF
(Token Bucket Flow), CBQ (Class Based Queuing), RED (Random Early
Detection) and TEQL (True Link Equalizer). A Queuing Discipline may
have a complex structure; a set of Classes and Filters may be associated to a
root Queuing Discipline where Filters are used to assign packets to Classes.
In this case, a Class must have associated a new Queuing Discipline that,
in tum, may have more Classes and Filters. This principle allows the
combination of Queuing Disciplines, Classes and Filters in an arbitrary
hierarchical structure. Each network interface may own one of these
complex structures.
Filters are then installed on Queuing Disciplines to direct packets to
Classes on that Queuing Discipline. The following filters may be used in
LINUX: u32, rsvp, jW, route and tcindex. The u32 generic filter allows
classification of packets based on header fields, such as IPv4, IPv6, TCP or
UDP. The rsvp filter allows the classification of packets based on the
parameters that define an RSVP flow such as the IP destination address and
either the port or the Flow Label.
The traffic control may be configured in user space using the tc command
available in the iproute2 package. A good description of this command as
well as on scheduling, queuing disciplines, filters and traffic control in
general may be found in [7], [8].
5. TRAFFIC CONTROL ON THE UMTS

TERMINAL
After a PDP context has been negotiated and the associated RAB
established, the terminal may start communicating. However, the terminal
may have more than one PDP context activated and more than one RAB
established, each with its own QoS parameters. It is, therefore, necessary to
direct the packets to the proper RAE, schedule the packets according to their
priorities and shape the traffic so that the flow sent to a RAE is compliant
with the QoS previously negotiated for that RAE.
5.1 Scheduling
Scheduling is usually required when there is the need to share a link with
limited bandwidth. Flows with higher priorities must be scheduled first,
taking care that flows with lower priorities do not starve. Thus, the concepts
of sharing and priority hold when dealing with scheduling. A flow with high
priority may require less bandwidth than another flow with lower priority.
Sharing is about bandwidth, priority about delay and jitter.
The services introduced in Section 2 can be ordered in ascending priority
as Email, Web Browsing, Video streaming and Videoconference. However,
a Video streaming session may require a larger bandwidth than a pure voice
seSSIOn.
The CBQ (Class Based Queuing) discipline [9] can be used to solve the
priority issue. This discipline is based on statistical scheduling and a
hierarchical tree of traffic classes. When a packet is received, it is classified
and associated to a leaf class. It is possible to associate bandwidth and a
priority to each class. The CBQ queuing discipline is delivered with two
schedulers: generic and link sharing. The generic scheduler must guarantee a
low delay to real time flows. The link sharing scheduler tries to avoid that
real time flows monopolize the use of the link.
5.2 Shaping
The rate of a flow can be regulated using shaping techniques. In this case,
the traffic passed to a RAE needs to conform to the bandwidth previously
negotiated for that RAE. The use of a CBQ class for this purpose is not
adequate, since none of its schedulers addresses this problem. Better results
can be achieved if a TBF (Token Bucket Flow) queuing discipline [10] is
associated to each leaf class.
The TBF consists of a buffer (bucket), filled with virtual pieces of
information (tokens) at a specific constant rate (token rate). An important
parameter of the bucket is its size, that is, the number of tokens it can store.
Each token in the bucket lets one incoming data octet to be sent out of the
queue and is then deleted from the bucket.
Associating this algorithm with the two flows (token and data) gives
three possible scenarios: 1) the data arrives into TBF at a rate equal to the
token rate, which means that each incoming packet has a matching token and
passes the queue without delay. 2) the data arrives into TBF at a rate lower
than the token rate and therefore only some tokens are deleted when data
packets are sent out. Tokens accumulate up to the bucket size and can be
used to send data above the token rate, if this situation arises. 3) the data
arrives into TBF at a rate higher than the token rate. In this case, incoming
data can be sent out immediately, while the token bucket is not empty. The
accumulation of tokens allows short bursts of data to be passed without
delay and loss, but any lasting overload will cause packets to be constantly
dropped (or delayed).
In any case, the average data rate is bounded by the token rate.
5.3 Proposed configuration
A generic traffic control configuration proposed for an IP based UMTS

terminal is shown in "Figure 3". Scheduling and shaping of the flows are
implemented with CBQ and TBF queuing disciplines, respectively. The
location of the filters is also shown. For each RAB, one leaf class on the
CBQ queuing discipline (e.g., 1: 11) is created. Each class has one TBF
queuing discipline associated instead of the generic one installed by default.
For the sake of generality and to prove the flexibility ofthe solution, four
leaf classes are drawn. To support the services presented in Section 2, a finer
assignment is required. Videoconference requires two classes, corresponding
to two RABs and serving two flows, one for audio and the other for video.
Video streaming also requires two classes, for the base and enhancement
flows. The other flows, corresponding to Web Browsing, Email and generic
signalling, such as RSVP, may be assigned to a fifth class which, in tum, can
be mapped to the primary PDP context. No values are presented for
configuring the buckets (token rate and bucket size) since they will depend
mainly on the cost of the radio channels. However, the proposed solution is
flexible enough to fully support dynamic configuration of these values.
Once the packet is sent to the network interface (NAS module, umtO)
how can it know to which RAB the packet belongs to? The solution
proposed is to use the Flow Label, in IPv6 or the Type of Service (ToS) in
IPv4 to distinguish the RABs. The TFT associated to a RAB can be given a
list of ToS or Flow Label values that belong to each RAB. In this case a
given ToS or Flow Label would always belong to the same RAB.
I;~h.d·--- -----1
i r····;~·~······l
I I
___________.
filters
r------- ___ ~·~··~·~·r~·~·~i

:---r------r~?-~!~Si-~:!----r-- 1
I I
r--~--'r--~--'r--~--'r--~--'
I II II II
: CBQ :: CBQ :: CBQ :: CBQ
: Class :: Class : : Class :: Class
I II II II
L_!:.1 ~_1L _!:~~ _lL _!:.~~_1 L_!: 1_~_ ~
TBF TBF TBF TBF

QDis QDis QDis Qdis
11: 12: 13: 14:
Figure 3 - UMTS terminal traffic control
5.4 Configuration Example
The traffic control may be configured either using netlink sockets, if it is

an application configuring it, or with the use of Ic command from iproute2
package. An example of a script used to configure a hierarchy similar to
"Figure 3", but for device ethl, is shown below.
Only two classes are configured, with token rates 64 and 32 kbitls,
respectively. The classes have the same priority and are configured for very
small IP datagrams - 72 octets, on average. This parameter is required to
calculate the variables of the CBQ algorithm. The bucket size is 1500 octets.
Finally, two u32 filters are installed in the root CBQ queuing discipline.
Classification is based on the ToS value - packets with ToS values of Ox20
and OxOO are classified into the 64 and the 32 kbitls classes, respectively.
#!/binlsh
case "$1" in
start)
tc qdisc add dey ethl root handle 1: cbq bandwidth 10Mbit allot 9200 cell 16
avpkt 72 mpu 64
tc class add dey eth1 parent 1: classid 1:1 cbq bandwidth 10Mbit rate 10Mbit
avpkt 72 prio 2 allot 9200 bounded
tc class add dey ethl parent 1: 1 classid 1: 11 cbq bandwidth 10Mbit rate 64kbit
avpkt 72 prio 2 allot 9200
tc class add dey eth1 parent 1: 1 classid 1: 12 cbq bandwidth lOMbit rate 32kbit
avpkt 72 prio 2 allot 9200
tc qdisc add dey eth1 parent 1: 11 handle 11: tbflimit 20k rate 64kbit burst 1500
mtu 1500
tc qdisc add dey ethl parent 1: 12 handle 12: tbflimit 10k rate 32kbit burst 1500
mtu 1500
tc filter add dey eth1 parent 1: protocol ip prio 5 handle 1: u32 divisor 1
tc filter add dey ethl parent 1: prio 5 u32 match ip tos OxOO Oxffflowid 1: 12
tc filter add dey eth1 parent 1: prio 5 u32 match ip t05 Ox20 Oxff flowid 1: 11
stop)
tc filter del dey ethl parent 1: prio 5
tc class del dey eth1 classid 1: 12
tc class del dey ethl classid 1: 11
tc class del dey eth1 classid 1: 1
tc qdisc del dey ethl root
..
"
6. TOKEN BUCKET FLOW PERFORMANCE
A set of experiments using the Token Bucket Flow queuing discipline

was carried out with the purpose of evaluating the differences between the
ideal shaping and the one offered in practice by the LINUX kernel. The
physical configuration used for the tests consisted of two PCs directly
connected on a 10 Mbitls Ethernet. One PC generated traffic and the other
received it and gathered statistics about the received traffic. The tool used for
generating traffic was rude 0.61 and, for gathering statistics, its counterpart
crude 0.61 was used. This tool allows creating UDP over IPv4 CBR flows
with a good accuracy. To obtain the highest accuracy in the generated traffic
and measurements, the processes were run with maximum priority. The
duration of the tests was always 60 seconds.
Application data is encapsulated with UDP and IP headers, which
contribute with 28 octets of overhead per IP packet. Therefore, it is
necessary to relate the data rate generated by the application with the raw IP
rate. Since it is the latter that is controlled by the TBF discipline, it will be
used when comparing the expected and the detected (measured) rates at the
destination (the same relation would hold at application level).
In the tables of results, the labels have the following means: 1) Generated
Data Rate represents a constant bit rate generated by the rude tool (data
only); 2) Expected IP Rate is the expected raw IP rate at the detection side.
This value is limited by the Token Bucket rate; 3) Detected IP Rate is the
raw IP rate measured at the destination. The latter must be compared with
the Expected IP Rate value and should not exceed it.
For a Token Bucket rate of 64 kbit/s and a datagram of 100 octets (data
size -72 octets), the results are shown in "Table I".
Table J - Packet size 100 octets, Token rate 64 kbitls
Generated Expected Detected

Data Rate IP Rate IP Rate
28800 40000 40022
46080 64000 58967
57600 64000 58967
The critical data rate that corresponds to an IP rate equal to the Token
Bucket rate is 46080 bit/so When the data rate is below the critical value
(first line), the detected rate is similar to the expected one. When the data
rate is above that value (third line), the TBF controller is active; in this case
the detected rate is below the expected one. The error is -7.9%. The second
line, which corresponds to the critical value, exhibits the same behaviour.
The datagram size was then increased to 1228 octets (data size - 1200
octets) maintaining the same Token Bucket rate. The results are presented in
"Table 2".
Table 2 - Packet size 1228 octets; Token rate 64 kbitls
Generated Expected Detected

Data Rate IP Rate IP Rate
28800 29472 29644
57600 58944 59058
76800 64 000 65403
96000 64000 65395
The critical data rate is now 62540 bitls, which means that the last two
lines correspond to an effective rate control.
The detected rates are now above the expected ones, but the errors are
small. For example, considering the last line, the error is 2.2%. It seems that
the packet size does have influence on the accuracy of the Token Bucket
Flow and that there should be a packet size that optimises its performance.
The impact of Token Bucket rates on the performance of the TBF was
evaluated, as well. Packet sizes of 100 and 1228 octets and Token Bucket
rates of 8 and 128 kbitls were tested. The results are shown in "Table 3". In
all cases the data rates are above the critical values and therefore the
expected IP rates are the corresponding Token Bucket rates.
Table 3 - Packet size 100, 1228 octets; Token rate 8, 128 kbitls
Generated Packet Expected Detected

Data Rate Size IP Rate IP Rate
8000 100 8000 7477
8000 1 228 8000 8301
128000 100 128000 117778
128000 1228 128000 130 610
Although absolute errors increase with the token rate, relative errors are
similar. For a packet size of 100 octets, errors of -6.5%, -7.8 % and -8.0%
were obtained for data rates of 8, 64 and 128 kbit/s, respectively. For a
packet size of 1228 octets, errors of 3.8%,2.2% and 2.0% were obtained for
the same data rates. While with small packets, increasing the token rate
increases the relative error, with larger packets, increasing the token rate
decreases the relative error. For confirmation, another test, with a packet size
of 1228 octets, was carried out with a Token Bucket rate of 1 Mbitls. The
detected IP rate for this configuration was 1 013 344 bit/s, which means a
relative error of 1.33%, thus confirming the previous results.
Only one flow at a time was used in the previous tests. In a final test, two
flows classified into two different classes, one for 64 kbitls limit and the
other for a 32 kbitls limit were used. The results are shown in "Table 4".
Table 4 - Two simultaneous flows:

Packet size 100 octets; Token rates 32 and 64 kbitls
Flow Token Detected

Bucket Rate IP Rate
1 32000 29567
2 64000 58967
The results for flow 2 are the same as those obtained on the first test. The
existence of more than one Queuing Discipline/Class does not seem to affect
the result. For flow 1, the relative error is -7,6 %, lower than for 64 kbitls
and higher than for 8 kbitls, in conformance with what has been previously
said about the effect of the token rate.
7. CONCLUSIONS
This paper describes the architecture of an UMTS terminal that supports

IP based services and is implemented on a LINUX PC. The terminal is
required to support 4 services (Videoconference, Video streaming, Web
Browsing and Email) that are representative of UMTS traffic classes.
Videoconference and Video streaming use the RTPIUDPIIP protocol stack
and each service generates two real-time flows with QoS requirements. The
other services are less QoS demanding and use the traditional TCPIIP stack.
The UMTS protocol stack, implementing the Non-Access Stratum, is
expected to be available as a module with two interfaces: the umtO network
interface for IP packets (user plane) and a character device driver used to
establish an terminate RABs (control plane).
The mapping ofIPIRSVP flows into UMTS PDP contexts and the control
plane in general are not considered. The paper mainly addresses the problem
of classifying and shaping the packets passing from the IP layer to the
UMTS interface so that the following goals are fulfilled: 1) packets are
delivered in time and the application QoS requirements are satisfied but, on
the other hand, 2) these flows do not violate the QoS contracts previously
established between IP and NAS UMTS.
The solution proposed for this problem relies on the LINUX IP traffic
control capabilities for classifying and shaping the flows - a CBQ queuing
discipline is defined as a tree of classes, one for each RABIPDP Context.
Each leaf class is configured as a TBF queuing discipline that shapes the
flow.
The tests carried out to validate the proposed solution show that errors
between the ideal and the real shaping vary with packet size and token
bucket rate in a range from 0 to 8%. These errors, however, seem to be
constant and predictable. By biasing the token bucket rate, accurate results
seem to be possible.
The main advantage of this solution is to rely on well known and widely
available scheduling and shaping functions at IP level.
8. ACKNOWLEDGEMENTS
The authors wish to thank the support given by the 1ST research
programme of the European Union and their partners within the ARROWS
consortium: Universitat Politecnica de Catalunya, University of Limerick,
Telef6nica I+D and Telecom Italia Lab.
9. REFERENCES
[1] 1ST ARROWS project, http://www.arrows-ist.upc.es
[2] N.P. Magnani, F. Casadevall, A. Gelonch, S. McGrath, M. Ricardo, I. Berberana,
"Overview of the ARROWS Project", 1ST Mobile Communications Summit 2001,
Barcelona, Spain, September 9-12, 2001.
[3] 3GPP TS 23.107 V5.0.0, "QoS Concept and Architecture", April 2001.
[4] Paul White, "RSVP and Integrated Services in the Internet: A Tutorial", IEEE Comm.
Magazine, Vol. 35, No.5, May 1997, pp. 100-106.
[5] 3GPP TS 23.207 V5.0.0, "End-to-end QoS Concept and Architecture", June 2001.
[6] 3GPP TS 23.060 V5.0.0, "General Packet Radio Service (GPRS); Service description;
Stage 2", January 2002.
[7] iproute2+tc notes, http://snafu.freedom.org/linux2.2Iiproute-notes.html
[8] Linux Iproute2, http://defiant.coinet.comliproute2/
[9] Sally Floyd, "Notes on CBQ and Guaranteed Service", July1995.
[10] K. Dovrolis, M. Vedam, P. Ramanathan, "The Selection of the Token Bucket
Parameters in the IETF Guaranteed Service Class", Technical Report, Department of
ECE, University of Wisconsin-Madison, November 1997.
An Approach for Managing Networks and Services in
a Diversified Radio Environment!
P.Demestichas(I), G.Vivier(2), G.Martinez(2), F.Galliano(3), L.Papadopoulou(I),

V.Stavroulaki(1), and M.Theologou(1)
(1) National Technical University ofAthens, Electrical and Computer Engineering Department,
Computer Science Division, Telecommunications Laboratory, 9 Heroon Polytechneiou Street,
Zographou 15773, Athens, GREECE. Tel: + 30 I 7721493, Fax: + 30 I 7722534, E-mail:
louisa@telecom.ntua.gr
(2) Motorola Labs, Paris, FRANCE
(3) Motorola Technology Center ofItaly, Torino, ITALY
Abstract: This paper builds on the assumption that in the Fourth Generation (4G)
wireless system context UMTS, MBS and Digital Broadcasting Systems
(DBS) can be three co-operating components that enable wireless access to IP-
based services. Managing the resources of this powerful infrastructure in an
aggregate manner and multi-operator scenario is a complex task. This paper
presents an approach to the overall UMTS, MBS and DBS network and
service management problem. Key points addressed are the development of
open interfaces with Service Provider mechanisms and the heterogeneous
managed infrastructure, performance monitoring, joint optimisation of the
UMTS, MBS and DBS resources in accordance with the service provider
requests and the changing with time environment conditions. The architecture
of a corresponding UMTS, MBS and DBS network and service management
platform is presented, in terms of logical blocks, their functionality,
implementation options and validation approach.
Key words: UMTS, MBS, Digital Broadcasting Systems, Network and Service
Management Platforms
I This work is partially funded by the Commission of the European Communities, under the
Fifth Framework Program, within the 1ST (Information Societies Technology) project
MONASIDRE (IST-2000-26144: Management of Networks and Services in a Diversified
Radio Environment).
337
X. Lagrance and B. labbari (eds.).
Multiaccess. Mobility and Teletraffic for Wireless Communications. Volume 6. 337-352.
338 P.Demestichas, G. Vivier, G.Martinez, F. Galliano, L.Papadopoulou,
v.Stavroulaki, M. Theologou
1. INTRODUCTION
Wireless systems continue to attract immense research and development

effort [1,2,3], especially in the following areas. First, the enhancement of
legacy infrastructures [4], mainly through the General Packet Radio Service
(GPRS), the Enhanced Data rates for GSM Evolution (EDGE), and the
High-Speed Circuit Switched Data (HSCSD) concepts. Second, the gradual
introduction of third generation systems like the Universal Mobile
Telecommunications System (UMTS) [5] and the development of the IMT-
2000 framework [6]. Third, the standardisation, development and
introduction of Mobile Broadband Systems (MBS), which support radio
access to broadband services, with limited mobility; a pertinent promising
example is the HIPERLAN (High Performance LAN) initiative [3]. Fourth,
the advent of Digital Broadcasting Systems (DBS) like the Digital Video
Broadcasting (DVB) and the Digital Audio Broadcasting (DAB) initiatives
[7].
In addition to the above, European sponsored research already addresses
the joint utilization of heterogeneous access networks, in order to improve
service quality, access, and availability [8,9,10]. For instance, in [8] cellular
and broadcast networks are combined to deliver interactive multimedia
services to the car. In [9] the combination ofUMTS and BRAN is envisaged
as a solution for providing users with high capacity in hot spots and [10]
aims at designing a system architecture where heterogeneous access
networks are integrated and can dynamically have their spectrum assigned to
match user demand.
In the light of the aspects above, it can be anticipated that the wireless
world of the (near) future will comprise (among others) the UMTS, MBS
and DBS technologies, while the fixed network will be (almost exclusively)
IP-based. Moreover, a recent trend in the direction of defining the Fourth
Generation (4G) wireless systems' era is to assume that UMTS, MBS and
DBS will be three co-operating wireless access components. In other words,
these systems can be seen as parts of a wider infrastructure through which
their operators will be enabled to provide users and service providers with
alternatives regarding the efficient (in terms of cost and QoS) wireless access
to IP-based services. The motivation for this is twofold. On the one hand, for
a network provider it can be preferable, instead of rejecting traffic, to serve it
through an alternate radio technology (and an affiliated network provider).
On the other hand, for users and service providers the supporting radio
technology can be irrelevant as long as the terminal capabilities allow multi-
mode operation (in accordance with the software radio concepts [11,12]),
and cost and quality criteria are met.
An Approach for Managing Networks and Services in a Diversified 339
Radio Environment
Servic...... Fiudne....rk
do_
IPN~
Paper scope
---.: - - - - - - -...----~/
Figure 1. A view of the future wireless-access world. Diverse transmission technologies offer
wireless access to IP-based services. This paper presents a corresponding network and service
management system
The exploitation of the powerful UMTS, MBS and DBS infrastructure

poses new requirements for mechanisms that promote cost-efficiency and
therefore competition. In this respect, there are two concepts that will assist
in the commercial success of UMTS, MBS and DBS. The first is the use of
advanced network-planning methodologies, which will assist in the cost-
effective deployment (introduction) of these systems. Pertinent work
samples conducted in the recent past can be found in [13,14,15,16]. The
second, and currently more challenging, is the implementation of
sophisticated network and service management platforms, which will enable
the cost-efficient operation of the systems (once they are deployed). This is
the main topic in this paper as also depicted in Figure 1. Specifically, this
paper presents a UMTS, MBS and DBS network and service management
system capable of:
Monitoring and analysing the statistical performance and QoS levels
provided by the network elements (segments) of the managed
infrastructure, and the associated requirements originating from the
service area (environment conditions, e.g., traffic load, mobility levels,
etc.).
Inter-working with service provider mechanisms, so as to allow service
providers to dynamically request the reservation (release, etc.) of network
resources.
Performing dynamic reconfigurations of the overall managed UMTS,
MBS and DBS infrastructure, as a result of resource management
V. Stavroulaki, M. Theologou
strategies, for handling new environment conditions and service provider

requests in a cost-efficient manner.
It should be clarified that resource management strategies are seen as
means for improving statistical performance indicators (e.g., blocking or
outage probability, average delay, etc.). Schemes that are targeted to the
accommodation of individual events (e.g., channel requests corresponding to
a new call or handover) are seen as part of the control domain. The
management system acts in a manner that is complementary to the control
domain, by incorporating (management) strategies that will be targeted to the
activation (cessation, etc.) of resources and resource control schemes within
time periods of the system operation. Nevertheless, no constraint is imposed
on the time-scale on which management actions may be applied. Hence,
these may be sought in the short, or medium, term depending on the specific
problem that is addressed. Both automatic (in accordance with the self-
organised system paradigm) and human-driven reconfiguration of the
network is envisaged. Emphasis is also placed to the realisation of an open,
distributed, component-based management architecture, by means of
CORBA (Common Object Request Broker Architecture) compliant
middleware platforms [17,18].
The rest of this paper is organised as follows. Section 2 presents the
overall management system architecture. Specifically, the separation into
three logical entities is introduced. These entities are presented in sections 3,
4 and 5. Section 6 discusses implementation and validation aspects and
section 7 includes concluding remarks and future work.
2. HIGH LEVEL ARCHITECTURE OF THE

MANAGEMENT SYSTEM
2.1 Structure
Figure 2 depicts the high level structure of the UMTS, MBS and DBS
management system. The management system is split in three logical blocks.
The first logical block is called managed system performance Monitoring
and Assessment and Service Provider mechanism Interworking (MASPI). In
essence, its role can be summarised in the following two general points. The
first is to capture the (changing with time) conditions that originate from the
environment (service area) of the managed UMTS, MBS and DBS
infrastructure; this is accomplished by monitoring and assessing the relevant
network and service level performance of the managed network elements
and segments. The second is to interwork with service provider mechanisms,
Radio Environment
so as to allow service providers to request the reservation of resources

(establishment of virtual networks) over the managed network infrastructure.
Virtual networks are seen as the realisation of contracts that the managed
system should maintain with service providers.
As will be discussed in a later section, work items addressed for enabling
the accomplishment of the MASPI role are the following. First the definition
of open interfaces with the service provider mechanisms. Second, the
development of components for the processing, establishment and
maintenance of contracts (or Service Level Agreements - SLAs) with service
providers. Third, the definition of common interfaces among the
management system and the (highly heterogeneous) elements or segments of
the managed system. Fourth, the integration of sophisticated monitoring
procedures.
Managed System
Monitoring and Assessment & Resource Management
Service Provider Mechanism Strategies (RMS)
Interworking (MASPI)
UMTS, MBS and DBS

Network & Environment Simulator
(UMD-NES)
UMTS, MBS and DBS Network and ServIce Management System
Figure 2. High-level architecture of the UMTS, MBS and DBS management system
The second logical block of the management system is in charge of the

Resource Management Strategies (RMS). In general terms, its role is to
apply resource management strategies, so as to dynamically find and impose
the appropriate UMTS, MBS, and DBS infrastructure reconfigurations,
through which the service provider requests, and/or the (new) service area
conditions, will be handled in the most cost-efficient manner. This
component includes a suite of efficient (in terms of computational time and
space) optimisation tools that can handle the UMTS, MBS, and DBS
resources in an aggregate manner. The target air-interfaces are the Frequency
Division Duplex (FDD) and Time Division Duplex (TDD) components of
UMTS, HIPERLAN II (among the various alternatives that fall in the MBS
category), and DVB-T and DAB (jointly addressed as DBS in this paper).
The third component is the UMTS, MBS, and DBS Network and
Environment Simulator (UMD-NES). It is motivated from the need of
342 P.Demestichas, G. Vivier, G.Martinez, F.Galliano, L.Papadopoulou,
validating some management decisions prior to their application in the real

network. Moreover, the component enables off line testing, validation and
demonstration of the management software in a wide range of test cases. In
this respect, it covers the absence oftestbeds.
2.2 Sample functionality
A sample scenario according to which the components above co-operate

is provided in Figure 3. The scenario consists of the following phases.
1. A service provider issues a virtual network establishment request through
the respective Service Provider Mechanisms (SPM).
2. The MASPI component intercepts the requests, processes it and forwards
it, together with the current managed system state to the RMS
component. Relevant strategies of the RMS component will be applied so
as to satisfy the received request.
3. The RMS component requests from the UMD-NES component to
validate (perhaps, parts of) the proposed solutions.
4. After the validation the appropriate reconfiguration commands are issued
towards the managed sub-network network and the core network
management and control systems.
r - - - - - UMIS
. ;" MB~ DBS Netwo'k,
1. Virtual Network
Establishment Request
2. Virtual Network
Establishment Reauest
Managed System State
3. Virtual Network
Validation ReQUest
3 . Reply em Virtual Netwo

<_.
f.qb@tiQfI.Re.,!«c.j'L
. Reply on Virtual Netwo

./f,§f.fl.bJl$.hnwntRe.qu,fNil..
a. Reply on Virtual Netwo k

, H.s!gQl!~b!,!.~.n(B~gy_~~·!.
4a. Reconf/f!uration Command
4b Rec:onfiguration Comma
Figure 3. Sample scenario of co-operation among the components of the UMTS, MBS and
DBS network and service management system
Radio Environment
Naturally, several variations of the procedure above may be devised and

specified in a finer level of detail. Moreover, it should be noted that the
procedure above would be similar if the MASPI component observed a new
environment condition (e.g., alteration in the traffic demand, mobility and
interference levels, etc.) in the managed network.
2.3 Deployment pattern
The previous subsections presented the high-level structure of the

management system and a sample of its functionality. This subsection goes
one step beyond and refines the architecture by taking into account the fact
that different operators will control the different wireless access systems. In
this context, a pattern for distributing the software components of the
management system is depicted in Figure 4. Specifically, it is assumed that
replicas are distributed in each domain. Each replica contains only the
appropriate subset functionality for managing the specific (UMTS, MBS or
DBS) network. However, these replicas can co-operate for handling service
provider requests and/or new environment conditions through the aggregate
reconfiguration of the UMTS, MBS or DBS resources.
o 0 0
IP NETWORK
Figure 4. Decentralised management system deployment pattern in a multi-operator scenario.

The Service Providers are assumed to have an interface to the MASPI elements of one (e.g.,
UMTS) or more radio networks
More specifically in this case, the joint management of the radio systems
is not straightforward since the different operators would not share strategic
information on their network. In this respect, the management system
introduced an important aspect. It presents a unified interface to the service
provider thanks to the MASPI element; from that point the dialog to get the
required QoS is handled in a manner that is transparent to the service
providers and their users. This dialog involves the management system
replicas that control each radio network. In this sense, some of the replicas
act as resource brokers, which means entities that control, offer and allocate
resources of a radio network segment. In this way all the radio systems are
involved and the appropriate radio technology for the required service is
selected. It should be noted that in the discussed case each UMD-NES
replica has a role limited to one radio technology, and that the management
system has the potential to simplify the relationship (particularly the billing
aspect) between service providers and the network operators.
The sample scenario of Figure 3 can be refined based on the discussion
above. A detailed description is omitted for brevity. Moreover, less
distributed scenarios can be defined, where for instance the management
system has the control of several (or all of the) radio systems and dialog with
the remaining ones through management system replicas.
3. MONITORING, ASSESSMENT AND SP

INTERWORKING
The internal structure of the MASPI block can be depicted as in Figure 5.

The following tasks are framed in this component.
Interface with the service provider mechanisms. This interface enables
service providers to request the reservation of resources from the managed
network, or in other words, to request the establishment of virtual networks
over the managed system infrastructure. It is reminded that combined
UMTS, MBS and DBS resources can be allocated in response to the virtual
network establishment request. Some of the information that should be
specified by the service provider through this interface is the service volume,
the target quality levels, the area and time-zone for which the request is
made, cost related constraints. This is a novel work area that results in the
formal definition of the interface, which should be at the same time adequate
and as less complex as possible.
Contracts with service providers processing, establishment and
maintenance. In literature these contracts are often called Service Level
Agreements (SLAs). The outcome of the processing of the service provider
Radio Environment
request may be the establishment, modification, rejection, cessation, etc. , of

a contract. Framed in the context of processing a service provider request is
the translation of the service provider view to a corresponding network view.
The trade-off associated with the translation has on the one hand the pursued
efficiency, in terms of the required (potentially combined UMTS, MBS and
DBS) network resources, and on the other hand the compliance with the
service provider requests. The associated problems are addressed jointly by
all the components of the management system. However, this part of the
management system has an important role, since it has the functionality for
providing information on the co-relation among the managed system
performance and the quality perceived at the service level.
Interfaces with the (highly heterogeneous) elements or segments of the
managed network. In principle, a typical UMTS, MBS, or DBS
configuration will consist of highly heterogeneous equipment, coming from
different vendors, and segments that are based on different technologies. The
anticipated heterogeneity of typical UMTS, MBS, DBS network
configurations may be addressed by common standards on the interfaces that
the managed elements, or network segments, should provide. The possibility
of inter-working with proprietary (perhaps legacy) element management
systems is included.
f-·~~~=-=~~~~~-=··~··=~=~~=~~·~~·~~~=~;;~~~::~:~-:~~;;i~~;~;;;~~;~-" !
I
: i !
i, Ii ! i
! •
I I
i
II II !
I•
!'
I
! ii
II
I !
i iI !
~ ~ ! •
!L ~I
l~ ~j
i, I, Ii' !t
f i I ;
f I i
! I !
Il---- - ~M!~~l'!"!'_"!~~l~~'''''Ao~'m~J
;t ..... __..______. .- ___________ -. - --_....._-- ---- ... _.. _-- --- --- .. ~ ... . .:. _.... __MASPI
.... --..-.... ....... Component
~
Structure
.. ---- --_
,
.. _- ... _--_ ..i
Figure 5. High-level internal structure of the MASPI component

Sophisticated managed system monitoring methodologies and

procedures. Traffic, mobility and propagation conditions in UMTS, MBS,
and DBS environment (service area) will be highly complex. For example,
the probability distribution of parameters that characterise the traffic and
mobility conditions in the service area may be changing with time. Efficient
resource exploitation may be achieved if these alterations are monitored and
the managed system is adapted to the new environment conditions.
Sophisticated procedures for (readily and accurately) obtaining, or detecting
changes in, the probability distribution of parameters are means for coping
with this aspect.
Graphical User Interface (GU!). The GUI role is to visualise aggregated
information (statistics, history) allowing the user to check the managed
system performance at the element, network and service level, the effect
(short or long term) of management actions, information on the contracts
established with service providers, etc. The Java technology is used for
ensuring portability on a wide variety of computing platforms.
Collaboration with the RMS component. The MASPI component assists
the RMS component (which is presented in the sequel) in selecting the
managed system reconfigurations that are required for handling the new
environment conditions and/or the service provider requests. It should be
noted that, ultimately, the RMS component proposes the appropriate
reconfigurations. The role of the MASPI component is to generate triggers
and to supply all the required information (e.g., the current managed system
state, service provider constraints, etc.).
4. RESOURCE MANAGEMENT STRATEGIES
The internal structure of the RMS component can be depicted as in

Figure 6. A rough categorisation of management strategies derives from their
relevance to the radio segment or the fixed network.
Radio Environment
Figure 6. High-level internal structure of the RMS component
4.1 Radio resource management

An indicative categorisation of problems solved by the radio resource
management strategies is the following.
Service mapping on radio technologies. This part involves the
distribution of traffic over the different air-interfaces of the managed UMTS,
MBS and DBS systems. A sample action, in this respect, is the assignment
of portion of the traffic (e.g., corresponding to certain service types) to the
DBS network, which in principle will enable large bit rates and have a
significant amount of spectrum available. In other words, the strategies in
this category aim at finding the most appropriate radio component (e.g.,
FDD, TDD, HIPERLAN, Broadcast, etc.) among those available for certain
user or service classes. In this respect, heterogeneous technologies, as those
considered in this paper, are made to co-exist and complement each other.
Different management strategies are associated to the different air interface
types.
Radio access point reconfiguration and spectrum assignment. These may
be done in response to significantly new environment (traffic and mobility)
conditions (with respect to those anticipated), or as proactive actions
motivated by a new set of contracts with service providers that should be
satisfied. In future, radio access points (i.e., base stations) will be generic
and could be reconfigured with the radio access technology and portion of
spectrum that best fit the service demand in a particular area. This can also
be called dynamic radio planning.
Radio resource control scheme selection. This part implies the
modification of the radio resource configuration pattern, and/or of the
parameter values that drive the radio resource control schemes that are
applied in the network, so as to satisfy the (new and established) contracts
with service providers and the environment conditions in a cost efficient
manner. Sample actions that fall in this category are the modification of the
number of connections that can be admitted per cell, the modification of
values related to power control strategies or of the transmission power of
certain base stations, etc. Such strategies may also be invoked in order to
relief hot spots, respond to network element failures, etc. Another action of
this category is the separation of radio resources to various concepts, e.g.,
virtual networks, paths within the virtual networks, service classes, new and
handover traffic, etc.
The list above can be refined in several dimensions. Each problem in the
list is an optimisation problem whose solution is integrated in the context of
the overall management tool.
4.2 Fixed network resource management

Handling service provider requests or new environment conditions may
require reconfigurations in the access network, or in the interface with the
core network. As all-IP fixed network segments are considered, bridges to
existing solutions of the access or the core network (e.g., IP control concepts
like Differentiated or Integrated Services, etc.) are implemented to address
transition from today's solutions. A relevant goal is mapping the radio
network traffic into appropriate IP classes.
An indicative categorisation of problems solved by the fixed network
resource management strategies is the following. First, the separation of
access network resources to virtual networks, paths or flows within the
virtual networks, service classes, etc., given the access network
configuration (e.g., element interconnections, etc.) and capabilities, and the
state of the interface with the core network. Second, the modification of the
access network configuration (e.g., restructure of the network element
interconnections) given the access network layout capabilities. Third, the
distribution of traffic over the appropriate IP service classes offered by the
access network segments of the managed UMTS, MBS and DBS systems.
4.3 Optimisation engine

The strategies identified in the previous subsections can be
mathematically formulated and be reduced to optimisation problems. This
part of the management system is a suite of tools for solving these problems
that relies in implementation work realised in the past [19]. In general, these
Radio Environment
procedures optimise functions including for instance cost, network

performance criteria, etc. under a set of constraints related to target QoS
levels, resource utilisation, fault tolerance, etc. Innovative approaches trying
to solve optimise the selection in a distributed manner will be particularly
addressed because of their inherent scalability properties.
5. NETWORK AND ENVIRONMENT SIMULATOR
This part can be seen as an integral component of the management

system as there is a need for tools that will assist operators in the validation
of management actions, prior to their application in the network. From a
different perspective, the current lack of large-scale experimental test-beds
during the project lifetime, demand the development of this component for
validating the efficiency of the management system.
Figure 7. High-level internal structure of the UMD-NES component
The structure of this component is as depicted in Figure 7. Through the

appropriate initialisation of the environment and network-modelling libraries
an extended set of test cases can be realised. For example, fault or
exceptional scenarios (e.g., hot spots, failing network element, etc.) can be
simulated by appropriately tuning certain input parameters. In more detail,
the following tasks are framed in this component.
Interface with MASPI and RMS components. This part of the UMD-NES
component supports the standard interface among the management system
and the managed infrastructure. This enables extensive testing of the
interface. This is so because it includes the full functionality for enabling the
management system to monitor and assess the managed system performance
at the element, network and service levels, and to reconfigure the managed
(and in this case simulated) system in accordance with the outcome of the
V. Stavroulaki, M. Theologou
management strategies. The information exchange between the simulated

network configuration and the MASPI and RMS components of the
management platform can be query-driven and event-driven.
Environment modelling library. This part is an extended software library
capable of representing user behaviour, service characteristics, and service
area and radio environment characteristics will be made available. User
behaviour covers aspects like service preferences (e.g., rate at which users of
a certain class invoke a service that is offered by the system), service usage
characteristics (e.g., duration of service usage or rate at which users invoke a
service feature) and mobility. Mobility models with different levels of detail
are included. Service modelling provides the generic characteristics that may
be encountered in a service usage (e.g., source or information flow
requirements, service features, etc.). User behaviour and service modelling
provide the traffic and mobility conditions that the network will have to
handle. Modelling of radio environment is needed to predict noise and
interference experienced in different subsets of the service area.
Network modelling library. This part is a software library of models that
can represent the radio interface (e.g., UMTS FDD and TDD, HIPERLAN,
etc.), and the relevant network elements. For example network element
models included correspond to UMTS Base Transceiver Stations and Radio
Network Controllers, HIPERLAN Access Point Transceivers and Access
Point Controllers, DBS Transceivers and Gateways, etc. Each network
element model comprises functionality for handling operations and
procedures of the respective system, e.g., call set-up, packet call, handover,
location update, etc.
6. SOFTWARE IMPLEMENTATION AND

VALIDATION
The target software architecture of the management system is illustrated

in Figure 8. The service provider mechanisms (which however in their full
detail are beyond the scope of the management system) are assumed Web-
based. The SPI component acts as an application server with respect to the
service provider mechanisms. The information exchange between that
MASPI, RMS and UMD-NES components is based on CORBA-compliant
platforms. The same approach has been followed for the RMS and core
network control (management) platforms. In terms of computing hardware
the RMS and the MASPI (apart from the GUI that is platform independent)
runs on UNIX or Linux machines, while the UMD-NES runs on Windows
machines.
Radio Environment
.... ... . -... _.. ...••...•..•.. ..__ ._ - --- -.. - - - --.... , r····-····. ·.·•• ·....... .... ··___·.· ··__ ·.·"
I~~: liB '--__- - ' ,--_~=, . ~I _ ~..1,,; : _ _
MIDDLEWAIlE PlATFORM
UMD-NESPIa!fonn ~
Figure 8. Target software architecture of the UMTS, MBS and DBS network and service
management system
Validation of the management system will be against the following major

points.
- The adequacy of the interfaces between the management platform and the
managed system and service provider mechanisms. These interfaces
should include all the necessary information and at the same time be as
easy as possible to integrate and use.
The efficiency of the management strategies in proposing cost-effective
reconfigurations of the overall UMTS, MBS and DBS infrastructure that
are suitable for handling service provider requests and new environment
conditions. This constitutes the most important validation point.
The efficiency of the monitoring procedures in capturing changes in the
environment conditions, and hence, assisting the management system in
improving the network performance and QoS levels.
The response time of the management system. This is a complex point as
the overall response time relies on the performance of the individual
components of the management system. For example, affecting factors
are the processing of the service provider requests, the required
computation effort of the optimisation algorithms, the simulation-time (in
case a validation phase is conducted), the software distribution pattern,
etc. Moreover, a complete answer to this point should also involve real
network experiments.
7. CONCLUSIONS
This paper presented an approach to the overall UMTS, MBS and DBS
network and service management problem. The starting point was the
presentation of the architecture of a corresponding platform, and its splitting
into the MASPI, RMS and UMD-NES logical units. The internal structure of
these components, and implementation and validation issues were discussed.
In general terms, future work is targeted to the following aspects. First,
the completion of the ongoing implementation. Second, the realisation of
validation studies and the refinement of the architecture based on the
obtained results. Third, the dissemination of results that can prove the
efficiency of each component and of the overall platform.
REFERENCES
"IP-based mobile telecommunications networks", Special issue in the IEEE Personal
Commun., Vol. 7, No.4, Aug. 2000
2 "Next generation wireless broadband networks", Special issue in the IEEE Network, Vol.
14, No.5, Sep./Oct. 2000
3 U. Varshney, R. Vetter, "Emerging mobile and broadband wireless networks", Commun.
ofthe ACM, Vol. 43, No.6, June 2000
4 "The evolution ofTDMA to 3G", Special issue in IEEE Personal Commun., Vol. 6, No.3,
June 1999
5 "Wideband CDMA", Feature topic in IEEE Commun. Mag., Vol. 36, No.9, Sept. 1998
6 "IMT-2000: Standards effort of the ITU", Special issue on IEEE Personal Commun., Vol.
4, No.4, Aug. 1997
7 Digital Video Broadcasting Web site, www.dvb.org, Jan. 2001
8 Multimedia Car Platform, mcp.fantastic.ch, Jan. 2001
9 Broadband Radio Access for IP based Networks, www.ist-brain.org, Jan. 2001
10 Dynamic Radio for IP-Services in Vehicular Environments, www.ist-drive.org, Jan. 2001
11 "Software and DSP in radio", Feature topic in the IEEE Commun. Mag., Vol. 37, No.8,
Aug. 1999
12 "Software Radio", Special issue in IEEE Personal Commun., Vol. 6, No.4, Aug. 1999
13 "Design of broadband multiservice networks", Feature topic in the IEEE Commun. Mag.,
Vol. 36, No.5, May 1998
14 "Tools and techniques for the performance evaluation of communication networks",
Feature topic in the IEEE Commun. Mag., Vol. 36, No.8, Aug. 1998
15 E. Berruto, M. Gudmundson, R. Menolascino, W. Mohr, M. Pizarroso, "Research
activities, on UMTS radio interface, network architectures and planning", IEEE Commun.
Mag., Vol. 36, No.2, Feb. 1998
16 P. Demestichas, E. Tzifa, V. Demesticha, M. Louta, D. Karagiannis, M. Theologou,
"Issues in the design offuture mobile communications systems", In Proc. IEEE
International Conference on Communications (ICC'99), pp. 364-368, Vancouver, Canada,
June 1999
17 S. Vinoski, "COREA: Integrating diverse applications within distributed heterogeneous
environments", IEEE Commun. Mag., Vo1.35, No.2, Feb. 1997
18 "Distributed real-time systems", Feature topic in the IEEE Computer Journal, Vol. 33, No.
6, June 2000
19 R. Menolascino, P. Cullen, P. Demestichas, S. Josselin, P. Kuonen, Y. Markoulidakis, M.
Pizarroso, D. Zeghlache, "A realistic UMTS planning exercise", in Proc. 3rd ACTS Mobile
Communications Summit 98, Rhodes, Greece, June 1998

Multiaccess, Mobility and Teletraffic For Wireless Communications

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Multiaccess, Mobility and Teletraffic For Wireless Communications

Uploaded by

Copyright:

Available Formats

MULTIACCESS, MOBILITY

AND TELETRAFFIC FOR

SPRINGER SCIENCE+BUSINESS MEDIA, B.V.

ISBN 978-1-4419-5290-5 ISBN 978-1-4757-5918-1 (eBook)

Printed on acid-free paper

All Rights Reserved

Capacity Analysis of Voice over IP over GERAN with Statistical 25

Uplink RRM for Conversational and Interactive Services in UTRA-FDD 43

Rate and Power Adaptation for Downlink Shared Channel in WCDMA 61

Capacity and CII Performance of Different Cell Clusters in a Cellular 75

Performance Study of Soft Handover with CDMA Heterogeneous Cellular 87

Terminal Migration Model in which Cell Dwell Time is Defined by 133

Concatenated Location Management 143

Handoff Scheme Improvement in Wireless Networks 155

Hierarchical Mobility Controlled by the Network 167

Gallager Codes for Asynchronous Multiple Access 205

Interactions between the TCP and RLC Protocols in UMTS 251

Impact of SR-ARQ with Finite Buffer on TDD/TDMA Wireless LAN 263

Traffic Performance Analysis of Multimedia Applications in Evolved GSM 279

VoicelVideo over IP with Multi-class QoS in 3G Mobile Networks 295

Establishment of Mobile Extranets through Mobile Ipv6 and GPRS: 309

IP Traffic Control on UMTS Terminal Equipment 323

An Approach for Managing Networks and Services in a Diversified Radio 337

We express our gratitude to the reviewers: Bardia Alavi, Thorsten Benkner,

Mohamed Hassan, Marwan Krunz, and William Ryan

Abstract Successful provisioning of multimedia services over wireless networks hinges

signal-to-noise ratio (SNR). Due to their mathematical tractability, finite-

Figure 1. Partitioning of the received SNR space.

Several researchers have addressed the modeling of wireless links through

packet buffering at the transmitter side of a wireless link introduces vari-

2. Wireless Link Model

wireless link by minimizing the number of retransmissions. In practice,

Figure 2. Wireless-link model.

riving at a node can be viewed as an alternating sequence of active and

Figure 3. Markov chain in Mitra's producer-consumer model.

code block (packet), the probability of an uncorrectable error in a received

where p ~ E[r]. An important parameter that reflects the behavior of

where f m = ~ is the Doppler frequency, w is the speed of the electro-

2.2 Mitra's Producers-Consumers Fluid Model

Note that in the producer-consumer model, the consumption rate in state

Ti+l = -pln(e -;; -7ri), i = 1,2, ... , N - 2 (8)

Dividing (9) by (10), we get the following expression for ~:

The above equation relates *'

(recall that ro ~ 0). From (12) and (13), we have

Similarly, the steady-state probability that the channel is in state N is

1I'N= (~)N (l-e 7 ) (15)

Taking the logarithm of both sides, we end up with:

To explicitly express N in terms of rl and rN, we replace * in (14)

-In ( l-e .::.!:l.)

where 0 < E* « 1 is a predefined control parameter. In (21) the expression

Parameterize-FSMC takes as inputs the parameters of the coding

Parameterize-FSMC(p, Pe(.), EO, n, T)

Figure 4. Algorithm to calculate the parameters of the FSMC model.

The evolution of the buffer content follows the following first-order

dIT(x)D = IT(x)M (22)

D = diag[-co, ... , -CN, r - co, ... , r - CN] (23)

IT (x ) = [7fo,o ... 7fO,N 7f1,0 ... 7f1,N] (24)

7f s ,i(X) = Pr{ Q :S x, system is in state (8, in

IT(x) = L aie-ZiXc/>i (25)

4.2 Wireless Effective Bandwidth

AA(Z) ~ lim sup C 1 log E[exp(zA(t))]

G(x) ~Pr[Q > x] = e-z*x (27)

For a two-state arrival process, AA(Z) is given by [17]: