SEPTEMBER 2011 VOLUME 2 NUMBER 3
september 2011 volume 2 number 3 - Advances in Electronics and ...
september 2011 volume 2 number 3 - Advances in Electronics and ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong>SEPTEMBER</strong> <strong>2011</strong> <strong>VOLUME</strong> 2 <strong>NUMBER</strong> 3
Journal Editorial Board<br />
KRZYSZTOF WESOŁOWSKI, Editor-in-Chief<br />
Poznan University of Technology<br />
Piotrowo 3A, 60-965 Poznań, Poland<br />
krzysztof.wesolowski@et.put.poznan.pl<br />
ANNA PAWLACZYK, Secretary<br />
Poznan University of Technology<br />
Piotrowo 3A, 60-965 Poznań, Poland<br />
anna.pawlaczyk@et.put.poznan.pl<br />
ADRIAN LANGOWSKI, Technical Editor<br />
Poznan University of Technology<br />
Piotrowo 3A, 60-965 Poznań, Poland<br />
adrian.langowski@et.put.poznan.pl<br />
WOJCIECH BANDURSKI<br />
Poznan University of Technology<br />
ANNA DOMAŃSKA<br />
Poznan University of Technology<br />
MACIEJ STASIAK<br />
Poznan University of Technology<br />
HANNA BOGUCKA<br />
Poznan University of Technology<br />
MAREK DOMAŃSKI<br />
Poznan University of Technology<br />
RYSZARD STASIŃSKI<br />
Poznan University of Technology<br />
ANDRZEJ DOBROGOWSKI<br />
Poznan University of Technology<br />
WOJCIECH KABACIŃSKI<br />
Poznan University of Technology<br />
PAWEŁ SZULAKIEWICZ<br />
Poznan University of Technology<br />
Advisory Board<br />
FLAVIO CANAVERO<br />
Politecnico di Torino<br />
Italy<br />
TADEUSZ CZACHÓRSKI<br />
Polish Academy of Science<br />
Institute of Theretical and Applied<br />
Informatics<br />
Gliwice, Poland<br />
PIERRE DUHAMEL<br />
CNRS - Supélec<br />
France<br />
LAJOS HANZO<br />
University of Southampton<br />
UK<br />
MICHAEL LOGOTHETIS<br />
University of Patras<br />
Greece<br />
JÓZEF MODELSKI<br />
Warsaw University of Technology<br />
Poland<br />
MACIEJ OGORZAŁEK<br />
AGH Technical University<br />
Jagiellonian University<br />
Cracow, Poland<br />
JOHN G. PROAKIS<br />
University of California<br />
San Diego, USA<br />
RALF SCHÄFER<br />
Fraunhofer Heinrich-Hertz-Institut<br />
Berlin, Germany<br />
Cover design Barbara Wesołowska<br />
c○ Copyright by POZNAN UNIVERSITY OF TECHNOLOGY, Poznań, Poland, 2010<br />
Edition based on ready-to-print materials submitted by authors<br />
Materials published without further editing at the responsibility of the authors<br />
ISBN 978-83-7143-899-8<br />
ISSN 2081-8580<br />
PUBLISHING HOUSE OF POZNAN UNIVERSITY OF TECHNOLOGY<br />
60-965 Poznań, pl. M. Skłodowskiej-Curie 2<br />
tel. +48 (61) 6653516, fax +48 (61) 6653583<br />
e-mail: office_ed@put.poznan.pl<br />
www.ed.put.poznan.pl<br />
ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS is a peer-reviewed journal published at Poznań University of Technology, Faculty<br />
of Electronics and Telecommunications. It publishes scientific papers addressing crucial issues in the area of contemporary electronics and<br />
telecommunications. Detailed information about the journal can be found at: www.advances.et.put.poznan.pl.
<strong>SEPTEMBER</strong> <strong>2011</strong> <strong>VOLUME</strong> 2 <strong>NUMBER</strong> 3<br />
Radio Communication Series:<br />
Recent Advances in Teletraffic<br />
Issue Editors: Grzegorz Danilewicz and Mariusz Głąbowski<br />
Note from the Issue Editors<br />
Grzegorz Danilewicz and Mariusz Głabowski ˛ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1<br />
Algorithm for queueing networks with multi-rate traffic<br />
Villy B. Iversen and King-Tim Ko . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3<br />
Retry Loss Models Supporting Elastic Traffic<br />
Ioannis D. Moscholios, Vassilios G. Vasilakis, John S. Vardakas and Michael D. Logothetis . . . . . . . . . . . . 8<br />
Damming the Torrent: Adjusting BitTorrent-like Peer-to-Peer Networks to Mobile and Wireless Environments<br />
Philipp M. Eittenberger, Seungbae Kim, and Udo R. Krieger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14<br />
Analysis of OBS Burst Assembly Queue with Renewal Input<br />
Tomasz Hołyński and Muhammad Faisal Hayat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23<br />
Scheduling and Capacity Estimation in LTE<br />
Olav Østerbø . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31<br />
Multi-Service Load Balancing in a Heterogeneous Network with Vertical Handover<br />
Jie Xu, Yuming Jiang, Andrew Perkis and Elissar Khloussy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43<br />
Resources Management and Services Personalization in Future Internet Applications<br />
Paweł Światek, Piotr Rygielski and Adam Grzech . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50<br />
Compact node-link formulations for the optimal single path MPLS Fast Reroute layout<br />
Cezary Żukowski, Artur Tomaszewski, Michał Pióro, David Hock, Matthias Hartmann and Michael Menth . . . . 55<br />
Enhancing Data Transmission Reliability with Multipath Multicast Rate Allocation<br />
Matin Bagherpour, Mehrdad Alipour and Øivind Kure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61<br />
Analytical Model for Virtual Link Provisioning in Service Overlay Networks<br />
Piotr Krawiec, Andrzej Bęben and Jarosław Śliwiński . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Decisive Factors for Quality of Experience of OpenID Authentication Using EAP<br />
Charlott Lorentzen, Markus Fiedler, and Peter Lorentzen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79<br />
Agent based VoIP Application with Reputation Mechanisms<br />
Grzegorz Oryńczak and Zbigniew Kotulski . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong> 1<br />
Note from the Issue Editors<br />
This volume of Advances in Electronics and Telecommunications contains<br />
papers submitted for “The First European Teletraffic Seminar (ETS)” held<br />
in Poznań, Poland, in February <strong>2011</strong>. The ETS has come into being as an<br />
extension to the Polish German Teletraffic Symposium, which was initiated<br />
in 2000, and the Nordic Teletraffic Seminar, initiated in 1977, with kind<br />
support of French Teletraffic Community. The European Teletraffic Seminar<br />
is intended to maintain a regular series of international seminars providing<br />
a forum for discussions related to the issues of teletraffic, a discipline<br />
covering phenomena in control and transport of information within communications<br />
and computer networks, for researchers, practitioners, young<br />
scientists, and students.<br />
Traffic theory and engineering have been an inseparable part of the<br />
development of telecommunications and ICT infrastructure from their very<br />
beginning. Teletraffic problems have changed substantially over the recent<br />
years as a result of the shift in the telecommunications area towards integrated<br />
digital networks, data services, Internet and mobile communications.<br />
Each and every newly introduced network technology is followed by a<br />
major increase in both the number and the complexity of problems that need<br />
to be resolved by theoreticians and traffic engineers. No matter what these<br />
developing changes may bring, the essential task for traffic theory remains<br />
the same – to determine and evaluate the relationship between the quality of<br />
service parameters, the parameters that determine the intensity of calls, and<br />
the amount of resources demanded by such calls as well as the parameters<br />
that describe available network resources. These relationships provide a<br />
basis to develop engineering algorithms used for designing, analysis, and<br />
optimization of systems and networks.<br />
This issue of Advances in Electronics and Telecommunications contains<br />
extended versions of selected ETS papers presenting a broad range of<br />
teletraffic usage in modern telecommunications. The topics cover subjects related to teletraffic issues in next<br />
generation and new generation networks, e.g. Future Internet architectures and technologies, operation of modern<br />
telecommunication and computer networks; broadband and mobile communication systems; integration of a<br />
broad spectrum of services; computer and communications systems applications; methods and tools for networks<br />
and services modeling issues; networks and services planning; forecasting and management; performance<br />
evaluation, etc. We believe that all articles presented in the journal clearly prove the importance and justify<br />
the presence of traffic theory and engineering in providing solutions to problems we face in all modern networks,<br />
telecommunications and computer networks alike, even (or especially) today when transmission offers high<br />
bandwidth and switching is performed within extremely short time. We hope that readers will find the ETS<br />
papers selected for this issue of Advances interesting, as we believe that teletraffic is equally important for<br />
researchers and practitioners.<br />
We would like to thank all the authors for their contributions to this issue of Advances in Electronics and<br />
Telecommunications. Our thanks extend also to ETS’ Technical Programme Committee Chairs: Paul J. Kühn<br />
and Michał Pióro, as well as to ETS’ General Conference Chairs: Prosper Chemouil, Markus Fiedler, Wojciech<br />
Kabaciński, and Maciej Stasiak, whose involvement and commitment was critical to the successful completion<br />
of the efforts to integrate European conferences focused on traffic theory and traffic engineering.<br />
Grzegorz Danilewicz<br />
Mariusz Głąbowski<br />
Issue Editors
ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong> 3<br />
Algorithm for queueing networks with<br />
multi-rate traffic<br />
Villy B. Iversen and King-Tim Ko<br />
Abstract—In this paper we present a new algorithm for<br />
evaluatingqueueingnetworkswithmulti-ratetraffic.Thedetailed<br />
state space of a node is evaluated by explicit formulæ. We<br />
consider reversible nodes with multi-rate traffic and find the<br />
state probabilities by taking advantage of local balance. Theory<br />
of queueing networks in general, presumes that we have product<br />
form betweenthenodes.Otherwise,wehave thestatespaceexplosion.<br />
Even so, the detailed state space of each node may become<br />
very large because there is no product form between chains<br />
inside a node. A prerequisite for product form is reversibility<br />
which implies that the arrival process and departure process<br />
are identical processes, for example state-dependent Poisson<br />
processes. This property is equivalent to reversibility. Due to<br />
product form, an open network with multi-rate traffic is easy to<br />
evaluate by convolution algorithms because the nodes behave as<br />
independent nodes. For closed queueing networks with multiple<br />
servers inevery nodeandmulti-rate services wemay applymultidimensional<br />
convolution algorithm to aggregate the nodes so that<br />
we endupwith twonodes, the aggregated node and asingle node,<br />
for which we can calculate the detailed performance measures.<br />
Index Terms—multi-rate traffic, queueingnetworks, reversibility,<br />
insensitivity, product form, convolution algorithm<br />
I. INTRODUCTION<br />
IN 1957, J.R. Jackson who was working at production<br />
planning and manufacturing systems, published a paper<br />
[1] showing that a queueing network of M/M/n–nodes has<br />
product form. Knowing the fundamental theorem of Burke<br />
(1956 [2]) Jackson’s result is obvious. Historically, the first<br />
paper on queueing systems in series was by another Jackson,<br />
R.R.P. Jackson (1954 [3]).<br />
The key point of Jackson’s theorem is that each node can<br />
be considered to be independent of all other nodes, and that<br />
the state probabilities are given by Erlang’s waiting time<br />
modelM/M/n.Thissimplifiesthecalculationofthestate space<br />
probabilities significantly. The proof of the theorem was given<br />
byJacksonin1957byshowingthatthenodebalanceequations<br />
are fulfilled under the assumption of statistical equilibrium.<br />
Jackson’s first model thus only deals with open queueing<br />
networks.<br />
In Jackson’s second model (1963 [4]) the arrival intensity<br />
from outside may depend on the current number of customers<br />
in the network. Furthermore, the service rates may depend on<br />
the number of customers k in the nodes. In this way, we can<br />
model queueing networks which are either closed, open, or<br />
mixed. In all three cases, the state probabilities have product<br />
Villy B. Iversen is with Department of Photonic Engineering Technical<br />
University of Denmark, 2800 Kongens Lyngby, Denmark. Email:<br />
vbiv@fotonik.dtu.dk<br />
King-Tim Ko is with Department of Electronic Engineering City University<br />
of Hong Kong, Hong Kong. Email: eektko@cityu.edu.hk<br />
form between nodes. The model by Gordon & Newell from<br />
1967 which is often cited in the literature can be treated as a<br />
special case of Jackson’s second model.<br />
The theory of queueing networks assumes that a customer<br />
samples a new service time in every node. This is a necessary<br />
assumption for having product form. This assumption was<br />
investigated by Kleinrock (1964 [5]) and it turns out to be<br />
a good approximation in real life.<br />
In 1975 the second model of Jackson was further generalizedbyBaskett,Chandy,MuntzandPalacios(1975[6])tosocalled<br />
BCMP–networks. These authors showed that queueing<br />
networks with K nodes and more than one type of customers<br />
also have product form, provided that:<br />
a) The customers are classified into N chains. Each chain<br />
j ∈ N is in each node i ∈ K characterized by<br />
its own mean service time s j i and routing probabilities<br />
p ik , {i, k ∈ K}. A customer may change from one chain<br />
to another chain with a certain probability after finishing<br />
service at a node. If the queueing system of a node is<br />
a classical M/M/n system (including M/M/1), then the<br />
average service time in a node must be identical for all<br />
chains.<br />
b) Each node is a symmetric (= reversible) queueing system<br />
mentioned below (Sec. II-B): for each chain a Poisson<br />
arrival process implies a Poisson departure process.<br />
BCMP–networks can be evaluated by the multi-dimensional<br />
convolution algorithm for multi-server systems. The famous<br />
MVA (mean value) algorithm by Lavenberg & Reiser [7]<br />
is applicable only if all nodes are single server systems<br />
or infinite server systems. This paper is based on models<br />
of Kingman (1969 [8]) and Sutton (1980 [9]), which are<br />
generalizations of Erlang’s approach based on the assumption<br />
of statistical equilibrium. All derivations are mathematically<br />
very simple. Similar models are dealt with by Bonald &<br />
Proutière (2003 [10]) and Bonald & Virtamo (2005 [11]).<br />
They also consider multi-rate queueing nodes, but only with<br />
infinite buffer. They present expressions for the average flow<br />
throughput. Serfozo (1999 [12]) presents the general theoretical<br />
background.<br />
Our approach is algorithmic and directed to engineering<br />
applications. For open networks we achieve an algorithm<br />
which is linear in both number of traffic streams and number<br />
of channels and has a very small memory requirement<br />
(Iversen, 2007 [13]). For loss (buffer-less) systems we obtain<br />
an algorithm for BPP (Binomial–Poisson–Pascal) traffic with<br />
individual performancemeasures for each stream. For systems<br />
with buffers (finite or infinite) we obtain both mean virtual
4 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
· · ·<br />
· · ·<br />
.<br />
✗ ✔<br />
✗<br />
x 1 −d 1 , x 2<br />
x 1 , x 2<br />
✖<br />
..<br />
.<br />
✕<br />
x 1 µ 1<br />
✖<br />
..<br />
..<br />
. .<br />
. .<br />
..<br />
..<br />
..<br />
..<br />
✔<br />
. · · ·<br />
.<br />
✕· · ·<br />
For type j customers the service rate in state<br />
x = (x 1 , x 2 , . . . , x j , . . . , x N ) is reduced by a factor<br />
g j (x). The reduction factors g j (x) are chosen so that we<br />
maintain reversibility, and they can be specified for various<br />
parts of the state transition diagram as follows, where (d) is<br />
the only non-trivial case.<br />
· · ·<br />
· · ·<br />
.<br />
..<br />
..<br />
. ✗ ✔ λ 1<br />
✗<br />
x 1 −d 1 , x 2 −d 2 x 1 , x 2 −d 2<br />
✖<br />
.<br />
✕<br />
✖<br />
..<br />
λ 1<br />
λ 2<br />
x 2 µ 2 λ 2<br />
x 2 µ 2<br />
x 1 µ 1<br />
..<br />
.<br />
..<br />
.<br />
..<br />
.<br />
..<br />
.<br />
✔<br />
. · · ·<br />
.<br />
✕· · ·<br />
Fig. 1. State-transition diagram for four neighboring states in a reversible<br />
nodewith two multi-rate trafficstreams and infinite capacity. State x j denoters<br />
that x j channels are occupied by type j calls. For type j with slot-size d j we<br />
choose the service rate d j µ j .<br />
waiting times, virtual mean queue lengths, and loss (overflow)<br />
probability for each stream.<br />
In this paper we derive explicit formulæ for detailed state<br />
probabilities of multi-rate closed queueing networks.<br />
II. REVERSIBLE MULTI-SERVER MULTI-SERVICE NODES<br />
We consider a system with n servers (channels, bandwidth<br />
units) and infinite queue. The system is offered N different<br />
traffic streams. Customers of type j arrive to the system<br />
according to a Poisson arrival process with intensity λ j . A<br />
customer of type j attempts to occupy d j servers, and if all<br />
these are obtained the service time is exponentially distributed<br />
with intensity d j µ j (j = 1, 2, . . . , N) where the factor d j<br />
only is chosen for convenience. Later we may choose µ j = µ<br />
for all services, then the service rate in a state with a total<br />
of x busy channels will be equal to x µ, independent of the<br />
actual mix of services. The state of the system is defined<br />
by x = (x 1 , x 2 , . . . , x j , . . . , x N ), where x j is the number<br />
of channels and/or queueing positions occupied by type j<br />
customers. Thus, the number of type j customers is equal<br />
to x j /d j . If the total demand for channels is bigger than the<br />
number of channels available, then the customers share the<br />
capacity and queueing positions in some particular way which<br />
is specified below. When the number of servers is infinite, we<br />
get the state transition diagram shown for two traffic streams<br />
in Fig. 1. This diagram is reversible and has a simple product<br />
form solution.<br />
However, the capacity is limited to n servers, so we have<br />
to reduce the service rates in all states requiring more than<br />
n servers (overload). In the following we illustrate the theory<br />
for two services, but also mention the general case with N<br />
different services.<br />
A. Reduction factors<br />
We consider a system with N traffic streams. Let:<br />
x = (x 1 , x 2 , . . . , x j−1 , x j , x j+1 , . . . , x N )<br />
x − d j = (x 1 , x 2 , . . . , x j−1 , x j − d j , x j+1 , . . . , x N )<br />
(a) x i ≤ 0 : g j (x) = 0.<br />
The reduction factors are undefined for these states for which<br />
the state probabilities are zero. By choosing the value zero,<br />
the recursion formulæ below are correctly initiated.<br />
(b) x i ≥ 0 ∧ 0 < ∑ N<br />
j=1 x i ≤ n : g j (x) = 1 .<br />
Every call is allocated the capacity required and there is no<br />
reduction.<br />
(c) x j = 0 for all j ≠ i, x i ≥ n : g i (x) = n/x i<br />
Along the axes we have a classical M/M/n–system with only<br />
one service, and the calls share the capacity equally.<br />
(d)Stateswith moretypesofcustomersin totalrequiringmore<br />
than n channels. If possible, we want to choose g i (x 1 , x 2 ) so<br />
that all capacity is used. This requirement implies:<br />
N∑<br />
x j · g j (x) = n ,<br />
j=1<br />
N∑<br />
x j ≥ n . (1)<br />
j=1<br />
We consider states x where x =<br />
∑ N<br />
i=1 x i > n and<br />
all capacity is used. We apply flow balance equations for<br />
Kolmogorov cycles.<br />
A necessary and sufficient condition for reversibility (Kingman,<br />
1969 [8], Sutton, 1980 [9]) is that all two-dimensional<br />
flow paths are in equilibrium. In total we may choose:<br />
( ) N<br />
=<br />
2<br />
N (N − 1)<br />
2<br />
different cycles and thus different balance equations.<br />
We assume that we know the reduction factors for states<br />
x − d j below state x. To find the N reduction factors in<br />
state x = {x 1 , x 2 , . . . , x N } we need N independent equations.<br />
Thus, we may choose Kolmogorov cycles for the twodimensional<br />
planes {1, j}, (j = 2, 3, . . . , N) which yields<br />
N −1 independent equations. Furthermore we have the normalization<br />
equation (1) requiring that the total capacity used<br />
during overload is n. We get the following flow balance<br />
equations for j = 2, 3, . . . N:<br />
or<br />
g 1 (x) · g j (x − d 1 ) = g j (x) · g 1 (x − d j )<br />
g j (x) = g 1 (x) · gj(x − d 1 )<br />
, j = 2, 3, . . . , N . (2)<br />
g 1 (x − d j )
INVERSEN AND KO: ALGORITHM FOR QUEUEING NETWORKS WITH MULTI-RATE TRAFFIC 5<br />
The capacity normalization equations is (1):<br />
n =<br />
=<br />
g 1 (x) =<br />
N∑<br />
x i · g i (x)<br />
i=1<br />
N∑<br />
i=1<br />
{<br />
x i · g 1 (x) · gi(x }<br />
− d 1 )<br />
,<br />
g 1 (x − d i )<br />
n<br />
N∑<br />
{<br />
x i · gi(x } . (3)<br />
− d 1 )<br />
g 1 (x − d i )<br />
i=1<br />
Thus, we find g 1 (x) from (3) and all other reduction factors<br />
in state x from (2). As we know, all reduction factors for<br />
global states x up to n where x = ∑ N<br />
i=1 x i and all reduction<br />
factors for states where only one service is active, then we can<br />
calculate all reductionfactors recursively.This is equivalent to<br />
calculating the relative state probabilities, and thus by global<br />
normalization the normalized state probabilities.<br />
For two traffic streams we get the reduction factors:<br />
n<br />
g 1 (x 1 , x 2 ) =<br />
x 1 + x 2 · g2(x1−d1,x2)<br />
g 2 (x 1 , x 2 ) =<br />
g 1(x 1,x 2−d 2)<br />
n<br />
x 2 + x 1 · g1(x1,x2−d2)<br />
g 2(x 1−d 1,x 2)<br />
We always find a unique solution when the offered traffic is<br />
less than the capacity. When we know the reduction factors, it<br />
is easy to find the relative state probabilities by local balance<br />
equations. Due to reversibility we have local balance for each<br />
service. We get<br />
λ j · p(x − d j ) = g j (x) x j µ j · p(x) , j = 1, . . . , N.<br />
Thus, we can recursivelyfind all reductionfactorsand all state<br />
probabilities expressed by state zero. By normalization, which<br />
should be done in each step to ensure numerical accuracy, we<br />
obtain the absolute state probabilitiesfor a single multi-rate n-<br />
server node. From the state probabilities we find performance<br />
measures as mean sojourn time and throughput.<br />
For low and normal load, each connection will almost get<br />
the required bandwidth as for proportional fair scheduling. It<br />
can be shown by studying the reduction factors for increasing<br />
load that if the system becomes highly overloaded, then in<br />
the limit we get fair scheduling where all connections will<br />
be allocated the same capacity independent of the required<br />
slot-sizes.<br />
B. Classical queueing networks as special cases<br />
When all traffic streams have the same bandwidth demand<br />
d j = 1, we get the following simple solution:<br />
g j (x) =<br />
n<br />
∑ N<br />
j=1 x ,<br />
j<br />
N∑<br />
x j ≥ n , j = 1, 2, . . . , N. (4)<br />
j=1<br />
Thus, the service rates of all customers are reduced by the<br />
same factor and during overload the customers share the capacity<br />
equally. The state transition diagram can be interpreted<br />
as modeling the following systems, illustrated by Fig. 2.<br />
.<br />
· · ·<br />
· · ·<br />
· · ·<br />
· · ·<br />
.<br />
.<br />
..<br />
..<br />
✗ ✔ λ 1 ✗<br />
✔<br />
. . · · ·<br />
x 1 −d 1 , x 2<br />
x 1 , x 2<br />
.<br />
.<br />
✖ ✕<br />
✖<br />
✕· · ·<br />
..<br />
..<br />
g 1 (x 1 , x 2 ) · x 1 µ 1<br />
λ 2 g 2 (x 1 −d 1 , x 2 ) · x 2 µ 2<br />
λ 2 g 2 (x 1 , x 2 ) · x 2 µ 2<br />
g 1 (x 1 , x 2 −d 2 ) · x 1 µ 1<br />
. .<br />
..<br />
..<br />
..<br />
✗ ✔ λ 1<br />
✗<br />
x 1 −d 1 , x 2 −d 2 x 1 , x 2 −d 2<br />
✖<br />
.<br />
✕<br />
✖<br />
. .<br />
.<br />
..<br />
..<br />
.<br />
.<br />
..<br />
.<br />
.<br />
..<br />
..<br />
.<br />
.<br />
..<br />
.<br />
✔<br />
. · · ·<br />
.<br />
✕· · ·<br />
Fig. 2. State-transition diagram for four neighboring states in a reversible<br />
system with two multi-rate traffic streams.<br />
C. Generalized processor sharing (GPS) system<br />
In states (x 1 , x 2 ) below saturation (x 1 + x 2 ≤ n) every user<br />
occupy one server. Above saturation all users share the available<br />
capacity equally. The model is insensitive to the service<br />
time distribution and each service may have individual mean<br />
service time. This model is called the GPS = Generalized<br />
Processor Sharing model. For states x 1 + x 2 > n, a customer<br />
typeone wantsatotal service rate µ 1 , andacustomertype two<br />
wants a service rate µ 2 . But the service rates of all customers<br />
are reduced by the same factor n/(x 1 + x 2 ).<br />
As a special case, the Σ j M j /G j /1–PS single-server system<br />
with processor sharing (PS) is reversible and insensitive to the<br />
service time distributions and each class may have individual<br />
mean service times.<br />
D. Classical multi-service multi-server system<br />
In classical queueing, a customer always get one server for<br />
exclusiveusage. Thento maintainreversibilityfor x 1 +x 2 > n<br />
wehavetorequirethatallcustomershavethesameservicerate<br />
µ j = µ, and thus the same exponentially distributed service<br />
time. The mathematical proof will be give elsewhere. This<br />
corresponds to an M/M/n system with total arrival rate λ =<br />
∑<br />
j λ j andservicerate µ.All customersinthesystemhavethe<br />
same probability of being the next one departing, independent<br />
of the call type.<br />
The system M/G/∞ is reversible, as the departure process<br />
is a Poisson process because it is a random translation of the<br />
arrival process.<br />
E. Σ j M j /G j /1–LCFS–PR single-server system<br />
From the very nature of the model it is obvious that it is<br />
reversible as we always follow the same path back to state<br />
zero as away from state zero. It is insensitive to the service<br />
time distribution and each service may have individual mean<br />
service time.<br />
In conclusion, multi-server queueing systems with more<br />
classes of customers, all having bandwidth demand d j =
6 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
1, (j = 1, 2, . . . , N), will only be reversible when the system<br />
is one of the following queueing system:<br />
• M/G/n–GPS , (for n = 1 PS - Processor Sharing).<br />
• M/G/1–LCFS–PR,<br />
• M/M/n, including M/M/1 with same service time for all<br />
customers.<br />
These systems are also called symmetric queueing systems.<br />
Reversibility implies that the departure processes are Poisson<br />
processes for all classes. These systems make up the nodes<br />
allowed in BCMP–queueing networks.<br />
III. MULTI-RATE MULTI-SERVER QUEUEING NETWORKS<br />
We now build a network of nodes of the above type. As in<br />
ordinary queueing networks we have a routing matrix which<br />
specifies the route followed by a given traffic stream, named<br />
a chain.<br />
Open Queueing Networks<br />
These are easy to deal with. By solving the flow balance<br />
equations(flow in = flow out for each node) for each chain we<br />
find the load (offered traffic) of each node. We then find the<br />
state probabilities of each node and due to product form we<br />
easily get the state probabilities of the total queueing network.<br />
Closed Queueing Networks<br />
From the flow balance equations we first find the relative<br />
load of each chain in each node. For a closed network we<br />
aggregate the nodes by multi-dimensional convolutions, keeping<br />
account of the number of customers in each chain in the<br />
aggregated node. All nodes except the target node are aggregated<br />
into one node. This aggregated node is convolved with<br />
the target node for which we find the performance measures<br />
during the convolution. The multi-dimensional convolution is<br />
defined as follows:<br />
p 1,2 (x 1 , x 2 , . . . , x N ) = p 1 ∗ p 2 =<br />
∑x 1<br />
∑x 2<br />
i 1=0 i 2=0<br />
. . .<br />
x N<br />
∑<br />
i N =0<br />
p 1 (x 1 −i 1 , x 2 −i 2 , . . . , x N −i N ) · p 2 (i 1 , i 2 , . . . , i N )<br />
The parameter x j is given in numberof channels. Both x j and<br />
i j belong to {0, d j , 2d j , . . . , d j S j }, where S j is the number<br />
of customers in chain j, and j = 1, 2, . . . , N. By changing<br />
the order of convolution we obtain the performance measures<br />
for each node. The performance measures are for example<br />
mean waiting time and mean queue length for each chain, and<br />
carried traffic for each stream in the node.<br />
IV. NUMERICAL EXAMPLE<br />
The convolution algorithm for closed multi-rate queueing<br />
networks has been inplemented in a master thesis project by<br />
Iliakis & Kardaras (2007 [14]).<br />
Let us consider a closed network with two nodes and<br />
two types of customers, alternating between the two nodes<br />
(a generalized machine-repair model). We have 10 type-1<br />
customers each requesting one channel for service in a node.<br />
The service time is 4 tu in node-1 and 1 tu in node-2 (tu =<br />
time units). We have 5 type-2 customers each requesting two<br />
channels for service in a node. The service time is 2 tu in<br />
node-1 and 0.5 tu in node-2. Thus, the packet size (bandwidth<br />
times mean service time) is the same for the two types.<br />
The capacity is 20 channels in node-1 so that we never<br />
experience delay. The sojourn time in node-1 is thus 4 tu,<br />
respectively 2 tu. In node-2 we have 5 channels and infinite<br />
queue. We find the sojourn times equal to 1.1929 tu, respectively<br />
0.6990 tu. Thus, the virtual mean waiting time (increase<br />
in sojourn time due to limited capacity) in node-2 is 0.1929<br />
tu for type-1, and 0.1990 for type-2. The waiting times of the<br />
two services are ofsame orderof size, but by allocatingbigger<br />
bandwidth to a type of traffic we can reduce the sojourn time<br />
(response time).<br />
Limitations of the algorithm are the number of states in the<br />
multi-dimensionalstate space of each node, and the numberof<br />
operations required during the multi-dimensional convolution<br />
of the nodes.<br />
V. FUTURE WORK<br />
The above model correspondsto a store-and-forwardpacket<br />
switched network with bottlenecks and to models considered<br />
in production systems. The model may be generalized in<br />
several ways. We way reserve at fixed minimum bandwidth<br />
to a certain type in each node this type visits, so that we have<br />
an end-to-end dedicated path with a guaranteed bandwidth as<br />
in ATM and MPLS networks. If the stream requires more<br />
bandwidth than guaranteed, then it has to compete with other<br />
streams in each node for additional bandwidth. The system<br />
will still be reversible.<br />
We may introduce an upper limit to the number of simultaneous<br />
connections of each type in a node. Then we may<br />
experience blocking in the nodes, and the system will only be<br />
reversible if we include blocked calls in the departure process.<br />
VI. CONCLUSIONS<br />
The convolution algorithm for the closed multi-rate queueing<br />
networks is a generalization of the classical convolution<br />
algorithm for queueing networks and has similar limitations<br />
in number of chains and customers in each chain.<br />
REFERENCES<br />
[1] J. R. Jackson, “Networks of waiting lines,” Operations Research, vol. 5,<br />
pp. 518–521, 1957.<br />
[2] P. J. Burke, “The output of a queueing system,” Operations Research,<br />
vol. 4, pp. 699–704, 1956.<br />
[3] R. R. P. Jackson, “Queueing systems with phase type service,” Operational<br />
Research Quarterly, vol. 5, pp. 109–120, 1954.<br />
[4] J. R. Jackson, “Jobshop–like queueing systems,” Management Science,<br />
vol. 10, no. 1, pp. 131–142, 1963.<br />
[5] L. Kleinrock, Communication nets: Stochastic message flow and delay.<br />
McGraw–Hill, 1964, dover Publications 1972.<br />
[6] F.Baskett, K.M.Chandy, R.R.Muntz, andF.G.Palacios, “Open, closed<br />
and mixed networks of queues with different classes of customers,”<br />
Journal of the ACM, pp. 248–260, Apr. 1975.<br />
[7] S. S. Lavenberg and M. Reiser, “Mean–value analysis of closed multichain<br />
queueing networks,” Journal of the Association for Computing<br />
Machinery, vol. 27, pp. 313–322, 1980.<br />
[8] J. F. C. Kingman, “Markov population processes,” J. Appl. Prob, vol. 6,<br />
pp. 1–18, 1969.<br />
[9] D. J.Sutton, “The application of reversible Markov population processes<br />
to teletraffic,” A.T.R., vol. 13, pp. 3–8, 1980.
INVERSEN AND KO: ALGORITHM FOR QUEUEING NETWORKS WITH MULTI-RATE TRAFFIC 7<br />
[10] T. Bonald and A. Proutière, “Insensitive bandwidth sharing in data<br />
networks,” Queueing Systems, vol. 44, pp. 69–100, 2003.<br />
[11] T. Bonald and J. Virtamo, “A recursive formula for multirate systems<br />
with elastic traffic,” IEEE Commun. Lett., vol. 9, pp. 752–755, Aug.<br />
2005.<br />
[12] R. Serfozo, Introduction to Stochastic Networks. Springer, Applications<br />
of Mathematics, 1999.<br />
[13] V. B. Iversen, “Reversible fair scheduling: the teletraffic theory revisited,”<br />
Springer Lecture Notes on Computer Science, vol. LNCS 4516,<br />
pp. 1135–1148, 2007, 20th International Teletraffic Congress, Ottawa,<br />
Canada.<br />
[14] E. Iliakis and G. Kardaras, “Resource allocation in next generation<br />
internet,” Master’s thesis, Technical University of Denmark, 2007.<br />
Villy Bæk Iversen received the M.Sc. degree in electrical engineering in<br />
1968 and the Ph.D. degree in teletraffic engineering in 1976, both from<br />
The Technical University of Denmark, where he is associate professor at<br />
Department of Photonics Engineering. He has worked in many developing<br />
countries as an ITU-expert. He is Professor Honoris Causa at Beijing<br />
University of Posts and Telecommunications, former vice-chairman of the<br />
International Advisory Committee of the International Teletraffic Congresses,<br />
and Danish member of IFIP TC6 (Digital Communication). His research<br />
interests include stochastic modelling, communication systems, and teletraffic<br />
engineering. He has published more than 150 papers and edited several<br />
conference proceedings. He is editor of the Teletraffic Engineering Handbook.<br />
King-Tim Ko received his B.Eng.(Hons) and Ph.D. degrees from The<br />
University of Adelaide, Australia. He has worked several years with the<br />
Telecom Australia Research Laboratories in Melbourne before joining the<br />
Department of Electronic Engineering of the City University of Hong Kong<br />
in 1986, and currently an Associate Professor of the same Department. In<br />
research, he is interested in the performance evaluation of communication<br />
networks and mobile networks.
8 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Retry Loss Models Supporting Elastic Traffic<br />
Ioannis D. Moscholios, Vassilios G. Vasilakis, John S. Vardakas and Michael D. Logothetis<br />
Abstract—We consider a single-link loss system of fixed capacity,<br />
which accommodates K service-classes of Poisson traffic with<br />
elastic bandwidth-per-call requirements. When a new call cannot<br />
be accepted in the system with its peak-bandwidth requirement,<br />
it can retry one or more times (single and multi-retry loss<br />
model, respectively) to be connected in the system with reduced<br />
bandwidth requirement and increased service time, exponentially<br />
distributed.Furthermore,ifitslast bandwidthrequirementisstill<br />
higher than the available link bandwidth, it can be accepted in<br />
the system by compressing not only the bandwidth of all inservice<br />
calls (of all service-classes) but also its last bandwidth<br />
requirement. The proposed model does not have a product form<br />
solution and therefore we propose an approximate recursive<br />
formula for the calculation of the link occupancy distribution<br />
and consequently call blocking probabilities. The accuracy of<br />
the proposed formula is verified by simulation and is found to<br />
be quite satisfactory.<br />
Index Terms—Markov chain, call blocking, recursive formula,<br />
retry, elastic services<br />
I. INTRODUCTION<br />
MULTI-RATE loss models are extensively used in the<br />
literature for the call-level QoS assessment of modern<br />
telecom networks. This assessment is critical not only for<br />
the bandwidth allocation among calls of different serviceclasses<br />
but also for the avoidance of over-dimensioning of<br />
a network. Despite of its importance, the call-level QoS<br />
assessment remains an open issue, due to the existence of<br />
elastictrafficinmoderntelecomnetworks.Bytheterm“elastic<br />
traffic” we mean calls whose assigned bandwidth can be<br />
compressed or expanded during their lifetime in the system.<br />
Modeling elastic traffic at call-level can be based on the<br />
classical Erlang Multirate Loss Model (EMLM) ([1], [2]])<br />
which has been widely used in wired (e.g. [3], [4], [5], [6]),<br />
wireless(e.g.[7],[8],[9],[10])andopticalnetworks(e.g.[11],<br />
[12],[13],[14],[15])tomodelsystemsthataccommodatecalls<br />
of differentservice-classeswith differenttrafficandbandwidth<br />
requirements.<br />
In the EMLM, calls of different service-classes arrive at a<br />
link of capacity C, following a Poisson process, and compete<br />
for the available link bandwidth under the complete sharing<br />
policy (all calls compete for all bandwidth resources). If upon<br />
arrival a call’s bandwidth requirement is not available, the call<br />
Ioannis D. Moscholios is with Dept. of Telecommunications Science and<br />
Technology, University of Peloponnese, 221 00 Tripolis, Greece. Email:<br />
idm@uop.gr<br />
Vassilios G. Vasilakis is with WCL, Dept. of Electrical and Computer<br />
Engineering, University of Patras, 265 04 Patras, Greece. Email: vasilak@wcl.ee.upatras.gr<br />
John S. Vardakas is with WCL, Dept. of Electrical and Computer<br />
Engineering, University of Patras, 265 04 Patras, Greece. Email: jvardakas@wcl.ee.upatras.gr<br />
Michael D. Logothetis is with WCL, Dept. of Electrical and Computer<br />
Engineering, University of Patras, 265 04 Patras, Greece. Email: m-<br />
logo@wcl.ee.upatras.gr<br />
is blocked and lost. Otherwise, it remains in the system for<br />
a generally distributed service time [1]. The analysis of the<br />
EMLM shows that the steady state distribution of in-service<br />
calls has a product form solution (PFS) [16]. Exploiting<br />
this fact, an accurate recursive formula (known as Kaufman-<br />
Roberts formula, KR formula) has been separately proposed<br />
by Kaufman [1] and Roberts [2] which determines the link<br />
occupancy distribution and simplifies the determination of<br />
call blocking probabilities (CBP). In [17], [18], the EMLM<br />
is extended to the retry models, in which blocked calls can<br />
immediately reattempt (one or more times – single-retry loss<br />
model (SRM) and multi-retry loss model (MRM), respectively)tobeconnectedbyrequiringlessbandwidthunits(b.u.),<br />
while increasing their service time which is exponentially<br />
distributed, so that the product (service time) by (bandwidth<br />
percall)remainsconstant.Aretrycallisblockedandlostfrom<br />
the system when its last bandwidth requirement is higher than<br />
the available link bandwidth.<br />
In this paper, we extend the models of [17], [18], by<br />
incorporatingelastictraffic.Wenametheproposedsingle-retry<br />
loss model, Extended SRM (E-SRM) and the multi-retry loss<br />
model, Extended MRM (E-MRM). In the proposed models,<br />
when a retry call attempts to be connected in the system and<br />
its last bandwidth requirement is higher than the available link<br />
bandwidth, the system accepts this call (contrary to [17], [18],<br />
where this call is lost) by compressing not only the bandwidth<br />
of all in-service calls (of all service-classes) but also the last<br />
bandwidth requirement of the retry call. The corresponding<br />
service times are increased so that the product (service time)<br />
by (bandwidth per call) remains constant. On the other hand<br />
when an in-service call, whose bandwidth is compressed,<br />
departs from the system then the remaining in-service calls<br />
(of all service-classes) expand their bandwidth. A retry call<br />
is blocked and lost from the system when the compressed<br />
bandwidth should be less than a minimum proportion (r min )<br />
of its required last-bandwidth. Note that r min is common<br />
for all service-classes. The compression/expansionmechanism<br />
together with the existence of retrials destroys reversibility in<br />
the proposed models and therefore no PFS exists. However,<br />
we propose approximate recursive formula for the calculation<br />
of the link occupancy distribution that simplifies the CBP<br />
determination. Simulation results validate the accuracy of the<br />
proposed formulas. In the case of no retrials for calls of all<br />
service-classes, the proposed models coincide with the model<br />
of [19] which has incorporated elastic traffic in the EMLM.<br />
We name this model, Extended EMLM (E-EMLM).<br />
The remainder of this paper is as follows: In Section II<br />
we review the SRM, MRM and E-EMLM. In Section III,<br />
we present the proposed E-SRM and E-MRM and provide<br />
formulasforthe approximatecalculationofthe linkoccupancy<br />
distribution and CBP. In Section IV, we present numerical and
MOSCHOLIOS et al.: RETRY LOSS MODELS SUPPORTING ELASTIC TRAFFIC 9<br />
simulation results in order to validate the models’ accuracy.<br />
We conclude in Section V.<br />
II. REVIEW OF THE RETRY LOSS MODELS AND THE<br />
E-EMLM<br />
A. Review of the single and multi-retry loss models<br />
Consider a link of capacity C b.u. that accommodates calls<br />
of K service-classes. Let j be the occupied link bandwidth,<br />
j = 0, 1, . . ., C. Calls of each service-class k (k = 1, . . ., K)<br />
arrive in the link according to a Poisson process with rate λ k<br />
and request b k b.u. If b k b.u. are available, a call of serviceclass<br />
k remains in the system for an exponentially distributed<br />
service-timewithmean µ −1<br />
k<br />
.Otherwise,thecallisblockedand<br />
retries immediately to be connected in the system with “retry<br />
parameters” (b kr , µ −1<br />
kr ) where b kr < b k and µ −1<br />
kr > µ−1 k<br />
. The<br />
SRM does not have a PFS and therefore the calculation of the<br />
link occupancy distribution, G(j), is based on an approximate<br />
recursive formula, [17], [18]:<br />
⎧<br />
1 for j=0<br />
1<br />
K∑<br />
⎪⎨ j<br />
α k b k G(j−b k )+<br />
G(j)=<br />
k=1<br />
for j=1, . . . , C ,<br />
K∑<br />
+ 1 j<br />
α kr b kr γ kr (j)G(j−b kr )<br />
⎪⎩ k=1<br />
0 otherwise<br />
(1)<br />
where α k = λ k µ −1<br />
k , α kr = λ k µ −1<br />
kr , γ kr(j) = 1 when j ><br />
C −(b k − b kr ), otherwise γ kr (j) = 0.<br />
Equation(1)isbasedontwo assumptions:1)theapplication<br />
of Local Balance (LB), which exists only in PFS models and<br />
2) the application of Migration Approximation (MA) which<br />
assumes that the occupied link bandwidth from retry calls is<br />
negligible when the link occupancy is below or equal to the<br />
retry boundary, i.e. when j ≤ C − (b k − b kr ). The existence<br />
of the MA in eq. (1) is expressed by the variable γ kr (j).<br />
The final CBP of a service-class k, denoted as B kr , is the<br />
probability of a call to be blocked with its retry bandwidth<br />
requirement and is given by:<br />
B kr =<br />
C∑<br />
G −1 G(j), (2)<br />
j=C−b kr +1<br />
where G = ∑ C<br />
j=0<br />
G(j) is the normalization constant and<br />
b kr > 0.<br />
In the MRM, a blocked call of service-class kcan have<br />
multiple retrials with “retry parameters” (b krl , µ −1<br />
kr l<br />
) for l =<br />
1, . . . , s(k), where b krs(k) < . . . < b kr1 < b k and µ −1<br />
kr s(k)<br />
><br />
. . . > µ −1<br />
kr 1<br />
> µ −1<br />
k<br />
. The MRM does not have a PFS and<br />
therefore the calculation of the link occupancy distribution,<br />
G(j), is based on an approximate recursive formula [18]:<br />
⎧<br />
1 for j = 0<br />
1<br />
K∑<br />
⎪⎨ j<br />
a k b k G(j−b k )+<br />
k=1<br />
G(j)= K∑<br />
+ 1 s(k)<br />
for j = 1, . . . , C ,<br />
∑<br />
j<br />
a krl b krl γ krl<br />
(j)G(j−b krl )<br />
⎪⎩ k=1 l=1<br />
0 otherwise<br />
(3)<br />
where: a krl = λ k µ −1<br />
kr l<br />
, γ krl (j) = 1, if C ≥ j > C − (b krl−1 −<br />
b krl ), otherwise γ krl (j) = 0.<br />
ThefinalCBP ofaservice-class k,denotedas B krs(k) ,isthe<br />
probabilityof a call to be blockedwith its last retry bandwidth<br />
requirement and is given by:<br />
B krs(k) =<br />
C∑<br />
G −1 G(j), (4)<br />
j=C−b krs(k) +1<br />
where G = ∑ C<br />
j=0 G(j) and b kr l<br />
> 0 for l = 1, . . ., s(k).<br />
B. Review of the E-EMLM<br />
Consider again a link of capacity C b.u. that accommodates<br />
Poisson arriving calls of K service-classes. A call of serviceclass<br />
k (k = 1, . . ., K) arrives in the system with rate λ k and<br />
requests b k b.u. (peak-bandwidth requirement). If j + b k ≤ C,<br />
the call is accepted in the system with its peak-bandwidth<br />
requirement and remains in the system for an exponentially<br />
distributed service time with mean µ −1<br />
k<br />
. If T ≥ j + b k > C<br />
the call is accepted in the system by compressing not only its<br />
bandwidthrequirementbut also the bandwidthofall in-service<br />
calls. The compressed bandwidth of the new service-class k<br />
call is:<br />
b ′ k = rb k = C j ′ b k, (5)<br />
where r ≡ r(n) = C/j ′ , j ′ = j + b k = nb + b k and T<br />
is the limit (in b.u.) up to which bandwidth compression is<br />
permitted.<br />
Similarly, the bandwidth of all in-service calls will be<br />
compressed and become equal to b ′ i = C j<br />
b ′ i for i = 1, . . ., K.<br />
After the compression of both the new call and the in-service<br />
callsthestateofthesystemis j = C.Theminimumbandwidth<br />
that a call of service-class k (either new or in-service) can<br />
tolerate is given by the expression:<br />
b ′ k,min = r minb k = C T b k, (6)<br />
where r min = C/T is the minimumproportionof the required<br />
peak-bandwidth and is common for all service-classes.<br />
Thismeansthatif uponarrivalofaservice-class k call, with<br />
peak-bandwidth requirement b k b.u., we have j = j + b k > T<br />
(or equivalently, j ′ > T or C/j ′ < r min ) then the call is<br />
blocked and lost without further affecting the system.<br />
After the bandwidth compression, calls increase their service<br />
time so that the product (service time) by (bandwidth per<br />
call) remains constant. Thus, due to bandwidth compression<br />
calls of service-class k may remain in the system more than<br />
µ −1<br />
k<br />
time units. Increasing the value of T, decreases r min and<br />
increases the delay of calls of service-class k (compared to<br />
the initial service time µ −1<br />
k<br />
). Therefore the value of T can be<br />
chosen so that this delay remains within acceptable levels.<br />
The compression/expansion of bandwidth destroys reversibility<br />
in the E-EMLM and therefore no PFS exists.<br />
However,in[19]anapproximaterecursiveformulaisproposed
10 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
which determines G(j)’s:<br />
⎧<br />
1 for j = 0<br />
⎪⎨<br />
1<br />
K∑<br />
G(j) =<br />
min(j,C)<br />
a k b k G(j − b k ) for j = 1, · · · , T ,<br />
⎪⎩<br />
k=1<br />
0 otherwise<br />
(7)<br />
where α k = λ k µ −1<br />
k .<br />
Equation (7) is based on a reversible Markov chain which<br />
approximates the bandwidth compression/expansion mechanism<br />
of the E-EMLM, described above. The LB equations<br />
of this Markov chain are of the form [19]:<br />
λ k P (n − k ) = n kµ k φ k (n)P (n), (8)<br />
where P (n) is the probability distribution of state n =<br />
n 1 , n 2 , . . ., n k , . . ., n K ), P (n − k<br />
) is the probability distribution<br />
of state n − k<br />
= (n 1, n 2 , . . . , n k−1 , n k − 1, n k+1 , . . . , n K ) and<br />
φ k (n) is a state dependent factor which describes: i) the<br />
compression factor of bandwidth and ii) the increase factor<br />
of service time of service-class k calls in state n, so that<br />
(service time) by (bandwidth per call) remains constant. In<br />
other words, φ k (n) has the same role with r(n) in eq. (5) or<br />
r min in eq. (6) but it may be different for each service-class.<br />
It is apparent now why the model of eq. (7) approximates the<br />
E-EMLM. The values of φ k (n) are given by:<br />
⎧<br />
⎪⎨ 1 , for nb ≤ C, n ∈ Ω<br />
x(n<br />
φ k (n) =<br />
− k )<br />
x(n)<br />
, for C < nb ≤ T, n ∈ Ω , (9)<br />
⎪⎩<br />
0 , otherwise<br />
where Ω = {n : 0 ≤ nb ≤ T and nb = ∑ K<br />
k=1 n kb k .<br />
In eq. (9), x(n) is a state multiplier, associated with state<br />
n, whose values, are chosen so that eq. (8) holds, [19]:<br />
⎧<br />
1 , for nb ≤ C, n ∈ Ω<br />
⎪⎨<br />
1<br />
K∑<br />
x(n) =<br />
C<br />
n k b k x(n − k<br />
⎪⎩<br />
) , for C < nb ≤ T, n ∈ Ω .<br />
k=1<br />
0 , otherwise<br />
(10)<br />
Having determined the values of G(j)’s we can calculate CBP<br />
according to the following formula:<br />
B k =<br />
T∑<br />
G −1 G(j), (11)<br />
j=T −b k +1<br />
where G = ∑ T<br />
j=0<br />
G(j) is the normalization constant.<br />
III. RETRY LOSS MODELS SUPPORTING ELASTIC TRAFFIC<br />
A. The extended single-retry loss model<br />
The proposed E-SRM is a non-PFS model that combines<br />
the characteristics of the SRM and the E-EMLM. In order<br />
to provide an approximate but recursive formula for the<br />
calculation of the link occupancy distribution we present the<br />
following simple example.<br />
Consider a link of capacity C b.u. that accommodates<br />
Poisson arriving calls of two service-classes with traffic<br />
parameters: (λ 1 , µ −1<br />
1 , b 1) f or the 1 st service-class and<br />
(λ 2 , µ −1<br />
2 , µ−1 2r , b 2, b 2r ) for the 2 nd service-class. Calls of the<br />
2 nd service-class have “retry parameters” with b 2r < b 2 and<br />
µ −1<br />
2r<br />
> µ −1<br />
2 . Let T be the limit up to which bandwidth<br />
compression is permitted for calls of both service-classes<br />
Although the E-SRM is a non-PFS model we will use the<br />
LB eq. (8), initially for calls of the 1 st service-class:<br />
λ 1 P (n − 1 ) = n 1µ 1 φ 1 (n)P (n), (12)<br />
for 1 ≤ nb ≤ T, where n = (n 1 , n 2 , n 2r ), n − 1 = (n 1 −<br />
1, n 2 , n 2r ) with n 1 ≥ 1 and<br />
⎧<br />
⎪⎨ 1 , for nb ≤ C, n ∈ Ω<br />
x(n<br />
φ 1 (n) =<br />
− 1 )<br />
x(n)<br />
, for C < nb ≤ T, n ∈ Ω (13)<br />
⎪⎩<br />
0 , otherwise<br />
with nb = j = ∑ 2<br />
k=1 (n kb k +n kr b kr ) = n 1 b 1 +n 2 b 2 +n 2r b 2r .<br />
Based on eq. (13) and multiplying both sides of eq. (12)<br />
with b 1 we have:<br />
a 1 b 1 x(n)P (n − 1 ) = n 1b 1 x(n − 1 )P (n), (14)<br />
where α 1 = λ 1 µ −1<br />
1 and the values of x(n) are given by:<br />
⎧<br />
1 , for nb ≤ C, n ∈ Ω<br />
⎪⎨ 1<br />
K∑<br />
x(n) = C<br />
n k b k x(n − k )+<br />
k=1<br />
, for C < nb ≤ T, n ∈ Ω .<br />
+n ⎪⎩ kr b kr x(n − kr )<br />
0 , otherwise<br />
(15)<br />
To derive the corresponding LB equations of 2 nd serviceclass<br />
calls consider that a call of the 2 nd service-class arrives<br />
in the system when the occupied link bandwidth is j b.u. with<br />
j = 0, 1, . . ., T. If j ≤ C − b 2 , the call will be accepted<br />
in the system with b 2 b.u. If j > C − b 2 , the call will be<br />
blocked with its b 2 requirement and will immediately try to<br />
be connected in the system with b 2r < b 2 . We consider three<br />
cases: 1) If j + b 2r ≤ C the retry call will be accepted in<br />
the system with b 2r . 2) If j + b 2r > T the retry call will be<br />
blocked and lost. 3) If C < j + b 2r ≤ T the retry call will be<br />
accepted in the system by compressing not only its bandwidth<br />
requirement b 2r but also the bandwidth of all in-service calls.<br />
The compressed bandwidth of the retry call is b ′ 2r = rb′ 2r =<br />
C<br />
j<br />
b ′ ′ 2r where r = C/j, j′ = j + b 2r = nb + b 2r . Similarly,<br />
the bandwidth of all in-service calls will be compressed (by<br />
the same factor) and become b ′ i = C j<br />
b ′ i for i = 1, 2. After the<br />
compression of both the new call and the in-service calls the<br />
state of the system is j = C. The minimum bandwidth that<br />
a call of the 2 nd service-class (either new or in-service) can<br />
tolerate is: b ′ 2r,min = r minb 2r = C T b 2r.<br />
Based on the previous discussion we consider the following<br />
LB equations for calls of the 2 nd service-class:<br />
a) λ 2 P (n − 2 ) = n 2µ 2 φ 2 (n)P (n), (16)<br />
for 1 ≤ nb ≤ C, where n = (n 1 , n 2 , n 2r ), n − 2 = (n 1, n 2 −<br />
1, n 2r ) with n 2 ≥ 1 and<br />
⎧<br />
⎪⎨ 1 , for nb ≤ C, n ∈ Ω<br />
x(n<br />
φ 2 (n) =<br />
− 2 )<br />
x(n)<br />
, for C < nb ≤ T, n ∈ Ω . (17)<br />
⎪⎩<br />
0 , otherwise
MOSCHOLIOS et al.: RETRY LOSS MODELS SUPPORTING ELASTIC TRAFFIC 11<br />
Based on eq. (17) and multiplying both sides of eq. (16)<br />
with b 2 we have:<br />
a 2 b 2 x(n)P (n − 2 ) = n 2b 2 x(n − 2 )P (n), (18)<br />
for 1 ≤ nb ≤ C, where α 2 = λ 2 µ 2 −1 and the values of x(n)<br />
are given by eq. (15).<br />
b) λ 2 P (n − 2r ) = n 2rµ 2r φ 2r (n)P (n), (19)<br />
for C − b 2 + b 2r < nb ≤ T, whereP (n − 2r ) is the probability<br />
distribution of state n − 2r = (n 1, n 2 , n 2r − 1) and<br />
⎧<br />
⎪⎨ 1 , for nb ≤ C, n ∈ Ω<br />
x(n<br />
φ 2r (n) =<br />
− 2r )<br />
x(n)<br />
, for C < nb ≤ T, n ∈ Ω . (20)<br />
⎪⎩<br />
0 , otherwise<br />
Based on eq. (20) and multiplying both sides of eq. (19)<br />
with b 2r we have:<br />
a 2r b 2r x(n)P (n − 2r ) = n 2rb 2r x(n − 2r )P (n), (21)<br />
for C − b 2 + b 2r < nb ≤ T, where α 2r = λ 2r µ −1<br />
2r and the<br />
values of x(n) are given by eq. (15).<br />
Equations (14), (18) and (21) lead to the following system<br />
of equations:<br />
a 1 b 1 x(n)P (n − 1 ) + a 2b 2 x(n)P (n − 2 ) =<br />
for 1≤ nb ≤ C − b 2 + b 2r ,<br />
(n 1 b 1 x(n − 1 ) + n 2b 2 x(n − 2 ))P (n), (22)<br />
a 1 b 1 x(n)P (n − 1 )+a 2b 2 x(n)P (n − 2 )+a 2rb 2r x(n)P (n − 2r )=<br />
(n 1 b 1 x(n − 1 )+n 2b 2 x(n − 2 )+n 2rb 2r x(n − 2r ))P (n), (23)<br />
for C − b 2 + b 2r < nb ≤ C,<br />
a 1 b 1 x(n)P (n − 1 ) + a 2rb 2r x(n)P (n − 2r ) =<br />
(n 1 b 1 x(n − 1 ) + n 2rb 2r x(n − 2r ))P (n) (24)<br />
or C < nb ≤ T.<br />
Equations (22)-(24) can be combined into one equation by<br />
assuming that calls with b 2r are negligible when 1 ≤ nb ≤<br />
C − b 2 + b 2r (MA) and calls with b 2 are negligible when<br />
C < nb ≤ T:<br />
a 1 b 1 x(n)P (n − 1 ) + a 2b 2 γ 2 (nb<br />
nb)x(n)P (n − 2 )+<br />
+a 2r b 2r γ 2r (nb<br />
nb)x(n)P (n − 2r ) =<br />
(n 1 b 1 x(n − 1 ) + n 2b 2 x(n − 2 ) + n 2rb 2r x(n − 2r ))P (n),(25)<br />
where γ 2 (nb<br />
nb) = 1 for 1 ≤ nb ≤ C, otherwise γ 2 (nb<br />
nb) = 0<br />
and γ 2r (nb<br />
nb) = 1 for C − b 2 + b 2r < nb ≤ T, otherwise<br />
γ 2r (nb<br />
nb) = 0.<br />
Note that the approximations introduced in eq. (25) are<br />
similar to those introduced in the single- threshold model of<br />
[18].<br />
Since x(n) = 1, when 0 ≤ nb ≤ C, it is proved in [18]<br />
that:<br />
a 1 b 1 G(j − b 1 ) + a 2 b 2 G(j − b 2 )+<br />
+a 2r b 2r γ 2r (j)G(j − b 2r ) = jG(j), (26)<br />
for 1 ≤ j ≤ C and γ 2r (j) = 1 for C −b 2 +b 2r < j, otherwise<br />
γ 2r (j) = 0.<br />
Toproveeq.(26),theMAisneeded,whichassumesthatthe<br />
population of retry calls of the 2 nd service-class is negligible<br />
in states j ≤ C − b 2 + b 2r .<br />
When C < nb ≤ T and based on eq. (15), eq. (25) can be<br />
written as:<br />
a 1 b 1 P (n − 1 ) + a 2rb 2r γ 2r (nb<br />
nb)P (n − 2r ) = CP (n). (27)<br />
To introduce the link occupancy distribution G(j) in eq.<br />
(27) we sum both sides of eq. (27) over the set of states Ω j =<br />
{n ∈ Ω |nb<br />
= j}:<br />
∑<br />
a 1 b 1 P (n − 1 ) + a 2rb 2r γ 2r (nb<br />
nb) ∑<br />
P (n − 2r ) =<br />
{n|nb<br />
nb=j}<br />
{n|nb<br />
nb=j}<br />
C ∑<br />
P (n). (28)<br />
{n|nb<br />
nb=j}<br />
Since by definition ∑ n∈Ω j<br />
P (n) = G(j), eq. (28) is written<br />
as:<br />
a 1 b 1 G(j − b 1 ) + a 2r b 2r γ 2r (j)G(j − b 2r ) = CG(j), (29)<br />
where γ 2r (j) = 1 for C < j ≤ T.<br />
Thecombinationofeq.(26)andeq.(29)givesthefollowing<br />
approximate recursive formula for the calculation of G(j)’s in<br />
the case of two service-classes when only calls of the 2 nd<br />
service-class have “retry parameters”:<br />
G(j) =<br />
1<br />
min(j,C) [a 1b 1 G(j − b 1 ) + a 2 b 2 γ 2 (j)G(j − b 2 )+<br />
+a 2r b 2r γ 2r (j)G(j − b 2r )] (30)<br />
for 1 ≤ j ≤ T, where γ 2 (j) = 1 for 1 ≤ j ≤ C, otherwise<br />
γ 2 (j) = 0 and γ 2r (j) = 1 for C −b 2 +b 2r < j ≤ T, otherwise<br />
γ 2r (j) = 0.<br />
In the case of K service-classes and assuming that all<br />
service-classes may have “retry parameters”, eq. (30) takes<br />
the general form:<br />
⎧<br />
1 , for j=0<br />
K∑<br />
1 ⎪⎨ min(j,C)<br />
α k b k γ k (j)G(j−b k )+<br />
G(j)=<br />
k=1<br />
, for j=1, . . . , T,<br />
K∑<br />
+ 1<br />
min(j,C)<br />
α kr b kr γ kr (j)G(j−b kr )<br />
⎪⎩<br />
k=1<br />
0 , otherwise<br />
(31)<br />
where γ k (j) = 1 for 1 ≤ j ≤ C, otherwise γ k (j) = 0 and<br />
γ kr (j) = 1 for C − b k + b kr < j ≤ T, otherwise γ kr (j) = 0.<br />
The final CBP of a service-class k, B kr , is the probability<br />
of a call to be blocked with its retry bandwidth requirement:<br />
T∑<br />
B kr = G −1 G(j), (32)<br />
where G = ∑ T<br />
b kr > 0.<br />
j=0<br />
j=T −b kr +1<br />
B. The extended multi-retry loss model<br />
G(j) is the normalization constant and<br />
Similar to the MRM, in the E-MRM a blocked call of<br />
service-class k can have more than one “retry parameters”<br />
(b krl , µ −1<br />
kr l<br />
) for l = 1, . . ., s(k), where b krs(k) < ... <<br />
b kr1 < b k and µ −1<br />
kr s(k)<br />
> ... > µ −1<br />
kr 1<br />
> µ −1<br />
k<br />
. The E-MRM
12 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Fig. 1. CBP for the 1 st service-class.<br />
Fig. 2. CBP for the 2 nd service-class.<br />
does not have a PFS and therefore the calculation of the<br />
occupancy distribution, G(j), is based on an approximate<br />
recursive formula whose proof is similar to that of eq. (31):<br />
⎧<br />
(<br />
1 ,for j=0<br />
∑ K<br />
1 ⎪⎨<br />
a k b k γ k<br />
(j)G(j−b k )+<br />
min(j,C)<br />
G(j)=<br />
∑<br />
+ K<br />
⎪⎩ k=1<br />
s(k)<br />
∑<br />
l=1<br />
k=1<br />
), for j=1, . . . , T,<br />
a krl b krl γ krl<br />
(j)G(j−b krl )<br />
0 , otherwise<br />
(33)<br />
where: a krl = λ k µ −1<br />
kr l<br />
, γ k (j) = 1 for 1 ≤ j ≤ C, otherwise<br />
γ k (j) = 0 and γ krl (j) = 1, if j > C − (b krl−1 − b krl ),<br />
otherwise γ krl (j) = 0.<br />
ThefinalCBP ofaservice-class k,denotedas B krs(k) ,isthe<br />
probabilityof a call to be blocked with its last retry bandwidth<br />
requirement and is given by:<br />
B krs(k) =<br />
T∑<br />
G −1 G(j), (34)<br />
j=T −b krs(k) +1<br />
where G = ∑ T<br />
j=0 G(j) and b krl > 0 for l = 1, . . ., s(k).<br />
IV. EVALUATION<br />
In this section, we present an application example and<br />
compare the analytical CBP probabilities with those obtained<br />
by simulation. The latter is based on SIMSCRIPT II.5 [20].<br />
Simulation results are mean values of 7 runs with 95%<br />
confidence interval. Since, the resultant reliability ranges of<br />
the measurements are small enough we present only mean<br />
values.<br />
Consider a link of capacity C = 80 b.u. that accommodates<br />
three service-classes of elastic calls which require b 1 = 1<br />
b.u., b 2 = 2 b.u. and b 3 = 6 b.u., respectively. All calls<br />
arrive in the system according to a Poisson process. The<br />
call holding time is exponentially distributed with mean value<br />
µ −1<br />
1 = µ −1<br />
2 = µ −1<br />
3 = 1. The initial values of the offered<br />
traffic-load are: α 1 = 20 erl, α 2 = 6 erl and α 3 = 2 erl.<br />
Fig. 3. CBP for the 3 rd service-class (retry calls with b 3r2 ).<br />
Calls of the 3 rd service-classmay retry two times with reduced<br />
bandwidth requirement: b 3r1 = 5 b.u. and b 3r2 = 4 b.u. and<br />
increased service time so that α 3 b 3 = a 3r1 b 3r1 = a 3r2 b 3r2 . In<br />
the x-axis of all figures, we assume that α 3 remains constant<br />
while α 1 , α 2 increase in steps of 1.0 and 0.5 erl, respectively.<br />
The last value of α 1 = 26 erl while that of α 2 = 9 erl.<br />
We consider three different values of T: a) T = C = 80<br />
b.u., where no bandwidth compression takes place. In that<br />
case,theproposedE-MRMgivesexactlythesameCBP results<br />
with the MRM of [18], b) T = 82 b.u. where bandwidth<br />
compression takes place and r min = C/T = 80/82 and c)<br />
T = 84 b.u. where bandwidth compression takes place and<br />
r min = C/T = 80/84.<br />
In Fig. 1, we present the analytical and simulation CBP<br />
results of the 1 st service-class for all values of T. Similar<br />
results are presented in Fig. 2, for the 2 nd service-class and in<br />
Fig. 3 for the 3 rd service-class (CBP of calls with b 3r2 ). All<br />
figures presented herein show that: i) the model’s accuracy<br />
is absolutely satisfactory compared to simulation and ii) the<br />
increase of T above C results in a CBP decrease due to the
MOSCHOLIOS et al.: RETRY LOSS MODELS SUPPORTING ELASTIC TRAFFIC 13<br />
existence of the compression mechanism.<br />
V. CONCLUSION<br />
We propose multirate loss models that support elastic traffic<br />
under the assumption that Poisson arriving calls have the<br />
ability, when blocked with their initial bandwidthrequirement,<br />
to retry to be connected in the system one (E-SRM) or more<br />
times(E-MRM)withreducedbandwidthandincreasedservice<br />
time requirements. Furthermore, if a retry call is blocked<br />
with its last bandwidth requirement, it can still be accepted<br />
in the system by compressing not only the bandwidth of<br />
all in-service calls (of all service-classes) but also its last<br />
bandwidth requirement. The proposed models do not have<br />
a PFS and therefore we propose approximate but recursive<br />
formulasforthe CBP calculation.Simulation results verifythe<br />
analytical results. As a future work, we will examine multirate<br />
retry loss models that support both elastic and adaptive traffic<br />
(e.g. adaptive video). Adaptive calls can tolerate bandwidth<br />
compression, but their service time cannot be altered.<br />
REFERENCES<br />
[1] J. S. Kaufman, “Blocking in a shared resource environment,” IEEE<br />
Trans. Commun., vol. 29, no. 10, pp. 1474–1481, Oct. 1981.<br />
[2] J. W. Roberts, Performance of Data Communications systems and their<br />
applications. North Holland, Amsterdam, 1981, ch. A service system<br />
with heterogeneous user requirements, pp. 423–431.<br />
[3] M. Logothetis and G. Kokkinakis, “Path Bandwidth Management for<br />
Large Scale Telecom Networks,” IEICE Trans. Commun., vol. E83-B,<br />
no. 9, pp. 2087–2099, Sep. 2000.<br />
[4] M. Stasiak and M. Głąbowski, “A simple approximation of the link<br />
model with reservation by a one-dimensional Markov chain,” Journal of<br />
Performance Evaluation, vol. 41, no. 2-3, pp. 195–208, Jul. 2000.<br />
[5] I. Moscholios, M. Logothetis, and G. Kokkinakis, “Call-burst blocking<br />
of ON-OFF traffic sources with retrials under the complete sharing<br />
policy,” Journal of Performance Evaluation, vol. 59, no. 4, pp. 279–<br />
312, Mar. 2005.<br />
[6] W. Bziuk, “Approximate state probabilities in large shared multirate<br />
loss systems with an application to trunk reservation,” European Trans.<br />
Telecom, vol. 16, no. 3, pp. 205–216, May 2005.<br />
[7] W. Choi and J. Y. Kim, “Joint Erlang Capacity of DS/CDMA Forward<br />
Link Based on Resource Sharing Algorithm,” IEICE Trans. Fundamentals,<br />
vol. E84-A, no. 6, pp. 1406–1412, Jun. 2001.<br />
[8] P. Fazekas, S. Imre, and M. Telek, “Modelling and Analysis of Broadband<br />
Cellular Networks with Multimedia Connections,” Telecommunication<br />
systems, vol. 19, no. 3-4, pp. 263–288, Mar. 2002.<br />
[9] D. Staehle and A. Mäder, “An Analytic Approximation of the Uplink<br />
Capacity in aUMTSNetwork with Heterogeneous Traffic,”in Proc.18th<br />
International Teletraffic Congress (ITC), Berlin, Sep. 2003, pp. 81–90.<br />
[10] M. Głąbowski, M. Stasiak, A. Wiśniewski, and P. Zwierzykowski,<br />
“Blocking Probability Calculation for Cellular Systems with WCDMA<br />
Radio Interface Servicing PCT1 and PCT2 Multirate Traffic,” IEICE<br />
Trans. Commun, vol. E92-B, pp. 1156–1165, Apr. 2009.<br />
[11] A. Washington and H. Perros, “Call blocking probabilities in a trafficgroomed<br />
tandem optical network,” Computer Networks, vol. 45, no. 3,<br />
pp. 281–294, Jun. 2004.<br />
[12] A. Sahasrabudhe and D. Manjunath, “Performance of optical burst<br />
switched networks: A two moment analysis,” Computer Networks,<br />
vol. 50, no. 18, pp. 3550–3563, Dec. 2006.<br />
[13] J. Vardakas, V. Vassilakis, and M. Logothetis, “Blocking Analysis in Hybrid<br />
TDM-WDM Passive Optical Networks,” in Proc. 5th Int. Working<br />
Conference onPerformance Modelling and Evaluation ofHeterogeneous<br />
Networks (HET-NETs 2008), Karlskrona, Sweden, Feb. 2008.<br />
[14] K. Kuppuswamy and D. Lee, “An analytic approach to efficiently<br />
computing call blocking probabilities for multiclass WDM networks,”<br />
IEEE/ACM Trans. Netw., vol. 17, no. 2, pp. 658–670, Apr. 2009.<br />
[15] J.S.Vardakas, I.D.Moscholios, M.D.Logothetis, andV.G.Stylianakis,<br />
“An Analytical Approach for Dynamic Wavelength Allocation in WDM-<br />
TDMA PONs Servicing ON-OFF Traffic,” IEEE/OSA J. Opt. Commun.<br />
Netw., vol. 3, no. 4, pp. 347–358, Apr. <strong>2011</strong>.<br />
[16] K. W. Ross, Multiservice loss models for broadband telecommunication<br />
networks. Springer, Berlin, 1995.<br />
[17] J.S.Kaufman, “Blocking in aCompletely Shared Resource Environment<br />
With State Dependent Resource and Residency Requirements,” in Proc.<br />
IEEE INFOCOM’92, 1992, pp. 2224–2232.<br />
[18] ——, “Blocking with retrials in a completely shared resource environment,”<br />
Journal of Performance Evaluation, vol. 15, no. 2, pp. 99–113,<br />
Jun. 1992.<br />
[19] G. Stamatelos and V. Koukoulidis, “Reservation – Based Bandwidth<br />
Allocation in a Radio ATM Network,” IEEE/ACM Trans. Netw., vol. 5,<br />
no. 3, pp. 420–428, Jun. 1997.<br />
[20] “Simscript II.5,” [online], http://www.simscript.com.<br />
Ioannis D. Moscholios was born in Athens, Greece, in 1976. He received the<br />
Dipl.-Eng. degree in Electrical & Computer Engineering from the University<br />
of Patras, Patras, Greece, in 1999, the M.Sc. degree in Spacecraft Technology<br />
& Satellite Communications from the University College London, UK, in<br />
2000 and the Ph.D. degree in Electrical & Computer Engineering from the<br />
University of Patras, in 2005. From 2005 to 2009 he was aResearch Associate<br />
at the Wire Communications Laboratory, Dept. of Electrical & Computer<br />
Engineering, University of Patras. Currently, he is a Lecturer in the Dept.<br />
of Telecommunications Science and Technology, University of Peloponnese,<br />
Greece. His research interests include simulation and performance analysis of<br />
communication networks. He has published over 65 papers in international<br />
journals/ conferences and has over 120 third-part citations. He is a member<br />
of the Technical Chamber of Greece (TEE).<br />
Vassilios G. Vassilakis was born in Sukhumi, Georgia, in 1978. He received<br />
his Dipl.-Eng. degree in Computer Engineering & Informatics and PhD in<br />
Electrical and Computer Engineering, both from the University of Patras,<br />
Greece, in 2003 and <strong>2011</strong>, respectively. Hi is involved in national research<br />
and R&D projects. His main research interests are in the areas of teletraffic<br />
engineering and QoS in wireless networks. He has published over 20 papers<br />
in international journals/ conferences and has over 40 thirs-part citations. He<br />
is a member of the Technical Chamber of Greece (TEE).<br />
John S.Vardakas was born in Alexandria, Greece, in 1979. He received his<br />
Dipl.-Eng. degree in Electrical & Computer Engineering from the Democritus<br />
University of Thrace, Greece, in 2004. Since 2005 he is a Ph.D student at the<br />
Wire Communications Laboratory, Department of Electrical and Computer<br />
Engineering, University of Patras, Greece. His research interests include<br />
teletraffic engineering in optical and wireless networks. He is a member of<br />
the Optical Society of America (OSA) and the Technical Chamber of Greece<br />
(TEE).<br />
Michael D. Logothetis was born in Stenies, Andros, Greece, in 1959. He<br />
received his Dipl.-Eng. degree and Doctorate in Electrical Engineering, both<br />
from the University of Patras, Patras, Greece, in 1981 and 1990, respectively.<br />
From 1982 to 1990, he was a Teaching and Research Assistant at the<br />
Laboratory of Wire Communications, University of Patras, and participated in<br />
many national and EU research programmes, dealing with telecommunication<br />
networks, as well as with office automation. From 1991 to 1992 he was<br />
Research Associate in NTT’s Telecommunication Networks Laboratories,<br />
Tokyo, Japan. Afterwards, he was a Lecturer in the Dept. of Electrical &<br />
Computer Engineering of the University of Patras, and recently he has been<br />
elected Professor in the same Department. His research interests include<br />
teletraffic theory and engineering, traffic/network control, simulation and<br />
performance optimization of telecommunications networks. He has published<br />
over 120 conference/journal papers and has over 220 third-part citations.<br />
He has published a teletraffic book in Greek. He has organised the 5th<br />
International Conference on Communications Systems, Networks and Digital<br />
Signal Processing, CSNDSP 2006. He served/is serving on the Technical<br />
Program Committee of several international conferences while he organizes<br />
and chairs several technical sessions. He has become a Guest Editor in three<br />
journals, while participates in the Editorial Board of journals. He is a member<br />
of theIARIA (Fellow), IEEE(Senior), IEICE,ETRI,FITCEand the Technical<br />
Chamber of Greece (TEE).
14 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Damming the Torrent: Adjusting BitTorrent-like<br />
Peer-to-Peer Networks to Mobile and Wireless<br />
Environments<br />
Philipp M. Eittenberger, Seungbae Kim, and Udo R. Krieger<br />
Abstract—Mobile peer-to-peer (P2P) traffic is rapidly growing,<br />
but present P2P applications are not specifically designed to<br />
operate under mobile conditions. To assess the performance of<br />
the prevalent file sharing application BitTorrent in a mobile<br />
WiMAX network, we performed a measurement and analysis<br />
campaign. In this study, we use the obtained measurement traces<br />
to further investigate specific characteristics of this P2P network.<br />
In particular, we analyze the distribution of its peer population<br />
under mobile conditions and present a general classification<br />
scheme for peer populations in BitTorrent-like P2P networks.<br />
Further, we propose a simple heuristic to bound the outdegree<br />
of BitTorrent-like P2P systems when operating in mobile<br />
environments.<br />
Index Terms—BitTorrent, P2P, Traffic Analysis, WiMAX.<br />
I. INTRODUCTION<br />
IN recent years P2P networks evolved to the dominant<br />
source of traffic in the Internet [1]. Along with the evolution<br />
of a new generation of wireless networks, like WiMAX and<br />
3GPP LTE, a shift to an increased user mobility can be<br />
witnessed. This development implies that users will have<br />
more and more opportunities to use all their accustomed<br />
applications wherever they are. Hence, mobile P2P traffic<br />
will continue to grow rapidly in the next years (cf. [2]), but<br />
current P2P networks have been designed to operate in wired<br />
networks under stable conditions. Thereby, a growing need for<br />
new traffic models supporting P2P applications in a mobile<br />
environment and redesigned, mobility aware P2P protocols is<br />
emerging.<br />
Kim et al. [3] conducted several measurements in Korea<br />
Telecom’s WiMAX network in Seoul, Korea, to investigate<br />
the behavior of BitTorrent, one of the most popular P2P<br />
file sharing networks under true mobile conditions. For this<br />
purpose several measurements have been carried out while<br />
driving by bus and by subway through Seoul at several days.<br />
At each single measurement, a popular torrent with video<br />
content has been chosen for download and all the packet<br />
headers of the whole transmission have been captured. To<br />
allow comparison, the same files have been downloaded under<br />
static conditions in the WiMAX network and exemplary with<br />
Manuscript received April 3, <strong>2011</strong>.<br />
Ph. M. Eittenberger and U. R. Krieger are with the Faculty of Information<br />
Systems and Applied Computer Science, Otto-Friedrich University, Bamberg,<br />
Germany (e-mail: philipp.eittenberger@uni-bamberg.de).<br />
S. Kim was with the School of Computer Science and Engineering, Seoul<br />
National University, Seoul, Korea. He is now with the Future Communications<br />
Team, KAIST Institute for Information Technology Convergence, Daejeon,<br />
Korea (e-mail: sbkim@itc.kaist.ac.kr).<br />
a host in a campus network connected to the Internet by<br />
an Ethernet link. In a WiMAX network disruptions occur<br />
during hand-overs and the wireless link conditions fluctuate<br />
due to signal fading. The main purpose of the measurement<br />
campaign was to investigate the impact of mobile conditions<br />
on the behavior of BitTorrent. In this paper we are extending<br />
our previous results presented at the 1st European Teletraffic<br />
Seminar (ETS<strong>2011</strong>) [4].<br />
The outline of the paper is as follows. We start with a<br />
discussion of related work in Section II. An introduction to<br />
the BitTorrent network, its operational behavior and a presentation<br />
of our measurement methodology follow in Section<br />
III. Subsequently, we present a classification scheme for peer<br />
populations in BitTorrent-like P2P systems and report the<br />
results of our analysis in Section IV. We use our insights into<br />
the BitTorrent peer population to propose a simple heuristic<br />
to limit the number of open connections in Section V. Finally,<br />
we conclude the paper with additional implications for the<br />
adaption of BitTorrent-like systems to wireless networks.<br />
II. RELATED WORK<br />
Recently, several analytical models have been proposed to<br />
yield a deeper understanding of the P2P data dissemination<br />
among peers (e.g. [5] or [6]). Analytical models can provide<br />
precious insights, but are typically based on unrealistic assumptions,<br />
like global system knowledge. Therefore, we performed<br />
a measurement campaign, to reveal the actual system<br />
characteristics of the complex dissemination network of Bit-<br />
Torrent and to yield novel insights, which can be incorporated<br />
into new analytical models. Sen and Wang [7] performed a<br />
large measurement campaign to analyze different P2P file<br />
sharing networks. In their paper they tried to characterize<br />
the peers of a particular mesh-pull P2P file sharing network.<br />
To examine their distribution, they plotted the traffic volume,<br />
duration of on-time and number of connections of the top 10<br />
% of the investigated peer population on a log-log scale. From<br />
the plot they concluded that the distribution is heavy-tailed, but<br />
does not follow a power law. Subsequently, they concluded<br />
“that P2P traffic does not obey power laws”. Despite the fact<br />
that there can be found numerous other measurement studies<br />
concerning the BitTorrent network (cf. [8] or [9] among many<br />
others), the first study, which analyzed the performance of<br />
such a file sharing network in a WiMAX environment under<br />
true mobile conditions, was performed by Kim et al. [3]. In<br />
this paper, we use the data traces obtained by Kim et al. to
EITTENBERGER et al.: DAMMING THE TORRENT 15<br />
reveal deeper insights into the complex dissemination network<br />
of BitTorrent. We investigate the data access patterns between<br />
a single peer and the remote peer population, and show that<br />
in this environment certain parts of the peer population can be<br />
modeled by a power law distribution.<br />
III. BITTORRENT<br />
As already mentioned, BitTorrent is one of the most popular<br />
P2P file sharing networks and itself a major source of the<br />
Internet traffic nowadays. The protocol specification is publicly<br />
available on the BitTorrent website [10]. Several distinct client<br />
applications, which implement the protocol, are available.<br />
Despite the fact that they differ in particular implementation<br />
details, they are able to exchange data with each other.<br />
BitTorrent is a representative of a mesh-pull P2P network, it<br />
builds an overlay on top of the transport network based on a<br />
mesh topology and the data dissemination is realized by a pull<br />
mechanism, i.e. on request. To enable fast data dissemination,<br />
BitTorrent uses the so called swarming technique, where each<br />
shared file is divided into smaller parts, called chunks, and<br />
then transmitted to (respectively received from) a multitude of<br />
peers (the swarm).<br />
A. BitTorrent Operations in Detail<br />
A torrent is the set of peers collaborating to share a single<br />
file. A peer can be in two different states, if the peer possesses<br />
already the complete file and uploads it to other peers, then it is<br />
called seeder. Otherwise, if it is still in the downloading phase,<br />
it is called leecher. To join a torrent, a peer needs some metainformation<br />
about it. This information is provided by a torrent<br />
file, containing all the information necessary to download the<br />
content, e.g. the number of chunks, hashes to verify their<br />
correctness, the IP of the tracker server etc. Typically this<br />
torrent file is retrieved from a website. Upon reception the<br />
peer contacts the tracker server in the bootstrapping process,<br />
which provides an initial list of the latest remote peers. The<br />
peers of a torrent exchange messages to indicate the chunks<br />
that they already possess.<br />
Two vital optimization problems have to be solved to achieve a<br />
high data throughput in a mesh-pull P2P network. At first, the<br />
choice which pieces should be requested from other peers, and<br />
subsequently, the selection which peers should be contacted for<br />
the data. BitTorrent addresses the first problem with a rarestfirst<br />
algorithm, i.e. each peer maintains a list of pieces, that<br />
each of the remote peers has and builds an index of the pieces<br />
with the least number of copies. The rarest pieces are then<br />
requested from the remote peers. However, when a download<br />
is almost completed, the peer does not use the rarest-first<br />
algorithm; instead it sends requests for all of its missing pieces<br />
to all of its remote peers to increase the throughput. This is<br />
called end-game mode. The peer selection strategy is handled<br />
by the so called choking algorithm. To encourage peers to<br />
contribute their resources for the data dissemination, a tit-fortat<br />
mechanism is implemented to impede free-riding, i.e. peers<br />
not contributing data to the network should not be able to<br />
achieve high download rates. Instead, this choking algorithm<br />
provides sharing incentives by rewarding peers who contribute<br />
data to the system. The algorithm determines the selection of<br />
peers to exchange data with. Peers that upload data at high<br />
rates are preferred. Once per choking period, usually every<br />
ten seconds, a peer evaluates the transmission rates from all<br />
the connected peers and selects a fixed number of the fastest<br />
ones, depending on its upload capacity. It will only upload<br />
data to these unchoked peers in this period. Data requests<br />
from other peers are denied in this period, i.e. those peers<br />
are choked. Another important part of the algorithm is the<br />
optimistic unchoking behavior: every 30 seconds one peer is<br />
randomly chosen and will be unchoked. This is meant to<br />
explore new peers with even higher upload capacities and<br />
as a side effect ensures data dissemination to low-capacity<br />
peers. Once a leecher has finished the download and enters<br />
the seeding state, it follows a different unchoke strategy. In<br />
most of the implementations peers in seed state unchoke peers<br />
with the highest download capacity to optimize the general<br />
dissemination performance of the network and to maintain<br />
high upload utilization.<br />
B. Measurement of BitTorrent in a WiMax Network<br />
The analyzed measurement traces have been captured in<br />
March, 2010 in Seoul, Korea and were firstly presented in [3].<br />
The measurements have been carried out on four different days<br />
in parallel, i.e. on each day all the measurement runs started<br />
at the same time, in four different scenarios, three WiMax<br />
settings and one in a Ethernet environment. The throughput<br />
of the WiMax network ranges from 30 to 50 Mbps, and a<br />
base station typically covers a radius between 1 and 5 km.<br />
Three laptops equipped with WiMax USB dongles where<br />
used for the WiMax measurements and one desktop computer<br />
for the reference Ethernet measurement. Vuze [11] was used<br />
as the BitTorrent client in all measurement runs. For the<br />
measurement some popular sitcoms served as torrents, which<br />
had at least 300 seeds, with a file size ranging from 300 to 400<br />
MB. An important fact to note is the over-provisioning with<br />
seeding capacity in the measurement runs to ensure a high<br />
data throughput. Of course, on each day the same torrent was<br />
chosen to download in all four different settings. One WiMax<br />
measurement was conducted stationary, i.e. the measurement<br />
peer was located statically about 800 meters away from its<br />
base station. Therefore, the signal strength was stable, but<br />
not strong. For the next scenario one peer traveled about 12<br />
km in a subway train through Seoul while conducting the<br />
measurement. The duration was about 20 minutes. In the last<br />
WiMax scenario the peer took a bus ride through Seoul, which<br />
lasted about 30 minutes and the distance of the route was about<br />
11 km. In both mobile scenarios the link quality fluctuated<br />
highly and it even occurred that the WiMax connection was<br />
completely lost due to hand-overs in between the base stations,<br />
and thereby, the peer got a new IP address in some of the<br />
measurement runs. An Ethernet measurement in a 100 Mbps<br />
LAN was conducted on the university campus as a reference<br />
to allow comparison. For a more detailed description, we refer<br />
to Kim et al.’s study [3]. The main results of this study will<br />
be also presented shortly. The WiMax peers suffer from poor<br />
connectivity, the connections to the peers are less stable and
16 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
the connection duration is shorter as opposed to the reference<br />
Ethernet measurement. In all WiMax settings the duration of<br />
the file download took 4 to 5 times longer than in the Ethernet<br />
measurement. The signaling traffic has been increased in all<br />
WiMax measurements by a factor of 100 %. The mean of the<br />
average download rates in the WiMax measurements ranges<br />
from roughly 240 to 400 Kbps, as opposed to a mean of<br />
1930 Kbps in the Ethernet runs. One reason for the poor<br />
performance of BitTorrent in the WiMax network is certainly<br />
the fluctuating signal strength, but the hand-overs in the mobile<br />
measurements have a negative impact on the performance of<br />
TCP transmissions too.<br />
IV. CLASSIFICATION OF THE PEER POPULATION<br />
A peer in a mesh-pull P2P network is typically connected<br />
to more than a thousand different peers in just tens of<br />
minutes. Hence, it is vital for the understanding of the inherent<br />
hierarchical structure of a mesh-pull topology to explore the<br />
preference relationship among the peer population. We have<br />
shown in our previous work (cf. [12]) that it is possible to<br />
classify the peer population of P2P streaming networks. In<br />
this study we investigate the aggregated conversation model<br />
of superimposed flows in inbound and outbound direction to<br />
a home peer, i.e. the traffic volume generated by the superimposed<br />
flows from and to the distinct feeder peers p i of the<br />
dissemination flow graph G V . To clarify our terminology, we<br />
regard a conversation as bidirectional data exchange between<br />
two endpoints, in this case peers. A flow represents the directed<br />
data transfer, which can be identified by the traffic direction,<br />
the IPs of both peers, the port numbers of the used transport<br />
protocol and a given time-out value to differentiate widely<br />
disparate flows. A stream φ represents the aggregated set of<br />
flows sent from one peer to another.<br />
Using the captured data traces, it allows us to describe the<br />
exchange of chunk sequences among the peers by appropriate<br />
teletraffic models. We use the amount of exchanged traffic,<br />
i.e. received and sent bytes, in all streams as a metric for<br />
the further classification of the peer population. Since we are<br />
interested in the contribution of each peer and the variance<br />
among different peers’ contribution, it is a convenient choice<br />
to rank all peers according to their contribution and then<br />
identify those with high contributions. Therefore, we sort<br />
each stream by the number of transferred bytes in descending<br />
order. If we arrange the streams φ(p i , p j ) according to their<br />
number of exchanged bytes on a log-log scale, where p j is<br />
representing the home peer, i.e. the measurement host, and<br />
p i , i ∈ {1, ...., n}, denotes the feeding peer population, we<br />
can realize a hierarchy of the peer population. To clarify our<br />
concept, we use the WiMax bus trace of March, 17 as an<br />
example. By using the rank ordering technique, we plotted the<br />
distribution of the incoming streams in Fig. 1. Several distinct<br />
regions can be spotted in this distribution. The profitable<br />
region consists of the top peers and it’s body resembles in<br />
all captured WiMAX traces asymptotically a straight line. In<br />
this example, the region ranges roughly from 4,000,000 bytes<br />
to 10,000,000 bytes. The upper limit of this region is given<br />
by the file size or for streaming P2P networks by the session<br />
length. Top peers contribute the largest share of the total data<br />
volume, i.e. most of the data is received in conversations with<br />
these peers. In the exemplary trace this region consists of 29<br />
peers, which sent 47.80 % of the total data volume.<br />
The next region, ranging from 20,000 bytes to 4,000,000 bytes,<br />
is called the productive region, because it is likely that the<br />
streams do not only consist of signaling overhead, but also of<br />
useful data. This means, that chunks have been transferred,<br />
but the ratio of signaling overhead is worse than before.<br />
This region is built by ordinary peers. In this trace the 129<br />
ordinary peers sent 52.03 % of the total data volume. Thus,<br />
99.84 % of the total data volume has been transferred by the<br />
158 peers of the profitable and productive region. All other<br />
streams below this region can be regarded as almost useless<br />
for the operations of the P2P network due to their minimal<br />
contribution towards the volume of useful data. They consist<br />
mainly of signaling overhead, e.g. connection establishment<br />
and maintenance. Hence, we call the next region unproductive,<br />
which is inhabited by futile peers. The horizontal lines at the<br />
end of this region mainly consist of unsuccessful connection<br />
attempts. Out of the 2060 incoming streams 1902 streams lie in<br />
the unproductive region. To investigate this inefficiency more<br />
thoroughly, we visualize the intensity of the incoming streams,<br />
i.e. the amount of bytes per 10 second intervals, with the help<br />
of traffic analyzer Atheris [13] (see Fig. 3). The effects of the<br />
hand-overs are clearly visible, precisely when the intensity<br />
of the incoming streams tends towards zero. Also, due to<br />
the end-game mode, at the end of the trace, the intensity<br />
increases dramatically. When we visualize only the streams<br />
of the unproductive region, see Fig. 4, one realizes that the<br />
incoming streams of the unproductive region continue over<br />
the whole trace. This is not a big problem, when BitTorrent<br />
is operating in a wired environment, like Ethernet, but in a<br />
WiMax scenario this can be a high burden for wireless access<br />
points, especially when multiple clients use simultaneously a<br />
BitTorrent-like application. For the adaption of the BitTorrent<br />
protocol to a wireless environment, we recommend to intensify<br />
the data exchange with peers in the profitable region and<br />
restrict it to peers in the productive region, and thereby,<br />
avoiding the many useless connections in the unproductive<br />
region.<br />
So far, we have explained our observations by one representative<br />
trace, but to allow a comparison, we have additionally<br />
plotted the distributions of the inbound and outbound streams<br />
in all the gathered traces, see Fig. 5 and 6. Apart from the<br />
Ethernet measurement on March 16, the bodies of all the<br />
data access distributions have the same shape on the loglog<br />
scale. We see very clearly the influence of the choking<br />
algorithm on the head of the distributions in all Ethernet<br />
measurements, but also in the outbound traffic distribution of<br />
the WiMax traces. Very few peers, between three and six,<br />
receive by far the biggest share of the data. Additionally, in the<br />
Ethernet scenario only a few peers sent most of the data to the<br />
measurement host. However, this pattern changes in all WiMax<br />
traces and the head of the inbound data distribution becomes<br />
flat. This implies that the received data volume is more evenly<br />
distributed among the peer population. This change is due to<br />
the limited upload capacity of the WiMax peers. They are
EITTENBERGER et al.: DAMMING THE TORRENT 17<br />
Fig. 1. Inbound data distribution (bus trace of March, 17).<br />
Fig. 2.<br />
17).<br />
Inbound data distribution in the profitable region (bus trace of March,<br />
Fig. 3. Intensity of the complete inbound traffic (Bus trace of March, 17).<br />
Fig. 4. Intensity of the inbound traffic in the unproductive region (Bus trace<br />
of March, 17).<br />
choked more often and rely mainly on the optimistic unchoke<br />
behavior, in order to successfully complete their download.<br />
Thereby, the download performance suffers in all WiMax<br />
scenarios. We interpret the number of transferred bytes of<br />
a stream φ(p i , p j ) as realization x i of an equivalent income<br />
X i ∈ R of the home peer p j . Considering the overhead for<br />
establishing and maintaining the connection to the feeded peer<br />
p j as costs, only the connections with top peers are really<br />
profitable. Thus, the main focus of interest is given by the<br />
distribution of the top peers. Since the feeding peer population<br />
of this region contributes the largest proportion (approx. 50<br />
% in the exemplary bus trace of March, 17) of useful data<br />
with the best signaling overhead ratio. The asymptotic straight<br />
line of the profitable region on the log-log scale (see Fig. 2)<br />
indicates that the head of the distribution follows a power law.<br />
Thereby, we can use a generalized Pareto distribution to model<br />
this region of the peer population with a random variable<br />
X and its sample {x 1 , ..., x n }. We denote the distribution<br />
function of this generalized Pareto model by<br />
F (x) = 1 − (1 + k x − µ<br />
σ )−1/k for k ≠ 0,<br />
with µ ≤ x ≤ µ−σ/k for k < 0. To investigate the distribution<br />
of the top peers, we set the minimum x min to the lower bound<br />
of the profitable region, i.e. x min = 3, 981, 064 bytes. x min is<br />
obtained with Clauset’s estimator (cf. [14]), which chooses<br />
the value of ˆx min such that the probability distribution of<br />
the measured data and the best-fit power law model is as<br />
similar as possible above ˆx min . Hereby, we consider only the<br />
flows from the top peers φ(p i , p j ), with p i , i ∈ {1, ...., n}<br />
and n = 29, feeding the home peer p j . Using the transferred<br />
amount of bytes x 1 ≤ x 2 ... ≤ x n , we can determine the<br />
scaling parameter ˆα = 4.281707 by Newman’s estimate [15]<br />
[ n<br />
] −1<br />
∑ x i<br />
ˆα ≃ 1 + n ln<br />
x min − 1 .<br />
2<br />
i=1<br />
It is obvious that the Pareto model can only be applied for<br />
the profitable region, since the distribution of the peers in the<br />
productive region is not a straight line on the log-log plot,
18 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Fig. 5.<br />
Inbound data access distribution.<br />
Fig. 6.<br />
Outbound data access distribution.<br />
but has a flat head and a steep tail. Such a rank distribution<br />
indicates a Weibull distribution. The cumulative distribution<br />
function (cdf) for the Weibull distribution is given by<br />
F (x) = 1 − e −(xα)k for x ≥ 0,<br />
where α is the scale parameter and k the shape parameter. Both<br />
parameters are constant and in the analyzed trace α = 2259901<br />
and k = 1.015688. To get the parameters, we used the maximum<br />
likelihood estimation method. Since the shape parameter<br />
k is close to 1, the increase of the data volume per peer<br />
over time is fairly constant. Fig. 10 shows the cumulative<br />
distribution function (cdf) of the received data volumes in<br />
this region and the dotted lines show the fitted Pareto model.<br />
Fig. 7-9 underline our observation, they show the empirical<br />
cumulative distribution function (ecdf) of all other traces and<br />
the fit to the Weibull distribution. The according parameters<br />
for the fitted Weibull distribution are presented in Table I.<br />
As a goodness-of-fit metric we use the Kolmogorov-Smirnov<br />
test, which is the largest vertical distance between the fitted<br />
and actual cumulative distribution functions, measured in<br />
percentiles. For the bus trace we obtain a P-value of 0.9134 for<br />
the Pareto model fit of the profitable region and a P-value of<br />
0.9563 for the fitting of the productive region to the Weibull<br />
distribution and regarding a significance level of α = 0.01<br />
in both cases the null hypothesis can not be rejected. This<br />
constitutes a strong indication that the observed sample data<br />
obey a generalized Pareto respectively a Weibull distribution.<br />
Regarding the distribution of the peers in the profitable and<br />
productive region, the same observations can be made in all<br />
the other captured WiMAX traces (cf. Table I regarding the<br />
fit to the Weibull distribution).<br />
V. CUTTING CONNECTIONS<br />
Jelasity et al. [16] have shown how BitTorrent causes noticeable<br />
problems through the sheer number of flows with ordinary<br />
network equipment, i.e. normal Cisco routers. Wireless access<br />
points have even less resources and when more and more<br />
users start to use their accustomed applications in mobile<br />
networks, the wireless infrastructure might encounter problems<br />
tackling the quantity of flows caused by such P2P applications.<br />
Therefore, we propose a small protocol adaption to prevent<br />
overloading the wireless network infrastructure without<br />
sacrificing performance. Inspired by Clauset’s estimator we<br />
developed a simple heuristic to limit the open connections<br />
of P2P clients when operating under mobile conditions. The<br />
proposed Algorithm 1 can be used to limit the number of<br />
contributing peers and thereby, reduce the load on the wireless<br />
infrastructure, since the distribution of the top peers is quite<br />
stable in all traces. Only very few peers enter respectively<br />
leave this region after the process has stabilized, quite often<br />
the ranking order of the top peers does not even change after<br />
stabilization. We use again the Bus trace of March, 17 to<br />
describe the idea of the proposed algorithm. Fig. 11 shows<br />
the inbound data distribution at different points in time. When<br />
the process is stable and shows a good fit to the Weibull<br />
distribution, in this exemplary case after two minutes, we use<br />
Clauset’s estimator to determine the cutting point (red line).<br />
The cutting point identifies x min of the Pareto model and<br />
the lowest part of the profitable region. Since the peer gets<br />
most of the data from peers of this region, it is appropriate<br />
to restrict data requests to such top peers without loosing too<br />
much performance.<br />
The algorithm needs initially only a vector with the amount<br />
of contributed data volume of each peer and a threshold.
EITTENBERGER et al.: DAMMING THE TORRENT 19<br />
TABLE I<br />
WEIBULL FITTING PARAMETERS OF THE CAPTURED TRACES, ALONG WITH THE CORRESPONDING P-VALUES.<br />
Trace n x min k α p<br />
Bus trace<br />
Static trace<br />
Subway trace<br />
15.03.2010 152 20000 1.009601 2,409,848 0.7195<br />
16.03.2010 188 10000 1.111717 2,029,553 0.7479<br />
17.03.2010 158 20000 1.015688 2,259,901 0.9563<br />
18.03.2010 368 20000 0.9513743 1,431,563 0.7307<br />
15.03.2010 137 100000 1.421969 2,812,421 0.9331<br />
16.03.2010 105 100000 1.543955 3,752,800 0.7834<br />
17.03.2010 110 100000 1.474172 3,539,219 0.8111<br />
18.03.2010 195 100000 1.411851 2,983,383 0.8527<br />
15.03.2010 166 300000 1.414346 2,235,072 0.63<br />
16.03.2010 106 200000 1.608155 3,692,439 0.695<br />
17.03.2010 205 20000 1.025806 1,746,502 0.9538<br />
18.03.2010 200 200000 1.274519 2,823,105 0.8357<br />
Fig. 7.<br />
Ecdf of the inbound data distribution and the Weibull model fit (Bus traces).<br />
Fig. 8.<br />
Ecdf of the inbound data distribution and the Weibull model fit (Static traces).<br />
Fig. 9.<br />
Ecdf of the inbound data distribution and the Weibull model fit (Subway traces).
20 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Fig. 10.<br />
Cdf of the inbound data distribution in the profitable region.<br />
Fig. 11. Inbound data distribution (bus trace of March, 17) with the cutting<br />
points of Clauset’s estimator (red line).<br />
The threshold determines the minimal data contribution of<br />
a peer needed to be considered. The pseudo code is given<br />
in Algorithm 1. The limit function returns a new vector<br />
containing all values greater or equal the given threshold. The<br />
algorithm estimates the scaling parameter α using maximumlikelihood<br />
estimation and computes the Kolmogorov-Smirnov<br />
distance D. The point where the measured data and the bestfit<br />
power law model is as similar as possible, i.e. the distance<br />
D is minimal, determines x min . When x min is computed,<br />
a list of all peers with a greater or equal contribution is<br />
returned. Subsequently, data requests should be restricted to<br />
these peers. Thereby, the amount of open connections is<br />
limited without loosing too much performance. The proposed<br />
algorithm is scalable, since no global knowledge is required<br />
and the computational overhead is manageable. However, the<br />
main field of application is tied to environments with sufficient<br />
seeding capacity. In the opposite case other measures are<br />
needed but we leave this open to future work.<br />
VI. CONCLUSION<br />
To the best of our knowledge, this is the first study which<br />
investigates the data access patterns of BitTorrent and tries to<br />
fit the distribution of the feeding peer population to a model.<br />
In our work we have additionally introduced a novel scheme<br />
to classify the peer population of BitTorrent-like P2P networks<br />
into different categories. We have shown that there are strong<br />
indications that a particular part of the peer population obeys<br />
a power law and that its distribution can be modeled by a<br />
generalized Pareto model. Furthermore, when we investigated<br />
the data distribution of the productive region, we found strong<br />
indications that the part of the peer population, which contributes<br />
nearly the complete data volume, can be modeled by<br />
a Weibull distribution. The limitation of this study is the single<br />
point of view regarding the measurements, but in future work<br />
we plan to support our results with a distributed measurement<br />
campaign. We have seen that the BitTorrent protocol and<br />
especially the choking algorithm is not well adapted to the<br />
fluctuating conditions in a WiMax environment. Therefore,<br />
we propose following recommendations for the adaption of<br />
BitTorrent-like systems to wireless networks. A very simple<br />
part of the solution could be using another client instead of<br />
Vuze. Iliofotou et al. [17] have shown that µTorrent achieves<br />
on average a 16 % better download performance compared to<br />
Vuze. One reason for the better performance might be the<br />
limitation of the amount of open connections to peers by<br />
µTorrent. In general, this might also be a good recommendation<br />
in wireless scenarios, and thereby, not to overload the<br />
base stations with too many open connections. To address this<br />
problem we have developed a simple heuristic to restrict the<br />
number of open connections. Another proposition would be<br />
the usage of UDP instead of TCP as a transport protocol,<br />
since especially TCP, as a connection oriented protocol, suffers<br />
from the fluctuating link conditions and by the hand-overs<br />
in a mobile wireless scenario. Finally, Lehrieder et al. [18]<br />
investigated the positive effect of caches in a BitTorrent<br />
network. As recommendation for network operators, supplying<br />
a dedicated infrastructure with local caches to support the data<br />
dissemination of BitTorrent-like networks can dramatically increase<br />
the data throughput, but at the same time reduce the load<br />
on the own infrastructure by reducing inter-domain traffic. This<br />
proposal is very important with regard to the limited upload<br />
capacity in a WiMax network, since the choking algorithm<br />
is not well adapted to conditions of wireless networks and<br />
the peers rely heavily on the optimistic unchoking behavior<br />
to complete their download. Therefore, supplying a dedicated
EITTENBERGER et al.: DAMMING THE TORRENT 21<br />
Algorithm 1 Cutting point algorithm<br />
Require: ⃗x, x thres<br />
Ensure: Process is stable.<br />
sort(⃗x)<br />
⃗y = ⃗x<br />
⃗x = limit(x thres , ⃗x)<br />
unique(⃗x)<br />
n = size_of(⃗x)<br />
for i = 1 to n do<br />
x min = ⃗x i<br />
⃗y = limit(x min , ⃗y)<br />
o = size_of(⃗y)<br />
/* Estimate α using MLE */<br />
α = o/( ∑ o<br />
l=1 (log( ⃗y l<br />
x min<br />
)))<br />
/* Compute the KS-distance */<br />
for j = 1 to o do<br />
⃗z j = j/o<br />
end for<br />
for k = 1 to o do<br />
⃗v k = 1 − ( xmin<br />
⃗y k<br />
) α<br />
end for<br />
⃗r i = max(abs(⃗v − ⃗z))<br />
end for<br />
/* Determine the minimal distance */<br />
D = min(⃗r)<br />
p = get_position(D, ⃗r)<br />
x min = ⃗x p<br />
return list of peers which sent more than x min bytes<br />
infrastructure with local caches could foster the dissemination<br />
performance of BitTorrent-like systems in wireless networks.<br />
[6] G. Dán and N. Carlsson, “Dynamic Swarm Management for Improved<br />
BitTorrent Performance,” in Proc. International Workshop on Peer-to-<br />
Peer Systems (IPTPS ’09), 2009.<br />
[7] S. Sen and J. Wang, “Analyzing peer-to-peer traffic across large networks,”<br />
IEEE/ACM Transactions on Networking, vol. 12, no. 2, pp.<br />
219–232, 2004.<br />
[8] L. Guo, S. Chen, Z. Xiao, E. Tan, X. Ding, and X. Zhang, “Measurements,<br />
analysis, and modeling of BitTorrent-like systems,” in IMC<br />
’05: Proceedings of the 5th ACM SIGCOMM conference on Internet<br />
Measurement. USENIX Association, 2005, pp. 35–48.<br />
[9] A. R. Bharambe, C. Herley, and V. N. Padmanabhan, “Analyzing and<br />
improving a BitTorrent networks performance mechanisms,” in 25th<br />
IEEE International Conference on Computer Communications (INFO-<br />
COM 2006). IEEE, 2006, pp. 1–12.<br />
[10] Bittorrent. [Online]. Available: http://www.bittorrent.org/beps/bep_0003.<br />
html<br />
[11] Vuze. [Online]. Available: http://www.vuze.com/<br />
[12] N. M. Markovich, A. Biernacki, P. M. Eittenberger, and U. R.<br />
Krieger, “Integrated measurement and analysis of peer-to-peer traffic,” in<br />
Wired/Wireless Internet Communications, 8th International Conference.<br />
Springer, 2010, pp. 302–314.<br />
[13] P. M. Eittenberger and U. Krieger, “Atheris: A first step towards a<br />
unified peer-to-peer traffic measurement framework,” in 19th Euromicro<br />
International Conference on Parallel, Distributed and Network-Based<br />
Computing (PDP <strong>2011</strong>). Euromicro, <strong>2011</strong>.<br />
[14] A. Clauset, C. R. Shalizi, and M. E. J. Newman, “Power-law distributions<br />
in empirical data,” SIAM Reviews, 2007.<br />
[15] M. E. J. Newman, “Power laws, pareto distributions and zipf’s law,”<br />
Contemporary Physics, vol. 46, pp. 323–351, 2005.<br />
[16] M. Jelasity, V. Bilicki, and M. Kasza, “Modeling network-level impacts<br />
of p2p flows,” in 19th Euromicro International Conference on Parallel,<br />
Distributed and Network-Based Computing (PDP <strong>2011</strong>). Euromicro,<br />
<strong>2011</strong>.<br />
[17] M. Iliofotou, G. Siganos, X. Yang, and P. Rodriguez, “Comparing bittorrent<br />
clients in the wild: the case of download speed,” in Proceedings<br />
of the 9th international conference on Peer-to-peer systems(IPTPS’10).<br />
USENIX Association, 2010.<br />
[18] F. Lehrieder, G. Dán, T. Hoßfeld, S. Oechsner, and V. Singeorzan, “The<br />
Impact of Caching on BitTorrent-like Peer-to-peer Systems,” in 10th<br />
IEEE International Conference on Peer-to-Peer Computing 2010 - IEEE<br />
P2P 2010, 2010, pp. 69–78.<br />
ACKNOWLEDGMENT<br />
The authors at Otto-Friedrich University acknowledge the<br />
partial financial suppport by the ESF project COST IC0703.<br />
REFERENCES<br />
[1] Cisco Systems, “Cisco visual networking index: Forecast and methodology,<br />
2009-2014,” White Paper, 2010.<br />
[2] ——, “Cisco visual networking index: Global mobile data traffic forecast<br />
update, 2009-2014,” White Paper, 2010.<br />
[3] S. Kim, X. Wang, H. Kim, T. T. Kwon, and Y. Choi, “Measurement<br />
and analysis of BitTorrent traffic in mobile WiMAX networks,” in 10th<br />
IEEE International Conference on Peer-to-Peer Computing 2010 - IEEE<br />
P2P 2010. IEEE, 2010, pp. 49–52.<br />
[4] P. M. Eittenberger, U. Krieger, and S. Kim, “A classification scheme<br />
for the peer population of bittorrent-like peer-to-peer networks.” in The<br />
First European Teletraffic Seminar (ETS <strong>2011</strong>), <strong>2011</strong>.<br />
[5] D. Qiu and R. Srikant, “Modeling and performance analysis of<br />
bittorrent-like peer-to-peer networks,” in Proceedings of the 2004 conference<br />
on Applications, technologies, architectures, and protocols for<br />
computer communications (SIGCOMM ’04). ACM, 2004, pp. 367–378.<br />
Philipp Eittenberger received the Diploma degree in information systems<br />
from the Otto-Friedrich University, Bamberg, Germany, in 2010.<br />
He is currently a Ph.D. candidate at the Communication Services, Telecommunication<br />
Systems, and Computer Networks Group, Otto-Friedrich University,<br />
Bamberg, Germany. His research areas include Internet traffic measurement<br />
and analysis and peer-to-peer networks.<br />
Seungbae Kim received the B.S. degree in computer science and engineering<br />
from the Chung-Ang University, Seoul, Korea, in 2009, the M.S. degree in<br />
computer science and engineering from Seoul National University, Seoul,<br />
Korea, in <strong>2011</strong>.<br />
He is currently a research engineer at the Future Communications Team,<br />
KAIST Institute for Information Technology Convergence, Daejeon, Korea.<br />
His recent research areas include Internet traffic measurement and analysis,<br />
peer-to-peer networks and content distribution networks.
22 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Udo Krieger received the M.S. degree in applied mathematics and the<br />
Ph.D. degree in computer science from the Technische Hochschule Darmstadt,<br />
Germany.<br />
In 1985 he joined the Research and Technology Center of Deutsche Telekom,<br />
now called T-Nova Technology Center, in Darmstadt, Germany. He joined the<br />
Otto-Friedrich-University, Bamberg, Germany in 2003, where he is currently<br />
an associate professor. Udo Krieger has participated in the European research<br />
projects COST 257, COST 279, COST IC0703 and EURESCOM P1112 and<br />
has served on numerous technical program committees of ITC and IEEE<br />
conferences including Infocom’98 and the European Conference on Universal<br />
Multiservice Networks 2000. Currently, he serves on the editorial board of the<br />
journal Computer Networks. His research interests include traffic management<br />
of IP and wireless networks, teletraffic theory, and numerical solution methods<br />
for Markov chains. He is a member of Gesellschaft fuer Informatik, Germany,<br />
and IEEE.
ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong> 23<br />
Analysis of OBS Burst Assembly Queue with<br />
Renewal Input<br />
Tomasz Hołyński and Muhammad Faisal Hayat<br />
Abstract—Ongoing research in Optical Burst Switching (OBS)<br />
requires more in-depth studies both in theory and in practice<br />
before the technology is realized. In OBS paradigm, traffic from<br />
access networks is groomed at edge OBS nodes in the forms<br />
of large chunks called bursts. This grooming called assembly is<br />
crucial in analyzing the overall performance of OBS networks<br />
as it affects the design of almost all major functions of OBS<br />
nodes. The characteristics of assembled traffic and its effects<br />
on OBS performance have been already extensively studied in<br />
literature. In this work, the assembled traffic is studied using a<br />
transform-based approach, since it is a natural way of analyzing<br />
such processes where random variables are summed. The main<br />
contribution of this paper is formulation of distributions of<br />
burst length and burst inter-departure time in form of Laplace<br />
transforms, which are valid for general independent lengths and<br />
inter-arrival times of assembled packets. The results can be<br />
subsequently inverted analytically or numerically to give full<br />
densities or serve as moment generating functions for partial<br />
characteristics. A simple method for the distribution of the<br />
number of packets in a burst based on discrete Markov chain is<br />
provided. Detailed analytical derivations with numerical results<br />
are presented for Erlangian traffic and verified by simulations<br />
to show good exactness of this approach.<br />
Index Terms—Optical burst switching, burst assembly, hybrid<br />
assembly, performance modelling, queueing theory, Laplace<br />
transform<br />
I. INTRODUCTION<br />
THERE is an ever increasing demand for transmission<br />
capacity due to increased popularity of new applications<br />
requiring large amounts of data exchange. Dense wavelength<br />
division multiplexing (DWDM) has promised to cater the<br />
needs of future Internet backbones providing huge bandwidth<br />
capacities. From the first generation of optical networks with<br />
point to point connections, through the second generation with<br />
DWDM ring networks, now we are heading towards the third<br />
generation with flexible mesh topologies. Therefore, optical<br />
networks demand a real change in transfer mode of data as the<br />
established packet switching is not realizable in optical domain<br />
in the near future with the current state of technology. Optical<br />
circuit switching in the form of wavelength-routed networks<br />
also cannot provide scalability required to achieve real flexible<br />
mesh networks.<br />
Optical burst switching has been proposed as a new<br />
paradigm a few years back [1] in attempts to pave the<br />
way for an all-optical backbone switching infrastructure. It<br />
incorporates prospects of both coarse-grained optical circuit<br />
switching and fine-grained optical packet switching and is<br />
T. Hołyński and M.F. Hayat are with the Institute of Telecommunications,<br />
Vienna University of Technology, Vienna, Austria, e-mail: {tomasz.holynski,<br />
muhammad.faisal.hayat}@tuwien.ac.at<br />
considered as implementable solution for future all-optical<br />
networks.<br />
Principles of OBS can be briefly summarized as follows.<br />
The network is divided into two functional domains, the edge<br />
and the core. At the edge, packetized traffic is buffered and<br />
assembled into bursts consisting of many packets. As soon as a<br />
burst is assembled, it is placed into a transmission queue and a<br />
burst control packet is sent out of band over the network along<br />
the path determined by a routing protocol. The burst control<br />
packet configures switching connections in core nodes just for<br />
the time of transmission of the incoming burst. Subsequently,<br />
the burst is transmitted over the core without any nodal delays<br />
and electronic conversion until it reaches an egress node where<br />
disassembly finally takes place. Due to possibility of time<br />
contention among different flows, bursts can be lost at the<br />
core nodes.<br />
The process of assembly results in a modified type of traffic,<br />
a good understanding of which is crucial in practical engineering<br />
of OBS networks as it affects many design parameters and<br />
functions of OBS nodes at both edge and core [2], [3], [4].<br />
From the viewpoint of performance evaluation, characteristics<br />
of this traffic are the main input parameters for the theoretical<br />
analysis of core switches (burst losses, optimization of fiber<br />
delay lines). Therefore, it is important to dispose with at least<br />
approximations of probability distributions of time intervals<br />
between bursts and burst length. In a predominant number<br />
of studies on OBS core (e.g. [5], [6], [7], [8], [9]) these<br />
distributions are assumed to be negative exponential. While<br />
this is true for the inter-burst times, due to superposition of<br />
many independent flows, it is rather unrealistic when burst<br />
length is considered.<br />
Since OBS is still in a pre-deployment phase, one does not<br />
know how large on average bursts should be and what degree<br />
of variability in their length can be tolerated in practice. Long<br />
bursts require more time to be assembled which results in<br />
greater delays for single packets. Moreover, they will certainly<br />
suffer higher losses and degrade performance of the upper<br />
layers due to the need of reordering of the packets delivered<br />
out of sequence [10]. On the other hand, shorter bursts,<br />
generated at higher rate, will cause more control traffic and<br />
unnecessarily load the network.<br />
The analytical tools that can be used to analyze assembly<br />
process are queueing theory, renewal theory or some complex<br />
models which can be evaluated numerically. Analyzing the<br />
assembly process with queueing theoretical approaches is not<br />
straightforward as it does not fall under classical queueing<br />
discipline. It is because an OBS assembly queue is not strictly<br />
a queueing system with a server but it acts rather like a delay
24 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
element without a server and passes accumulated customers in<br />
a batched manner when some criterion is met. Such behavior<br />
may have an analogy but it is not completely mappable to<br />
the batch service nor to gated vacation models. Therefore, we<br />
devise a simple probabilistic technique based on observation of<br />
the development of assembly of a single burst and determine<br />
probabilities that the n-th packet will complete the burst.<br />
Subsequently, referring to the trends from classical queueing<br />
literature, for the distributions of interest we formulate the<br />
solutions in the form of transforms.<br />
Previous studies in this regard have analyzed the assembly<br />
process in detail, with focus on the characteristics of assembled<br />
burst lengths [11], [12], [13], the burst inter-departure<br />
times [14], [13], its impact on different aspects of global<br />
network performance, such as link-utilization and blocking<br />
probability at intermediate nodes [3], [4], or a combination of<br />
some of these aspects. However, to the best of our knowledge,<br />
there is no study which have tried to analyzed this burstification<br />
process with transform-based approach and general input<br />
traffic conditions and only a few studies have paid attention to<br />
the distribution of number of packets in burst and actual delay<br />
distribution experienced by the packets especially for the most<br />
favorite hybrid burst assembly. For example, Zapata et al. [15]<br />
analyzed packet delays but only for non-hybrid mechanisms<br />
and only average and maximum delays have been evaluated<br />
but no other metrics, such as the variance, nor the actual delay<br />
distribution. Rodrigo de Vega et al. [16] also analyzed the<br />
packet delays and compute the delay of the first packet in the<br />
burst, which is upper bound for a packet in burst but do not<br />
mention the delay suffered by other packets in that burst and<br />
they have also not extended their analysis for hybrid assembly.<br />
Hernandez [17] used fixed packet lengths and Poisson arrivals<br />
to find out delay distribution of each packet in a burst. To<br />
our best knowledge, there is no study on the mentioned<br />
probability distributions for general input and general packet<br />
length. This work assumes this generality and aims at study<br />
of two performance metrics that are most relevant for further<br />
analysis of transmission queue and OBS core nodes, namely<br />
burst length and inter-departure time in case of hybrid scheme.<br />
At this stage of our development, we show exemplar solutions<br />
where Erlang distributions of packet length and inter-arrival<br />
time are assumed, mainly due to easiness of their transform<br />
inversion. However, the method can be used also for more<br />
complicated distributions, especially when powerful numerical<br />
inversion techniques are applied [18], [19]. The study of the<br />
delay of n-th packet can be also treated with this method.<br />
The rest of paper is organized as follows. In Section II, the<br />
OBS edge node architecture and burst assembly schemes are<br />
briefly described. The analytical model for burst assembly is<br />
presented in Section III. In Section IV numerical results with<br />
simulations are discussed and Section V concludes the paper.<br />
Fig. 1.<br />
General architecture of an OBS edge node.<br />
different destination/egress node. The classifier distributes the<br />
incoming packets, with respect to each packet’s destination<br />
address, into the respective queues of the burst assembly modules.<br />
Based on the burst assembly technique, burst assembler<br />
module then assembles bursts consisting of packets headed<br />
for a specific egress node. After a burst has been aggregated,<br />
the corresponding control packet is generated and sent on the<br />
control channel. The assembled bursts wait for transmission<br />
in the electronic transmission buffers called burst transmission<br />
queues. The decision about scheduling a wavelength channel<br />
and time on which a burst is going to be sent is taken by<br />
the scheduling unit at the edge node. There are three main<br />
schemes that have been categorized in literature for burst<br />
assembly: time-based assembly, length-based assembly and<br />
hybrid assembly.<br />
In time-based assembly [11], after receiving the first packet<br />
in an assembler queue, a timer is started. Packets are collected<br />
in the queue until a defined time-out expires. The collected<br />
packets are then assembled into a burst and sent to the<br />
transmission buffer. The timer is restarted when a new packet<br />
is received in the queue. Therefore in time assembly, bursts are<br />
produced in periodic intervals from a single assembler queue,<br />
however, their sizes may vary depending on the arrival rate. In<br />
length-based assembly [14], [20], packets are collected until<br />
the total length of packets exceeds a defined threshold. The<br />
last packet that makes the total length equal or greater than<br />
threshold triggers the assembly of packets into a burst. Therefore,<br />
this kind of assembly generates bursts of approximately<br />
equal lengths but variable inter-departures.<br />
The two mentioned have are simple to implement but have<br />
the following drawbacks. Monitoring only the time results<br />
in undesirably long bursts in high-load scenario, whereas<br />
II. OBS BURST ASSEMBLY<br />
Typically, the edge node consists of a classifier, burst assemblers,<br />
burst transmission queues, a routing and wavelength<br />
assignment modules and schedulers as shown in Fig.1. Each<br />
burst assembler module maintains one separate queue for each<br />
Fig. 2.<br />
Model of the hybrid assembly queue.
HOŁYŃSKI AND HAYAT: ANALYSIS OF OBS BURST ASSEMBLY QUEUE WITH RENEWAL INPUT 25<br />
huge packet delays arise in length-based assembly under low<br />
load condition. These problems are overcome with the hybrid<br />
mechanism that takes into account both criteria. On arrival of<br />
the first packet, the timer is started. The burst is assembled<br />
on the basis of the time-out or length exceedance depending<br />
upon which event happens first, as schematically presented in<br />
Fig. 2. In this work, we have considered the hybrid assembly<br />
as it encompasses the two other strategies as special cases.<br />
III. ANALYTICAL MODELLING<br />
In this section, we develop an analytical model for hybrid<br />
assembly in which we consider a single assembly queue. In<br />
the model, packets destined to this queue arrive from a renewal<br />
process with general gap distribution and the lengths of packets<br />
are general independent random variables. The analysis is<br />
based on the observation of arrivals of subsequent packets<br />
and summation of their lengths to find probability that the<br />
aggregate of n packets exceeds one of the thresholds. Because<br />
of the thresholds, the involved distributions and their Laplace<br />
transforms (LT) are subjected to truncations from the right,<br />
which are here indicated by an auxiliary operator [...] ∗ .<br />
First, we find the probability mass function (pmf) of the<br />
number of the packets in a burst and derive general Laplace<br />
transforms of burst length and inter-departure time. Then we<br />
proceeds with evaluations in case lengths and arrivals are<br />
Erlang distributed. The quantities and notation used are listed<br />
below.<br />
Symbol<br />
L<br />
T A<br />
l o<br />
t o<br />
f L (l)<br />
f A (t)<br />
ψ(s)<br />
φ(s)<br />
f L,n (l)<br />
f A,n (t)<br />
ψ n(s)<br />
φ n(s)<br />
[ψ(s)] ∗<br />
q n<br />
r n<br />
a n, b n<br />
p t<br />
p l<br />
π n<br />
f BL (l)<br />
f D (t)<br />
f ex(l)<br />
ψ BL (s)<br />
φ D (s)<br />
ψ ex(s)<br />
φ A (s)<br />
TABLE I<br />
NOTATION USED IN THE ANALYSIS.<br />
Description<br />
packet length (random variable)<br />
packet inter-arrival time (random variable)<br />
length threshold<br />
time threshold<br />
probability distribution function of packet length<br />
probability distribution function of inter-arrival time<br />
Laplace transform of f L (l)<br />
Laplace transform of f A (t)<br />
pdf of the length of n aggregated packets<br />
pdf of the time up to (n+1)th packet arrival<br />
Laplace transform of f L,n (l)<br />
Laplace transform of f A,n (t)<br />
Laplace transform of a truncated pdf<br />
prob. that n aggregated packets are shorter than l o<br />
prob. that time up to (n+1)th packet arrival is less than t o<br />
normalizing probabilities used for truncation of pdfs<br />
prob. that a burst is assembled due to time criterion<br />
prob. that a burst is assembled due to length criterion<br />
pmf of the number of packets in a assembled burst<br />
pdf of length of a assembled burst<br />
pdf of burst inter-departure, LT: φ D (s)<br />
pdf of the part of the burst part which exceeds l o<br />
Laplace transform of f BL (l)<br />
Laplace transform of f D (t)<br />
Laplace transform of f ex(t)<br />
Laplace transform of burst assembly time<br />
the burst assembly is a regenerative process. This happens<br />
due to the fact the timer is reset upon arrival of a first packet<br />
making the past realizations irrelevant for the way the current<br />
burst is completed. That means that analysis of assembly of<br />
a single burst is sufficient for characterization of the whole<br />
process.<br />
The development of hybrid assembly can be followed with<br />
the help of a discrete Markov chain shown in Fig. 3. The state<br />
number represents the number of packets currently aggregated<br />
and the absorbing state stands for the assembly completion.<br />
Pmf of the number of packets in a burst π n<br />
Starting with the state no 1 (first arrival and timer reset),<br />
the transition probabilities can be explained by the following<br />
narration (Fig. 4). The burst will consist of only one packet if<br />
either its length exceeds l o (with probability P {L > l o }) or<br />
the time up to the next packet arrival is greater than t o (with<br />
probability P {T A > t o }), as shown in Fig. 4a. Since both<br />
events are not disjoint, the probability of their union is<br />
p 1 = P {L > l o } + P {T A > t o }<br />
− P {L > l o }P {T A > t o }.<br />
With the probability (1 − p 1 ), the burst will consist of at<br />
least two packets. The same reasoning is repeated up to the<br />
nth arrival, upon which one of the thresholds will be exceeded,<br />
as depicted in Fig. 4b. Then, the probability π n that burst<br />
comprises exactly n packets can be read out from the Markov<br />
chain. Then<br />
n−1<br />
∏<br />
π n = p n (1 − p i ), (1)<br />
i=1<br />
with p n expressed in general by<br />
p n = (1 − q n ) + (1 − r n ) − (1 − q n )(1 − r n ), (2)<br />
whereby the auxiliary probabilities q n and r n are calculated<br />
by the following integrals<br />
q n =<br />
∫ l o<br />
0<br />
f L,n (l) dl, r n =<br />
∫ t o<br />
0<br />
f A,n (t) dt. (3)<br />
Derivation of the densities f A,n (t) and f L,n (l) requires summations<br />
of independent random variables that are equivalent to<br />
multiplications of their Laplace transforms. The summations<br />
related with occurrence of the nth packet is done in such a way<br />
that the length of the nth packet (or the inter-arrival between<br />
the nth and (n+1)-th packet) is added to the already aggregated<br />
A. Model for general independent traffic<br />
First observation to be made is that under assumption of<br />
stationary renewal input and independence of packet lengths,<br />
Fig. 3.<br />
Markov chain describing the assembly process
26 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Fig. 5.<br />
Description of hybrid burst assembly with Laplace transforms.<br />
Fig. 6. Phase diagram for formulation of the Laplace transform of burst<br />
length ψ BL (s).<br />
Fig. 4. Examples of hybrid assembly with lengths and time thresholds: (a)<br />
two possible realizations of assembly of a burst comprising one packet, (b)<br />
assembly of a burst comprising n packets triggered by length exceedance.<br />
portion (or elapsed time) which is truncated at l o (or t o ). Using<br />
the Laplace transforms ψ(s) = L { f L (l) } , φ(s) = L { f A (t) } ,<br />
φ n (s) = L { f A,n (t) } and ψ n (s) = L { f L,n (l) } the summation<br />
can be expressed by the following recursive relations<br />
ψ 1 (s) = ψ(s)<br />
ψ n (s) = [ψ n−1 (s)] ∗ ψ(s) for n = 2, 3, ..., ∞ (4)<br />
φ 1 (s) = φ(s)<br />
φ n (s) = [φ n−1 (s)] ∗ φ(s) for n = 2, 3, ..., ∞, (5)<br />
where the operator [...] ∗ denotes the fact that the Laplace<br />
transform is calculated from the density truncated at l o or t o .<br />
Understanding of Eq. 4 and 5 is supported by Fig. 5. Usage<br />
of these relations can be greatly simplified by the observation<br />
that [ψ n−1 (s)] ∗ = [ψ n−1 (s)] ∗ and [φ n−1 (s)] ∗ = [φ n−1 (s)] ∗<br />
that leads to the non-recursive expressions<br />
φ n (s) = [φ n−1 (s)] ∗ φ(s) for n = 1, 2, ..., ∞ (6)<br />
ψ n (s) = [ψ n−1 (s)] ∗ ψ(s) for n = 1, 2, ..., ∞. (7)<br />
Finally, calculation of the probabilities q n and r n requires either<br />
analytical or numerical inversion of the transforms ψ n (s)<br />
and φ n (s). Note that this operation needs to be performed only<br />
on the intervals [0, l o ] or [0, t o ], respectively. Knowledge of the<br />
probabilities q n and r n enables formulation of the transforms<br />
of burst length and inter-departure time.<br />
Laplace transform of burst length ψ BL (s)<br />
Now, we are not interested in the probability for a concrete<br />
number of packets but rather in finding the probabilities that<br />
the assembly is completed by exceeding either of the length- or<br />
time threshold. We observe that in the former case the length<br />
is composed by a constant portion l o with transform e −slo and<br />
a random exceeding part with transform ψ ex (s). In the latter<br />
case, the burst is formed by sum of n packets with distribution<br />
truncated at l o with the transform ψ n (s). Considering arrivals<br />
of subsequent packets, we employ the probabilities q n and r n<br />
to construct a phase diagram for the Laplace transform of burst<br />
length shown in Fig. 6. The final result is a weighted sum of<br />
all possible ways the diagram can be traversed:<br />
∞∑<br />
[<br />
∏ n<br />
ψ BL (s) = q i−1 r i−1<br />
](1 − q n )ψ ex (s)e −slo<br />
+<br />
n=1<br />
i=1<br />
∞∑<br />
[<br />
∏ n<br />
q i r i−1<br />
](1 − r n )[ψ n (s)] ∗ , (8)<br />
n=1<br />
i=1<br />
whereby we define q 0 = 1 and r 0 = 1. The transform ψ ex (s)<br />
is not easy to derive from the original packet distribution, but<br />
if a burst consists of many packets, it can be successfully<br />
approximated by the well-known transform of residual lifetime<br />
interpreted in the length domain<br />
ψ ex (s) ≈ 1 − ψ (s)<br />
. (9)<br />
sE[L]<br />
If the probability that the burst consists of few packets is<br />
relatively small, the effect of this approximation negligible.<br />
Moments of the burst length can be computed in the standard<br />
way<br />
E[BL k ] = (−1) k dk ψ BL (s)<br />
ds k ∣<br />
∣∣s=0<br />
. (10)<br />
Laplace transform of burst inter-departure time φ D (s)<br />
The inter-departure time is equal to the sum of two periods:<br />
the burst assembly time and the period separating the start<br />
of the current timer and the departure of the previous burst.
HOŁYŃSKI AND HAYAT: ANALYSIS OF OBS BURST ASSEMBLY QUEUE WITH RENEWAL INPUT 27<br />
where<br />
φ A (s) =<br />
∞∑<br />
[<br />
∏ n ]<br />
q i−1 r i−1 (1 − q n )[φ n−1 (s)] ∗<br />
n=1<br />
i=1<br />
Fig. 7. Phase diagram for formulation of the Laplace transform of burst<br />
assembly time φ A (s).<br />
∞∑<br />
[<br />
∏ n<br />
+ q i r i−1<br />
](1 − r n )e −sto , (13)<br />
n=1 i=1<br />
where again q 0 = 1 and r 0 = 1.<br />
E[T k D] = (−1) k dk φ D (s)<br />
ds k ∣<br />
∣∣s=0<br />
. (14)<br />
All the above considerations are valid for general conditions.<br />
Fig. 8.<br />
Two possible realisations of burst inter-departure time.<br />
B. Solutions for Erlangian traffic<br />
In the sequel, packet length and inter-arrival time have<br />
Erlang k and Erlang m densities, respectively.<br />
(kε) k<br />
( ) k kε<br />
f L (l) =<br />
(k − 1)! lk−1 e −kεl ψ(s) =<br />
kε + s<br />
( ) m<br />
f A (t) = (mλ)m<br />
mλ<br />
(m − 1)! tm−1 e −mλt φ(s) =<br />
,<br />
mλ + s<br />
where λ is the mean arrival rate and ε is reciprocal of the<br />
mean packet length<br />
λ = 1<br />
E[T A ]<br />
Calculation of the probabilities q n , r n , π n<br />
ε = 1<br />
E[L] . (15)<br />
Formulation of the transform of the assembly time, φ A (s), is<br />
very similar to that of burst length and is shown by Fig. 7.<br />
If a n-packet-burst is completed due to length criterion, with<br />
probability 1 − q n , its assembly time is a sum of n − 1 interarrival<br />
times truncated at t o ( [φ n−1 (s)] ∗ ). A burst completed<br />
due to timer expiry with probability 1 − r n , has obviously the<br />
assembly time equal to the time threshold ( e −sto ). Duration of<br />
the second period depends upon the fact whether the previous<br />
burst departed due to length or time exceedance. In the former<br />
case, the separating period is a full inter-arrival time, in the<br />
latter it can be again approximated by residual life time<br />
of T A ( [φ res (s)] ). Probabilities associated with both events<br />
explained in Fig. 8 equal p l and p t , respectively.<br />
p l =<br />
p t =<br />
∞∑<br />
[<br />
∏ n<br />
q i−1 r i−1<br />
](1 − q n )<br />
n=1<br />
i=1<br />
∞∑<br />
[<br />
∏ n<br />
q i r i−1<br />
](1 − r n ).<br />
n=1<br />
i=1<br />
Then the transform of burst inter-departure time is<br />
(11)<br />
[<br />
]<br />
φ D (s) = φ A (s) p l φ(s) + p t φ res (s) , (12)<br />
Subsequent derivations are shown for the length, that is their<br />
concern the probabilities q n . The procedure is identical for<br />
time and r n . To start with, we express the density resulting<br />
from addition of n − 1 Erlang k variables as<br />
f L,n−1 (l) =<br />
(kε)k(n−1)<br />
(k(n − 1) − 1)! lk(n−1)−1 e −kεl . (16)<br />
The Laplace transform of the truncated density f L,n−1 (l) can<br />
be calculated as follows 1<br />
∫ l o<br />
[ψ n−1 (s)] ∗ =<br />
0<br />
1<br />
a n−1<br />
(kε) k(n−1)<br />
(k(n − 1) − 1)! lk(n−1)−1 e −kεl e −sl dl<br />
= 1 (kε) k(n−1)<br />
(17)<br />
a n−1 (kε + s) k(n−1)<br />
[ (kε)<br />
k(n−1)<br />
k(n−1)−1<br />
∑ (l o ) i ]<br />
e −kεlo<br />
−<br />
e −slo ,<br />
a n−1 i!(kε + s) k(n−1)−i<br />
i=0<br />
1 In this section, calculation of definite integrals of the type ∫ a<br />
0 xn e −bx dx<br />
is required. Applying multiple integrations by parts, one obtains<br />
∫ a<br />
0<br />
x n e −bx dx = n!<br />
b n+1 [<br />
1 −<br />
See for example [21] on page 670.<br />
n∑ (ab) i ]<br />
e −ab i!<br />
i=0
28 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
where a n−1 is the normalizing factor needed due to truncation<br />
at l o<br />
∫ l o<br />
(kε) k(n−1)<br />
a n−1 =<br />
(k(n − 1) − 1)! lk(n−1)−1 e −kεl dl<br />
Then<br />
0<br />
k(n−1)−1<br />
∑<br />
= 1 −<br />
i=0<br />
(kεl o ) i<br />
e −kεlo . (18)<br />
i!<br />
ψ n (s) = [ψ n−1 (s)] ∗ ψ(s) (19)<br />
=<br />
1 (kε) kn<br />
a n−1 (kε + s) kn<br />
−<br />
[ (kε)<br />
kn<br />
k(n−1)−1<br />
∑ (l o ) i ]<br />
e −kεlo<br />
i!(kε + s) kn−i e −slo<br />
a n−1<br />
i=0<br />
To find the probability q n according to Eq. 3, the inversion<br />
of ψ n (s) is needed on the interval [0, l o ]. By observing that<br />
the second complicated term in Eq. 19 has no contribution to<br />
the inversion below l o (due to the transform shift theorem),<br />
we invert only the first one:<br />
and finally<br />
{ 1<br />
f L,n (l) [0,lo] = L −1 (kε) kn }<br />
a n−1 (kε + s) kn<br />
1 (kε) kn<br />
=<br />
(kn − 1)! lkn−1 e −kεl (20)<br />
a n−1<br />
q n = 1 [<br />
kn−1<br />
∑<br />
1 − e −kεlo<br />
a n−1<br />
i=0<br />
(kεl o ) i<br />
i!<br />
]<br />
. (21)<br />
By identical procedure we find r n together with the associated<br />
normalizing factor b n−1 .<br />
r n = 1 [<br />
mn−1<br />
∑<br />
1 − e −mλto<br />
b n−1<br />
i=0<br />
(mλt o ) i ]<br />
i!<br />
(22)<br />
inversion involving the shift property:<br />
∞∑<br />
[<br />
∏ n<br />
f BL (l) = q i−1 r i−1<br />
](1 − q n )<br />
n=1 i=1<br />
[ k−1 ∑<br />
j=0<br />
ε [ kε(l − l<br />
×<br />
o<br />
) ] ]<br />
j<br />
e −kε(l−l o ) u(l − l o )<br />
j!<br />
[<br />
∞∑ [<br />
∏ n ] ]<br />
(1 − rn ) (kε) kn<br />
+ q i r i−1<br />
a n (kn − 1)! lkn−1 e −kεl ,<br />
n=1<br />
i=1<br />
where u(l) is the unit step function.<br />
Pdf of burst inter-departure time f D (t)<br />
(24)<br />
This derivation is done by analogy to that of the burst length<br />
pdf, but because of the multiplication of transforms in Eq.12<br />
the final inversion gives rather complicated expression, Eq. 25.<br />
However, there is no practical need for detailed knowledge<br />
of this function. Usually, in an edge node, output streams of<br />
many assembly queues are merged before they are directed to<br />
the transmission buffer(s). Since the single departure stream<br />
is nearly renewal, we infer that the total departure process<br />
tends to a Poisson process as the number of merged streams<br />
increases.<br />
This effect was proved in simulation where a number of<br />
independent assembly queues fed by uncorrelated renewal<br />
inputs was implemented. Fig. 10 shows the results for 10<br />
queues with nearly negative exponential density irrespectively<br />
of the type of packet inter-arrival density assumed.<br />
IV. NUMERICAL EXAMPLES<br />
Fig. 11 shows how the number of packets in a burst varies<br />
when configuration of thresholds is changed in case of purely<br />
Poisson traffic. The probability mass is symmetrically concentrated<br />
around the mean. Fig. 12 depicts the situation when the<br />
thresholds allow much more packets to be assembled. With<br />
The pmf of the number of packets in a burst is now found by<br />
Eq. 1 and 2.<br />
Pdf of burst length f BL (l)<br />
According to Eq. 8, this derivation involves the transforms<br />
ψ ex (s) and [ψ n (s)] ∗ . The first one we approximate by the<br />
transform of residual life which is:<br />
ψ ex (s) ≈ 1 k<br />
k∑<br />
[ ] j+1<br />
kε<br />
(23)<br />
(kε + s)<br />
j=0<br />
and the second we readily obtain from Eq. 17 substituting<br />
n−1 by n. After insertion of the transforms into Eq. 8, we can<br />
obtain the pdf of burst length by means of simple analytical<br />
Fig. 9. Superposition of the departure streams from multiple assembly queues<br />
in the edge node.
HOŁYŃSKI AND HAYAT: ANALYSIS OF OBS BURST ASSEMBLY QUEUE WITH RENEWAL INPUT 29<br />
[<br />
∑ ∞ [<br />
∏ n<br />
f D (t) = p l<br />
n=1<br />
i=1<br />
[ 1 (mλ)<br />
q i−1r i−1<br />
](1−q mn<br />
n)<br />
b n−1 (mk−1)! tmn−1 e −mλt − (mλ)mn m(n−1)−1 ∑<br />
b n−1<br />
i=0<br />
[ (λ)<br />
m<br />
] [<br />
∑ ∞ [<br />
∏ n<br />
+ p l p t<br />
(m−1)! e−mλ(t−to) u(t−t o) + p t q i−1r i−1<br />
](1−q n)<br />
n=1 i=1<br />
m∑<br />
[ 1<br />
j=1<br />
b n−1<br />
t i ] ]<br />
o<br />
i!(mn−i−1)! (t−to)mn−i−1 e −mλ(t−to) u(t−t o)<br />
(mλ) m(n−1)+j<br />
(m(n−1) + j−1)! tm(n−1)+j−1 e −mλt (25)<br />
−<br />
(mλ)m(n−1)+j<br />
mb n−1<br />
m(n−1)−1 ∑<br />
i=0<br />
t i o e−mλto<br />
i!(m(n−1) −i+j−1)! (t−to)m(n−1)−i+j−1 e −mλ(t−to) u(t−t o)<br />
] ] m−1<br />
+ p 2 t λ ∑<br />
[ mλ(t−to) i<br />
i=0<br />
i!<br />
]<br />
e −mλ(t−to) u(t−t o)<br />
Fig. 10. Simulation results of the superposition of inter-departures processes<br />
from 10 independent assembly queues with thresholds l o=5 and t o=5 for<br />
different distributions of packet inter-arrival times T A , whereby ε=1 and λ=1<br />
Fig. 11. Pmf of the number of packets in a burst for various values of<br />
thresholds. Packet lengths and inter-arrival times exponentially distributed<br />
with ε=1 and λ=1.<br />
decreasing variance of the input distributions, the analyzed<br />
pmf tends to be more and more deterministic.<br />
In Fig.13 density of burst length is plotted for various<br />
settings. All three pdfs exhibit discontinuities at the point<br />
equal to the length threshold. The parts of the curves below<br />
l o represent realizations of assembly due to time-out expiry.<br />
The smoothly decaying peaks, which are approximated by<br />
mixtures residual lifetimes of Erlang distributions, are slightly<br />
inexact compared to simulations in its initial region but this<br />
error vanishes rather fast along the tail.<br />
Fig. 14 presents densities for relatively high number of<br />
packets collected. If l o = 50 and t o = 15, we have practically<br />
only time-based assembly and expected manifestation of the<br />
central limit theorem is observed regarding the pdf. In the<br />
converse case, nearly no burst is smaller than l o resulting with<br />
a sharp peak at this point.<br />
Fig. 12. Pmf of the number of packets in a burst for Erlangian traffic (length<br />
and time) with different coefficients of variation, ε =1, λ=1, t o=25, l o=25.<br />
V. CONCLUSIONS<br />
We have provided a nearly exact analysis of hybrid burst<br />
assembly. Although the output traffic preserves the renewal<br />
properties of the input, the distributions of interest turned out<br />
to be complex functions of the involved parameters. Nevertheless,<br />
they give an important insight into the characteristics of<br />
the assembled traffic and could be approximated by simpler<br />
tractable distributions. The presented results show mainly<br />
that the assembled traffic is highly sensitive to the defined<br />
thresholds and if their values are not properly adjusted, the<br />
resulting large variances of lengths can severely degrade the<br />
performance of OBS core nodes. Finally, the distribution of<br />
burst length is very far from negative exponential as it is<br />
commonly assumed in the literature of the subject.<br />
REFERENCES<br />
[1] C. Qiao and M. Yoo, “Optical Burst Switching (OBS) - a New Paradigm<br />
for an Optical Internet,” Journal of High Speed Networks, vol. 8, no. 1,<br />
pp. 69–84, 1999.<br />
[2] X. Yu, J. Li, X. Cao, Y. Chen, and C. Qiao, “Traffic Statistics and Performance<br />
Evaluation in Optical Burst Switching Networks,” IEEE/OSA<br />
Journal of Lightwave Technology, vol. 22, no. 12, pp. 2722–2738, 2004.
30 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Fig. 13. Pdf of burst length for Erlang 2 traffic (length and time) for various<br />
values of thresholds, ε=1 and λ=1.<br />
Conversion,” in Proc. of First Int. Conf. on Broadband Networks<br />
(BROADNETS04), 2004.<br />
[9] S. Sarwar, S. Aleksic, and K. Aziz, “Optical Burst Switched (OBS)<br />
Systems with Segmentation-Based Dropping,” Eletrotechnik und Informationstechnik,<br />
vol. 125, no. 7-8, pp. 296–300, 2008.<br />
[10] S. Gunreben, “An Optical Burst Reordering Model for Time-Based<br />
and Random Selection Assembly Strategies,” Performance Evaluation,<br />
vol. 68, pp. 237–255, 2010.<br />
[11] J. Choi, J. Choi, and M. Kang, “Dimensioning Burst Assembly Process<br />
in Optical Burst Switching Networks,” IEICE Trans. Commun., vol. E88-<br />
B(10), pp. 3855–3863, 2005.<br />
[12] K. Dolzer and C. Gauger, “On Burst Assembly in Optical Burst<br />
Switching Networks - a Performance Evaluation of Just-Enough-Time,”<br />
Proc. of 17th Int. Teletraffic Congress, 2001.<br />
[13] A. Rostami and A. Wolisz, “Modeling and Synthesis of Traffic in Optical<br />
Burst-Switched Networks,” Journal of Lightwave Technology, vol. 25,<br />
2007.<br />
[14] K. Leavens, “Traffic Characteristics Inside Optical Burst Switching<br />
Networks,” in Proc. of SPIE/IEEE OPTICOMM, 2002.<br />
[15] A. Zapata and P. Bayvel, “Impact of Burst Aggregation Schemes<br />
on Delay in Optical Burst Switched Networks,” in Proc. IEEE/LEOS<br />
Annual Meeting, Tucson, Arizona, 2005.<br />
[16] M. de Vega Rodrigo and J. Gotz, “An Analytical Study of Optical Burst<br />
Switching Aggregation Strategies,” in Proc. of Broadnets (Workshop on<br />
OBS), San Jose, California, 2004.<br />
[17] J. Hernandez, J. Aracil, V. Lopez, and J. L. de Vergara, “On the Analysis<br />
of Burst-Assembly Delay in OBS Networks and Applications in Delay-<br />
Based Service Differentiation,” Photonic Network Commun., vol. 14,<br />
no. 1, pp. 49–62, 2007.<br />
[18] A. Cohen, Numerical Methods for Laplace Transform Inversion.<br />
Springer, 2007.<br />
[19] J. Abate and W. Whitt, “The Fourier-Series Method for Inverting<br />
Transforms of Probability Distributions,” Queueing Systems, vol. 10,<br />
pp. 5–87, 1992.<br />
[20] X. Yu, Y. Chen, and C. Qiao, “Performance Evaluation of Optical Burst<br />
Switching with Assembled Burst Traffic Input,” in Proc. of IEEE Global<br />
Telecommunications Conference (Globecom), Taipei, 2002.<br />
[21] P. Pfeiffer, Probability for Applications. Springer, 1990.<br />
Fig. 14. Pdf of burst length for Erlang 2 traffic (length and time) for various<br />
values of thresholds, ε=1 and λ=1.<br />
[3] J. Choi, H. Vu, G. Cameron, M. Zukerman, and M. Kang, “The Effect of<br />
Burst Assembly on Performance of Optical Burst Switched Networks,”<br />
in ICOIN 2004, Busan, Korea, 2004.<br />
[4] J. Liu and N. Ansari, “The Impact of the Burst Assembly Interval on the<br />
OBS Ingress Traffic Characteristics and System Performance,” in Proc.<br />
of IEEE ICC, Paris, France, 2004.<br />
[5] N. Barakat and E. H. Sargent, “On Teletraffic Applications to OBS,”<br />
IEEE Commun. Lett., vol. 8, no. 1, pp. 119–121, 2004.<br />
[6] M. Zukerman, E. W. Wong, Z. Rosberg, G. M. Lee, and H. L. Vu, “An<br />
Accurate Model for Evaluating Blocking Probabilities in Multi-Class<br />
OBS Systems,” IEEE Commun. Lett., vol. 8, no. 2, pp. 116–118, 2004.<br />
[7] J. Teng and G. N. Rouskas, “Wavelength Selection in OBS Networks<br />
Using Traffic Engineering and Priority-Based Concepts,” IEEE J. Sel.<br />
Areas Commun., vol. 23, no. 8, pp. 1658–1669, 2005.<br />
[8] N. Akar and E. Karasan, “Exact Calculation of Blocking Probabilities<br />
for Bufferless Optical Burst Switched Links with Partial Wavelength<br />
Tomasz Hołyński received MSc degree in Telecommunications and Computer<br />
Science from the International Faculty of Engineering (IFE) at the Technical<br />
University of Lodz, Poland, in 2009. Since 2007 he has been working at<br />
the Institute of Telecommunications (former the Institute of Broadband Communications)<br />
at the Vienna University of Technology as a project assistant.<br />
His master thesis on queueing theoretical modelling and analysis of data<br />
link protocols was distinguished by the Austrian Electrotechnical Association<br />
(ÖVE) with the GIT-award. His current doctoral research concerns transformbased<br />
methods and related complex variable techniques in selected areas of<br />
queueing theory and performance evaluation.<br />
Muhammad Faisal Hayat received BSc (Hons) and MSc degrees in Computer<br />
Engineering from University of Engineering and Technology, Lahore,<br />
Pakistan in 2000 and 2005, respectively. In 2001 he joined the abovementioned<br />
university where he worked as a lecturer (2001-2005) and as an assistant<br />
professor (2005-2008). Since 2008 he has been pursuing his PhD at the<br />
Institute of Telecommunications at the Vienna University of Technology,<br />
Austria. His research focus is modelling, analysis and simulation of all-optical<br />
networks.
ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong> 31<br />
Scheduling and Capacity Estimation in LTE<br />
Olav Østerbø<br />
Abstract—Due to the variation of radio condition in LTE<br />
the obtainable bitrate for active users will vary. The two<br />
most important factors for the radio conditions are fading and<br />
pathloss. By considering analytical analysis of the LTE conditions<br />
including both fast fading and shadowing and attenuation due<br />
to distance we have developed a model to investigate obtainable<br />
bitrates for customers randomly located in a cell. In addition<br />
we estimate the total cell throughput/capacity by taking the<br />
scheduling into account. The cell throughput is investigated for<br />
three types of scheduling algorithms; Max SINR, Round Robin<br />
and Proportional Fair where also fairness among users is part<br />
of the analysis. In addition models for cell throughput/capacity<br />
for a mix of Guaranteed Bit Rate (GBR) and Non-GBR greedy<br />
users are derived.<br />
Numerical examples show that multi-user gain is large for the<br />
Max-SINR algorithm, but also the Proportional Fair algorithm<br />
gives relative large gain relative to plain Round Robin. The Max-<br />
SINR has the weakness that it is highly unfair when it comes to<br />
capacity distribution among users. Further, the model visualize<br />
that use of GBR for high rates will cause problems in LTE due<br />
to the high demand for radio resources for users with low SINR,<br />
at cell edge. Persistent GBR allocation will be a waste of capacity<br />
unless for very thin streams like VoIP. For non-persistent GBR<br />
allocation the allowed guaranteed rate should be limited.<br />
Index Terms—LTE, scheduling, capacity estimation, GBR.<br />
I. INTRODUCTION<br />
THE LTE (Long Term Evolution) standardized by 3GPP is<br />
becoming the most important radio access technique for<br />
providing mobile broadband to the mass marked. The introduction<br />
of LTE will bring significant enhancements compared<br />
to HSPA (High Speed Packet Access) in terms of spectrum<br />
efficiency, peak data rate and latency. Important features of<br />
LTE are MIMO (Multiple Input Multiple Output), higher order<br />
modulation for uplink and downlink, improvements of layer 2<br />
protocols, and continuous packet connectivity [1].<br />
While HSPA mainly is optimized data transport, leaving the<br />
voice services for the legacy CS (Circuit Switched) domain,<br />
LTE is intended to carry both real time services like VoIP in<br />
addition to traditional data services. The mix of both real time<br />
and non real time traffic in a single access network requires<br />
specific attention where the main goal is to maximize cell<br />
throughput while maintaining QoS and fairness both for users<br />
and services. Therefore radio resource management will be a<br />
key part of modern wireless networks. With the introduction<br />
of these mobile technologies, the demand for efficient resource<br />
management schemes has increased.<br />
The first issue in this paper is to consider the bandwidth<br />
efficiency for a single user in cell for the basic unit of radio<br />
resources, i.e. for a RB (Resource Block) in LTE. Since LTE<br />
uses advanced coding like QPSK, 16QAM, and 64QAM, the<br />
O. Østerbø is with Telenor, Corporate Development, Fornebu, Oslo, Norway,<br />
(phone: +4748212596; e-mail: olav-norvald.osterbo@telenor.com)<br />
obtainable data rate for users will vary accordingly depending<br />
on the current radio conditions. The average, higher moments<br />
and distribution of the obtainable data rate for a user either<br />
located at a given distance or randomly located in a cell, will<br />
give valuable information of the expected cell performance.<br />
To find the obtainable bitrate we chose a truncated and<br />
downscaled version on Shannon formula which is in line with<br />
what is expected from real implementations and also comply<br />
with the fact that the maximal bitrate per frequency or symbol<br />
for 64 QAM is at most 6 [2].<br />
For the bandwidth efficiency, where we only consider a<br />
single user, the scheduling is without any significance. This<br />
is not the case when several users are competing for the<br />
available radio resources. The scheduling algorithms studied in<br />
this paper are those only depending on the radio conditions, i.e.<br />
opportunistic scheduling where the scheduled user determined<br />
by a given metrics which depends on the SINR (signalto-interference-plus-noise<br />
ratio). The most commonly known<br />
opportunistic scheduling algorithms are of this type like PF<br />
(Proportional Fair), RR (Round Robin) and Max-SINR. The<br />
methodology developed will, however, will apply for general<br />
scheduling algorithms where the scheduling metrics for a user<br />
is given by a known function of SINR, however, now the SINR<br />
may vary in different scheduling intervals taking rapid fading<br />
into account. The cell capacity distribution is found for cases<br />
where the locations of the users all are known or as an average<br />
where all the users are randomly located in the cell [3].<br />
Also the multi user gain (relative increase in cell throughput)<br />
due to the scheduling is of main interest. The proposed models<br />
demonstrate the magnitude of this gain. As for Max SINR<br />
algorithm this gain is expected to be huge, however, the gain<br />
comes always at a cost of fairness among users. And therefore<br />
fairness has to be taken into account when evaluating the<br />
performance of scheduling algorithms.<br />
It is likely that LTE will carry both real time traffic and<br />
elastic traffic. We also analyze scenarios where a cell is loaded<br />
by two traffic types; high priority CBR (Constant Bit Rate)<br />
traffic that requires a fixed data-rate and low priority (greedy)<br />
data sources that always consume the leftover capacity not<br />
used by the CBR traffic. This is actual a very realistic traffic<br />
scenario for future LTE networks where we will have a mix of<br />
both real time traffic like VoIP and data traffic. We analyse this<br />
case by first estimate the RB usage of the high priority CBR<br />
traffic, and then subtract the corresponding RBs to find the<br />
actual numbers of RBs available for the (greedy) data traffic<br />
sources. Finally we then estimate the cell capacity as the sum<br />
of the bitrates offered to the CBR and (greedy) data sources.<br />
The remainder of this paper is organized as follows. In<br />
section II the basic radio model is given and models for<br />
bandwidth efficiency are discussed. Section III gives an outline<br />
of the multiuser case where resource allocation and scheduling
32 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
is taking into account. Some numerical examples are given in<br />
section IV and in section V some conclusions are given.<br />
II. SPECTRUM EFFICIENCY<br />
A. Obtainable bitrate per symbol rate as function of SINR<br />
For LTE the obtainable bitrate per symbol rate will depend<br />
on the radio signal quality (both for up-and downlink). The<br />
actual radio signal quality is signaled over the radio interface<br />
by the so-called CQI (Channel Quality Indicator) index in<br />
the range 1 to 15. Based on the CQI value the coding rate<br />
is determined on basis of the modulation QPSK, 16QAM,<br />
64QAM and the amount of redundancy included. The corresponding<br />
bitrate per bandwidth is standardized by 3GPP [4]<br />
and is shown in Table 1 below. For analytical modeling the<br />
actual CQI measurement procedures are difficult to incorporate<br />
into the analysis due to the time lag, i.e. the signaled CQI is<br />
based measurements taken in earlier TTIs (Transmission Time<br />
Interval). To simplify the analyses, we assume that this time<br />
lag is set to zero and that the CQI is given as a function of the<br />
momentary SINR, i.e. CQI=CQI(SINR). This approximation<br />
is justified if the time variation in SINR is significantly slower<br />
than the length of a TTI interval. Hence, by applying the CQI<br />
table found in [4] we get the obtainable bitrate per bandwidth<br />
as function of the SINR as the step function:<br />
B = fc j , for SINR ∈ [g j , g j+1 ) ; j = 0, 1, ..., 15, (1)<br />
where f is the bandwidth of the channel, c j is the efficiency<br />
for QCI equal j (as given by Table 1) and [g j , g j+1 ) are the<br />
corresponding intervals of SINR values. (We also take c 0 = 0,<br />
g 0 = 0 and g 16 = ∞.)<br />
To fully describe the bitrate function above we also have to<br />
also specify the intervals [g j , g j+1 ). Several simulation studies<br />
e.g. [5] suggest that there is a linear relation between the CQI<br />
index and the actual SINR limits in [dB]. With this assumption<br />
we have SINR j [dB] = 10 log 10 g j = aj + b or g j = 10 aj+b<br />
10<br />
for some constants a and b. It is also argued that the actual<br />
range of the SINR limits in [dB] is determined by the<br />
following (end point) observations: SINR[dB]=-6 corresponds<br />
to QCI=1, while SINR[dB]=20 corresponds to CQI=15. Hence<br />
we then have −6 = a + b and 20 = 15a + b or a = 13/7 and<br />
b = −55/7.<br />
For extensive analytical modelling the step based bandwidth<br />
function is cumbersome to apply. An absolute upper bound<br />
yields the Shannon formula B = f log 2 (1+SINR), however,<br />
we know that the Shannon upper limit is too optimistic.<br />
First of all the bandwidth function should never exceed the<br />
highest rate c 15 = 5.5547. We therefore suggest downscaling<br />
and truncating the Shannon formula and take an alternative<br />
bandwidth function as:<br />
B = d min[T, ln(1 + γSINR)], (2)<br />
with d = f C<br />
c15 ln 2<br />
ln 2<br />
and T =<br />
C<br />
where C is the downscaling<br />
constant (relative to the Shannon formula) and γ is a constant<br />
less than unity. By choosing C and γ that minimize the square<br />
distances between the CQI based and the truncated Shannon<br />
formula (2) above we find C = 0.9449 and γ = 0.4852.<br />
(Upper and lower estimates of the CQI based zigzagging<br />
N ormalised T hroughpu t @ b i t êsê H z D<br />
TABLE I<br />
TABLE 1 CQI TABLE.<br />
CQI index modulation code rate x 1024 efficiency<br />
0 out of range<br />
1 QPSK 78 0.1523<br />
2 QPSK 120 0.2344<br />
3 QPSK 193 0.3770<br />
4 QPSK 308 0.6016<br />
5 QPSK 449 0.8770<br />
6 QPSK 602 1.1758<br />
7 16QAM 378 1.4766<br />
8 16QAM 490 1.9141<br />
9 16QAM 616 2.4063<br />
10 64QAM 466 2.7305<br />
11 64QAM 567 3.3223<br />
12 64QAM 666 3.9023<br />
13 64QAM 772 4.5234<br />
14 64QAM 873 5.1152<br />
15 64QAM 948 5.5547<br />
6<br />
5<br />
4<br />
3<br />
2<br />
1<br />
---- Shannon<br />
---- Modified Shannon<br />
---- LTE CQI table<br />
0 5 10 15 20 25<br />
SINR @dBD<br />
Fig. 1. Normalized throughput as function of the SINR based on: 1.-QCI<br />
table, 2.-Shannon and 3.-Modified Shannon.<br />
bitrate function is obtained by taking γ u = γ10 a/20 = 0.6008<br />
and γ l = γ10 −a/20 = 0.3918).<br />
We observe that a downscaling of the Shannon limit is very<br />
much in line with the corresponding bitrates obtained by the<br />
CQI table as shown in Figure 1 and hence we believe that (2)<br />
yields a quite accurate approximation. In fact the approximated<br />
CQI values c app<br />
j follow the similar logarithmic behaviour:<br />
c app<br />
j = C log 2 (1 + αβ j ), (3)<br />
where now have α = γ10 a/20+b/10 = 0.0984 and β =<br />
10 a/10 = 1.5336.<br />
B. Radio channel models<br />
Generally, the SINR for a user will be the ratio of the<br />
received signal strength divided by the corresponding noise.<br />
The received signal strength is the product of the power P w<br />
times path loss G and divided by the noise component N,<br />
i.e. SINR = PwG<br />
N<br />
. Now the path loss G will typical be a<br />
stochastic variable depending on physical characteristics such
ØSTERBØ: SCHEDULING AND CAPACITY ESTIMATION IN LTE 33<br />
as rapid and slow fading, but will also have a component<br />
that are dependent on distance (and possible also the sending<br />
frequency). Hence, we first consider variations that are slowly<br />
varying over time intervals that are relative long compared<br />
with the TTIs (Transmission Time Intervals). Then the path<br />
loss is usually given in dB on the form:<br />
G = 10 L/10 with L = C − A log 10 (r) + X t , (4)<br />
where C and A are constants, A typical in the range 20-40, and<br />
X t a normal stochastic process with zero mean representing<br />
the shadowing (slow fading). The other important component<br />
determining the SINR is the noise. It is common to split<br />
the noise power into two terms: N = N int + N ext where<br />
N int is the internal (or own-cell) noise power and N ext is<br />
the external (or other-cell) interference. In a CDMA (Code<br />
Division Multiple Access) network, the lack of orthogonality<br />
induces own-cell interference. In an OFDMA (Orthogonal<br />
Frequency Division Multiple Access) network, however, there<br />
is a perfect orthogonality between users and therefore the<br />
only contribution to N int is the terminal noise at the receiver.<br />
The interference from other cells depends on the location of<br />
surrounding base stations and will typically be largest at cell<br />
edges. In the following we shall assume that the external noise<br />
is constant throughout the cell or negligible, i.e. we assume<br />
the noise N to be constant throughout the cell.<br />
Hence, with the assumptions above, we may write SINR on<br />
the form S t /h(r, λ) where S t represent the stochastic variations<br />
which we assume to be distance independent capturing<br />
the slowly varying fading, and h(r, λ) represent the distance<br />
dependant attenuation (which we also allow to depend on the<br />
sending frequency). Most commonly used channel models as<br />
described above have attenuation that follows a power law, i.e.<br />
we chose to take h(r, λ) on the form<br />
h(r, λ) = h(λ)r α , (5)<br />
where α = A/10 is typical in the range 2-4 and h(λ) =<br />
N<br />
P w<br />
10 −C/10 with Z = 10 log 10 (N) − 10 log 10 (P w ) − C given<br />
dB, where we also indicate that h(λ) may depend of the<br />
(sending) frequency. With the description above the stochastic<br />
variable S t = 10 Xt/10 with S t • ln 10<br />
=<br />
10 X t, and hence S t is<br />
a lognormal process with E[S t • ln 10<br />
] = 0 and σ =<br />
10 σ(X t)<br />
where σ(X t ) is the standard deviation (given in dB) for<br />
the normal process X t . With these assumptions we have<br />
the Probability Density Function (PDF) and Complementary<br />
Distribution Function (PDF) of S t as:<br />
s ln (x) =<br />
1<br />
(ln x)2<br />
−<br />
√ e 2σ 2<br />
2πσx<br />
where erfc(y) = 2 √ π<br />
function.<br />
∞∫<br />
x=y<br />
C. Including fast fading<br />
and ˜S ln (x) = 1 2 erfc ( ln x<br />
σ √ 2<br />
)<br />
,<br />
(6)<br />
e −x2 dx is the complementary error<br />
There are several models for (fast) fading in the literature<br />
like Rician fading and Rayleigh fading [6]. In this paper we<br />
restrict ourselves to the latter mainly because of its simple<br />
negative exponential distribution.<br />
It is possible to include fast fading into the description<br />
above. To do so we assume that the fast fading effects are on<br />
a much more rapid time scales than slow fading. We therefore<br />
assume that the slow fading actual is constant during the<br />
rapid fading variations. Hence, condition on the slow fading<br />
to be y then for a Rayleigh faded channel the SINR will be<br />
exponentially distributed with mean y/g(r, λ) Hence, we may<br />
therefore take SINR as S t /g(r, λ) where S t = X ln X e is the<br />
product of a Log-normal and a negative exponential distributed<br />
variables. The corresponding distribution often called Suzuki<br />
distribution have PDF ad CDF given as the integrals:<br />
s su (x) =<br />
∫ ∞<br />
t=0<br />
1<br />
t e− x t sln (t)dt and ˜S su (x) =<br />
∫ ∞<br />
t=0<br />
e − x t sln (t)dt,<br />
(7)<br />
where s ln (t) is the lognormal PDF above by (6). Since<br />
s ln ( 1 t ) = t2 s ln (t) it is possible to express the integrals above in<br />
terms of the Laplace transform of the Log-normal distribution<br />
and therefore the CDF (and PDF) of the Suzuki distribution<br />
may be written as: ˜Ssu (x) = Ŝln(x) and s su (x) = −Ŝ′ ln (x)<br />
where<br />
Ŝ ln (x) =<br />
∫ ∞<br />
t=0<br />
e −xt s ln (t)dt = 1 √<br />
2πσ<br />
∫∞<br />
t=0<br />
(ln(t/x))2<br />
−t−<br />
e 2σ 2<br />
dt (8)<br />
t<br />
is the Laplace transform of the Log-normal distribution. If we<br />
define the truncated transform:<br />
˜S su (x, M) = 1 x<br />
=<br />
∫ M<br />
t=0<br />
1<br />
√<br />
2πσ<br />
e −t s ln (t/x)dt<br />
∫M<br />
t=0<br />
(ln(t/x))2<br />
−t−<br />
e 2σ 2<br />
dt, (9)<br />
t<br />
then ˜S su (x) = lim ˜S su (x, M) and further the corresponding<br />
M→∞<br />
error is exponentially small. An attempt to expand the integral<br />
(8) in terms of the series of the exponential function e −t =<br />
∑ ∞ (−1) k t k<br />
k=0 k!<br />
yields a divergent series; however, this is not<br />
the case for the truncated transform (9). We find the following<br />
series expansion:<br />
˜S su (x, M) = 1 ∞∑ (−1) k<br />
(<br />
x k e k2 σ 2 kσ<br />
2 erfc √ + ln(x/M) )<br />
2 k!<br />
k=0<br />
2 σ √ 2<br />
(10)<br />
Similar the PDF of the Suzuki random variable may be<br />
found from (8) by differentiation:<br />
s su (x) = − ˜S ′ su(x) =<br />
=<br />
∫ ∞<br />
t=0<br />
e −xt ts ln (t)dt<br />
1<br />
√<br />
2πσx<br />
∫∞<br />
t=0<br />
(ln(t/x))2<br />
−t−<br />
e 2σ 2 dt, (11)<br />
and for the PDF we now we take the corresponding truncated<br />
integral to be:<br />
s su (x, M) =<br />
∫M<br />
1<br />
√<br />
2πσx<br />
t=0<br />
(ln(t/x))2<br />
−t−<br />
e 2σ 2 dt (12)
34 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
L og @ 1 0,SHxLD<br />
0<br />
-1<br />
-2<br />
-3<br />
-4<br />
-5<br />
σ = 0.2<br />
σ = 0.6<br />
σ = 5.0<br />
σ = 2.0<br />
σ = 1.0<br />
5 10 15 20 25 30 35 40<br />
x<br />
Fig. 2. Logarithmic plot of the CDF for the Suzuki distribution as function<br />
of x for some values of σ.<br />
σ2 −M+<br />
In this case we find 0 ≤ s su (x) − s su (x, M) = e 2 as a<br />
bound of the truncation error.<br />
By expanding the integral (12) in terms of the exponential<br />
function as above, we now obtain a similar (convergent) series:<br />
s su (x, M)= 1 ∞∑ (−1) k<br />
(<br />
x k e (k+1)2 σ 2 (k + 1)σ<br />
2 erfc √ + ln(x/M) )<br />
2 k!<br />
k=0<br />
2 σ √ 2<br />
(13)<br />
In Figure 2 we have plotted the CDF of the Suzuki distribution<br />
for σ equals 0.2, 0.6, 1.0, 2.0 and 5.0. (The CDF<br />
Suzuki distribution is calculated by applying the series (10)<br />
with M = 20.0 which secure an accuracy of 2.0x10-9 in<br />
the computation.) Note that the k’th moment of the Suzuki<br />
distribution is k! that of the Log-normal.<br />
D. Distribution of the obtainable bitrate for channel of a<br />
certain bandwidth for a user located at a given distance from<br />
the sender antenna<br />
Below we express the distribution of the possible obtainable<br />
bitrate according to the distribution of the stochastic part of<br />
the SINR; namely S t . From (1) we get the bit-rate B t (r) for<br />
a channel occupying a bandwidth f located at distance r as:<br />
B t (r) = fc j when S t ∈ [h(r, λ)g j , h(r, λ)g j+1 ) ,<br />
for j = 0, 1, ..., 15. (14)<br />
Hence, the DF (Distribution Function) of the bandwidth distribution<br />
for a user located at distance r; B(y, r) = P (B t (r) ≤<br />
y) may be written:<br />
B(y, r) = S(h(r, λ)g j+1 ), for y ∈ (fc j , fc j+1 ]<br />
for j = 0, 1, ..., 15, (15)<br />
where S(x) is the DF of the variable fading component.<br />
Hence, we obtain the k’moment of the obtainable bitrate for<br />
a user located at a distance r from the antenna as the (finite)<br />
sum:<br />
∑15<br />
m k (r) = f k (<br />
c<br />
k<br />
j − c k )<br />
j−1 ˜S(h(r, λ)gj ), (16)<br />
j=1<br />
where ˜S(x) = 1 − S(x) is the CDF of the variable fading<br />
component.<br />
Rather than applying the discrete modeling approach above<br />
we may prefer to apply the smooth (continuous) counterpart<br />
defined by relation (2). The bit-rate B t (r) for a channel<br />
occupying a bandwidth f located at distance r is then given<br />
by<br />
B t (r) = d min[T, ln(1 + S t /g(r, λ))], (17)<br />
with d = f C<br />
c15 ln 2<br />
ln 2<br />
and T =<br />
C<br />
and where C is the<br />
downscaling constant (relative to the Shannon formula) and<br />
where we also define g(r, λ) = γ −1 h(r, λ). For the continuous<br />
bandwidth case the DF of the bandwidth distribution for a user<br />
located at distance r is given by:<br />
{<br />
S(g(r, λ)(e<br />
B(y, r) =<br />
y/d − 1)) for y/d < T<br />
(18)<br />
1 for y/d ≥ T<br />
Based on (18) we may write the k’moment of the obtainable<br />
bitrate for a user located at a distance r from the antenna:<br />
m k (r) = d k<br />
∫<br />
g(r,λ)(e T −1)<br />
y=0<br />
(ln (1 + y/g(r, λ))) k s(y)dy +<br />
+d k T k ˜S(g(r, λ)(e T − 1) (19)<br />
E. Distribution of the obtainable bitrate for channel of a<br />
certain bandwidth for a user that is randomly placed in a<br />
circular cell with power-law attenuation<br />
Since the bitrate/capacity for a user will strongly depend of<br />
the distance from the sender antenna, a better measure of the<br />
capacity will be to find the distribution of bitrate for a user that<br />
is randomly located in the cell. This is done by averaging over<br />
the cell area and therefore the distribution of the ∫ corresponding<br />
averaging bitrate B t is given as B(y) = 1 A<br />
B(y, r)dA(r)<br />
A<br />
where A is the cell area. For circular cell shape and power law<br />
attenuation on the form h(r, λ) = h(λ)r α (where we also take<br />
g(λ) = γ −1 h(λ) i.e. g(r, λ) = g(λ)r α ) the corresponding<br />
integral may be partly evaluated. By defining an α-factor<br />
averaging variable S α with DF S α (x) = P (S α ≤ x) given<br />
by<br />
S α (x) = 2 α x− 2 α<br />
and with PDF<br />
∫ x<br />
t=0<br />
s α (x) = 2 α x− 2 α −1<br />
t 2 α −1 S(t)dt = 2 α<br />
∫x<br />
t=0<br />
t 2 2<br />
α s(t)dt =<br />
α<br />
∫ 1<br />
t=0<br />
∫ 1<br />
t=0<br />
t 2 α −1 S(tx)dt (20)<br />
t 2 α s(tx)dt (21)<br />
the bitrate distribution will have the exact same form as (15)<br />
for the discrete bandwidth case and (18) for the continuous<br />
bandwidth case, and with moments given by (16) and (19) by<br />
changing r → R and S(x) → S α (x) (and s(x) → s α (x)).<br />
1) Distribution of the stochastic variable S α for Lognormal<br />
and Suzuki distribution: Based on the definition we<br />
may derive the CDF and PDF of stochastic variable S α for<br />
the Log-normal and Suzuki distributed fading models. For the<br />
Log-normal distribution we have<br />
˜S lnα (x) = 1<br />
αx 2/α<br />
∫<br />
x<br />
t=0<br />
( ) ln t<br />
t 2/α−1 erfc<br />
σ √ dt.<br />
2
ØSTERBØ: SCHEDULING AND CAPACITY ESTIMATION IN LTE 35<br />
By changing variable according to y = ln t in the integral we<br />
find:<br />
˜S lnα (x) = 1 ( ( ) ln x<br />
erfc<br />
2 σ √ +<br />
2<br />
(<br />
+x −2/α e 2σ2 /α 2 2σ 2 ))<br />
− α ln x<br />
erfc<br />
ασ √ (22)<br />
2<br />
and further the PDF is found by differentiation:<br />
s lnα (x) = 1 (<br />
α x−(2/α+1) e 2σ2 /α 2 2σ 2 )<br />
− α ln x<br />
erfc<br />
ασ √ 2<br />
(23)<br />
For the Suzuki distribution we have the CDF given by the<br />
integral ˜Ssu (x) = x ∫ ∞<br />
t=0 t−2 e −t s ln (x/t)dt and therefore we<br />
have:<br />
˜S suα (x) = 2 α<br />
= x<br />
∫ 1<br />
t=0<br />
∫ ∞<br />
t=0<br />
t 2/α−1 ˜Ssu (xt)dt<br />
t −2 e −t s lnα (x/t)dt (24)<br />
where s lnα (x) is given by (23) above for the Lognormal<br />
distribution. As for the Suzuki distribution approximation to<br />
any accuracy is possible to obtain of ˜S suα (x) by truncating<br />
the integral above:<br />
˜S suα (x, M) = x<br />
∫ M<br />
t=0<br />
t −2 e −t s lnα (x/t)dt (25)<br />
and also for this case we find that the truncation error is<br />
exponentially small. By expanding e −t = ∑ ∞ (−1) k t k<br />
k=0 k!<br />
and<br />
integrating term by term we find:<br />
∞∑ (−1)<br />
˜S k x k (<br />
suα (x, M)=<br />
(2 + kα)k! e k2 σ 2 kσ<br />
2 erfc √ + ln(x/M) )<br />
k=0<br />
2 σ √ +<br />
2<br />
( ) (<br />
/α 2<br />
2 2σ<br />
+ e2σ2 γ<br />
α α , M x −2/α 2 )<br />
−α ln(x/M)<br />
erfc<br />
ασ √ 2<br />
(26)<br />
where γ(a, x) = ∫ x<br />
t=0 ta−1 e −t dt is the incomplete gamma<br />
function. (Observe the similarity with the corresponding expansion<br />
for ˜S su (x) by (10).)<br />
The corresponding integral for the PDF is given by:<br />
s suα (x) =<br />
∫ ∞<br />
t=0<br />
t −1 e −t s lnα (x/t)dt (27)<br />
and we take the truncated approximation of the PDF as the<br />
integral:<br />
s suα (x, M) =<br />
∫ M<br />
t=0<br />
t −1 e −t s lnα (x/t)dt (28)<br />
and we find the following error bound: 0 ≤ s suα (x) −<br />
σ2 −M+<br />
s suα (x, M) ≤ e 2 . By the similar approach as for the<br />
CDF we find the following series expansion of the truncated<br />
PDF:<br />
s suα (x, M) =<br />
∞∑ (−1) k x k<br />
(<br />
=<br />
(2+(k+1)α)k! e (k+1)2 σ 2 (k+1)σ<br />
2 erfc √ + ln( x M ) )<br />
k=0<br />
2 σ √ 2<br />
+ e ( ) ( 2σ2<br />
α 2 2 2σ<br />
α γ α +1,M x −(1+2 α) 2 −α ln( x<br />
erfc<br />
M ) )<br />
ασ √ 2<br />
(29)<br />
III. ESTIMATION OF CELL CAPACITY<br />
In the following we assume that the cell is loaded by two<br />
traffic types:<br />
• High priority CBR traffic sources that each requires to<br />
have a fixed data-rate and<br />
• Low priority (greedy) data sources that always consumes<br />
the leftover capacity not used by the CBR traffic.<br />
This is actually a very realistic traffic scenario for future LTE<br />
networks where we actual will have a mix of both real time<br />
traffic like VoIP and typical elastic data traffic. Below, we<br />
first estimate the RB usage of the high priority CBR traffic,<br />
and then we may subtract the corresponding RBs to find the<br />
actual numbers of RBs available for the (greedy) data traffic<br />
sources. Then finally we estimate the cell throughput/capacity<br />
as the sum of the bitrates offered to the CBR and (greedy)<br />
data sources.<br />
A. Estimation of the capacity usage for GBR sources in LTE<br />
The reservation strategy considered simply allocate recourses<br />
on a per TTI bases and allocate RBs so that the<br />
aggregate rate equals the required GBR (Guaranteed Bit Rate)<br />
rate (Non-Persistent scheduling).<br />
1) Capacity usage for a single GBR source : We first<br />
consider the case where we know the location of the CBR<br />
user in the cell, i.e. at a distance r from the antenna. We take<br />
B as the bitrate obtainable for a single RB and consider a GBR<br />
source that requires a fixed bit-rate of b CBR . We assumes that<br />
this is achieved by offering n RBs for every k-TTI interval.<br />
A way of reserving resources to GBR sources is to allocate<br />
RBs so that n k B will be close to the required rate bCBR over<br />
a given period. We take N CBR = n k<br />
to be the number of<br />
RBs granted to a GBR connection in a TTI as (the stochastic<br />
variable):<br />
N CBR =<br />
{ αb<br />
CBR<br />
B<br />
if CQI > 0<br />
0 if CQI = 0 , (30)<br />
where we have introduced a scaling factor α so that on the<br />
long run we obtain the desired GBR-rate b CBR . By choosing<br />
α = p −1<br />
CQI where p CQI = P (CQI > 0) = ˜S(h(r, λ)g 1 ) then<br />
E [N CBR B] = b CBR and hence we also have:<br />
E [N CBR |CQI > 0] = bCBR<br />
p CQI<br />
E [ B −1 |CQI > 0 ] . (31)<br />
The mean numbers of RBs is therefore:<br />
β = β(r, b CBR ) = b CBR m CQI<br />
−1 (r), (32)
36 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
where the conditional moments m CQI<br />
k<br />
(r) =<br />
E [ B k |CQI > 0 ] is found as<br />
⎛<br />
m CQI<br />
k<br />
(r) =<br />
f k<br />
˜S(h(r, λ)g 1 )<br />
⎝c k 1 ˜S(h(r, λ)g 1 )+<br />
⎞<br />
∑15<br />
(<br />
+ c<br />
k<br />
j − c k )<br />
j−1 ˜S(h(r, λ)gj ) ⎠ , (33)<br />
j=2<br />
for the discrete bandwidth case and by<br />
m CQI<br />
k<br />
(r) =<br />
⎛<br />
d k ⎜<br />
⎝<br />
˜S(h(r, λ)g 1 )<br />
g(r,λ)(e<br />
∫<br />
T −1) (<br />
y=h(r,λ)g 1<br />
⎞<br />
(<br />
ln<br />
1+ y<br />
g(r,λ)<br />
)) k<br />
s(y)dy+<br />
+T k ˜S(g(r, λ)(e T ⎟<br />
− 1) ⎠ (34)<br />
for the continuous bandwidth case. Note that by conditioning<br />
on having CQI > 0 we exclude the users that are unable<br />
to communicate due to bad radio conditions and avoid the<br />
problems due to division of zero in the calculation of the mean<br />
of 1/B.<br />
For circular cells and power law attenuation we obtain the<br />
corresponding result as above by changing r → R and S(x) →<br />
S α (x).<br />
2) Estimation of RBs usage for several CBR sources: We<br />
first estimate the RB usage for a fixed number of M CBR<br />
sources located at distances r j from the antenna and with bitrate<br />
requirements b CBR<br />
j j = 1, ..., M. The total usage of RBs<br />
β CBR will be the sum the individual contribution from each<br />
source as given by (32):<br />
β CBR =<br />
M∑<br />
j=1<br />
β(r j , b CBR<br />
j ). (35)<br />
For the case with random location the expression gets even<br />
simpler:<br />
M∑<br />
β CBR = β(R, b CBR<br />
j ), (36)<br />
j=1<br />
i.e. we may add the CBR rates from all the sources in the cell.<br />
The corresponding throughput for the CBR sources is taken<br />
as the sum of the individual rates i.e.<br />
b CBR =<br />
M∑<br />
j=1<br />
b CBR<br />
j (37)<br />
B. Estimation of the capacity usage for a fixed number of<br />
greedy sources<br />
We shall estimate the capacity usage for a fixed number of<br />
greedy sources under the following assumptions:<br />
• There are totally K active (greedy) users that are placed<br />
random in the cell which always have traffic to send, i.e.<br />
we consider the cell in saturated conditions.<br />
• There is totally N available RBs and the scheduled user<br />
is granted all of them in a TTI interval.<br />
1) Scheduling of based on metrics: In the following we<br />
consider the case with K users that are located in a cell with<br />
distances from the sender antenna given by a distance vector<br />
r = (r 1 , ...., r K ) and we assume that the user scheduled in a<br />
TTI is based on:<br />
i schedul = arg max<br />
i=1,..,K {M i}, (38)<br />
where M i = M i (r) is the scheduling metric which also may<br />
depend on the location of all users (through the location vector<br />
r = (r 1 , ...., r K )). Hence, for the scheduler to choose user i,<br />
the metric M i must be larger than all the other metrics (for<br />
the other users), i.e. we must have M i > U i where<br />
U i =<br />
max<br />
k=1,..,K<br />
k≠i<br />
M k . (39)<br />
Since we assume that a user is granted all the RBs when<br />
scheduled, this gives the cell throughput when user is scheduled<br />
(located at distance r i ) to be NB(r i ), where B(r i ) is the<br />
corresponding obtainable bit-rate per RB. Hence, cell bit-rate<br />
distribution (with K users located in the cell with distance<br />
vector r = (r 1 , ...., r K )) may then be written as:<br />
B g (y, r) =<br />
K∑<br />
B i (y, r), where (40)<br />
i=1<br />
B i (y, r) = P (NB(r i ) ≤ y, M i (r) > U i (r)) (41)<br />
is bitrate distribution when user i is scheduled. Unfortunately,<br />
for the general case exact expression of the probabilities<br />
B i (y, r) is difficult to obtain mainly due to the involvement of<br />
the scheduling metrics. However, for some cases of particular<br />
interest closed form analytical expression is possible to obtain.<br />
For many scheduling algorithms the scheduling metrics is only<br />
function of the SINR for that particular user (and does not<br />
depend of the SINR for the other users) and for this case<br />
extensive simplification is possible to obtain. In the following<br />
we therefore assume that the metrics M i only are functions<br />
of their own SINR i and the location r i for that particular<br />
user, i.e. we have M i = M(S i , r i ), where we (for simplicity)<br />
also assume that M(x, r i ) is an increasing function of x<br />
with an unlikely defined inverse M −1 (x, r i ). The distribution<br />
functions for M i and U i =<br />
max<br />
k=1,..,K<br />
k≠i<br />
M k are then<br />
M i (x, r i ) = P (M i ≤ x) = S(M −1 (x, r i )) and (42)<br />
K∏<br />
U i (x, r) = P (U i ≤ x) = S(M −1 (x, r k )) (43)<br />
k=1,k≠i<br />
If we now condition on the value of S i = x in (41), we find<br />
the distribution of the cell capacity when user i is scheduled<br />
as:<br />
B i (y, r) =<br />
∫ ∞<br />
x=0<br />
(<br />
P B(r i ) ≤ y )<br />
∣ S i = x U i (M(x, r i ), r)s(x)dx.<br />
N<br />
(44)
ØSTERBØ: SCHEDULING AND CAPACITY ESTIMATION IN LTE 37<br />
By using (14) as the obtainable bit-rate per RB for the discrete<br />
case we find:<br />
B i (y, r) =<br />
∫<br />
h(r i,λ)g j+1<br />
x=0<br />
F i (x, r)s(x)dx, if y/N ∈ (fc j , fc j+1 ]<br />
for j = 0, 1, ..., 15, (45)<br />
where we now have defined the multiuser “scheduling” function<br />
F i (x, r) by:<br />
F i (x, r) = U i (M(x, r i ), r) =<br />
K∏<br />
k=1,k≠i<br />
S(M −1 (M(x, r i ), r k ))<br />
(46)<br />
Similar for the continuous case based on (17) as the<br />
obtainable bit-rate per RB gives:<br />
⎧<br />
g(r ⎪⎨ i,λ)(e<br />
∫<br />
y/dN −1)<br />
B i (y, r) =<br />
F i (x, r)s(x)dx for y/dN < T<br />
,<br />
x=0<br />
⎪⎩<br />
p i (r)<br />
for y/dN ≥ T<br />
(47)<br />
where p i (r) = ∫ ∞<br />
x=0 F i(x, r)s(x)dx is the probability that user<br />
is scheduled in a TTI (and therefore ∑ K<br />
i=1 p i(r) = 1).<br />
Finally, by assuming that all users are randomly located<br />
throughout the cell the corresponding bit-rate distribution<br />
is found by performing a K-dimensional averaging<br />
over all possible distance vectors r, over the<br />
cell; B g (y) = 1<br />
A K ∫A . . . ∫ A r 1 . . . r K B cell (y, r)dA 1 · · · dA K ,<br />
where A here is the cell area. Due to the special form<br />
of the function F i (x, r) = ∏ K<br />
k=1,k≠i S(M −1 (M(x, r i ), r k ))<br />
the “cell averaging” over the K − 1 dimension variables<br />
r 1 , . . . , r i−1 , r i+1 , . . . , r K (not including the variable r i )<br />
[ ⌢S(M(x, ] K−1<br />
yields the product ri ) where<br />
∫<br />
⌢ 1 S(y) =<br />
A<br />
A<br />
uS(M −1 (y, u))dA(u) (48)<br />
Hence, for the case when user i is located at distance r i<br />
and all the K − 1 other users located at random, then we find<br />
for the discrete case:<br />
∫<br />
B i (y, r i ) =<br />
h(r i,λ)g j+1<br />
x=0<br />
[ ⌢S(M(x,<br />
ri ))] K−1<br />
s(x)dx, if y/N ∈ (fc j , fc j+1 ]<br />
for j = 0, 1, ..., 15 (49)<br />
and for the continuous case:<br />
⎧<br />
g(r i,λ)(e<br />
∫<br />
y/dN −1)<br />
[<br />
⎪⎨<br />
⌢S(M(x, K−1<br />
ri ))]<br />
s(x)dx<br />
B i (y, r i )=<br />
x=0<br />
for y/dN < T<br />
⎪⎩<br />
p i (r i ) for y/dN ≥ T,<br />
(50)<br />
where p i (r i ) = ∫ [<br />
∞ ⌢S(M(x, K−1<br />
ri ))]<br />
x=0 s(x)dx is the probability<br />
that user i is scheduled. (Observe that the p i (r) = p(r)<br />
and B i (y, r) = B(y, r) only depend on the location r i and<br />
hence are equal for all the users.)<br />
For circular cell size the cell bit-rate distribution integrals<br />
above is reduced to:<br />
B g (y) = 2 ∫R<br />
R 2<br />
r=0<br />
r<br />
∫<br />
h(r,λ)g j+1<br />
x=0<br />
[ ⌢S(M(x, ] K−1s(x)dxdr<br />
K r))<br />
if y/N ∈ (fc j , fc j+1 ] ; for j = 0, 1, ..., 15<br />
(51)<br />
for the discrete case and<br />
⎧<br />
⎨<br />
R∫<br />
2<br />
B g (y) = R<br />
rL(y, r)dr for y/dN < T<br />
2<br />
(52)<br />
⎩<br />
r=0<br />
1 for y/dN ≥ T<br />
where L(y, r) = ∫ [<br />
g(r,λ)(e y/dN −1) ⌢S(M(x, ] K−1s(x)dx.<br />
K<br />
x=0 r))<br />
For the continuous case where we now have<br />
⌢<br />
S(y) =<br />
2<br />
R 2<br />
∫R<br />
r=0<br />
uS(M −1 (y, u))du (53)<br />
The moments of the capacity (when the users are located<br />
according to the vector r = (r 1 , ...., r K ) may be written as:<br />
E[B g (r) k ] = f k N k<br />
∫<br />
∑<br />
K h(r<br />
∑15<br />
i,λ)g j+1<br />
c k j<br />
i=1 j=1<br />
x=h(r i,λ)g j<br />
for the discrete bandwidth case and<br />
E[B g (r) k ] =<br />
⎛<br />
∑<br />
K<br />
= d k N k ⎜<br />
⎝<br />
i=1<br />
∫∞<br />
+T k<br />
g(r i,λ)(e T −1)<br />
F i (x, r)s(x)dx (54)<br />
∫<br />
(ln(1+x/g(r i , λ))) k F i (x, r)s(x)dx<br />
x=0<br />
x=g(r i,λ)(e T −1)<br />
⎞<br />
⎟<br />
F i (x, r)s(x)dx⎠ , (55)<br />
for the continuous case.<br />
The corresponding moments for the case where the users<br />
are randomly located in a circular cell are given by:<br />
E[Bg k ]= 2f k N k ∑15<br />
R 2<br />
∫ R<br />
c k j r<br />
j=1<br />
r=0<br />
⎧<br />
⎪⎨<br />
⎪⎩<br />
⎫<br />
∫ [ ⌢S(M(x, ] K−1<br />
⎪⎬<br />
K r)) s(x)dx<br />
⎪⎭ dr<br />
h(r,λ)g j+1<br />
x=h(r,λ)g j<br />
for the discrete bandwidth case and<br />
E[B k g ] =<br />
= 2dk N k<br />
R 2<br />
⎛<br />
∫R<br />
⎜<br />
r⎝<br />
r=0<br />
∫∞<br />
+T k<br />
x=g(r,λ)(e T −1)<br />
∫<br />
g(r,λ)(e T −1)<br />
x=0<br />
(56)<br />
( (<br />
ln 1 + x )) k<br />
L(x, r)s(x)dx<br />
g(r, λ)<br />
⎞<br />
⎟<br />
L(x, r)s(x)dx⎠ dr (57)<br />
for the continuous case, where L(x, r) =<br />
[ ⌢S(M(x, ] K−1.<br />
K r))
38 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
2) Examples: Below we consider and compare three of the<br />
most commonly known scheduling algorithms, namely Round<br />
Robin (RR), Proportional Fair (PF) and Max SINR by applying<br />
the cell capacity models described above.<br />
a) Round Robin : For the Round Robin algorithm each<br />
user is given the same amount of bandwidth and hence this<br />
case corresponds to taking K = 1 i.e. the results in section II<br />
may be applied by to find the cell capacity with f → Nf and<br />
S(x) → S su (x) and also S α (x) → S suα (x).<br />
b) Proportional Fair (in SINR) : Normally, the shadowing<br />
is varying over a much longer time scale than the TTI<br />
intervals, and hence we may assume that the slow fading is<br />
constant during the updating of the scheduling metric M i and<br />
therefore should only account for the rapid fading component.<br />
This means that the shadowing effect may be taken as constant<br />
that may be included in the non varying part of the SINR<br />
over several TTI intervals. Hence, we take SINR as S t /g(r; λ)<br />
where S t = zX e conditioned that the shadowing X ln = z. By<br />
assuming that X ln = z is constant over the short TTI intervals<br />
zXe/h(ri,λ)<br />
zE[X =<br />
e]/h(r i,λ)<br />
the scheduling metrics will be M i =<br />
In the final result we then “integrate over the Log-normal<br />
slow fading component”. We find that the probability of being<br />
scheduled is p(r) = 1 K<br />
and that the conditional bandwidth<br />
distribution for a user at located at distance r (and the K − 1<br />
users random located) is given by the results in section II-D<br />
with f → Nf and S(x) → S K (x) with:<br />
S K (x) =<br />
s K (x) =<br />
∫ ∞<br />
t=0<br />
∫ ∞<br />
t=0<br />
S e<br />
( x<br />
t<br />
) K<br />
sln (t)dt and<br />
Xe<br />
E[X . e]<br />
K<br />
( x<br />
) K−1 ( x<br />
)<br />
t S e se s ln (t)dt, (58)<br />
t t<br />
where S e (x) = 1 − e −x and s e (x) = e −x .<br />
Further, the distribution of the cell capacity is given by<br />
the results in section II-E with f → Nf and further the α-<br />
averaging is given by the integrals:<br />
˜S Kα (x) =<br />
s Kα (x) =<br />
∫ ∞<br />
t=0<br />
∫ ∞<br />
t=0<br />
KS e (t) K−1 s e (t) ˜S lnα<br />
( x<br />
t<br />
KS e (t) K−1 s e (t)t −1 s lnα<br />
( x<br />
t<br />
)<br />
dt and (59)<br />
)<br />
dt (60)<br />
c) Max SINR algorithm.: For this algorithm the scheduling<br />
metric is M i = S i /h(r i , λ). By assuming circular cell size<br />
and radio signal attenuation on the form h(r, λ) = h(λ)r α<br />
gives:<br />
⌢<br />
S(M(x, r)) =<br />
2<br />
R 2<br />
∫R<br />
uS(x(u/r) α )du = S α (x(R/r) α ). (61)<br />
r=0<br />
We find that the probability of being scheduled<br />
p(r) =<br />
∫ ∞<br />
x=0<br />
[S α (x(R/r) α )] K−1 s(x)dx (62)<br />
and that the conditional bandwidth distribution for a user<br />
located at distance r (and the K − 1 users random located)<br />
is given by the results in section II-D with f → Nf and<br />
S(x) → S c (x; r) with:<br />
S c (x; r) = 1<br />
p(r)<br />
∫ x<br />
y=0<br />
[S α (y(R/r) α )] K−1 s(y)dx (63)<br />
It turns out that extensive simplifications occur for the case<br />
where all the users are randomly located in the cell and we find<br />
that the distribution of the cell capacity is given by the results<br />
in section II-E with f → Nf and further the α-averaging<br />
is given by taking S α (x) → S α (x) K i.e. is simply the K’th<br />
power of the α-averaging of S(x).<br />
C. Combining real-time and non real time traffic over LTE<br />
We are now in the position to combine the analysis in<br />
sections III.A and III.B to obtain complete description of the<br />
resource usage in a LTE cell. The combined modeling is based<br />
on the following assumptions:<br />
• There are M CBR sources applying one of the allocation<br />
options described in section III.A.<br />
• There are totally K active (greedy) data sources which<br />
always have traffic to send, i.e. we consider the cell in<br />
saturated conditions.<br />
• The number of available RBs is taken to be N.<br />
Since the CBR sources have “absolute” priority over the data<br />
sources, they will always get the number of RBs they need<br />
and hence the leftover RBs will be available for the Non-<br />
GBR data sources. By conditioning on the RB usage of the<br />
GBR sources we may apply all the results derived in section<br />
III.B with available RBs taken to be the leftover RBs not used<br />
by the CBR sources. Then we may find the average usage of<br />
RBs for the CBR traffic as done in section III.A.<br />
We consider first the case where the location of the sources<br />
is given:<br />
• CBR sources are located at distances s j from the antenna<br />
with bit-rate requirements b CBR<br />
j ; j = 1, ..., M.<br />
• The greedy data sources are located at distance r i (i =<br />
1, ..., K).<br />
With these assumptions the mean cell throughput is given<br />
as:<br />
B cell =<br />
⎛<br />
⎞<br />
M∑<br />
K∑ ∑15<br />
=f⎝N−<br />
β(s j , b CBR ⎠<br />
+<br />
M∑<br />
j=1<br />
j=1<br />
j )<br />
c j<br />
h(r i,λ)g j+1<br />
i=1 j=1<br />
x=h(r i,λ)g j<br />
∫<br />
F i (x, r)s(x)dx +<br />
b CBR<br />
j , (64)
ØSTERBØ: SCHEDULING AND CAPACITY ESTIMATION IN LTE 39<br />
for the discrete bandwidth case and<br />
B cell =<br />
⎛<br />
⎞ ⎛<br />
M∑<br />
K∑<br />
= d ⎝N − β(s j , b CBR ⎠ ⎜<br />
⎝V i (x, r) +<br />
T<br />
∫ ∞<br />
x=g(r i,λ)(e T −1)<br />
j=1<br />
j )<br />
j=1<br />
i=1<br />
⎞<br />
⎟<br />
M∑<br />
F i (x, r)s(x)dx⎠ + b CBR<br />
j , (65)<br />
for the continuous bandwidth case; where V i (x, r) =<br />
∫ g(ri,λ)(e T −1)<br />
x=0<br />
ln(1 + x/g(r i , λ))F i (x, r)s(x)dx, β(r, b CBR )<br />
is given by by (32) and further F i (x, r) is defined by (46). For<br />
circular cells and power law attenuation on the form h(r, λ) =<br />
h(λ)r α and randomly placed sources the corresponding cell<br />
throughput is found to:<br />
where<br />
B cell =<br />
⎛<br />
M∑<br />
= f ⎝N − β(R,<br />
+<br />
V j (x, r) =<br />
M∑<br />
j=1<br />
∫ R<br />
r=0<br />
b CBR<br />
j )<br />
j=1<br />
⎞<br />
⎠ 2 ∑15<br />
R 2 c j V j (x, r)<br />
j=1<br />
b CBR<br />
j (66)<br />
⎧<br />
⎪⎨<br />
r<br />
⎪⎩<br />
h(r,λ)g<br />
∫ j+1<br />
x=h(r,λ)g j<br />
K<br />
for the discrete bandwidth case and<br />
B cell =<br />
= d(N − β(R,<br />
+T<br />
∫ ∞<br />
x=g(r,λ)(e T −1)<br />
M∑<br />
j=1<br />
[ ⌢S(M(x, r))<br />
] K−1s(x)dx<br />
⎫<br />
⎪⎬<br />
⎪ ⎭<br />
dr<br />
b CBR<br />
j )) 2 ∫R<br />
R 2<br />
r=0<br />
r<br />
⎧<br />
⎪⎨<br />
V (x, r)<br />
⎪⎩<br />
⎫<br />
[ ⌢S(M(x, ] K−1<br />
⎪⎬ M<br />
K r)) s(x)dx<br />
⎪⎭ dr + ∑<br />
j=1<br />
b CBR<br />
j<br />
(67)<br />
for the continuous bandwidth case; where V (x, r) =<br />
∫ [ g(r,λ)(e T −1)<br />
⌢S(M(x, ] K−1<br />
ln(1 + x/g(r, λ))K<br />
x=0 r)) s(x)dx,<br />
β = β(r, b CBR ) is given by (32) and further ⌢ S(M(x, r)) is<br />
defined by (53). Observe that the CBR traffic only will affect<br />
the cell throughput by the sum ∑ M<br />
j=1 bCBR j of the rates and<br />
not the actual number of CBR sources.<br />
IV. DISCUSSION OF NUMERICAL EXAMPLES<br />
In the following we give some numerical example of<br />
downlink performance of LTE. Before describing the results<br />
we first rephrase some of the main assumptions:<br />
• The fading model includes lognormal shadowing (slow<br />
fading) and Rayleigh fast fading.<br />
• The noise interference is assumed to be constant over the<br />
cell area.<br />
Parameters<br />
TABLE II<br />
INPUT PARAMETERS FOR THE NUMERICAL CALCULATIONS<br />
Bandwidth per Resource Block<br />
Total Numbers of Resource Blocks<br />
(RB)<br />
Distance-dependent path loss. (The<br />
actual model is found in [4].)<br />
Lognormal Shadowing with standard<br />
deviation<br />
Rayleigh fast fading<br />
Noise power at the receiver<br />
Total send power<br />
Numerical values<br />
180 kHz=12x 15 kHz<br />
100 RBs for 2Ghz<br />
L = C + 37.6 log 10 (r),<br />
r in kilometers and<br />
C=128.1 dB for 2GHz,<br />
8 dB (in moust of the cases)<br />
-101 dBm<br />
46.0 dBm=(40W)<br />
Radio signaling overhead 3/14<br />
• The cell shape is circular.<br />
Basically, there are three different cases we would like to<br />
investigate. First and foremost is of course the actual efficiency<br />
of the LTE radio interface. We choose the bitrate obtainable for<br />
the smallest unit available for users, namely a Resource Block<br />
(RB). Since different implementation may chose different<br />
bandwidth configurations the performance based on RBs will<br />
give a good indication of the overall capacity/throughput<br />
for the LTE radio interface. Secondly, we know that the<br />
scheduling also will affect the overall throughput for a LTE<br />
cell. Based on the modeling we are able to investigate the<br />
performance of the three basic scheduling algorithms: Round<br />
Robin (RR), Proportional Fair (PF) and Max-SINR. All these<br />
three algorithms have their weaknesses and strengths, like<br />
Max-SINR that try to maximize the throughput but at the cost<br />
of fairness among users. Thirdly, we would also investigate<br />
the effect on overall performance by introducing GBR traffic<br />
in LTE. Normally, GBR traffic will higher priority than Non-<br />
GBR or “best effort” traffic and to guarantee a particular rate<br />
the number of radio resources required may vary depending<br />
on the radio conditions. For users with bad radio conditions<br />
i.e. located at cell edge the resource usage to maintain a<br />
fixed guaranteed rate may be quite high so an investigation<br />
of the cell performance with both GBR and Non-GBR will be<br />
important.<br />
A. LTE spectrum efficiency<br />
First, we consider bitrate that is possible to obtainable for<br />
the basic resource unit in LTE namely a RB. In the examples<br />
we have considered sending frequency of 2 GHz. The aim is<br />
to predict the bandwidth efficiency, i.e. the obtainable bitrate<br />
per RB. The rest of the input parameters are given in Table 2.<br />
The mean obtainable bitrate per RB is depicted in Figure<br />
3. With our assumptions the maximum bitrate is just below<br />
0.8 Mbit/s for excellent radio conditions. The mean bit-rate is<br />
as expected a decreasing function of the cell size both for a<br />
randomly placed user and for a user at the cell edge. The mean<br />
bitrate have decreased to 0.1 Mbit/s per RB for cell sizes of<br />
approximately 2 km for shadowing std. equals 8 dB and when<br />
users are random located. The corresponding bit-rate for users<br />
at the cell edge is proximately 0.04 Mbit/s.
40 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
0.8<br />
80<br />
Mean<br />
throughpu<br />
t pe<br />
r R B @ Mbi<br />
t êsD<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
Located at cell<br />
edge<br />
Random<br />
location<br />
C ell c apacit y @ M bi t êsD<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
-- PF<br />
-- Max-SINR<br />
-- RR<br />
0.1<br />
10<br />
1 2 3 4 5<br />
Distance @KmD<br />
1 2 3 4 5<br />
Distance @KmD<br />
Fig. 3. Mean throughput right and std. left per RB for a user random located,<br />
and fixed located as function of cell radius with 2 GHz sending frequency and<br />
Suzuki distributed fading with std. of fading σ=0dB, 2dB, 5dB, 8dB, 12dB<br />
from below.<br />
B. Cell capacity and scheduling<br />
Below we examine the downlink performance in an LTE cell<br />
with the input parameters given by Table 2, however, with the<br />
following additional input parameters:<br />
• Type of scheduling algorithm i.e. RR, PF or Max-SINR,<br />
• number of RB available i.e. 100,<br />
• number of active (greedy) users.<br />
In Figure 4 the mean downlink cell capacity is depicted<br />
as function of the cell radius for RR, PF and the Max<br />
SINR scheduling algorithms. As expected the Round Robin<br />
algorithm gives the lowest cell throughput while the Max<br />
SINR algorithm give the highest throughput. For the latter<br />
the multiuser gain is huge and may be explained from the<br />
fact that when the number of users increases those who by<br />
chance are located near the sender antenna will with high<br />
probability obtain the best radio condition and will therefore be<br />
scheduled with high bit-rate. For users located at the cell edge<br />
the situation is opposite and those users will normally have low<br />
SINR and surely obtain very little of the shared capacity. This<br />
explains why the Max SINR will increase the throughput but<br />
is highly un-fair. For the PF only the relative size of the SINR<br />
is important and in this case each user has equal probability<br />
of sending in each TTI. The multiuser gain for this algorithm<br />
is much lower than for the Max SINR algorithm but is not<br />
negligible. When it comes to actual cell downlink throughput<br />
the expected values lays in the range 26-48 Mbit/s for cells<br />
with radius of 1 km while if the radius is increased to 2 km<br />
the cell throughput is reduced to approximately 10-20 Mbit/s.<br />
As seen from Figure 4 the Max-SINR algorithm will over<br />
perform the PF algorithm when it comes to cell throughput.<br />
But if we consider fairness among users the picture is complete<br />
different. When considering the performance of users located<br />
at cell edge the Max-SINR algorithm actually performs very<br />
badly. While PF give equal probability of transmitting in a<br />
TTI for all active users the Max-SINR strongly discriminate<br />
the user close to cell edge. As seen from Table 3 below; if there<br />
are totally 10 active users in a cell the PF fairly give each user<br />
10% chance of accessing radio resource while the Max-SINR<br />
Fig. 4. Multiuser gain as function of cell radius for Max-SINR (red), PF<br />
(blue) and RR (black) scheduling, 2GMHz frequency with 100 RB and with<br />
Suzuki distributed fading with std. σ = 8 dB. The number of users is 1, 2,<br />
3, 5, 10, 25, 100 from below.<br />
TABLE III<br />
PROBABILITY THAT A USER IS SCHEDULED AS FUNCTION OF <strong>NUMBER</strong>S OF<br />
USERS AND LOCATION FOR PF AND MAX-SINR SCHEDULING<br />
ALGORITHMS, SUZUKI DISTRIBUTED FADING WITH STD. OF 8DB.<br />
Number of users PF MAX-SINR<br />
r/R=1 r/R=0.5 r/R=0.25 r/R=0.1<br />
2 0.50 0.308708 0.594756 0.82579 0.96119<br />
3 0.33 0.147869 0.414839 0.71126 0.92784<br />
5 0.20 0.055113 0.245871 0.56102 0.87130<br />
10 0.10 0.012690 0.104912 0.36531 0.76418<br />
25 0.04 0.001356 0.025222 0.16326 0.56989<br />
100 0.01 0.000019 0.001293 0.02453 0.24325<br />
only give 1.2% chance of accessing the radio resources if a<br />
user is located at cell edge. As the number of user increases<br />
this unfairness increases even more.<br />
Table 3 demonstrates one of the unfortunate properties of<br />
the MAX-SINR scheduling algorithm. While the PF algorithm<br />
distribute the capacity among the users with equal probability<br />
the MAX-SINR algorithm is far more unfair when it comes to<br />
the distribution of the available radio resources. For instance,<br />
the users located at the cell edge e.g. r/R=1 will suffer from<br />
extremely poor performance if the numbers of users is higher<br />
than 10. The Max-SINR algorithm will also be unfair for<br />
small cell sizes where users actually may have so high signal<br />
quality that most of them may use coding with high data rate<br />
i.e. 64 QAM with high rare and there should be no need for<br />
scheduling according highest SINR to obtain high throughput.<br />
C. Use of GBR in LTE<br />
It is likely the LTE in the future will carry both real time<br />
type traffic like VoIP and elastic data traffic. This is possible<br />
by introducing GBR bearers where users are guaranteed the<br />
possibility to send at their defined GBR rate. The GBR traffic<br />
will have priority over the Non-GBR traffic such that the RBs<br />
scheduled for GBR bearers will normally not be accessible<br />
for other type of traffic. However, the resource usage over<br />
the radio interface in LTE will strongly depends on the radio
ØSTERBØ: SCHEDULING AND CAPACITY ESTIMATION IN LTE 41<br />
conditions. This means that the amount of radio resources a<br />
user occupies (to obtain a certain bit rate) will vary according<br />
to the local radio conditions and a user at the cell edge must<br />
seize a larger number of resource blocks (RBs) to maintain a<br />
constant rate (GBR bearer) than a user located near the antenna<br />
with good radio signals.<br />
An interesting example is to see the effect of multiplexing<br />
traffic with both greedy and GBR users and observe the effect<br />
on the cell throughput. In Figure 5 we consider the cases where<br />
10 greedy users are scheduled by the PF algorithm together<br />
with a GBR user with guaranteed rate of 3, 1, 0.3 or 0.1 Mbit/s.<br />
We consider the cases where either the GBR user is located<br />
at cell edge or have random location throughout the cell.<br />
We observe that thin GBR connections do not have big<br />
impact on the cell throughput. From the figures it seems that<br />
GBR bearers up to 1 Mbit/s should be manageable without<br />
influencing the cell performance very much. But a 3 Mbit/s<br />
GBR connection will lower the total throughput by a quite big<br />
factor especially if the user is located at cell edge. For instance<br />
we observe for both cases that the effective reduction in cell<br />
throughput is approximately 20 Mbit/s for a user requiring a<br />
3 Mbit/s GBR connection when located at cell edge. As a<br />
consequence we recommend limiting GBR connection to less<br />
than 1 Mbit/s.<br />
We therefore recommend using high GBR values with<br />
particular caution. The GBR should be limited to a maximum<br />
rate to avoid that a particular GBR user consumes a too large<br />
part of the radio resources (too many RBs). A good choice of<br />
the actual maximum GBR value seems to be around 1 Mbit/s.<br />
V. CONCLUSIONS<br />
With the introduction of LTE the capacity in the radio<br />
network will increase considerably. This is mainly due to the<br />
efficient and sophisticated coding methods developed during<br />
the last decade. However, the cost of such efficiency is that<br />
the variation due to radio conditions will increase significantly<br />
and hence the possible capacity for users in terms of bitrate<br />
will vary a lot depending on the current radio conditions.<br />
The two most important factors for the radio conditions are<br />
fading and attenuation due to distance. By extensive analytical<br />
modeling where both fading and the attenuation due the<br />
distance are included we obtain performance models for:<br />
• Spectrum efficiency through the bitrate distribution per<br />
RB for customers that are either randomly or located at<br />
a particular distance in a cell.<br />
• Cell throughput/capacity and fairness by taking the<br />
scheduling into account.<br />
• Specific models for the three basic types of scheduling<br />
algorithms; Round Robin, Proportional Fair and Max<br />
SINR.<br />
• Cell throughput/capacity for a mix of GBR and Non-GBR<br />
(greedy) users.<br />
Numerical examples for LTE downlink show results which are<br />
reasonable; in the range 25-50 Mbit/s for 1 km cell radius at<br />
2GHz with 100 RBs. The multiuser gain is large for the Max-<br />
SINR algorithm but also the Proportional Fair algorithm gives<br />
relative large gain relative to plain Round Robin. The Max-<br />
SINR has the weakness that it is highly unfair in its behaviour.<br />
C ell c apacit y @ M bi t êsD<br />
C ell c apacit y @ M bi t êsD<br />
C ell c apacit y @ Mbi<br />
t êsD<br />
C ell c apacit y @ M bi t êsD<br />
80<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
80<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
80<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
80<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
0.5 1 1.5 2 2.5<br />
Distance @KmD<br />
0.5 1 1.5 2 2.5<br />
Distance @KmD<br />
GBR=0.3 Mbit/s<br />
-- Non-Persistent, cell edge<br />
-- Non-Persistent, random<br />
-- mean PF 10 users<br />
GBR=1 Mbit/s<br />
0.5 1 1.5 2 2.5<br />
Distance @KmD<br />
GBR=0.1 Mbit/s<br />
-- Non-Persistent, cell edge<br />
-- Non-Persistent, random<br />
-- mean PF 10 users<br />
GBR=3 Mbit/s<br />
-- Non-Persistent, cell edge<br />
-- Non-Persistent, random<br />
-- mean PF 10 users<br />
-- Non-Persistent, cell edge<br />
-- Non-Persistent, random<br />
-- mean PF 10 users<br />
0.5 1 1.5 2 2.5<br />
Distance @KmD<br />
Fig. 5. Mean cell throughput for PF, 10 users and a GBR user of 3.0,<br />
1.0, 0.3, 0.1 Mbit/s using non-persistent scheduling, for 2 GHz and 100 RB<br />
and Suzuki distributed fading with std. σ = 8dB. Red curves corresponds to<br />
random location and blue for user located at cell edge.
42 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
User at cell edge with poor radio condition will obtain very<br />
little data throughput. It turns out that the grade of unfairness<br />
increases with the numbers of active users. This unfortunate<br />
property is not found for the Proportional Fairs scheduling<br />
algorithm.<br />
The usage of GBR with high rates may cause problems in<br />
LTE due to the high demand for radio resources if users have<br />
low SINR i.e. at cell edge. For non-persistent GBR allocation<br />
the allowed guaranteed rate should be limited. It seems that a<br />
limit close to 1 Mbit/s will be a good choice.<br />
REFERENCES<br />
[1] H. Holma and A. Toskala, LTE for UMTS, OFDMA and SC-FDMA Based<br />
Radio Access. Wiley, 2009.<br />
[2] R.-. 3GPP TSG-RAN1#48, “LTE physical layer framework for performance<br />
verification,” 3GPP, St. Louis, MI, USA, Tech. Rep., Feb. 2007.<br />
[3] H. Kushner and P. Whiting, “Asymptotic Properties of Proportional-<br />
Fair Sharing Algorithms,” in Proc. of 2002 Allerton Conference on<br />
Communication, Control and Computing, Oct. 2002.<br />
[4] 3GPP TS 36.213 V9.2.0, “Physical layer procedures, Table 7.2.3-1: 4-bit<br />
CQI Table,” 3GPP, Tech. Rep., Jun. 2010.<br />
[5] M. C, M. Wrulich, J. C. Ikuno, D. Bosanska, and M. Rupp, “Simulating<br />
the Long Term Evolution Physical Layer,” in Proc. of 17th European<br />
Signal Processing Conference (EUSIPCO 2009), Glasgow, Scotland, Aug.<br />
2009.<br />
[6] B. Sklar, “Rayleigh Fading Channels in Mobile Digital Communication<br />
Systems Part I: Characterization and Part II: Mitigation,” IEEE Commun.<br />
Mag., Jul. 1997.<br />
Olav N. Østerbø received his MSc in Applied Mathematics from the<br />
University of Bergen in 1980 and his PhD from the Norwegian University<br />
of Science and Technology in 2004. He joined Telenor in 1980. His main<br />
interests include teletraffic modeling and performance analysis of various<br />
aspects of telecom networks. Activities in recent years have been related<br />
to dimensioning and performance analysis of IP networks, where the main<br />
focus is on modeling and control of different parts of next generation IPbased<br />
networks.
ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong> 43<br />
Multi-service Load Balancing in a Heterogeneous<br />
Network with Vertical Handover<br />
Jie Xu, Yuming Jiang, Andrew Perkis, and Elissar Khloussy<br />
Abstract—In this paper we investigate multi-service load balancing<br />
mechanisms in an overlay heterogeneous WiMAX/WLAN<br />
network through vertical handover. Considering the service characteristics<br />
of the overlay heterogeneous network together with<br />
the service requirements of different applications, all streaming<br />
applications are served in WiMAX while elastic applications<br />
are distributed to WiMAX and WLAN. Two load balancing<br />
mechanisms are compared which switch the elastic application<br />
with maximum (MAX) and minimum (MIN) remaining size<br />
respectively to WLAN. Simulation results indicate that MIN<br />
outperforms MAX at the cost of significantly increased number<br />
of load balancing actions. Furthermore, it is discovered that both<br />
load balancing granularity and proper integration of streaming<br />
and elastic applications in WiMAX determine the whole system<br />
performance.<br />
Index Terms—vertical handover, wireless heterogeneous networks,<br />
load balancing, multi-service<br />
I. INTRODUCTION<br />
AFTER decades of research, it is commonly believed that<br />
future wireless network will employ multiple techniques.<br />
Especially, for the access part, multiple radio access technologies<br />
(RATs) will coexist in terms of both space and time. For<br />
example, nowadays there are already lots of WLAN networks<br />
which are also covered by other 3G mobile networks at the<br />
same time.<br />
The coexistence of heterogeneous networks brings up both<br />
challenges and opportunities for providing better wireless<br />
service [1]. On the one hand, since wireless communications<br />
are intrinsically limited by interference, activities of different<br />
networks could interfere with each other and may result in<br />
severe service degradation. On the other hand, multiple overlay<br />
heterogeneous networks can provide more robust communication<br />
guarantee if they could cooperate instead of competition.<br />
Therefore, how to integrate coexisted multiple networks is<br />
of fundamental importance to the success of future wireless<br />
networks.<br />
To take advantage of multiple heterogeneous wireless networks,<br />
vertical handover [1] has been proposed as a means<br />
for enhancing end users’ service quality. Traditionally, due<br />
to its operation difficulty and introduced time delay, vertical<br />
handover is used as a reactive measure to prevent severe service<br />
degradation. Normally vertical handover is only triggered<br />
Jie Xu, Yuming Jiang and Andrew Perkis are with the Centre for Quantifiable<br />
Quality of Service in Communication Systems, Norwegian University of<br />
Science and Technology in Trondheim.<br />
Elissar Khloussy is with Department of Telematics, Norwegian University<br />
of Science and Technology in Trondheim.<br />
Center for Quantifiable Quality of Service in Communication Systems,<br />
Center of Excellence, is appointed by the Research Council of Norway, and<br />
funded by the Research Council, NTNU and UNINETT.<br />
when the served mobile user are about to move out of the<br />
coverage range of current serving network. To this end, various<br />
vertical handover mechanisms have been proposed to improve<br />
the performance of handover user [2][3].<br />
Recently, proper vertical handover are also used as a proactive<br />
means to improve the system performance. In [4][5],<br />
vertical handover is adopted as a tool for joint resource management<br />
in heterogeneous networks. The objective of vertical<br />
handover has been extended to include the whole system<br />
performance instead of the performance of handover user. In<br />
particular, the main idea is to distribute traffic load among<br />
heterogeneous networks in a balanced manner by designing<br />
vertical handover protocols. However, only streaming users<br />
are considered in these studies although the current wireless<br />
network normally serve multi-service applications. In [6], both<br />
streaming and elastic applications are considered. The authors<br />
preferably distribute streaming applications to cellular network<br />
because of its larger coverage and finer QoS guarantee, and the<br />
remaining capacities in cellular/WLAN networks are utilized<br />
for serving elastic applications. While the scheme in [6]<br />
performs well compared to random dispatch, there are still<br />
chances that streaming applications are distributed to WLAN<br />
and vertical handover is triggered whenever there are enough<br />
free capacities in the cellular network.<br />
In this study, we consider system performance of an overlay<br />
heterogeneous wireless network where elastic applications<br />
share network capacity with prioritized streaming applications.<br />
Specifically, we consider the WiMAX/WLAN heterogeneous<br />
network and assume all the traffic firstly arrives to<br />
the WiMAX network. Streaming applications are given strict<br />
preemptive priority over elastic applications in WiMAX. Then<br />
according to the comparison result of expected finish time<br />
in WiMAX and WLAN, vertical handover of certain elastic<br />
applications on their arrivals or during their service to WLAN<br />
is conducted. We compare the performance of two different<br />
handover mechanisms which selected the file with maximum<br />
and minimum remaining size for handover respectively. The<br />
results indicate that selection of files with minimum remaining<br />
size outperforms the other mechanism at the cost of significant<br />
increased number of handovers. Furthermore, based on<br />
analysis of simulation results, we conclude that both the load<br />
balancing granularity and integration of elastic and streaming<br />
applications in WiMAX determine the performance of the<br />
whole system.<br />
The remaining of this paper is arranged as follows. The<br />
system model is described in the next section. In section II, due<br />
to the complexity of exact analysis, theoretical approximations<br />
are given. In Section IV, we describe the simulation results,
44 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Fig. 1.<br />
WiMAX<br />
WLAN<br />
(a) typical application scenario.<br />
System model.<br />
Streaming applications<br />
Elastic applications<br />
Scheduling<br />
WiMAX<br />
Load balancing<br />
Scheduling<br />
(b) abstraction.<br />
WLAN<br />
followed by discussions in Section V. Finally, we conclude<br />
the paper in Section VI.<br />
II. SYSTEM MODEL<br />
A. WiMAX/WLAN Heterogeneous Network<br />
The network scenario of this study is shown in Fig. 1(a).<br />
Within the coverage area of the WiMAX network, there are<br />
multiple WLAN networks as well. WiMAX is infrastructurebased<br />
and can provide guaranteed service with centralized<br />
control. In addition, as one option for the future mobile<br />
networks, WiMAX has a large coverage area but limited rate<br />
to each end user. On the other hand, WLAN is contention<br />
based and incapable of providing service guarantee. However,<br />
since WLAN uses free frequency it can provide higher data<br />
rate to end users if the network is not congested. Therefore,<br />
WiMAX and WLAN networks are complementary in terms<br />
of service characteristics. It is beneficial to design the overlay<br />
heterogeneous network in a cooperative way.<br />
One possible way of cooperation is to distribute different<br />
types of applications to specific network by taking account of<br />
both the service requirement of applications and the network<br />
service characteristic. In this study, the cooperation scheme is<br />
shown in Fig. 1(b). Specially, all the streaming applications are<br />
served in WiMAX due to their stringent service requirements.<br />
Then the remaining service capacity of WiMAX and the total<br />
service capacity of WLAN are devoted to elastic applications.<br />
Furthermore, to prevent the disturbance of elastic applications,<br />
streaming applications are served with higher preemptive<br />
priority. Namely the streaming applications arrives and leaves<br />
without any disturbance of elastic applications.<br />
For streaming applications, the system can be modeled as<br />
an G/G/K loss-queue system. Specifically, if we assume the<br />
streaming applications arrive according to Poisson process and<br />
the duration follows minus exponential distribution, then the<br />
system can be seen as M/M/K Erlang system for which lots<br />
results have been obtained. Suppose the capacity of WiMAX is<br />
C wimax , then the number of channels K can be calculated as<br />
⌊C wimax /B s ⌋ where B s is the bandwidth requirement of each<br />
streaming application. In this study, when there are already<br />
K streaming applications in the system, the new streaming<br />
arrivals will be simply rejected and discarded.<br />
For elastic applications, both the remaining capacity of<br />
WiMAX and the total capacity of WLAN can be utilized. Due<br />
to the dynamic state of streaming applications, the remaining<br />
capacity of WiMAX varies. For WLAN, we assume the total<br />
capacity is fixed to simplify the analysis. Therefore, there are<br />
actually two servers for elastic applications. In each server,<br />
residing elastic applications share the capacity in processorsharing<br />
(PS) manner. In particular, no requirements on the<br />
maximum or minimum bandwidth for each elastic application<br />
is specified.<br />
B. Load Balancing Mechanisms<br />
Real-time load balancing is performed on application arrivals<br />
and departures as the network state only changes on<br />
these occasions. Once the network state changes, whether load<br />
balancing action should be conducted is checked in order to<br />
improve the performance of elastic applications. According to<br />
the application arrival assumption, in this study we perform<br />
unidirectional handover check only from WiMAX to WLAN.<br />
Two types of load balancing actions could be triggered<br />
depending on the system states. One kind of action is vertical<br />
handover which switches elastic applications that have<br />
already been served partly by WiMAX to WLAN. To proceed<br />
handover, both criteria for handover check and handover<br />
candidate selection need to be clearly defined. In fact, lots of<br />
efforts have been devoted to propose efficient algorithms since<br />
they actually determine the handover performance [3][2]. We<br />
conduct handover check based on the expected finish time of<br />
elastic applications. Since existing elastic applications share<br />
the service in PS way, the expected finish time can be linearly<br />
represented by service rate. Then these mechanisms actually<br />
belong to the bandwidth-based handover decision mechanisms.<br />
The service rates in two networks under current condition<br />
are compared and a handover decision is made if the service<br />
rate could be increased after handover. Handover candidate is<br />
selected from all the elastic applications in WiMAX including<br />
the newly arrival. Specifically, we select handover application<br />
based on remaining size. Two mechanisms which select the<br />
file with maximum and minimum remaining size respectively<br />
are tested. Later we refer the two mechanisms as MAX<br />
and MIN for convenience. In real applications the remaining<br />
size is difficult to get and may introduce lots of complexity.<br />
However, since we determine remaining size on flow level, the<br />
complexity introduced can be seen as affordable.<br />
The other kind of action is dispatching which switches the<br />
elastic application on its arrival to WLAN. Because we assume<br />
all traffic initially arrives to WiMAX, dispatching can be seen<br />
as one special case of vertical handover. However, it should<br />
be noted that the actual cost for dispatching is much lighter<br />
compared with vertical handover since the application has not<br />
been served yet.<br />
C. Performance Metrics<br />
For the aforementioned system model, we are only interested<br />
in the performance of elastic applications since the performance<br />
of streaming applications does not change with the<br />
adopted load balancing mechanism. Specifically, we consider<br />
three common performance metrics for elastic applications as<br />
follows.<br />
• Average sojourn time: The sojourn time defines the<br />
time interval of elastic application from its arrival to<br />
its departure. As users are very sensitive to duration of
XU et al.: MULTI-SERVICE LOAD BALANCING IN A HETEROGENEOUS NETWORK WITH VERTICAL HANDOVER 45<br />
elastic application, this metric is highly related to the user<br />
experience of service quality.<br />
• Time-average throughput: The time-average throughput<br />
defines the ratio of total service amount to total service<br />
time. This metric can be seen as an indicator of system<br />
performance in terms of service capability.<br />
• Call-average throughput: The call-average throughput defines<br />
the mean of individual ratios of service size to<br />
its service time. This metric integrates the influence<br />
of service size on users’ expectation of service time.<br />
Therefore, it is believed to be the best metric representing<br />
the users’ quality of experience (QoE) [7].<br />
In addition, another important metric for load balancing<br />
mechanisms is the number of load balancing actions. Although<br />
no specific cost is considered in this paper, it is still beneficial<br />
to compare the number of load balancing actions as normally<br />
costs of these actions do not vary with different applications.<br />
In addition, due to different costs of vertical handover and<br />
dispatch, we record both of them in simulations respectively.<br />
III. THEORETICAL APPROXIMATION<br />
It is usually fairly complicated to exactly analyze system<br />
performance when integrated services are involved [8][9][10].<br />
Therefore, in this section, we provide a simple approximation<br />
for theoretical analysis of our system model. The simplified<br />
analysis results provide a reference for further comparisons<br />
with our simulation results.<br />
First, the two servers are approximated as one server with<br />
capacity C = C wimax + C wlan . With this approximation,<br />
the system becomes the traditional integrated service system<br />
where elastic applications share the server capacity with prioritized<br />
streaming applications. However, even for this simplified<br />
system, the calculation of stationary results is still very difficult<br />
since no closed formulation can be derived [7].<br />
However, approximate results can be calculated with either<br />
of the two quasi-stationary assumptions. One quasi-stationary<br />
condition assumes elastic applications evolve much faster than<br />
streaming applications. This assumption is reasonable since<br />
the streaming applications usually last longer than elastic<br />
applications. However, theoretical derivation based on this<br />
assumption can only be obtained for rather light elastic traffic<br />
since it requires uniform stationarity [8].<br />
In our analysis, we take the other quasi-stationary condition<br />
which assumes streaming applications evolve much faster than<br />
elastic applications. While this assumption is not quite reasonable,<br />
it can provide upper performance bounds for elastic<br />
applications [8]. In this case, the service devoted to elastic<br />
application can be approximated as the total capacity minus<br />
the average aggregated service rates of streaming applications.<br />
Based on the former simplification, we could get approximate<br />
results for elastic applications. Specifically, since elastic<br />
applications share the capacity with processor-sharing policy,<br />
insensitivity of PS policy to service distribution can be applied.<br />
The average throughput can be expressed as C e ∗(1−ρ e ), and<br />
the mean sojourn time can be expressed as<br />
E[T ] =<br />
E[x]<br />
C e ∗ (1 − ρ e )<br />
(1)<br />
Average sojourn time (s)<br />
Fig. 2.<br />
1.4<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
MAX<br />
MIN<br />
Theoretical approximation<br />
0.2 0.4 0.6 0.8 1<br />
ρ e<br />
Average sojourn time with minus exponential distributed file size.<br />
where ρ s and ρ e represent the load of streaming and elastic<br />
applications respectively with respect to the capacity of<br />
WiMAX. In addition, C e = C(1−(1−P b )∗ρ s ), E[x] denotes<br />
the average size of elastic applications and P b is the blocking<br />
probability of streaming applications.<br />
IV. NUMERICAL RESULTS<br />
To evaluate the performance of the two handover mechanisms,<br />
we have conducted extensive simulations with varying<br />
parameters. Each simulation result is obtained based on the<br />
average of 30 runs with different seeds. In each run, we simulate<br />
10 6 files and remove the initial stage of the first 5 × 10 3<br />
files. The results of 30 runs are checked with Skewness and<br />
Kurtosis which ensure these runs follow a normal distribution.<br />
This guarantees the validation of our simulation results.<br />
The values of simulation parameters are listed in Tab. I.<br />
Specifically, we calculate ρ e based on the capacity of WiMAX.<br />
A. Exponential Distribution<br />
First we assume the size of elastic applications follows exponential<br />
distribution, which has been assumed and analyzed<br />
extensively for traffic modeling.<br />
In Fig. 2, the average sojourn time of finished files is shown.<br />
It can be seen that for exponential distributed files, the average<br />
sojourn time with MAX is sightly longer than that with MIN.<br />
TABLE I<br />
SIMULATION PARAMETERS<br />
Parameters Meaning Values<br />
C wimax WiMAX capacity 1000 kbps<br />
C wlan WLAN capacity 600 kbps<br />
B s Bandwidth of streaming applications 50 kbps<br />
ρ s Load of streaming applications 0.3<br />
ρ e Load of elastic applications 0.1-1.1<br />
1/µ s Average length of streaming applications 140 s<br />
1/µ e Average size of elastic applications 64 kbits
46 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Average throughput (kbps)<br />
1200<br />
1000<br />
800<br />
600<br />
400<br />
MAX, call−average throughput<br />
MIN, call−average throughput<br />
MIN, time−average throughput<br />
MAX, time−average throughput<br />
Theoretical approximation<br />
Times<br />
10<br />
5<br />
MAX, dispatch<br />
MAX, handover<br />
MIN, dispatch<br />
MIN, handover<br />
200<br />
0<br />
0.2 0.4 0.6 0.8 1<br />
ρ e<br />
0<br />
15 x 105 ρ e<br />
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1<br />
Fig. 3.<br />
Average throughput with minus exponential distributed file size.<br />
Fig. 4.<br />
size.<br />
Dispatch and handover times with minus exponential distributed file<br />
However, compared with the theoretical approximation, the<br />
performance of MAX and MIN are worse. In particular, when<br />
the traffic load is heavy, the average sojourn time of finished<br />
files with MAX and MIN increases very rapidly.<br />
In Fig. 3, the average throughput performance is shown.<br />
It can be seen that the throughput performance of MIN is<br />
better than MAX as well. The time-average and call-average<br />
throughputs are similar for low-load elastic traffic and the gap<br />
increases as the load of elastic applications becomes heavier.<br />
The main reason for the fast decay of call-average throughput<br />
of heavy-loaded elastic applications is the fast increase of<br />
average sojourn time as shown in Fig. 2. However, the averagetime<br />
throughput is not heavily influenced by sojourn time as<br />
it only depends on the system throughput. In addition, when<br />
the load of elastic applications is low, the simulation results<br />
of both MAX and MIN are fairly worse than the theoretical<br />
approximation. However, with the increase of elastic traffic<br />
load, the performance of MAX and MIN first gets closer to and<br />
then the time-average throughput outperforms the theoretical<br />
approximation. The reason for this performance alternation is<br />
due to the use of the service capacity of WLAN for heavyloaded<br />
elastic applications. When the load of elastic applications<br />
is low, almost all the elastic applications are served in<br />
WiMAX according to the handover criteria. However, with<br />
increasing load of elastic applications, more and more files<br />
are dispatched or handovered to WLAN. Thus all the service<br />
capacities of WiMAX and WLAN are utilized when the load<br />
of elastic applications are heavy enough.<br />
In Fig. 4 the dispatch and handover times are shown to<br />
illustrate the cost of each load balancing mechanism. It can<br />
be seen that MIN introduces much more load balancing actions<br />
(dispatch/handover) than MAX and even two times more when<br />
the load of elastic applications is heavy. Moreover, MIN<br />
adopts much more handovers than MAX while the numbers<br />
of dispatch for two mechanisms are comparable.<br />
In summary, for exponential distributed files, the performance<br />
results of MAX and MIN are similar. However, MIN<br />
introduces much more operation costs with a large number of<br />
Average sojourn time (s)<br />
Fig. 5.<br />
0.4<br />
0.35<br />
0.3<br />
0.25<br />
0.2<br />
0.15<br />
0.1<br />
MAX<br />
MIN<br />
0.05<br />
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1<br />
ρ e<br />
Average sojourn time with Pareto distributed file size.<br />
load balancing actions. Therefore, we prefer MAX to MIN<br />
with consideration of overall performance in this context.<br />
B. Pareto Distribution<br />
For realistic applications, it is believed that the size of<br />
elastic applications follows heavy-tailed distribution. Normally<br />
Pareto distribution is chosen as the representative heavy-tailed<br />
distribution especially for file sizes. Therefore, we also present<br />
results with Pareto distributed file sizes. Specifically, we take<br />
the shape parameter as 1.2 which is a typical value for Pareto<br />
distribution.<br />
In Fig. 5, the average sojourn time is shown. Compared<br />
with Fig. 2, there are two major differences. First, the average<br />
sojourn time with Pareto distributed file sizes is shorter than<br />
that with exponential distribution. This phenomenon has been<br />
stated in [11] and the reason is that Pareto distribution has<br />
higher variability than exponential distribution. Second, the<br />
difference between MAX and MIN is more visible with
XU et al.: MULTI-SERVICE LOAD BALANCING IN A HETEROGENEOUS NETWORK WITH VERTICAL HANDOVER 47<br />
Average throughput (kbps)<br />
1200<br />
1000<br />
800<br />
600<br />
400<br />
MAX, call−average throughput<br />
MIN, call−average throughput<br />
MIN, time−average throughput<br />
MAX, time−average throughput<br />
Theoretical approximation<br />
Times<br />
12<br />
10<br />
8<br />
6<br />
4<br />
MAX, dispatch<br />
MAX, handover<br />
MIN, dispatch<br />
MIN, handover<br />
200<br />
2<br />
0<br />
0.2 0.4 0.6 0.8 1<br />
ρ e<br />
0<br />
14 x 105 ρ e<br />
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1<br />
Fig. 6.<br />
Average throughput with Pareto distributed file size.<br />
Fig. 7.<br />
Dispatch and handover times with Pareto distributed file size.<br />
Pareto distributed file sizes. Moreover, for heavy-loaded elastic<br />
applications, MIN provides shorter average sojourn time than<br />
the theoretical approximation. In fact, since MAX and MIN<br />
actually take advantage of the size information of files, there<br />
are chances that the average sojourn time is shorter than the<br />
theoretical approximation based on PS.<br />
In Fig. 6, the throughput results are shown. It can be seen<br />
that the difference of call-average throughputs with MAX and<br />
MIN is much larger compared with Fig. 3. The difference<br />
complies with the visible difference of average sojourn time<br />
in Fig. 5. These results indicate that the performance difference<br />
of MAX and MIN is much larger with Pareto distributed<br />
files. In addition, there is a crossover between call-average<br />
throughput with MIN and the theoretical approximation. This<br />
is consistent with the result in Fig. 5.<br />
In addition, by comparing Fig. 2 with Fig. 5 and Fig. 3<br />
with Fig. 6, we obtain following observations. For the lightloaded<br />
area, the performance results with Pareto or exponential<br />
distributions are quite similar. However, for the heavy-loaded<br />
area, namely when the load of elastic applications is larger<br />
than 0.9, the performance results with Pareto distribution are<br />
much better than those with exponential distribution.<br />
Fig. 7 presents the numbers of load balancing actions.<br />
Compared with Fig. 4, the difference between MAX and<br />
MIN is even larger. Specifically, for MIN, the total number<br />
of load balancing actions is comparable for both Pareto and<br />
exponential distribution while more handovers are adopted<br />
with Pareto distribution. For MAX, much less load balancing<br />
actions are taken with Pareto distribution especially in the<br />
heavy-loaded area.<br />
C. Comparison<br />
In this subsection, we compare our size-based load balancing<br />
mechanisms with an intuitive load balancing method<br />
which admits elastic applications to certain networks without<br />
help of size information. We refer this intuitive method as<br />
INT. Specifically, whenever there is an elastic application<br />
arrival, INT admits the application into WiMAX or WLAN<br />
Average sojourn time (s)<br />
Fig. 8.<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
EXP − MIN<br />
EXP − INT<br />
Pareto − MIN<br />
Pareto − INT<br />
0<br />
0 0.2 0.4 0.6 0.8 1<br />
ρ e<br />
Performance comparison in terms of average sojourn time.<br />
based on the expected average bandwidth received. Suppose<br />
the number of active streaming applications in WiMAX is s,<br />
the numbers of elastic applications in WiMAX and WLAN<br />
are d wimax and d wlan , respectively. The expected bandwidth<br />
b e wimax in WiMAX for the arrival elastic application would<br />
be b e wimax = (C wimax − s ∗ B s )/(d wimax + 1). Similarly,<br />
the expected bandwidth b e wlan in WLAN would be be wlan =<br />
C wlan /(d wlan + 1). Based on the expected bandwidth, INT<br />
makes the routing decision. Specifically, we can express INT<br />
decision as<br />
{ WiMAX, if b<br />
e<br />
Route to<br />
wimax ≥ b e wlan<br />
(2)<br />
WLAN, others.<br />
Accordingly, we simulate INT in the same settings as for<br />
MIN. The comparison results with MIN are shown in Fig.<br />
8 and 9. Generally, it can be seen that for both exponential<br />
and Pareto distributions, MIN outperforms INT in terms<br />
of both average sojourn time and average throughput. This<br />
performance advantage of MIN suggests that with help of<br />
size information and vertical handover, the performance of
48 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Average throughput (kbps)<br />
Fig. 9.<br />
800<br />
600<br />
400<br />
200<br />
EXP − MIN, call−average throughput<br />
0<br />
EXP − MIN, time−average throughput<br />
EXP − INT, call−average throughput<br />
EXP − INT, time−average throughput<br />
Pareto − MIN, call−average throughput<br />
−200<br />
Pareto − MIN, time−average throughput<br />
Pareto − INT, call−average throughput<br />
Pareto − INT, time−average throughput<br />
−400<br />
0 0.2 0.4 0.6 0.8 1<br />
ρ e<br />
Performance comparison in terms of average throughput.<br />
heterogeneous networks could be greatly improved. In particular,<br />
we notice that the performance of INT with exponential<br />
distributed elastic applications becomes much worse when<br />
the traffic load exceeds certain level. However, with Pareto<br />
distributed elastic applications, INT still performs comparable<br />
to MIN. One possible explanation for this difference is that<br />
with higher size variability, Pareto distribution compensates<br />
some performance loss with INT since higher size variability<br />
leads to better performance in WiMAX [11]. For exponential<br />
distribution, the major problem with INT would be that the<br />
performance in WiMAX becomes much worse when the traffic<br />
load exceeds certain level.<br />
V. DISCUSSION<br />
Based on the simulation results and theoretical approximations,<br />
we present the following discussions on the impact<br />
of size of elastic applications and further thinking on load<br />
balancing.<br />
According to the performance advantage of MIN over INT,<br />
we should utilize the size information if the handover cost<br />
is neglectable. This conclusion complies with the advantage<br />
of size-based scheduling over other non-size-based scheduling<br />
policies. By choosing handover candidate based on the<br />
remaining size, we actually conduct some kind of prioritized<br />
scheduling with help of size information.<br />
The simulation results suggest that MIN outperforms MAX<br />
in terms of all the performance metrics for elastic applications.<br />
However, this performance overwhelm comes with cost of<br />
significantly increased number of load balancing actions. Besides<br />
the increased number of handover times, another possible<br />
reason for the better performance of MIN is that MIN lets<br />
large files stay in WiMAX. This moves the system towards<br />
the quasi-stationary assumption that streaming applications<br />
evolves much faster than elastic applications in WiMAX.<br />
However, MAX drives the system towards another quasistationary<br />
assumption that elastic applications evolve much<br />
faster. As stated in [8], the former quasi-stationary assumption<br />
leads to better average performance. However, this statement<br />
applies with condition of uniform stationarity where the load<br />
of elastic applications needs to be fairly low.<br />
Another possible reason for the better performance of MIN<br />
over MAX is the better load balancing granularity. Much more<br />
files need to be distributed to WLAN with MIN to achieve load<br />
balancing since those files are relatively small. Then both the<br />
frequency and size of load balancing actions ensure better load<br />
balancing granularity with MIN. However, as shown in Figs.<br />
4 and 7, this also introduces a large proportion of handover<br />
which is much costly than dispatching.<br />
The dilemma between MAX and MIN inspires us to think<br />
how to keep the advantage of MIN while in the meanwhile<br />
reducing the number of load balancing actions. To this aim,<br />
we point out that two aspects should be taken into consideration.<br />
First, sufficient load balancing granularity needs to be<br />
provided. Second, the characteristics of integrated services in<br />
WiMAX need to be explored and utilized.<br />
It is worth highlighting that in this study we have made<br />
several assumptions in order to investigate the fundamental<br />
effect of different load balancing mechanisms. Specifically, we<br />
assume the capacity of WLAN is fixed and does not depend<br />
on the number of users in the network. While the assumptions<br />
may not always hold, we believe that the trends discovered in<br />
this study will remain under released assumptions. Moreover,<br />
the results in the paper could also be helpful for designing<br />
load balancing mechanisms for other overlay heterogeneous<br />
networks.<br />
VI. CONCLUSION<br />
In this paper we study load balancing for multi-service<br />
in an overlay heterogeneous network. Based on the analysis<br />
of simulation results of two load balancing mechanisms, we<br />
draw the conclusion that both load balancing granularity and<br />
integration of elastic and streaming applications in WiMAX<br />
affect the whole system performance. This knowledge could<br />
provide guidance for further developing better load balancing<br />
mechanisms.<br />
REFERENCES<br />
[1] N. Nasser, A. Hasswa, and H. Hassanein, “Handoffs in Fourth Generation<br />
Heterogeneous Networks,” IEEE Commun. Mag., vol. 44, pp.<br />
96–103, 2006.<br />
[2] M. Kassar, B. Kervellaa, and G. Pujolle, “An Overview of Vertical<br />
Handover Decision Strategies in Heterogeneous Wireless Networks,”<br />
Computer Communications, vol. 31, pp. 2607–2620, 2008.<br />
[3] X. Yan, Y. A. Şekercioğlu, and S. Narayanan, “A Survey of Vertical<br />
Handover Decision Algorithms in Fourth Generation Heterogeneous<br />
Wireless Networks,” Computer Networks, vol. 54, no. 11, pp. 1848–<br />
1863, Aug. 2010.<br />
[4] L. Xiaoshan, V. Li, and Z. Ping, “Joint Radio Resource Management<br />
through Vertical Handoffs in 4G Networks,” in Proc. of Global Telecommunications<br />
Conference, 2006. GLOBECOM, 2006, pp. 1–5.<br />
[5] A.-E. M. Tahaa, H. S. Hassaneina, and H. T. Mouftah, “Vertical Handoffs<br />
as a Radio Resource Management Tool,” Computer Communications,<br />
vol. 31, pp. 950–961, 2008.<br />
[6] W. Song and W. Zhuang, “Multi-Service Load Sharing for Resource<br />
Management in the Cellular/WLAN Integrated Network,” IEEE Trans.<br />
Wireless Commun., vol. 8, pp. 725–735, 2009.<br />
[7] R. Litjens, H. van den Berg, and R. J. Boucherie, “Throughputs in<br />
Processor Sharing Models for Integrated Stream and Elastic Traffic,”<br />
Performance Evaluation, vol. 65, no. 2, pp. 152–180, Feb. 2008.<br />
[8] F. Delcoigne, A. Proutièreb, and G. Régnié, “Modeling Integration of<br />
Streaming and Data Traffic,” Performance Evaluation, vol. 55, pp. 185–<br />
209, 2004.
XU et al.: MULTI-SERVICE LOAD BALANCING IN A HETEROGENEOUS NETWORK WITH VERTICAL HANDOVER 49<br />
[9] R. Malhotra and J. L. v. d. Berg, “Flow Level Performance Approximations<br />
for Elastic Traffic Integrated with Prioritized Stream Traffic,”<br />
in Proc. of 12th Int. Telecommun. Network Strategy and Planning<br />
Symposium, NETWORKS 2006, 2006, pp. 1–9.<br />
[10] S. Borst and N. Hegde, “Integration of Streaming and Elastic Traffic in<br />
Wireless Networks,” in Proc. of IEEE Int. Conf. on Computer Commun.<br />
INFOCOM 2007, 2007, pp. 1884–1892.<br />
[11] R. Litjens and R. J. Boucherie, “Elastic Calls in an Integrated Services<br />
Network: the Greater the Call Size Variability the Better the QoS,”<br />
Performance Evaluation, vol. 52, pp. 193–220, 2003.<br />
He is recipient of a fellowship from the European Research Consortium for<br />
Informatics and Mathematics (ERCIM). He was Co-Chair of IEEE Globecom<br />
2005 General Conference Symposium, TPC Co-Chair of 67th IEEE Vehicular<br />
Technology Conference (VTC) 2008, and General/TPC Co-Chair of International<br />
Symposium on Wireless Communication Systems (ISWCS) 2007-2010.<br />
He is author of the book Stochastic Network Calculus published by Springer<br />
in 2008. His research interests include network measurement and the provision<br />
and analysis of quality of service guarantees in communication networks. In<br />
the area of network calculus, his focus has been on developing fundamental<br />
models and investigating their basic properties for stochastic network calculus<br />
(snetcal), and recently also on applying snetcal to performance analysis of<br />
wireless networks.<br />
Jie Xu received B.E. degree in Electronic Information Engineering from<br />
Beijing University of Aeronautics and Astronautics in 2003, his Ph.D.<br />
in Communication and Information Systems from Graduate University of<br />
Chinese Academy of Sciences, Beijing, China in 2010. During his Ph.D.<br />
study, he had involved in several research projects which covered a wide<br />
range of multimedia communications. He is currently postdoctoral fellow in<br />
Center of Quantifiable Quality of Service at Norwegian University of Science<br />
and Technology (NTNU), Trondheim, Norway, conducting research on the<br />
next generation wireless networks and on multimedia communications. He<br />
is also leading the development of a serious multiplayer game which is to<br />
be used for both scientific research and for university recruitment. He is the<br />
author or co-author of more than 10 research publications; he is a member of<br />
IEEE.<br />
Andrew Perkis is a full professor of Digital Image Processing at the Department<br />
of Electronics and Telecommunications at the Norwegian university<br />
of Science and Technology (NTNU) in Trondheim, Norway. He received<br />
his Siv.Ing and Dr.Techn. degrees in 1985 and 1994, respectively. He is<br />
member of the management team of the National Centre of Excellence -<br />
Q2S - Quantifiable Quality of Service in Communication Systems, where he<br />
is responsible for ”Networked Media Handling”. He is Vice Chair of COST<br />
action IC1003 QUALINET European Network on Quality of Experience in<br />
Multimedia Systems and Services (End date: November 2014). Currently he<br />
is focusing on Multimedia Signal Processing, specifically within methods and<br />
functionality of content representation, quality assessment and its use within<br />
the media value chain in a variety of applications. He is a senior member of<br />
the IEEE. He has more than 150 publications at international conferences and<br />
workshops and more than 50 contributions to International standards bodies.<br />
Dr. Yuming Jiang is presently a Professor in the Department of Telematics<br />
and the Centre for Quantifiable Quality of Service in Communication Systems<br />
(Q2S), at Norwegian University of Science and Technology (NTNU), Norway.<br />
He received his B.S. degree in electronic engineering from Peking University<br />
in 1988, M.E. degree in computer science and engineering from Beijing<br />
Institute of Technology in 1991, and Ph.D. degree in electrical and computer<br />
engineering (ECE) from the National University of Singapore (NUS) in 2001.<br />
From 1996 to 1997, he worked with Motorola. He was a Research Engineer<br />
at the ECE Department, NUS, from 1999 to 2001. From 2001 to 2003, he<br />
was with the Institute for Infocomm Research (I2R), Singapore as a Member<br />
of Technical Staff / Research Scientist. He joined NTNU in 2004. He visited<br />
Northwestern University, USA from 2009 to 2010.<br />
Elissar Khloussy received the B.S degree and M.S degree in computer<br />
engineering from the Conservatoire Nationale Des Arts et Métiers, Paris, in<br />
2001 and 2003 respectively. She is now a PhD candidate at the Telematics<br />
department, Norwegian Univeristy of Science and Technology, Norway. Her<br />
current research interests include the interworking of cellular networks and<br />
wireless local area networks, and cognitive networks.
50 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Resources Management and Services<br />
Personalization in Future Internet Applications<br />
Paweł Świątek, Piotr Rygielski and Adam Grzech<br />
Abstract—The aim of this paper is to introduce a problem of<br />
e-health services quality management. The process of delivering<br />
e-health services to users consists of two major tasks: service<br />
personalization and resources allocation. In this paper we introduce<br />
a use-cases of e-health system and distinguish services that<br />
can be offered. In order to satisfy user requirements one has<br />
to manage resources properly especially when communication<br />
conditions change i.e. in an ambulance. An exemplary solution<br />
has been presented and conclusions for further work have been<br />
formulated.<br />
Patient<br />
data<br />
aquisition<br />
Supervising<br />
physician<br />
PDA<br />
consultation<br />
access<br />
to data<br />
data<br />
transmision<br />
consultation<br />
consultation<br />
Digital<br />
Health Library<br />
access<br />
to data<br />
Ambulance<br />
Consulting<br />
physician<br />
I. INTRODUCTION<br />
ADVANCES in information and communication technologies<br />
allow health service providers to offer e-services in<br />
virtually every aspect of health care. The collection of e-health<br />
services facilitated by current and future ICT architectures<br />
ranges from remote health monitoring, through computeraided<br />
diagnosis, video-consultation, health education, remote<br />
surgery and many others [1].<br />
Each of possible e-health services is a composition of<br />
atomic services provided by either ICT infrastructure or<br />
medical personnel. In some scenarios medical personnel can<br />
be both service provider and service consumer. Delivery of<br />
complex services as a composition of atomic services is the<br />
key feature of service oriented architecture (SOA) paradigm<br />
[2]. Application of SOA approach in e-health services delivery<br />
allows to personalize and flexibly adjust services to individual<br />
needs of service consumers.<br />
According to the SOA paradigm each e-health service is<br />
composed of a set of atomic services providing required<br />
functionalities, which are delivered in certain predefined order.<br />
Each atomic service, which delivers certain functionality may<br />
be provided in different versions, which vary in non-functional<br />
characteristics such as: response time, security, availability, etc<br />
[2], [3]. Composition of complex services from different versions<br />
of atomic services allows to guarantee required quality<br />
of e-health services, which is very often critical in medical<br />
applications.<br />
In this paper we present a general framework for QoSaware<br />
SOA-based composition of e-health complex services.<br />
The presented approach is explained with use of an illustrative<br />
example — remote monitoring of patients health parameters.<br />
The remote monitoring example includes: definition of the remote<br />
monitoring problem as a business process, identification<br />
of complex services playing major roles in the monitoring<br />
Paweł Świątek, Piotr Rygielski and Adam Grzech are with Institute of Computer<br />
Science, Wroclaw University of Technology, Wybrzeze Wyspianskiego<br />
27, 50-370 Wroclaw, Poland<br />
Fig. 1.<br />
Remote monitoring of patients health - an overview.<br />
process, identification of the set of atomic services necessary<br />
to deliver required functionalities and composition of an<br />
exemplary complex service.<br />
II. REMOTE MONITORING OF PATIENT’S HEALTH<br />
As an example of e-health business process consider the<br />
process of remote monitoring of patient’s health (see Fig. 1).<br />
In this scenario it is assumed, that a patient is equipped with<br />
a mobile communication device (e.g. smartphone or PDA),<br />
which collects monitored data from sensors placed on the<br />
patient’s body. Collected data is preprocessed on the patient’s<br />
mobile device and sent for further processing and storage to<br />
Digital Health Library. Further processing of collected data<br />
may involve among others: modelling and identification of<br />
physiological processes, updating of medical knowledge base,<br />
expansion of medical case study database and decision making<br />
for computer-assisted diagnosis. In the last case decision<br />
making algorithm may recognize any abnormal situations<br />
in patient’s health and notify medical personnel about such<br />
events.<br />
Being alarmed, a physician should have an access to<br />
patients’ health records, their current health state, medical<br />
knowledge bases and additional services such as consultation<br />
with patient or other physician. After gathering necessary<br />
information the physician decides whether detected abnormal<br />
situation poses a threat to patient’s health and takes relevant<br />
action, e.g. sends an ambulance to the patient. In the lifethreatening<br />
situation the physician monitors patient’s status<br />
and instructs ambulance crew on how to proceed with the<br />
patient.<br />
The whole process of remote monitoring of patients’ health<br />
can be divided into three phases presented on Fig. 2: basic<br />
monitoring, supervised monitoring and monitoring in the ambulance.<br />
Each of presented monitoring phases is in fact a separate<br />
business process composed of one or more complex services.
ŚWIATEK ˛ et al.: RESOURCES MANAGEMENT AND SERVICES PERSONALIZATION IN FUTURE INTERNET APPLICATIONS 51<br />
Fig. 2.<br />
U1:<br />
Basic<br />
monitoring<br />
NO<br />
YES<br />
Alarm<br />
U2:<br />
Supervised<br />
monitoring<br />
NO<br />
YES<br />
Threat<br />
U3:<br />
Monitoring in<br />
ambulance<br />
The process of remote monitoring of patients health.<br />
a) U2: Supervised monitoring<br />
b)<br />
U2.1:<br />
Detailed<br />
monitoring<br />
U2.3:<br />
Access to<br />
monitored<br />
data<br />
U2.4:<br />
Access to<br />
patients<br />
record<br />
U2.2:<br />
Physician<br />
notification<br />
U2.5:<br />
Access to<br />
Digital<br />
Library<br />
U3.6:<br />
Monitoring from<br />
ambulance<br />
U3.2:<br />
Detailed<br />
monitoring<br />
U3: Monitoring in ambulance<br />
U3.7:<br />
Access to<br />
monitored<br />
data<br />
U3.1:<br />
Ambulance<br />
dispatch<br />
U3.3:<br />
Access to<br />
monitored data<br />
U3.5:<br />
Admittance to<br />
ambulance<br />
U3.8:<br />
Access to<br />
patients<br />
record<br />
U3.4:<br />
Access to<br />
patients record<br />
U2.6:<br />
Teleconsultation<br />
U2.7:<br />
Videoconsultation<br />
U3.9:<br />
Teleconsultation<br />
U3.10:<br />
Videoconsultation<br />
B. Monitoring in ambulance<br />
The process of monitoring in an ambulance consists of two<br />
stages. The first starts when the supervising physician finds<br />
the patient’s health to be at risk and sends an ambulance to<br />
take the patient to the hospital (U3.1). From now on the crew<br />
of the ambulance has on-line access to monitored parameters<br />
of the patient’s health (U3.3) as well as his/her health record<br />
(U3.4).<br />
After being taken by the ambulance (U3.5) patients monitoring<br />
mode changes again. All required patient’s health parameters<br />
are recorded now. Moreover, audio-video monitoring<br />
of the patient in the ambulance is performed as well (U3.6).<br />
Besides the previously accessible services (U3.7 and U3.8)<br />
the ambulance crew can communicate with the supervising<br />
physician through tele- and video-consultation services (U3.9<br />
and U3.10).<br />
Fig. 3. The process of supervised monitoring (a) and the process of<br />
monitoring in an ambulance (b).<br />
Basic monitoring covers routine day to day monitoring of<br />
patient’s vital parameters (e.g. ECG for post-cardiac patients).<br />
Supervised monitoring refers to a case when abnormal<br />
behaviour of monitored parameters is detected. In such a<br />
case an alarm is activated and other parameters (e.g.: body<br />
temperature, pulse and GPS coordinates) are monitored by<br />
the physician providing him/her with information and services<br />
necessary to make decisions concerning further actions.<br />
Monitoring in ambulance applies to life-threatening situations<br />
in which an ambulance was dispatched to collect (possibly<br />
unconscious) patient. In this case both the ambulance crew and<br />
the physician should be provided with necessary information<br />
and communication services [1].<br />
Processes of supervised monitoring and monitoring in ambulance<br />
can be decomposed into separate complex services.<br />
Such an exemplary decomposition of considered processes is<br />
presented on Fig. 3 a) and b) respectively.<br />
A. Supervised monitoring<br />
The process of supervised monitoring starts at the moment<br />
of generation of an alarm concerning abnormal behaviour of<br />
patient’s health parameters. There are two immediate implications<br />
of such an alarm. The first is the change in patient’s<br />
monitoring mode - more parameters are being recorded, possibly<br />
with denser time resolution. The second consequence<br />
is the notification of the patient’s physician about possible<br />
health risks. From this moment the physician has access to<br />
personalised services allowing him/her to gather information<br />
about current condition of the patient and make decision about<br />
further actions. These services may include (see Fig. 3 a)): online<br />
access to monitored health parameters (U2.3), access to<br />
the patient’s health record (U2.4), access to Digital Health<br />
Library containing knowledge about similar cases (U2.5),<br />
and tele- and video-consultation with patient and/or medical<br />
experts (U2.6 and U2.7).<br />
III. REMOTE MONITORING AS BUSINESS PROCESS<br />
Business process is a series of interrelated activities or tasks<br />
that solve a particular problem or lead to achievement of specific<br />
goal. In the SOA paradigm each of activities constituting<br />
in business process is represented as a complex service which<br />
delivers certain predefined functionality (see Fig. 4). Complex<br />
services, in turn, are composed of atomic services, which<br />
provide basic indivisible functionalities. The functionality of a<br />
complex service is an aggregation of functionalities of atomic<br />
services [4]. Similarly, the goal of a business process (its<br />
functionality) is an aggregation of functionalities of performed<br />
complex services.<br />
The difference between business process and complex service<br />
lies in that the former is defined and composed by service<br />
consumer, while the latter is delivered by service provider<br />
as a whole. Service consumer may influence the choice of<br />
particular complex services by specification of Service Level<br />
Agreement (SLA) containing functional and non-functional<br />
requirements. Service provider, on the other hand, composes<br />
complex services from available atomic services basing on<br />
requirements stated in the SLA [5].<br />
In special cases the whole business process may be specified<br />
by single complex service. Such situations may occur for<br />
example when available processes are managed by single<br />
entity basing on certain regulations (e.g. medical processes in<br />
health care). In general, however, this approach is inefficient<br />
and inelastic, since it does not allow consumers to modify their<br />
requirements.<br />
Fig. 4.<br />
Consumers<br />
layer<br />
Providers<br />
layer<br />
Infrastracture<br />
layer<br />
Business<br />
process<br />
Complex<br />
services<br />
Atomic<br />
services<br />
Transport<br />
services<br />
Composition of business processes.
52 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
TABLE I<br />
THROUGHPUT REQUIREMENTS FOR DIFFERENT SIGNALS AND QUALITY<br />
LEVELS CONSIDERED IN THE EXAMPLE.<br />
Required<br />
throughput<br />
[kbps]<br />
Signal Video Voice ECG EMG Heart<br />
sound<br />
High<br />
quality<br />
Low<br />
quality<br />
5000 256 24 600 120 5<br />
640 25 12 100 24 2<br />
IV. SERVICES PERSONALIZATION AND RESOURCES<br />
MANAGEMENT<br />
Heart<br />
rate<br />
Depending on individual needs of service consumers and<br />
capabilities of execution environment each of available complex<br />
services can be delivered in many versions which differ in<br />
non-functional characteristics [6], [7]. As an example consider<br />
monitoring from ambulance service (U3.6) which allows to<br />
transmit monitored signals, voice and video from ambulance to<br />
supervising physician. Depending on individual case different<br />
health parameters are recorded and transmitted through the<br />
network (e.g.: ECG and pulse for cardiac patients or glucose<br />
for diabetics). Additionally, voice and video transmission may<br />
be required by physician. Moreover, individual requirements<br />
and amount of available system resources may influence the<br />
number and the quality of transmitted signals [8], [3].<br />
Assume, that there are six signals possible to be measured<br />
and transmitted from ambulance to physician: video, voice,<br />
ECG, EMG, heart sound (HS) and heart rate (HR). Each signal<br />
can be measured and transmitted in two modes - high and<br />
low quality. High and low quality of health parameters can be<br />
reflected by higher and lower sampling rates. A transmission<br />
of high and low quality signals has different requirements for<br />
communication resources (see table I).<br />
In general, preferences concerning the quality of requested<br />
complex service can be defined by penalty matrix P l = [p ij ],<br />
where each element p ij (i = 1, ..., I; j = 1, ..., J) represents a<br />
penalty for not delivering j-th signal in higher than i-th quality<br />
levels. The exemplary matrix for six signals considered above<br />
and three quality levels is defined by:<br />
⎡<br />
P l = ⎢<br />
⎣<br />
p HI<br />
V ID<br />
p LO<br />
V ID<br />
p NO<br />
V ID<br />
pHI V OI<br />
pLO V OI<br />
pNO V OI<br />
p HI<br />
ECG<br />
p LO<br />
ECG<br />
p NO<br />
ECG<br />
pHI EMG<br />
pLO EMG<br />
pNO EMG<br />
pHI HS<br />
pLO HS<br />
pNO HS<br />
p HI<br />
HR<br />
p LO<br />
HR<br />
p NO<br />
HR<br />
⎤<br />
⎥<br />
⎦ , (1)<br />
where for example p 21 = p LO<br />
V ID is the penalty for not<br />
delivering video signal in high quality and p 11 = p NO<br />
EMG is<br />
the penalty for not delivering EMG signal at all. If penalty<br />
p ij = 0 then i-th quality level of j-th signal is not required.<br />
Denote by D l a binary matrix of quality level delivery,<br />
where each element d ij (i = 1, ..., I; j = 1, ..., J) is defined<br />
as follows:<br />
{ 1 i-th quality level for j-th signal is delivered<br />
d ij =<br />
0 i-th quality level for j-th signal is not delivered .<br />
(2)<br />
Exemplary matrix D l for a complex service in which video<br />
and voice signals are delivered at low quality, ECG signal<br />
is delivered at high quality, and remaining signals are not<br />
transmitted at all is presented below:<br />
⎡<br />
D l = ⎣ 0 0 1 0 0 0<br />
⎤<br />
1 1 0 0 0 0⎦ . (3)<br />
0 0 0 1 1 1<br />
Given the penalty matrix P l and quality delivery matrix<br />
D l for certain complex service request req l it is possible<br />
to calculate overall penalty p l for not satisfying consumers<br />
preferences as follows:<br />
p l =<br />
J∑<br />
p lj · d T lj, (4)<br />
j=1<br />
where p lj and d lj are j-th columns of matrices P l and D l<br />
respectively.<br />
Let R = [r ij ] (i = 1, ..., I; j = 1, ..., J) denote the matrix<br />
of resources consumption, where each element r ij represents<br />
the amount of resources required to deliver j-th signal at i-<br />
th quality level. In the example considered above resources<br />
requirements are stated in terms of required throughput (see<br />
table I). Therefore, the exemplary matrix R is defined as<br />
follows:<br />
⎡<br />
⎤<br />
5000 256 24 600 120 5<br />
R = ⎣ 640 25 12 100 24 2⎦ . (5)<br />
0 0 0 0 0 0<br />
Note, that elements of the last row of exemplary matrix R are<br />
equal to zero since the lowest quality level represents situation<br />
in which signals are not delivered at all.<br />
The amount of resources r l necessary to deliver complex<br />
service at quality level represented by certain matrix D l can<br />
be calculated as follows:<br />
r l =<br />
J∑<br />
r j · d T lj, (6)<br />
j=1<br />
where r j and d lj are j-th columns of matrices R and D l<br />
respectively.<br />
For the model presented above a number of resource<br />
management tasks can be formulated. The goal of each task<br />
may be different. For example, one may want to minimize<br />
the average or maximal penalty caused by violation of<br />
consumers individual preferences or to maximize consumers<br />
satisfaction for each incoming complex service request. The<br />
aforementioned tasks can be formulated as follows.<br />
Task 1: Average penalty minimization<br />
Given: set L(t) of service requests currently being served,<br />
capacity C of the system and matrices of penalties P l and<br />
resources consumption R.<br />
Find: set of quality delivery matrices {D ∗ l<br />
: l ∈ L(t)}<br />
such that average penalty caused by violation of consumers<br />
individual preferences is minimized:<br />
{D ∗ l : l ∈ L(t)} = arg min<br />
∑<br />
{D l :l∈L(t)}<br />
l∈L(t) j=1<br />
J∑<br />
p lj · d T lj (7)
ŚWIATEK ˛ et al.: RESOURCES MANAGEMENT AND SERVICES PERSONALIZATION IN FUTURE INTERNET APPLICATIONS 53<br />
with respect to system capacity constraints:<br />
∑<br />
l∈L(t) j=1<br />
J∑<br />
r j · d T lj ≤ C. (8)<br />
Task 2: Maximal penalty minimization<br />
This task is similar to Task 1 and can be derived by<br />
substitution of objective function in (7) by following formula:<br />
{D ∗ l : l ∈ L(t)} = arg min max<br />
{D l :l∈L(t)} l∈L(t)<br />
J∑<br />
p lj · d T lj (9)<br />
j=1<br />
Task 3: Maximization of service consumer satisfaction<br />
Given: the amount of resources C(t) currently available in the<br />
system, matrix of preferences P l for incoming service request<br />
req l and resources consumption matrix R.<br />
Find: composition of complex service defined by quality<br />
delivery matrix D ∗ l<br />
such that penalty caused by violation of<br />
consumers preferences is minimized:<br />
D ∗ l = arg min<br />
D l<br />
J∑<br />
p lj · d T lj (10)<br />
j=1<br />
with respect to system capacity constraints:<br />
J∑<br />
r j · d T lj ≤ C(t). (11)<br />
j=1<br />
The goal of Task 3 is to compose complex services according<br />
do consumers preferences. It can be performed each time<br />
when new complex service request arrives to the system. It<br />
can also be used for request admission control and Service<br />
Level Agreement renegotiation. Tasks 1 and 2, on the other<br />
hand, are performed in order to optimize utilization of systems<br />
resources and to guarantee average or minimal service<br />
consumer satisfaction.<br />
A. Numerical example<br />
Consider an example in which two monitoring service<br />
requests req 1 and req 2 arrive to the system. Requests req 1<br />
and req 2 are characterized by following preferences matrices:<br />
⎡<br />
P 1 = ⎣ × 0 × 0 0 ×<br />
⎤<br />
× 100 × ∞ 200 × ⎦<br />
× ∞ × ∞ ∞ ×<br />
⎡<br />
P 2 = ⎣ 0 × × 0 0 ×<br />
⎤ (12)<br />
1000 × × 100 10 × ⎦<br />
∞ × × ∞ 200 ×<br />
which mean that request req 1 requires voice and heart sound<br />
signal to be delivered at least at low quality and EMG signal<br />
to be delivered at high quality. Similarly request req 2 requires<br />
video and EMG signal at least at low quality and additional<br />
heart sound signal would improve consumers satisfaction.<br />
Assume, that systems capacity is equal to C = 2510kbps<br />
and that capacity requirements of available signals are given<br />
by matrix R defined in (5). As a result of minimizing average<br />
penalty for not satisfying consumers preferences (Task 1<br />
defined by (7) and (8)) following quality delivery matrices<br />
are calculated:<br />
⎡<br />
× 1 × 1 1<br />
⎤<br />
×<br />
D 1 = ⎣× 0 × 0 0 × ⎦<br />
× 0 × 0 0 ×<br />
⎡<br />
⎤, (13)<br />
0 × × 1 1 ×<br />
D 2 = ⎣1 × × 0 0 × ⎦<br />
0 × × 0 0 ×<br />
resulting in overall penalty p = 1000 for not delivering high<br />
quality video signal for second service request req 2 . Delivery<br />
of high quality video signal in this example is impossible<br />
because HQ video capacity requirements are higher than<br />
overall capacity C of the system. Services composition and<br />
resources allocation represented by matrices D 1 and D 2 (13)<br />
are illustrated on Fig. 5 (state before moment t 1 ).<br />
Assume, that at certain moment t 1 a third service request<br />
req 3 characterized by matrix P 3 arrives to the system. In<br />
order minimize average penalty for not satisfying consumers<br />
preferences systems resources have to be reallocated. Final service<br />
composition and resources allocation depends on penalty<br />
matrices P 1 , P 2 and P 3 .<br />
In order to show the difference between final resources<br />
allocation assume two alternate matrices P 3 :<br />
⎡<br />
P a 3 = ⎣ 0 0 × 0 0 × ⎤<br />
1000 250 × ∞ ∞ × ⎦<br />
∞ ∞ × ∞ ∞ ×<br />
⎡<br />
P b 3 = ⎣ 0 0 × 0 0 × ⎤, (14)<br />
1000 150 × ∞ ∞ × ⎦<br />
∞ ∞ × ∞ ∞ ×<br />
which differ in the penalty for not delivering voice signal in<br />
the high quality. Two different solutions of are represented by<br />
quality level matrices D a 1, D a 2, D a 3 and D b 1, D b 2, D b 3:<br />
⎡<br />
× 0 × 1 0<br />
⎤<br />
×<br />
D a 1 = ⎣× 1 × 0 1 × ⎦<br />
× 0 × 0 0 ×<br />
⎡<br />
0 × × 0 0<br />
⎤<br />
×<br />
D a 2 = ⎣1 × × 1 1 × ⎦, (15)<br />
0 × × 0 0 ×<br />
⎡<br />
0 1 × 1 1<br />
⎤<br />
×<br />
D a 3 = ⎣1 0 × 0 0 × ⎦<br />
0 0 × 0 0 ×
54 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
Final services composition and resources allocation<br />
<br />
represented<br />
by the above matrices are illustrated on figures Fig. 5a)<br />
<br />
<br />
<br />
<br />
depending on the penalty for<br />
not delivering HQ voice for<br />
<br />
request req<br />
<br />
3 , HQ voice in req 3 is traded for HQ heart sound<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
V. CONCLUSIONS<br />
In this paper a general problem of e-health services management<br />
was introduced. It consists of two major tasks: service<br />
<br />
personalization and resources allocation. Service personalization<br />
allows to flexibly adjust delivered services based on<br />
individual needs of service consumers. Resources management<br />
allows to reserve and allocate resources necessary to deliver<br />
requested services and satisfy consumers preferences. In the<br />
presented approach both tasks of personalization and resource<br />
allocation are solved simultaneously as single optimization<br />
problem in which certain parameters concern the personalization<br />
task (penalty matrices), while other parameters regard<br />
allocation tasks (resources consumption matrix). Unfortunately<br />
formulated tasks are in general NP-hard, therefore heuristic<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
<br />
Fig. 5. Services composition and resources reallocation for two versions<br />
of third request preferences differing in the value of penalty p 322 for not<br />
delivering HQ voice for request req 3 : (a) p 322 = 250, (b) p 322 = 150.<br />
and<br />
⎡<br />
× 0 × 1 1<br />
⎤<br />
×<br />
D b 1 = ⎣× 1 × 0 0 × ⎦<br />
× 0 × 0 0 ×<br />
<br />
<br />
<br />
<br />
⎡<br />
0 × × 0 1<br />
⎤<br />
×<br />
D b 2 = ⎣1 × × 1 0 × ⎦. (16)<br />
0 × × 0 0 ×<br />
⎡ <br />
⎤<br />
0 0 × 1 1 ×<br />
<br />
D b 3 = ⎣1 1 × 0 0 × ⎦<br />
0 0 × 0 0 ×<br />
and Fig. 5b) respectively. Note that in the presented example,<br />
in req 1 and req 2 . The threshold value of this penalty is<br />
p 322 = p 125 + p 225 = 210.<br />
<br />
algorithms should be applied to effectively control processes<br />
of service composition, personalization and resources management.<br />
<br />
<br />
<br />
ACKNOWLEDGEMENTS<br />
This work was partially supported by the European Union<br />
from the European Regional Development Fund within the<br />
Innovative Economy Operational Programme project number<br />
POIG.01.01.02-00-045/09-00 “Future Internet Engineering”.<br />
<br />
REFERENCES<br />
[1] D. Niyato, E. Hossain, and J. Diamond, “IEEE 802.16/WiMax-Based<br />
Broadband Wireless Access and its Application for Telemedicine/e-health<br />
Services,” IEEE Trans. Wireless Commun., vol. 14, no. 1, pp. 72–83, 2007.<br />
[2] P. Rygielski and P. Świątek, SOA Infrastructure Tools: Concepts and<br />
Methods. Springer-Verlang, Berlin, 2010, ch. QoS-aware Complex<br />
Service Composition in SOA-based Systems.<br />
[3] K. Shahadatand, F. L. Kin, and E. G. Manning, “The Utility Model For<br />
Adaptive Multimedia Systems,” in Proc. of Int. Conf. on Multimedia<br />
Modeling, 1997, pp. 111–126.<br />
[4] A. Grzech and P. Świątek, “Modeling and Optimization of Complex<br />
Services in Service-Based Systems,” Cybernetics and Systems, vol. 40,<br />
pp. 706–723, 2009.<br />
[5] A. Grzech, P. R. P., and P. Świątek, “QoS-Aware Infrastructure Resources<br />
Allocation in Systems Based on Service-Oriented Architecture Paradigm,”<br />
in HET-NETs, 2010, pp. 35–48.<br />
[6] M. Alrifai and T. Risse, “Combining Global Optimization with Local<br />
Selection for Efficient QoS-Aware Service Composition,” in Proc. of 18th<br />
Int. Conf. on World Wide Web WWW’09. ACM, 2009, pp. 881–890.<br />
[7] Y. Tao, Z. Yue, and K.-J. Lin, “Efficient Algorithms for Web Services<br />
Selection with End-to-End QoS Sonstraints,” ACM Trans. Web, vol. 1,<br />
no. 1, 2007.<br />
[8] A. Grzech and P. Świątek, “Parallel Processing of Connection Streams in<br />
Nodes of Packet-Switched Computer Communication Systems,” Cybernetics<br />
and Systems, vol. 39, no. 2, pp. 155–170, 2008.<br />
Paweł Światek ˛ received his MSc and PhD degrees in computer science from<br />
Wrocław University of Technology, Poland, in 2005 and 2009, respectively.<br />
From 2009 he is with Institute of Computer Science, Wrocław University of<br />
Technology, where from 2010 he works as an assistant professor. His main<br />
scientific interests are focused on services optimization and personalization,<br />
optimization of service-based systems, resources allocation, QoS delivery in<br />
heterogeneous networks and mobility management in wireless networks.<br />
Piotr Rygielski is a Ph.D. student at the Wrocław University of Technology<br />
(WUT), Poland. He received his MSc degree in computer science<br />
from WUT<br />
<br />
in 2009. Since 2009 he works as a young researcher in two<br />
<br />
Innovative Economy European Union projects: "Future Internet Engineering"<br />
<br />
and "New information technologies for electronic economy and information<br />
<br />
society based on service-oriented architecture". His main research interests<br />
include resource allocation issues in distributed computer systems, serviceoriented<br />
architecture, resources virtualization, Quality of Service provisioning<br />
in computer networks, discrete optimization and combinatorial algorithms.<br />
Adam Grzech, Ph.D., D.Sc . He is a professor in the Institute of Computer<br />
Science, Department of Computer Science and Management, Wrocław University<br />
of <br />
Technology (WUT). He obtained M.Sc. degree from Department<br />
of Electronics, WUT in 1977, Ph.D. degree from Institute of Technical<br />
Cybernetics, WUT in 1979, D.Sc. degree form Department of Electronics,<br />
Wrocław University of Technology in 1989 and professor title in 2003.<br />
His research interests include design and analysis of computer systems<br />
and networks, requirement analysis, modeling and identification of computer<br />
networks, design and application of local and wide area computer networks,<br />
flow control and congestion avoidance in computer networks, migration<br />
and integration of heterogeneous computer systems and networks, services<br />
integration in networks, intelligent networks, quality of service âĂŞ aware<br />
networks, SOA-paradigm based systems, engineering of the future internet,<br />
security of computer systems and networks, agent-based systems and its<br />
application in optimization and control and intelligent information systems.<br />
He is an author and co-author of nearly 300 research papers published in<br />
books, journals and conference proceedings, supervisor of 8 completed Ph.D.<br />
thesis and supervisor of more than 200 completed M.Sc. thesis in computer<br />
communication engineering.
ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong> 55<br />
Compact node-link formulations for the optimal<br />
single path MPLS Fast Reroute layout<br />
Cezary Żukowski, Artur Tomaszewski, Michał Pióro, David Hock, Matthias Hartmann and Michael Menth<br />
Abstract—This paper discusses compact node-link formulations<br />
for MPLS fast reroute optimal single path layout. We<br />
propose mathematical formulations for MPLS fast reroute local<br />
protection mechanisms. In fact, we compare one-to-one (also<br />
called detour) local protection and many-to-one (also called<br />
facility backup) local protection mechanisms with respect to minimized<br />
maximum link utilization. The optimal results provided by<br />
the node-links are compared with the suboptimal results provided<br />
by algorithms based on non-compact linear programming (path<br />
generation) approach and IP-based approach.<br />
I. INTRODUCTION<br />
MULTIPROTOCOL Label Switching (MPLS) technology<br />
enables configuration of end-to-end virtual connections<br />
in communication networks, especially in networks without<br />
connection-oriented capabilities. Labeled packets can be sent<br />
over the connections and forwarded according to the labels<br />
over so-called LSPs (Label Switched Paths).<br />
MPLS is able to detect network failures (link failures)<br />
locally and thus a failure-detecting router can quickly switch<br />
all packets from failing primary LSP path to a backup LSP<br />
path just after a failure is detected. This is so-called fast reroute<br />
(FRR) capability and the failure-detecting router is the socalled<br />
point of local repair (PLR).<br />
The way the backup LSPs are rerouted (from the PLR) depends<br />
on the FRR mechanism. Two mechanisms are possible:<br />
one-to-one backup (OOB) [1] and many-to-one backup (MOB)<br />
[2]. Many-to-one backup is also called facility backup as in<br />
[3]. In OOB and MOB backup LSP paths are rerouted over<br />
the next hop router (NHR) and terminated in NHR if and only<br />
if the failing link is the last one on the failing primary LSP<br />
path.<br />
For instance, in Figure 1 the primary path originates in<br />
router A, it goes through routers A, B, C, D, and terminates<br />
in router D. Link A-B fails, router A is the PLR, router B is<br />
the NHR, and router C is the next next hop router (NNHR).<br />
In the MOB a backup path is rerouted from the PLR router<br />
to the NNHR. On the other hand, in the OOB a backup path<br />
is rerouted from PLR to router D.<br />
When the OOB is used, backup LSP paths originate in the<br />
PLR and terminate in the destination node of the corresponding<br />
primary LSP path.<br />
When the MOB is used, backup LSP paths originate in<br />
the PLR and terminate in the NNHR. In MOB, all primary<br />
paths that go exactly through the same PLR, NHR, NNHR<br />
are rerouted from the PLR to the NNHR on a single LSP<br />
backup path.<br />
Individual demands can be sent (between any pair of network<br />
nodes) on single primary and single backup LSP paths<br />
(single path layout) or split on multiple primary and multiple<br />
backup LSP paths (multipath layout). The single or multipath<br />
layout we select, impacts on the minimized maximum link<br />
utilization value and network configuration complexity as<br />
explained in our previous paper [1].<br />
In this paper, we focus on compact node-link (NL) formulations<br />
for the single LSP paths layout as they provide the<br />
optimal solutions for this layout, they can be easily implemented<br />
(e.g. with CPLEX package), and they haven’t been<br />
presented before. We show and describe the NL formulation<br />
for the one-to-one backup as well as the NL formulation for<br />
many-to-one backup.<br />
We use applicable size networks to show the efficiency of<br />
OOB and MOB. We provide example results related to running<br />
times, the number of used continuous and binary variables to<br />
show the performance of OOB and MOB formulations according<br />
to the network sizes. We provide results for minimized<br />
maximum link utilization values and networks configuration<br />
complexities to compare OOB and MOB solutions qualities.<br />
Additionally, we use the same networks to generate suboptimal<br />
solutions for the single path layout. To do this, noncompact<br />
linear programming (LP) based approach and IPbased<br />
approach were applied. Detailed explanations of these<br />
methods and related work can be found in [1], [4], [5], [6],<br />
[7] and [8].<br />
Then, we discuss the gap between optimal and suboptimal<br />
solutions. Exactly, we compare the minimized maximum link<br />
Cezary Żukowski,Artur Tomaszewski and MichałPióro are with Institute<br />
of Telecommunications, Warsaw University of Technology, 00-665 Warsaw,<br />
Poland, Emails: czukowsk@mion.elka.pw.edu.pl, artur@tele.pw.edu.pl,<br />
mpp@tele.pw.edu.pl<br />
MichałPióro is also with Department of Electrical and Information Technology,<br />
Lund University, 221-00 Lund, Sweden, Email: mpp@eit.lth.se<br />
David Hock and Matthias Hartmann are with Institute of Computer<br />
Science, University of Würzburg, D-97074 Würzburg, Germany, Email:<br />
hock@informatik.uni-wuerzburg.de, hartmann@informatik.uni-wuerzburg.de<br />
Michael Menth is with Dept. of Computer Science, University of Tübingen,<br />
D-72076 Tübingen, Germany, Email: menth@informatik.uni-tuebingen.de<br />
Fig. 1.<br />
Explanation of one-to-one backup and many-to-one backup
56 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
utilization values to show optimal and suboptimal solutions<br />
qualities. Finally, we present conclusions summarizing the<br />
results.<br />
II. NODE-LINK FORMULATIONS<br />
The section discusses the common part of OOB and MOB<br />
formulations.<br />
A. Used symbols<br />
An MPLS/IP network is modeled as a graph G = (V,E)<br />
comprising a set V of nodes and a set E of directed edges (E ∈<br />
V 2 \{(v,v) : v ∈ V }). The nodes correspond to the MPLS/IP<br />
routers and edges correspond to IP links. Symbols a(e) and<br />
b(e) denote the source and the destination node of link e ∈ E.<br />
Sets δ − (v) and δ + (v) denote the incoming and the outgoing<br />
edges for node v ∈ V . A constant C e is the capacity of link<br />
e ∈ E.<br />
Set D denotes the set of demands. Symbols o(d) and t(d)<br />
denote source and destination node of demand d. A constant<br />
h d is the rate of demand d ∈ D.<br />
Set S denotes the set of failure states. In the paper only the<br />
failures of single links are considered, thus S ≡ E.<br />
In both OOB and MOB formulations, the paths of the LSP<br />
connections (in normal state and in each failure state) are<br />
modeled as unitary non-bifurcated flows between appropriate<br />
pairs of nodes.<br />
Variables x ed indicate whether link e belongs to the primary<br />
path for demand d.<br />
The Z denotes the objective function value. It minimizes<br />
maximum link utilization.<br />
B. Feasible and infeasible solutions provided by node-links<br />
Notice that there is no capacity constraints in the formulations<br />
for OOB and MOB. Instead, we use relative link<br />
utilization by addition of the constant (1/C e ) to the NL<br />
formulations.<br />
In the case when we get an optimal solution with Z > 1 it<br />
means the solution is infeasible in terms of required capacity.<br />
On the other hand, when we get some optimal solution with<br />
Z ≤ 1 it means the solution is feasible in terms of required<br />
capacity.<br />
III. NODE-LINK FORMULATION FOR ONE-TO-ONE BACKUP<br />
In this section, a compact NL formulation for one-to-one<br />
backup is described. For each s ∈ S and each d ∈ D, both<br />
the primary and its backup path that is used in the network<br />
state corresponding to the failure of link s can be viewed as<br />
consisting of two subpaths – the path between the source node<br />
of d and the PLR (which is the originating node a(s) of link s),<br />
and the path between the PLR and t(d) – the destination node<br />
of d. Since, according to the FRR mechanism, the primary<br />
and the backup paths share the subpaths that is upstream from<br />
the PLR and differ in the subpaths that are downstream from<br />
the PLR, the backup path can be described in terms of the<br />
primary path and those two downstream subpaths between the<br />
PLR and the destination node of the demand.<br />
The formulation is as follows:<br />
min Z (1a)<br />
Z ≥ ∑ (h d /C e )x ed ,∀e ∈ E (1b)<br />
d∈D<br />
Z ≥ ∑ (h d /C e )(x ed − y des + z des ),s ≠ e,∀e ∈ E,∀s ∈ E<br />
d∈D<br />
(1c)<br />
∑<br />
e∈δ + (v)<br />
x ed −<br />
∑<br />
e∈δ − (v)<br />
and for each d ∈ D, s ∈ E:<br />
y des ≤ x ds ,∀e ∈ E<br />
∑<br />
e∈δ + (v)<br />
y des −<br />
∑<br />
e∈δ − (v)<br />
y des ≤ x de ,∀e ∈ E<br />
z des ≤ x ds ,∀e ∈ E<br />
∑<br />
e∈δ + (v)<br />
z dss = 0<br />
z des −<br />
∑<br />
e∈δ − (v)<br />
⎧<br />
⎨<br />
x ed =<br />
⎩<br />
1, v = o(d)<br />
−1, v = t(d),<br />
0, otherwise<br />
⎧<br />
⎨ x ds , v = a(s)<br />
y des = −x<br />
⎩ ds , v = t(d)<br />
0, otherwise<br />
⎧<br />
⎨ x ds , v = a(s)<br />
z des = −x<br />
⎩ ds , v = t(d)<br />
0, otherwise<br />
z des = 0,∀e ∈ δ + (b(s)),b(s) ≠ t(d)<br />
∀d ∈ D<br />
(1d)<br />
(2a)<br />
(2b)<br />
(2c)<br />
(2d)<br />
(2e)<br />
(2f)<br />
(2g)<br />
We cannot write conservation constraints for variables y de<br />
with flow equal to 1 (the right-hand side of (2b)) because if<br />
x ds is equal to 0 then from (2c) all y de are equal to 0 and we<br />
would arrive at a contradiction. It only makes sense to look<br />
for the backup path described with y ds (and also with variables<br />
z ds described in the sequel) if the primary path fails in state<br />
s. That we know from x ds : if x ds is equal to 1 the primary<br />
path uses link s and thus fails in state s. Thus, putting x ds in<br />
conservation constraints (2b) allows us to look for the backup<br />
path and write the constraints somewhat ‘conditionally’ upon<br />
the failure of the primary path.<br />
The subpath of the backup path that is downstream from<br />
the PLR can be modeled as a unitary flow between the PLR<br />
and the destination node of the demand that must not use<br />
neither the failing link nor the terminating node of that link.<br />
For each e ∈ E, let z des be a binary variable that equals 1 if,<br />
and only if, link s belongs to the primary path of demand<br />
d, and link e belongs to the segment of the backup path that<br />
this downstream from the PLR. This is satisfied by constraints<br />
(2d), (2e), (2f) and (2g).<br />
Similarly to conditions (2a)-(2c) that describe subpath y,<br />
conditions (2d)-(2e) say that we must find subpath z between<br />
the PLR and the destination node; the form and the role of<br />
(2e) is the same as that of (2b). And the conditions for both<br />
types of subpaths paths – y must be embedded into the primary<br />
path and z must detour the failure – are given by (2c), (2f) and<br />
(2g). Constraints (2f) and (2g) say that we cannot use either<br />
the failed link or the links that originate at the terminating<br />
node of the failed link. Although we therefore cannot transit<br />
the terminating node of the failed link, this does not mean
ŻUKOWASKI et al.: COMPACT NODE-LINK FORMULATIONS FOR THE OPTIMAL SINGLE PATH MPLS FAST REROUTE LAYOUT 57<br />
that we cannot enter such a node. Thus, in particular, the last<br />
hop of the primary path is also protected since the destination<br />
node of the demand need not be used as a transit node for the<br />
backup path.<br />
It should be noted that although only link failures are<br />
assumed explicitly in the formulation, the determined backup<br />
paths provide protection of the primary paths against both link<br />
and node failures. But only in terms of flow routing; network<br />
capacity is not sufficient and the flows in fact are not protected.<br />
The meaning of this can be explained with the following<br />
example. Consider the following links, all with capacity equal<br />
to 1: k and l between nodes A and B; m and n between nodes<br />
B and C; and o between nodes A and C. Consider two primary<br />
paths k-m and l-n between nodes A and C, both with flows<br />
equal to 1. Assume that each of those primary paths has path<br />
o as its backup path. Then, theoretically each primary path<br />
is protected with its backup path against the failure of node<br />
B. But there is not enough capacity on link o to reroute both<br />
primary paths at the same time, so in fact the flows are not<br />
protected. Still, if either link k or l fails the affected path can<br />
be rerouted.<br />
IV. NODE-LINK FORMULATION FOR MANY-TO-ONE<br />
BACKUP<br />
In this section, a compact NL formulation for many-to-one<br />
backup is described. The MOB is a restricted version of OOB.<br />
The restrictions hold for all d ∈ D and consist in:<br />
• backup paths terminate in NNHRs (except the case the<br />
NNHR is the terminating node of demand – then terminate<br />
in NHRs),<br />
• backup paths originating in node a(s) and terminating in<br />
node b(q) are rerouted the same path, on the single path<br />
going from node a(s) to node b(q).<br />
For each s ∈ S the formulation can be shown according to<br />
its two cases. The case when b(s) ≠ t(d) and the case when<br />
b(s) = t(d). In the first case, backup LSP paths terminate in<br />
NNHRs and in the second in NHRs. The common part of the<br />
cases is formulated below:<br />
min Z (3a)<br />
⎧<br />
⎨ 1, v = o(d)<br />
∑ x ed − ∑ x ed = −1, v = t(d), ∀d ∈ D<br />
⎩<br />
e∈δ + (v) e∈δ − (v) 0, otherwise<br />
(3b)<br />
Z ≥ ∑ (h d /C e )x ed ,∀e ∈ E (3c)<br />
d∈D<br />
Z ≥ ∑ (h d /C e )x ed + ∑ ∑ (h d /C e )z desq ,∀e,s ∈ E<br />
d∈D<br />
d∈D q∈δ + (b(s))<br />
(3d)<br />
In the case when b(s) ≠ t(d), the constraints for each s ∈ E,<br />
d ∈ D are formulated as follows:<br />
z sdq ≤ x sd ,∀q ∈ δ + (b(s))<br />
z sdq ≤ x qd ,∀q ∈ δ + (b(s))<br />
z sdq ≥ x sd + x qd − 1,∀q ∈ δ + (b(s))<br />
z sdq ≥ 0,∀q ∈ δ + (b(s))<br />
(4a)<br />
(4b)<br />
(4c)<br />
(4d)<br />
f sq ≥ z sdq ,∀q ∈ δ + (b(s))<br />
⎧<br />
⎨<br />
f esq − f esq =<br />
⎩<br />
∑<br />
e∈δ + (v)<br />
∑<br />
e∈δ + (v)<br />
z desq −<br />
∑<br />
e∈δ − (v)<br />
∑<br />
e∈δ − (v)<br />
⎧<br />
⎨<br />
z desq =<br />
⎩<br />
z desq ≤ f esq ,∀e ∈ E,∀q ∈ δ + (b(s))<br />
z dess = 0,∀e ∈ E<br />
f sq ,v = a(s)<br />
− f sq ,v = b(q),<br />
0,otherwise<br />
z dsq ,v = a(s)<br />
−z dsq ,v = b(q),<br />
0,otherwise<br />
(4e)<br />
∀q ∈ δ + (b(s))<br />
z desq = 0,∀e ∈ δ + (b(s)),∀e ∈ δ − (b(s)),∀q ∈ δ + (b(s))<br />
(4f)<br />
∀q ∈ δ + (b(s))<br />
(4g)<br />
(4h)<br />
(4i)<br />
(4j)<br />
The constraints (3b)-(3d) correspond to constraints (1b)-<br />
(1d). The constraint (3c) concerns links load in nominal state<br />
(without failures) and constraint (3d) concerns link load in<br />
each failure state s, where s ∈ S.<br />
Constraints (4a)-(4e) model in fact logical ‘and’ operator.<br />
For each demand d ∈ D, operator takes as an input x sd and x qd<br />
variables and sets f sq variable as a result. If a primary path<br />
goes through link s and link q (x sd = 1 and x qd = 1), it means<br />
that in state s the f sq ( f sq = 1) backup path is selected for<br />
rerouting, for demand d. All primary paths that goes through<br />
link s and q in state s are rerouted on the same route f sq .<br />
The constraint (4f) forms a route f sq . The constraint is<br />
formulated similarly as (2b) and (2e). The route f sq originates<br />
in a(s) (PLR) and terminates in b(q) (NNHR).<br />
Backup paths, used in a failure state s ∈ E by demand d ∈<br />
D, are formed by z desq variables. In a feasible solution all<br />
variables z desq = 1 indicate edges that belong to the backup<br />
path used in state s by demand d. The constraint (4h) ensures<br />
that all primary paths that go through link s and q use f sq path<br />
for rerouting in the state s.<br />
The constraints (4i) and (4j) block links that should not<br />
be used in the state s. The constraints correspond to the<br />
constraints (2f) and (2g) in MOB formulation.<br />
In the case when b(s) = t(d) (s = q), the constraints for<br />
each s ∈ E, d ∈ D are formulated as follows:<br />
z sds ≥ x sd<br />
f ss ≥ z sds<br />
∑<br />
e∈δ + (v)<br />
∑<br />
e∈δ + (v)<br />
f ess −<br />
z dess −<br />
∑<br />
e∈δ − (v)<br />
∑<br />
e∈δ − (v)<br />
z dess ≤ f ess ,∀e ∈ E<br />
z dsss = 0<br />
⎧<br />
⎨ f ss , v = a(s)<br />
f ess = − f ss , v = b(s),<br />
⎩<br />
0, otherwise<br />
⎧<br />
⎨ z dss , v = a(s)<br />
z dess = −z<br />
⎩ dss , v = b(s),<br />
0, otherwise<br />
z desq = 0,q ≠ s,∀e ∈ E,∀q ∈ δ + (b(s))<br />
(5a)<br />
(5b)<br />
(5c)<br />
(5d)<br />
(5e)<br />
(5f)<br />
(5g)<br />
The ‘and’ operator is reduced to constraints (5a) and (5b)<br />
– when a primary path goes through the s link, then f ss path<br />
is selected for rerouting, for this path.
58 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
The constraints (5c) and (5d) simplify (4f) and (4g) constraints<br />
– there is no need to iterate over every q ∈ δ + (b(s)).<br />
Similarly (5e), (5f) and (5g) correspond to (4h), (4i) and<br />
(4j) respectively.<br />
V. SIMULATIONS<br />
In this section, experiments performed in the paper are<br />
explained in details. Network instances and settings used in<br />
the tests are described.<br />
The tests are performed on standard PC computer (2.8 GHz<br />
Intel, 2.5 GB RAM). CPLEX solver (version 12) is used<br />
with non-default mixed integer programming (MIP) parameters.<br />
MIP probing level and cuts generation level are set to<br />
aggressive and MIP emphasis is set to force optimality over<br />
feasibility.<br />
The network instances used in the tests are subnetworks<br />
of networks presented in [1]. They are randomly chosen to<br />
keep their size applicable for the tests. For example, network<br />
instance CO9 is a subnetwork of network cost-239-100 and is<br />
almost as large as cost-239-100.<br />
Each test for the optimal NL runs no more than 24 hours.<br />
Simulations are stopped after that time. There is no time limit<br />
for path generation as it runs no more than a few seconds for<br />
networks in Table I.<br />
Each subnetwork instance contains a full matrix of demands<br />
– for each pair of nodes (a,b) two demands exist: a directed<br />
demand from node a to b and a directed demand from node b<br />
to a. The number of demands is equal to |V | · (|V | − 1) for<br />
each network instance.<br />
VI. NODE-LINK FORMULATIONS COMPLEXITY<br />
In this section, NL formulations complexity for practical<br />
NL instances is described.<br />
The sizes of the applicable networks confirm that their MIP<br />
representations are hard to solve. Even with aggressive MIP<br />
settings for CPLEX, network instances CO9 and AT9 could<br />
not be solved in 24 hours time limit, as shown for MOB.<br />
Though presented NL formulations are compact in the<br />
number of variables, they prove to be unsolvable in reasonable<br />
time and computer memory. For example, the number of<br />
binary variables in OOB CO9 is equal to about 42000. In<br />
MOB CO9 the number of binary variables is equal to about<br />
13000 and number of continuous variables is equal to about<br />
50000. It shows a large size of tested MIP problems.<br />
Additionally, we solve NL formulations with cost239-100 –<br />
the network instance with the smallest number of nodes from<br />
[1]. The test is unsuccessful as it leads to a CPLEX “out of<br />
memory” error.<br />
VII. NUMERICAL RESULTS<br />
In this section, we present and discuss numerical results<br />
obtained in the tests.<br />
A. Optimal OOB and MOB link utilization<br />
In Table I we get the same value of minimized maximum<br />
link utilization for OOB and for MOB, for all tested networks.<br />
B. Optimal and suboptimal link utilization<br />
The suboptimal results for single path layout can be computed<br />
with a linear programming approach (path generation<br />
approach) and an IP-based approach. Though path generation<br />
has to solve several MIP problems it still provides solutions<br />
for large network instances in reasonable time as shown in<br />
[1]. The multipath suboptimal solutions provided by the path<br />
generation determine the lower bounds for the optimal single<br />
path layout solutions. On the other hand the results provided<br />
by the path generation for single path layout determine the<br />
upper bounds for the optimal single path layout solutions.<br />
In Table II we compare optimal and suboptimal solutions.<br />
The results show that a heuristic path generation provides<br />
significantly different values from the optimal one. The values<br />
of multipath solutions are better 32.05% (EX8) and 27.89%<br />
(EB8) than the optimal solution. For AT8 and CO9 networks,<br />
the suboptimal single path solutions are significantly worse<br />
than the optimal solutions.<br />
Due to the relatively small size of the network instances<br />
the maximum path lengths of the primary paths if an IP-based<br />
layout is used are rather small (in most cases not more than<br />
3 hops per path). For these short path lengths, the destination<br />
basically always equals either the NHR or NNHR. Therefore,<br />
all OOB and MOB layouts are supposed to be identical if IPbased<br />
layouts are used. It is thus an expected behavior that the<br />
IP-based values for OOB and MOB are equal for all network<br />
instances.<br />
C. OOB and MOB network configuration effort<br />
The configuration effort of OOB and MOB is related to the<br />
length of the LSP paths. In Table III the maximum and average<br />
configuration effort is shown for each network. Configuration<br />
effort is calculated for each node in the network as a sum of<br />
incoming and outgoing LSP paths. If LSP path goes through<br />
the node it is treated as incoming and outgoing at the same<br />
time. Maximum configuration relates to the node with the<br />
maximum sum of incoming and outgoing LSP paths.<br />
In Table III we observe that network configuration effort for<br />
backup paths is several times greater than for primary paths.<br />
We observe that there is no significant difference between<br />
average primary paths configuration effort of OOB and MOB.<br />
On the other hand, for EX8 and CO8 networks, there appear<br />
significant differences in OOB and MOB configuration effort<br />
for backup paths. The situation for EX8 can be described by<br />
fact that for OOB longer backup paths are used in the optimal<br />
solution. And similarly, for CO8 shorter paths are used.<br />
VIII. SUMMARY AND CONCLUSIONS<br />
The paper presents compact node-link formulations for<br />
MPLS Fast Reroute optimal single path layout. We test the<br />
formulations on network instances with practical sizes.<br />
We provide optimal solutions for the single path layout for<br />
two distinct MPLS local protection mechanisms: one-to-one<br />
backup and many-to-one backup. We obtain the same value of<br />
the minimized maximum link utilization for one-to-one backup<br />
and many-to-one backup, for all tested networks. This seems<br />
to be an interesting fact, taking into account that many-to-one
ŻUKOWASKI et al.: COMPACT NODE-LINK FORMULATIONS FOR THE OPTIMAL SINGLE PATH MPLS FAST REROUTE LAYOUT 59<br />
TABLE I<br />
MINIMIZED MAXIMUM LINK UTILIZATION<br />
Network Optimal NLs Path generation approach IP-based approach<br />
ID Name | V | | E | | D | MOB OOB OOB (multipath) OOB (single path) OOB MOB<br />
CO8 cost239-100_8 8 32 56 90.90% 90.90% 90.44% 102.27% 110.7% 110.7%<br />
CO9 cost239-100_9 9 38 72 - 70.16% 69.38% 89.37% 87.6% 87.6%<br />
GE8 geant_8 8 22 56 62.79% 62.79% 62.78% 71.68% 71.69% 71.69%<br />
GE9 geant_9 9 26 72 58.60% 58.60% 58.60% 71.27% 66.08% 66.08%<br />
EX8 exodus_8 8 36 56 65.15% 65.15% 33.10% 66.62% 74.84% 74.84%<br />
EB8 ebone_8 8 34 56 63.94% 63.94% 36.05% 65.85% 68.03% 68.03%<br />
AT8 atnt_8 8 34 56 52.40% 52.40% 44.43% 73.36% 91.01% 91.01%<br />
AT9 atnt_9 9 38 72 - 59.47% 59.47% 81.23% 81.74% 81.74%<br />
TABLE II<br />
GAPS BETWEEN OPTIMAL AND SUBOPTIMAL SOLUTIONS<br />
Path generation approach IP-based approach<br />
ID OOB (multipath) OOB (single path) OOB MOB<br />
CO8 0.46% 11.37% 19.8% 19.8%<br />
CO9 0.78% 19.21% 17.44% -<br />
GE8 0.01% 8.89% 8.9% 8.9%<br />
GE9 0.0% 12.67% 7.48% 7.48%<br />
EX8 32.05% 1.47% 9.69% 9.69%<br />
EB8 27.89% 1.91% 4.09% 4.09%<br />
AT8 7.97% 20.96% 38.61% 38.61%<br />
AT9 0.0% 21.76% 22.27% -<br />
TABLE III<br />
NETWORK CONFIGURATION EFFORT<br />
Primary paths Backup paths All paths<br />
avg. max. avg. max. avg. max.<br />
CO8 OOB 26.5 44 56.75 102 83.25 142<br />
CO8 MOB 26.75 38 65.25 94 92 132<br />
-0.25 6 -8.5 8 -8.75 10<br />
GE8 OOB 30 48 87 152 117 198<br />
GE8 MOB 30 44 83.5 166 113.5 208<br />
0 4 3.5 -14 3.5 -10<br />
GE9 OOB 34.67 70 106.22 212 140.89 282<br />
GE9 MOB 35.56 60 106.22 198 141.78 250<br />
-0.89 10 0 14 -0.89 32<br />
EX8 OOB 23.75 36 75.25 94 99 130<br />
EX8 MOB 22 30 53.5 72 75.5 102<br />
1.75 6 21.75 22 23.5 28<br />
EB8 OOB 25.75 32 58.5 86 84.25 118<br />
EB8 MOB 22.75 38 55.75 78 78.5 116<br />
3 -6 2.75 8 5.75 2<br />
AT8 OOB 29.75 48 75 150 104.75 194<br />
AT8 MOB 32 60 73.25 110 105.25 170<br />
-2.25 -12 1.75 40 -0.5 24<br />
ACKNOWLEDGMENT<br />
This work was supported by the Euro-FGI FP6 NoE as<br />
well as the Euro-NF FP7 NoE. The Polish authors were<br />
funded by Polish Ministry of Science and Higher Education<br />
under research grant N517 397334. The German authors<br />
were funded by Deutsche Forschungsgemeinschaft under grant<br />
TR257/23-2.<br />
REFERENCES<br />
[1] M. Pióro, A. Tomaszewski, C. Żukowski, D. Hock, M. Hartmann, and<br />
M. Menth, “Optimized IP-Based vs. Explicit Paths for One-to-One<br />
Backup in MPLS Fast Reroute,” in 14 th International Telecommunications<br />
Network Strategy and Planning Symposium, Warsaw, Poland, Sep. 2010.<br />
[2] R. Martin, M. Menth, and K. Canbolat, “Capacity Requirements for the<br />
Facility Backup Option in MPLS Fast Reroute,” in IEEE Workshop on<br />
High Performance Switching and Routing (HPSR), Poznan, Poland, Jun.<br />
2006, pp. 329 – 338.<br />
[3] A. Raj and O. C. Ibe, “A survey of ip and multiprotocol label switching<br />
fast reroute schemes,” Comput. Netw., vol. 51, pp. 1882–1907, June 2007.<br />
[4] D. Hock, M. Hartmann, M. Menth, and C. Schwartz, “Optimizing<br />
Unique Shortest Paths for Resilient Routing and Fast Reroute in IP-Based<br />
Networks,” Osaka, Japan, Apr. 2010.<br />
[5] R. Martin, M. Menth, and K. Canbolat, “Capacity Requirements for the<br />
One-to-One Backup Option in MPLS Fast Reroute,” in IEEE International<br />
Conference on Broadband Communication, Networks, and Systems<br />
(BROADNETS), San Jose, CA, USA, Oct. 2006.<br />
[6] S. Orlowski and M. Pióro, “On the complexity of column generation in<br />
survivable network design with path-based survivability mechanisms,” in<br />
International Network Optimization Conference (INOC), 2009.<br />
[7] M. Pióro, Á. Szentesi, J. Harmatos, A. Jüttner, P. Gajowniczek, and<br />
S. Kozdrowski, “On Open Shortest Path First Related Network Optimisation<br />
Problems,” Performance Evaluation, vol. 48, pp. 201 – 223, 2002.<br />
[8] M. Pióro and D. Medhi, Routing, Flow, and Capacity Design in Communication<br />
and Computer Networks. Morgan Kaufman, 2004.<br />
backup is a restricted version of one-to-one backup. It would<br />
be interesting if we could extend this observation for larger<br />
network instances.<br />
We compare optimal solutions for the single path layout<br />
with suboptimal solutions provided by algorithms based on<br />
path generation approach and IP-based approach. The values<br />
of the optimal solution are usually significantly better than<br />
suboptimal solutions.<br />
Though we are able to compute optimal solutions for a set of<br />
practical network instances, for larger networks more efficient<br />
methods have to be found.<br />
Cezary Żukowski is a master student in the Institute of Telecommunications<br />
at the Warsaw University of Technology, Poland. He received the B.S. degree<br />
in telecommunications in 2009. His studies concentrate on survivable networks<br />
design and routing problems.<br />
Artur Tomaszewski is an assistant professor at the Faculty of Electronics and<br />
Information Technologies. He received the MSc and the PhD degrees from<br />
the Warsaw University of Technology in 1990 and 1993, respectively, both in<br />
telecommunications. With his research interests focused on the architecture<br />
of telecommunications networks, network management and control systems<br />
and software, network planning methodologies and network design and<br />
optimization methods, he is an author or co-author of almost a hundred of<br />
journal and conference papers.
60 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Michał Pióro is a full professor and the head of the Computer Networks<br />
and Switching Division in the Institute of Telecommunications at the Warsaw<br />
University of Technology, Poland, and a full professor at Lund University,<br />
Sweden. He received the Ph.D. degree in telecommunications in 1979 and<br />
the D.Sc. degree in 1990, both from the Warsaw University of Technology.<br />
In 2002 he received a Polish State Professorship. His research interests<br />
concentrate on modeling, design and performance evaluation of telecommunication<br />
networks. He is an author of four books and around 150 technical<br />
papers presented in the telecommunication and OR journals and conference<br />
proceedings. He has lead many national and international research projects<br />
for telecom industry and EC in the field.<br />
David Hock studied computer science and mathematics at the University of<br />
Würzburg/Germany and at the BTH in Karlskrona/Sweden. He received his<br />
diploma degree in computer science in spring 2009. Since then he has been<br />
pursuing his PhD as a research assistant at the Chair of Distributed Systems at<br />
the Institute of Computer Science in Würzburg. His current research focuses<br />
on resilient IP networks and Fast Reroute.<br />
Matthias Hartmann studied computer science and mathematics at the<br />
University of Würzburg/Germany, the University of Texas at Austin/USA, and<br />
at the Simula Research Laboratory/Oslo, Norway. He received his diploma<br />
degree in computer science in 2007. Currently, he is a researcher at the<br />
Institute of Computer Science in Würzburg and pursuing his PhD. His<br />
current research focuses on IP Fast Reroute and Future Internet Routing in<br />
combination with performance evaluation and resilience analysis.<br />
Michael Menth is a full professor in Computer Science and the head of the<br />
Communication Networks chair in the Faculty of Science at the University of<br />
Tübingen/Germany. He received a Diploma and PhD degree in 1998 and 2004<br />
from the University of Würzburg/Germany. Prior he was studying computer<br />
science at the University of Texas at Austin and worked at the University of<br />
Ulm/Germany. His special interests are performance analysis and optimization<br />
of communication networks, resource management, resilience issues, and<br />
Future Internet. Prof. Menth holds numerous patents and received various<br />
scientific awards for innovative work.
ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong> 61<br />
Enhancing Data Transmission Reliability with<br />
Multipath Multicast Rate Allocation<br />
Matin Bagherpour, Mehrdad Alipour and Øivind Kure<br />
Abstract—In this paper, a multipath routing scheme is proposed<br />
for data transmission in a packet-switched network to<br />
improve the reliability of data delivery to multicast destinations,<br />
and to reduce network congestion. A multi-objective optimization<br />
model is presented that utilizes FEC (Forward Error Correction)<br />
across multiple multicast trees for transmitting packets toward<br />
the destinations. This model assigns the transmission rates over<br />
multicast trees so that the probability of irrecoverable loss for<br />
each destination and also the link congestion are minimized.<br />
We propose a genetic algorithm based on SPEA (Strength<br />
Pareto Evolutionary Algorithm) in order to approximate Pareto<br />
optimal solutions of this rate allocation problem with a nondominated<br />
solution set. Numerical results show that splitting data<br />
packets between multiple trees increases reliability and decreases<br />
network congestion when compared with the results obtained for<br />
transmitting data packets over a single tree.<br />
Index Terms—Forward Error Correction (FEC), load balancing,<br />
multicast communication, multipath routing, Quality of<br />
Service<br />
I. INTRODUCTION<br />
PATH diversity can be achieved by setting up multiple<br />
parallel paths between source and destination nodes. Multipath<br />
routing not only can reduce congestion in the network,<br />
but also can be considered as a tool for error resilience by<br />
providing higher bandwidth for each session. The idea of using<br />
multiple parallel paths for transmitting data was first proposed<br />
in [1]. In this work, a message is divided into a number of submessages<br />
and the sub-messages are transmitted over disjoint<br />
paths in the network.<br />
A comprehensive review of multipath routing for load balancing<br />
and traffic engineering considering Quality of Service<br />
(QoS) is presented in [2]. The authors introduced a general<br />
multi-objective optimization model to balance traffic load<br />
among multiple trees and optimize QoS measures such as<br />
average delay, and average delay jitter.<br />
Since transport protocols such as TCP favor reliability over<br />
timeliness, they are not appropriate for real time streaming<br />
applications. Therefore, many approaches have been proposed<br />
to deal with these kind of applications. Layered and errorresilient<br />
video coding are two approaches of this kind. Layered<br />
video codec adapts the internet bit rate to the available<br />
bandwidth and tries to deal with time-varying nature of the<br />
internet [3]. In error-resilient codec, the bit stream is modified<br />
M. Bagherpour is with the Norwegian University of Science and Technology,<br />
Trondheim, Norway (corresponding author to provide phone: 47-735-<br />
92775; fax: 47-735-92790; e-mail: matin@q2s.ntnu.no).<br />
M. Alipour is with Department of Industrial Engineering, University of<br />
Science and Culture, Tehran, Iran (e-mail: m.alipour@usc.ac.ir).<br />
Ø. Kure is with the Norwegian University of Science and Technology,<br />
Trondheim, Norway (e-mail: okure@q2s.ntnu.no).<br />
in a way that the decoded video degrades more smoothly<br />
in lossy environments [3]–[5]. It is shown that multipath<br />
transmission, when combined with error control schemes, can<br />
improve the quality of multimedia services in terms of packet<br />
loss and delay. There has recently been an increasing interest<br />
in using multipath routing for failure recovery of real-time<br />
multimedia applications. For example, Multiple Description<br />
Coding (MDC) and Forward Error Correction (FEC) are<br />
combined with multipath routing to improve data transmission<br />
in internet. In MDC approach, a video source is split into<br />
multiple descriptions and each of them is sent over a different<br />
channel. MDC has been studied in detail in [6]–[10]. FEC is<br />
a channel coding technique which increases reliability at the<br />
expense of bandwidth expansion [11]–[14].<br />
There are also some approaches based on multicasting to<br />
stream multimedia sessions over the internet [15]. Multicast<br />
reduces bandwidth consumption by not sending duplicate<br />
packets on the same physical link of the network [16].<br />
In this paper, we utilize path diversity over packet switched<br />
networks in transmitting data packets from a source node to<br />
multiple destination nodes of a multicast group. We integrate<br />
multicast routing and multipath transmission with failure recovery<br />
to improve reliability in data transmission and reduce<br />
congestion in a packet switched network. We suppose that<br />
multicast trees have the ability to send data packets from<br />
the source to destinations at different rates. In this work,<br />
each network link is modeled as a continuous Gilbert-Elliot<br />
channel as in [17]. A Gilbert-Elliot channel can have two<br />
states, namely “good” and “bad” states. During transmission<br />
of a packet, if the channel is in the good state, the packet<br />
will be delivered to the destination successfully; otherwise, it<br />
will be lost. We use FEC scheme to split data packets between<br />
multicast trees, i.e. the data is encoded into a block of N equal<br />
packets so that each destination is able to recover the video<br />
session by receiving at least K packets (K ≤ N); otherwise,<br />
an irrecoverable loss happens. A multi-objective optimization<br />
model is presented for the aforementioned problem which tries<br />
to minimize the probability of irrecoverable loss (reliability<br />
maximization) and network congestion (by minimizing maximum<br />
utilization of network links).<br />
By this model, we try to exploit both load balancing and<br />
failure recovery advantages of path diversity in transmitting<br />
data packets to receivers. We use simulation to estimate<br />
probability distribution of bad burst times of each path in the<br />
network in order to calculate the probability of irrecoverable<br />
loss for each destination. Since the multi-objective model is<br />
highly nonlinear and cannot be solved by common solvers,<br />
a genetic algorithm based on Strength Pareto Evolutionary
62 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Algorithm (SPEA) is presented to solve the multi-objective<br />
model for a sample network topology.<br />
The rest of this paper is organized as follows. Related works<br />
are surveyed in this section and motivation of this work is<br />
discussed. In Section II, the network model is introduced.<br />
Our mathematical programming model of the problem is given<br />
in Section III. In Section IV, the proposed genetic algorithm<br />
based on SPEA is presented. In Section V, numerical results<br />
of implementation of the algorithm for a sample network are<br />
presented and the advantages of using multiple trees over<br />
single tree are illustrated. In Section VI, conclusions and<br />
suggestions for future research are given.<br />
A. Related Works<br />
In previous works, multipath routing is combined with MDC<br />
in order to enhance resilience to loss in video streaming [18],<br />
or to reduce rate distortion of video streams [19]. Multipath<br />
routing of TCP packets is used to control the congestion<br />
in networks with minimum signaling overhead [20]. Path<br />
diversity is also used over IP voice streams to increase speech<br />
quality [21]. In addition, the problem of rate allocation over<br />
multiple paths is presented in [22] and [23]. In [22], the authors<br />
consider a leaky bucket model for network paths and try to<br />
minimize the end-to-end delay. The rate allocation problem<br />
with multiple senders to a single receiver is presented in [23].<br />
The authors suppose that the connection between each pair<br />
of source and receiver is a Gilbert-Elliot channel and propose<br />
an algorithm to solve packet partitioning and rate allocation<br />
problems. A rate allocation algorithm is used to minimize<br />
probability of irrecoverable loss in FEC approach.<br />
FEC is a system of error control for data transmission,<br />
where the sender adds redundant data to the messages to increase<br />
the chance of successful data recovering at the receiver.<br />
An irrecoverable loss occurs if the number of successfully<br />
delivered packets is smaller than the number of initial packets<br />
before adding redundant packets. Since complexity of the<br />
proposed model in [23] depends on the number of packets and<br />
increases exponentially by the number of paths, the authors<br />
proposed a brute-force search algorithm to solve the rate allocation<br />
problem for the special case of two disjoint paths. They<br />
considered disjoint paths to facilitate modeling of irrecoverable<br />
loss, but it should be noted that finding completely disjoint<br />
paths may only be possible in highly connected networks.<br />
The advantages of their approach in reducing probability of<br />
irrecoverable loss are investigated by implementing it for the<br />
actual internet in [24].<br />
A rate allocation problem is also presented in [17] where the<br />
authors take advantage of path diversity to send data packets<br />
from the source to destinations over general packet switched<br />
networks, like internet, in order to minimize the probability of<br />
irrecoverable loss. They assume that the paths are disjoint, and<br />
each path is modeled as a continuous Gilbert-Elliot channel.<br />
The authors calculate probability of irrecoverable loss by<br />
using a continuous approximation for probability distribution<br />
of the time each path spends in the bad state during a block<br />
time. However, in order to simplify calculation of probability<br />
distribution of bad state time, they assume that each path<br />
cannot have more than one bad burst time during a block<br />
time.<br />
A. Link Model<br />
II. NETWORK MODEL<br />
We model the network links with a two-state continuous<br />
time Markov process: Gillbert-Elliot. According to the Gilbert-<br />
Elliot model, a channel spends an exponentially distributed<br />
amount of time with mean 1/µ g in the good state. Then, it<br />
alternates to the bad state and stays in the bad state for another<br />
exponentially distributed amount of time with mean 1/µ b .<br />
Although this model is used for network paths in [17], we<br />
make use of it for each link in the network. This assumption is<br />
justified by the fact that the Gilbert-Elliot model has the ability<br />
to model a single transmission channel whereas network paths<br />
usually consist of several links and cannot be considered as<br />
single transmission channels. On the other hand, independent<br />
Gilbert-Elliot channel model is only applicable for disjoint<br />
paths. In this paper, each path consists of several links that<br />
each is modeled as an independent Gilbert-Elliot channel. It<br />
is also assumed that the good time mean 1/µ g of a link is<br />
much greater than its bad time mean 1/µ b , and the channel<br />
state does not change during transmission of a packet [17]. If<br />
a packet is transmitted during the bad state of a link, it will<br />
be lost before reaching the downlink node; otherwise, it will<br />
be delivered to the downlink node successfully. This model<br />
is widely used for transmission channels for the applications<br />
where delay is not a critical factor [17].<br />
B. Error Correction Model<br />
In this work, FEC is applied across multiple multi-rate<br />
trees to reduce probability of irrecoverable loss occurrence<br />
for each destination. In this scheme, data is encoded into<br />
a block of N equal packets so that each destination can<br />
recover the whole data by receiving at least K packets.<br />
K represents the number of initial packets before adding<br />
redundant data packets. In other words, an irrecoverable loss<br />
occurs for a destination if at least N − K packets are lost out<br />
of N packets that are transmitted toward that destination. By<br />
modeling network links as Gilbert-Elliot channels, we are able<br />
to formulate the irrecoverable loss for non-disjoint multicast<br />
trees for transmitting data packets to multicast destinations.<br />
Mathematical optimization formulation of the aforementioned<br />
problem is presented in section III.<br />
III. MATHEMATICAL FORMULATION<br />
In this model, MT represents the set of multicast trees, M<br />
the set of destinations, and L the set of network links. It is<br />
assumed that each multicast tree has the ability to transmit<br />
packets from the source to all the destinations in M at<br />
different rates. The problem is formulated as a multi-objective<br />
mathematical programming model as follows:<br />
⎛<br />
⎞<br />
|MT |<br />
min P j E = P ∑<br />
⎝ B ij X ij ≥ N−K ⎠∀j = 1, . . . , |M| (1)<br />
i=1
BAGHERPOUR et al.: ENHANCING DATA TRANSMISSION RELIABILITY WITH MULTIPATH MULTICAST RATE ALLOCATION 63<br />
Parameter<br />
TABLE I<br />
NOTATIONS: PARAMETERS AND DECISION VARIABLES<br />
Description<br />
N Total number of packets to be sent to each destination;<br />
K Minimum number of data packets needed in each destination<br />
to recover the whole data (number of information<br />
packets);<br />
T Total block time;<br />
path ij Path associated with the jth destination in the ith multicast<br />
tree;<br />
B ij Random variable representing the portion of time at least<br />
one of the links of path ij spends in the bad state during<br />
the total block;<br />
P j E Probability of irrecoverable loss for the jth destination;<br />
λ uv<br />
ij 1 if the link (u, v) exists in path ij , otherwise 0;<br />
Cuv<br />
max Maximum allowable capacity of the link (u, v) for data<br />
transmission;<br />
K ′ Maximum number of paths to transmit packets for each<br />
destination;.<br />
Decision variables<br />
X ij<br />
|MT |<br />
∑<br />
i=1<br />
Number of packets sent over path ij toward jth destination.<br />
|MT |<br />
∑<br />
i=1<br />
min max<br />
|MT |<br />
∑<br />
i=1<br />
|M|<br />
max<br />
j=1<br />
(u,v)∈L<br />
⌈<br />
Xij<br />
N<br />
⎛<br />
⎜<br />
⎝<br />
|MT<br />
∑<br />
|<br />
i=1<br />
|M|<br />
max<br />
j=1<br />
C max uv<br />
(<br />
Xijλ uv<br />
ij<br />
T<br />
)<br />
⎞<br />
⎟<br />
⎠<br />
(2)<br />
X ij = N ∀j = 1, . . . , |M| (3)<br />
(<br />
Xij λ uv )<br />
ij<br />
≤ Cuv max ∀(u, v) ∈ L (4)<br />
T<br />
⌉<br />
≤ K ′ ∀j = 1, . . . , |M| (5)<br />
0 ≤ X ij ≤ N, X ij Integer<br />
The model parameters and decision variables are defined in<br />
Table I.<br />
In this model, P j E<br />
, the probability of irrecoverable loss for<br />
the j th destination, is calculated by continuous approximation<br />
as in [17] i.e. the number of lost data packets in each path<br />
equals to the portion of time that the path spends in the<br />
bad state multiplied by the number of data packets that are<br />
transmitted over this path. This approximation is rational if<br />
we suppose that the packet inter-arrival time is much shorter<br />
than each typical bad burst of each link in the network,<br />
T<br />
N
64 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Means of these random variables are generated according to<br />
a uniform distribution. Simulation results will be presented in<br />
Section V.<br />
B. Multi-Objective Evolutionary Algorithm Approach<br />
We propose a multi-objective evolutionary algorithm to<br />
solve the traffic splitting problem modeled in Section III. A<br />
survey of the works successfully used multi-objective evolutionary<br />
algorithms for traffic engineering problems is available<br />
in [2].<br />
Genetic algorithm is one of the widely used successful<br />
evolutionary algorithms. In the genetic algorithm, an initial<br />
population of feasible solutions is generated as the starting<br />
point of the search. Individuals are evaluated based on the<br />
value of a fitness function. The fittest individuals are selected<br />
as parents according to a selection method to generate the next<br />
generation by using genetic operators such as crossover and<br />
mutation. This procedure continues repeatedly by replacing<br />
the old generation with the new one and keeping the best<br />
individuals of each generation until a terminating condition is<br />
satisfied.<br />
In this paper, a genetic algorithm based on SPEA [25] is<br />
proposed. In SPEA, in each generation an external set with the<br />
best Pareto solutions are held in addition to the evolutionary<br />
population. In the multi-objective optimization context, a solution<br />
x from the solution space dominates another solution y<br />
from that space if and only if x is as good as y regarding every<br />
objective, and is strictly better than y in at least one objective<br />
function. In the case that neither x nor y dominates the other,<br />
it is said that they are Pareto solutions or non-dominating<br />
solutions in contrast to each other. In SPEA, solutions in<br />
the external set are pruned and updated in each generation<br />
whenever a solution from the population dominates at least<br />
one solution in the external Pareto set.<br />
Our proposed multi-objective genetic algorithm is structured<br />
like SPEA with some modifications in generating the initial<br />
population and fitness assignment strategy. These changes<br />
are necessary because SPEA is constructed to solve nonconstrained<br />
multi-objective problems, whereas our problem is<br />
constrained. Since generating a feasible initial population in<br />
a way that satisfies all the constraints is not trivial, and also<br />
the population will not necessarily remain feasible after using<br />
crossover and mutation operators over the individuals, we use<br />
a penalty based approach to penalize the individuals that are<br />
not feasible based on their degree of infeasibility according<br />
to [26]. In this approach, an adaptive penalty function and a<br />
distance function are used to penalize infeasible individuals to<br />
reduce their chance of being selected as parents. Penalty based<br />
approaches have a good reputation in solving constrained optimization<br />
problems with evolutionary algorithms. This method<br />
is easy to implement and does not need any parameter tuning<br />
when compared with the other available approaches.<br />
The proposed multi-objective genetic algorithm has the<br />
following steps:<br />
Step 1: The initial population P is generated randomly in a<br />
way that each individual is feasible according to the constraints<br />
(3) and (5), but it can be infeasible regarding the capacity<br />
constraints represented by (4). Also, an empty set P ′ of nondominated<br />
solutions is created.<br />
Step 2: The non-dominated individuals of P are copied into<br />
P ′ and the solutions within P ′ which are covered by the other<br />
members of P ′ are removed. The solution x is said to cover<br />
the solution y, if and only if, x dominates y or x and y have<br />
the same fitness considering all the objective functions.<br />
Step 3: If the number of non-dominated solutions exceeds<br />
a given maximum number N ′ , thenP ′ is pruned by using<br />
a clustering method to limit the size of Pareto solutions<br />
to the specified size N ′ . Clustering is used to reduce the<br />
size of non-dominated solution set to its predefined size by<br />
keeping representative solutions that have specifications of all<br />
the solutions. We use average linkage method because it has<br />
proven to perform well with SPEA algorithm [25].<br />
Step 4: For each individual x, the values of |M| objective<br />
functions as formulated in (1), and link utilization as in (2)<br />
are calculated and preserved in the vector Obj x (Obj x is a<br />
vector of |M| + 1 scalar values). Then, the values of objective<br />
functions are modified for each solution according to the<br />
penalty based approach. The degree of violation of capacity<br />
constraints is calculated for each individual x as follows:<br />
v(x) = 1<br />
|L|<br />
∑<br />
2<br />
|L| 2<br />
C j (x)<br />
, (6)<br />
j=1<br />
Cmax<br />
j<br />
where C j (x) represents the degree of violation of the j th<br />
capacity constraint for individual x, and:<br />
C j max = max<br />
x<br />
C j (x) (7)<br />
C j (x) takes positive value if j th capacity constraint in equations<br />
(4) is violated by solution x; otherwise, it is 0.<br />
The distance value of the individual x in each dimension i<br />
of objective function (i th element of vector Obj x ) is calculated<br />
as follows:<br />
{ v(x) if rf = 0<br />
d i (x) = √<br />
Objx (i) 2 + v(x) 2<br />
(8)<br />
Otherwise<br />
where Obj x (i) represents the i th element of vector Obj x , and<br />
r f indicates the portion of feasible individuals in the current<br />
population. r f takes value from [0,1]. If there is no feasible<br />
individual in the current population, the individuals with<br />
smaller capacity constraint violation values will have smaller<br />
distance function in the i th dimension and will dominate the<br />
other individuals in this dimension. If there is at least one<br />
feasible individual in the population, those feasible individuals<br />
with smaller objective function value will dominate the other<br />
individuals in the i th dimension. In this case, among the<br />
infeasible individuals, the ones that are closer to the origin<br />
in Obj x (i) − v(x) space will have smaller distance function<br />
regarding the i th dimension [26].<br />
The penalty function for the i th objective function of<br />
individual x is calculated according to (9):<br />
where<br />
p i (x) = (1 − r f )X i (x) + r f Y i (x), (9)<br />
{ 0 if rf = 0<br />
X i (x) =<br />
v(x) Otherwise<br />
(10)
BAGHERPOUR et al.: ENHANCING DATA TRANSMISSION RELIABILITY WITH MULTIPATH MULTICAST RATE ALLOCATION 65<br />
{<br />
0 if x is a feasible individual<br />
Y i (x) =<br />
Obj x (i) Otherwise<br />
(11)<br />
This penalty function penalizes the infeasible individuals even<br />
more. The first part of this penalty function,(1 − r f )X i (x),<br />
has larger values for the individuals with large amount of<br />
constraint violation and will have more effect when r f tends<br />
to zero. In the second part of the penalty function, r f Y i (x),<br />
the infeasible individuals with larger objective function value<br />
will be penalized more. This penalty function has more impact<br />
when r f tends to one [26].<br />
Finally each dimension i of the modified objective function<br />
for the individual x is obtained by using (8) and (9) as follows:<br />
Modified Obj x (i) = d i (x) + p i (x) (12)<br />
Fig. 1.<br />
V-NSF network topology<br />
Step 5: Fitness function of individuals in P and P ′ is<br />
calculated as follows:<br />
Each individual i ∈ P ′ is assigned a strength value s i =<br />
n/N +1) ∈ [0, 1], where n represents the number of individuals<br />
in P which are dominated by i, and N represents the size<br />
of P . Fitness of each individual j is defined is follows:<br />
{<br />
sj j ∈ P ′<br />
f j = 1 + ∑ s i j ∈ P (13)<br />
i,i≥j<br />
Step 6: The individuals are selected from P +P ′ as parents to<br />
form the mating pool according to their fitness. In this study,<br />
the Roulette wheel selection is used to choose the parents.<br />
Step 7: Crossover and mutation operators are applied to<br />
the parents. In this work, we use flat and random operators<br />
which are proposed in [27] and [28] respectively. As a result,<br />
population of the next generation is generated.<br />
Step 8: The procedure terminates if the domination rate<br />
of the current generation is zero, otherwise it is continued<br />
from Step 2. Domination rate in a generation is defined as<br />
the portion of the individuals in the non-dominated set of<br />
the previous generation which are dominated by the nondominated<br />
individuals in the current generation.<br />
In the next section, numerical results of implementation of<br />
the multi-objective genetic algorithm for a sample network are<br />
presented.<br />
A. Simulation Results<br />
V. NUMERICAL RESULTS<br />
The 14-node NSF (National Science Foundation) network<br />
topology is chosen to study the performance of proposed<br />
algorithm. This network topology is shown in Fig. 1. We<br />
consider node N 0 as the source of the multicast transmission<br />
and nodes N 4 , N 5 , N 9 , and N 12 as the destination nodes.<br />
We also consider that three multicast trees are available to<br />
transmit data packets from the source to destinations. Let<br />
MT = {T 1 , T 2 , T 3 } be the set of multicast trees. The multicast<br />
trees T 1 , T 2 and T 3 are illustrated in Fig. 2 by using red, blue,<br />
and green colors respectively.<br />
As mentioned before, discrete event simulation is used to<br />
estimate probability distribution of the portion of time that<br />
each path of multicast trees spends in the bad state out of<br />
the total block time T . For this purpose, we assume that<br />
Fig. 2.<br />
V-NSF network multicast trees<br />
the total block time T for transmitting data packets is 1 s.<br />
We consider network links with uniformly distributed random<br />
good time and bad time means, and simulate the system to<br />
observe and record the behavior of network paths in order<br />
to derive distribution of the bad time portions. The good<br />
time mean for each network link is generated according to a<br />
uniform distribution with parameters 0.9 s. and 1 s. and the bad<br />
time mean of each network link is uniformly distributed with<br />
parameters 0.01 s. and 0.02 s. The range of these parameters<br />
is chosen based on the previous works [17] and [23].<br />
We replicate the simulation 1000 times for each path in<br />
order to have adequate number of observations to be able<br />
to find the shape of probability distribution functions, and<br />
estimate the parameters of the distribution. Three well-known<br />
statistical hypothesis tests are applied to data samples, namely<br />
Kolmogorov-Smirnov test, Anderson-Darling test, and Chisquared<br />
test. The results show that the distribution that fits<br />
best with the data samples of the bad time portion of each<br />
path and successfully passes all statistical tests is the shifted<br />
Log-Normal. Data samples for each path pass the hypothesis of<br />
following shifted Log-Normal distribution. The parameters of<br />
the shifted Log-Normal distribution for each path are obtained<br />
by using maximum likelihood estimation. Values of parameters<br />
obtained for the Log-Normal distribution associated of each<br />
path are given in Table II. In this table, P ath T i,Nj represents<br />
the path associated with destination N j in tree T i<br />
When a random variable X has shifted Log-Normal distribution<br />
with parameters (µ, σ, γ), variable (X − γ) has<br />
Log-Normal distribution with parameters (µ, σ), i.e. Y =
66 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Fig. 3. Data histogram of path T1,N4 and fitted Log-Normal (µ, σ, γ) =<br />
(−1.74, 0.124, −0.07)<br />
Fig. 5. Data histogram of path T1, N9 and fitted Log-Normal (µ, σ, γ) =<br />
(−2.037, 0.141, −0.054)<br />
Fig. 4. Data histogram of path T1, N5 and fitted Log-Normal (µ, σ, γ) =<br />
(−1.701, 0.114, −0.095)<br />
Fig. 6. Data histogram of path T1, N12 and fitted Log-Normal (µ, σ, γ) =<br />
(−1.098, 0.071, −0.206)<br />
log(X − γ) has normal distribution with parameters (µ,σ 2 ).<br />
Histograms of the observed bad time portions of the paths<br />
associated with destinations N 4 , N 5 , N 9 , and N 12 in T 1 are<br />
illustrated in Fig. 3 to Fig. 6 respectively. These figures also<br />
demonstrate Log-Normal distribution curve that is fitted on<br />
data samples.<br />
B. Genetic Algorithm Implementation Results<br />
After obtaining probability distribution of the bad time<br />
portion of paths, we use our proposed algorithm to find<br />
solutions for the optimization problem presented in Section<br />
III. In each generation of the proposed genetic algorithm, we<br />
need to calculate the irrecoverable loss probabilities according<br />
to (1) for each individual. To be able to calculate these<br />
probabilities, we need to have the cumulative probability<br />
distribution of the weighted sum of Log-Normal variables.<br />
P j E<br />
is calculated supposing that B ijs in (1) have shifted<br />
Log-Normal distribution. Since distribution of sum of Log-<br />
Normal variables cannot be found in the closed form, we<br />
use the well-known approximation of Fenton and Wilkinson<br />
[29]. They approximated summation of Log-Normal variables<br />
with another Log-Normal variable. If we assume that X j<br />
j = 1, ..., n has Log-Normal distribution with parameters<br />
(µ j , σ j ), then sum of X j s is approximated by a Log-Normal<br />
variable with parameters (µ, σ) as follows:<br />
⎛ n∑<br />
⎞<br />
e 2µj+σ2 j (e<br />
σ 2 j − 1)<br />
σ 2 j=1<br />
= log ⎜<br />
⎟<br />
⎝ ∑<br />
( n ⎠ (14)<br />
e µj+σ2 j /2 ) 2<br />
j=1<br />
⎛<br />
⎞<br />
n∑<br />
µ = log ⎝ e µj+σ2 j /2 ⎠ − σ2 j<br />
(15)<br />
2<br />
j=1<br />
By this approximation, fitness value of each individual in<br />
the genetic algorithm can be easily calculated from (1). We<br />
implemented the proposed genetic algorithm for the network<br />
and paths illustrated in Fig. 2 with random good and bad time<br />
for network links.<br />
The values of genetic algorithm parameters are presented<br />
in Table. III. We run the algorithm with two values for the<br />
maximum number of paths for each destination. Also, the<br />
proposed genetic algorithm is run for 12 different capacity<br />
sets for network links.<br />
Figs. 7-10 illustrate the average of irrecoverable loss probability<br />
versus K obtained by single-tree transmission and<br />
multi-tree transmission. In multi tree case, we are allowed<br />
to use all the aforementioned trees to send the data packets<br />
toward the destinations, whereas in single tree case, we can<br />
use only one specific tree to deliver the data packets to the
BAGHERPOUR et al.: ENHANCING DATA TRANSMISSION RELIABILITY WITH MULTIPATH MULTICAST RATE ALLOCATION 67<br />
TABLE II<br />
LOG-NORMAL DISTRIBUTION PARAMETERS<br />
Path Parameter µ Parameter σ Parameter γ<br />
Path T1,N4 -1.823 0.132 -0.065<br />
Path T2,N4 -1.917 0.147 -0.079<br />
Path T3,N4 -1.563 0.129 -0.106<br />
Path T1,N5 -2.322 0.146 -0.040<br />
Path T2,N5 -1.865 0.172 -0.049<br />
Path T3,N5 -1.412 0.103 -0.027<br />
Path T1,N9 -1.990 0.130 -0.063<br />
Path T2,N9 -0.658 0.046 -0.401<br />
Path T3,N9 -1.642 0.129 -0.078<br />
Path T1,N12 -1.530 0.108 -0.094<br />
Path T2,N12 -1.919 0.138 -0.082<br />
Path T3,N12 -2.284 0.174 -0.037<br />
Fig. 7. Average value of irrecoverable loss vs. K for single tree and multi-ree<br />
for the first destination<br />
TABLE III<br />
GENETIC ALGORITHM PARAMETERS<br />
Parameter<br />
Value<br />
Total packet number (N) 1000<br />
Initial packet number (K) 875, 880, 885, 890, 895, 900<br />
Maximum number of paths (K ′ ) 2, 3<br />
Network links capacity (C) 12 different capacity sets<br />
Population size 100<br />
Non-dominated set size 300<br />
Crossover rate 0.85<br />
Mutation rate 0.05<br />
destinations. Since we have multiple non-dominated solutions<br />
in each implementation of genetic algorithm, we calculate<br />
the average value of each objective function over different<br />
solutions in order to obtain one representative in each µσγ<br />
implementation of the algorithm.<br />
We can see from Fig. 7 that for NSF network topology, the<br />
average probability of irrecoverable loss for destination N 4 by<br />
using multiple trees is smaller than that by using only trees<br />
T 1 and T 3 . However, it is greater than the value obtained by<br />
using only T 2 . The similar observation can be seen for the<br />
other destinations in other figures.<br />
To compare multi-tree transmission and single-tree transmission,<br />
we can see from Fig. 7 to Fig. 10 that if we use<br />
only the tree T 1 to send data packets toward the destinations,<br />
probability of irrecoverable loss for the destinations N 9 are<br />
smaller as compared to the multi-tree case. However, probabilities<br />
of occurrence of irrecoverable loss for N 4 , N 5 , and<br />
N 12 in single-tree case are respectively more than 0.5, approximately<br />
0.3 and more than 0.8 that are much greater than the<br />
probability under multi-tree transmission. These probabilities<br />
of irrecoverable loss for these destinations can be considered<br />
very unacceptable. We can observe similar behavior for other<br />
destinations. In the presented results, we observe that sending<br />
all the data packets over only one tree results to enhance the<br />
probability of irrecoverable loss for some destinations, but<br />
can not necessarily keep the probability low enough for all<br />
destinations and make them much worse as compared to multi<br />
tree case.<br />
Fig. 8. Average value of irrecoverable loss vs. K for single tree and multi-ree<br />
for the second destination<br />
Fig. 9. Average value of irrecoverable loss vs. K for single tree and multi-ree<br />
for the third destination<br />
Fig. 10. Average value of irrecoverable loss vs. K for single tree and multiree<br />
for the fourth destination
68 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
TABLE IV<br />
NUMERICAL RESULTS<br />
Multicast<br />
trees<br />
Decision variables<br />
Objective<br />
functions<br />
T 1<br />
X 12 = 1000, PE 2 = 0.1172,<br />
X 11 = 1000, PE 1 = 0.4925,<br />
X 13 = 1000, PE 3 = 0.1072,<br />
X ij =0 for i=2, 3, j=1, 2, 3, 4. P 4 E = 0.8641.<br />
T 2<br />
X 22 = 1000, PE 2 = 0.1255,<br />
X 21 = 1000, PE 1 = 0.0041,<br />
X 23 = 1000, PE 3 = 0.4711,<br />
X ij =0 for i=2, 3, j=1, 2, 3, 4. P 4 E = 0.4980.<br />
T 3<br />
X 32 = 1000, PE 2 = 0.5024,<br />
X 31 = 1000, PE 1 = 0.1183,<br />
X 33 = 1000, PE 3 = 0.4905,<br />
X ij =0 for i=2, 3, j=1, 2, 3, 4. P 4 E = 0.1180.<br />
T 1 , T 1 , T 3<br />
X 14 = 2, X 21 = 692, X 22 = 488, PE 2 = 0.0547,<br />
X 11 = 61, X 12 = 508, X 13 = 977, PE 1 = 0.0019,<br />
X 23 = 19, X 24 = 4, X 31 = 247, PE 3 = 0.1071,<br />
X 32 = 4, X 33 = 4, X 34 = 994, P 3 E = 0.1190.<br />
Fig. 11. Domination rate, K = 950<br />
To show that sending data packets over multiple trees can<br />
improve reliability of transmission to multicast destinations in<br />
comparison with single-tree, we implemented the model with<br />
a different parameter setting. In this test, the means of good<br />
and bad times for all network links are equal to 0.1 s and<br />
0.01 s respectively. We relaxed the capacity constraint of links.<br />
This allows the decision variables associated with a destination<br />
take values independent from values of decision variables<br />
associated with other destinations. We used the first objective<br />
functions in (1) as the fitness function in the proposed<br />
genetic algorithm. By this setting, values of decision variables<br />
associated with each destination are independent from those<br />
of other destinations according to the mathematical model in<br />
Section III. The results of implementing the genetic algorithm<br />
for single-tree and multi-tree transmission and for K ′ = 950<br />
are illustrated in Table. IV. It can be observed that when we<br />
are allowed to use all the multicast trees to send data packets,<br />
probabilities of irrecoverable loss are significantly less thansingle<br />
tree transmission. Number of data packets transmitted<br />
over trees T 1 , T 2 , and T 3 are equal to the maximum number of<br />
packets that are being transmitted to different destinations over<br />
these trees and can be calculated having values of decision<br />
variables.<br />
As mentioned before, a domination rate is calculated at each<br />
iteration of the genetic algorithm that represents the portion<br />
of Pareto solutions in that iteration that dominate the Pareto<br />
solutions of the previous iteration. Fig. 11 to Fig. 13 show the<br />
domination rate at iterations of the proposed genetic algorithm<br />
for K = 950, 955, and 960. The genetic algorithm stops when<br />
there is only one non-dominated solution remained in the nondominated<br />
set in several subsequent iterations of the algorithm.<br />
We can see from these figures that the domination rate in the<br />
Pareto solutions at initial iterations is higher and it gradually<br />
Fig. 12. Domination rate, K = 955<br />
Fig. 13. Domination rate, K = 960<br />
converges to zero after around 100 generations. This rate can<br />
be pegged as a convergence sign of the genetic algorithm.<br />
Multi-tree transmission can also reduce network congestion<br />
in comparison to single-tree transmission. Fig. 14 represents<br />
the average value of maximum link utilization function versus<br />
link capacity for multi-tree and single-tree. The single-tree<br />
curve represents the average of results obtained from sending<br />
data packets over the first, the second and the third tree. We<br />
can see from Fig. 14 that the average of network congestion<br />
in multi-tree transmission is much smaller than this value for<br />
single-tree cases. It can also be inferred from this figure that<br />
when the links have lower capacity, the single tree solutions
BAGHERPOUR et al.: ENHANCING DATA TRANSMISSION RELIABILITY WITH MULTIPATH MULTICAST RATE ALLOCATION 69<br />
Fig. 14.<br />
Average value of maximum link utilization versus link capacity<br />
are not feasible considering capacity constraints, because the<br />
value of maximum link utilization is greater than 1. However,<br />
the average congestion for multi-tree solution is always below<br />
1. These results indicate that multi-tree transmission can be<br />
the better choice when we have strict capacity constraints.<br />
VI. CONCLUSION<br />
In this work, we proposed a multi-objective mathematical<br />
formulation for multipath multicast rate allocation problem in<br />
order to minimize the probability of irrecoverable loss and<br />
also minimize network congestion. For calculating probabilities<br />
of irrecoverable loss for each destination, we estimated<br />
distribution of the time that each path spends in the bad state<br />
out of the total block time. Discrete event simulation was<br />
used to derive this probability distribution for NSF network<br />
topology. We observed that the bad time portion of network<br />
paths follows shifted Log-Normal distribution. We proposed<br />
a multi-objective genetic algorithm based on SPEA to solve<br />
the model and to find the Pareto solutions. Numerical results<br />
show that multipath routing significantly decreases probability<br />
of irrecoverable loss in comparison to single-tree transmission.<br />
Deriving the relationship between parameters of Log-<br />
Normal distribution found for bad state time and number of<br />
links in each path and the parameters of Gilbert-Elliot model<br />
can be considered as a future research to the work presented in<br />
this paper. This work can further be expanded by considering<br />
other QoS measures such as end-to-end delay and delay jitter.<br />
ACKNOWLEDGEMENTS<br />
This work was supported by Centre for Quantifiable Quality<br />
of Service in Communication Systems, Centre of Excellence,<br />
appointed by the Research Council of Norway,<br />
and funded by the Research Council, NTNU and UNINETT<br />
(http://www.q2s.ntnu.no).<br />
REFERENCES<br />
[1] N. F. Maxemchuk, “Diversity Routing,” in Proc. of IEEE ICC, San<br />
Francisco, CA, pp. 10–41.<br />
[2] R. Fabregat, Y. Donoso, B. Baran, F. Solano, and J. L. Marzo, “Multi-<br />
Objective Optimization Scheme for Multicast Flows: A Survey, a Model<br />
and a MOEA Solution,” in Proc. of 3rd Int. IFIP/ACM Latin American<br />
Conf. on Networking, New York, USA, 2005, pp. 73–86.<br />
[3] W. Tan and A. Zakhor, “Real-Time Internet Video Using Error Resilient<br />
Scalable Compression and tcp-Friendly Transport Protocol,” IEEE<br />
Trans. Multimedia, vol. 1, pp. 172–186, Jun. 1999.<br />
[4] G. D. L. Reyes, A. Reibman, S. Chang, and J. Chuang, “Error-Resilient<br />
Transcoding for Video over Wireless Channels,” IEEE J. Sel. Areas<br />
Commun., vol. 18, pp. 1063–1074, Jun. 2000.<br />
[5] J. Robinson and Y. Shu, “Zerotree Pattern Coding of Motion Picture<br />
Residues for Error-Resilient Transmission of Video Sequences,” IEEE<br />
J. Sel. Areas Commun., vol. 18, pp. 1099–1110, Jun. 2009.<br />
[6] Reibman, H. Jafarkhani, Y.Wang, M. Orchard, and R. Puri, “Multiple<br />
Description Coding for Video Using Motion Compensated Prediction,”<br />
in Proc. of Int. Conf. Image Processing (ICIP), vol. 3, Oct. 1999, pp.<br />
837–841.<br />
[7] J. Apostolopoulos, “Reliable Video Communication over Lossy Packet<br />
Networks Using Multiple State Encoding and Path Diversity,” in Proc.<br />
of SPIE, vol. 4310, Jun. 2001, pp. 392–409.<br />
[8] R. Puri, K. Ramchandran, K. Lee, and V. Bharghavan, “Forward Error<br />
Correction (FEC) Codes Based Multiple Description Coding for Internet<br />
Video Streaming and Multicast,” Signal Process.: Image Commun.,<br />
vol. 6, no. 8, pp. 745–762, May 2001.<br />
[9] K. Goyal and J. Kovacevic, “Generalized Multiple Description Coding<br />
with Correlating Transforms,” IEEE Trans. Inf. Theory, vol. 47, pp.<br />
2199–2224, Apr. 2001.<br />
[10] Y. Wang, M. Orchard, V. Vaishampayan, and A. Reibman, “Multiple Description<br />
Coding Using Pairwise Correlating Transforms,” IEEE Trans.<br />
Image Process., vol. 10, pp. 351–366, Mar. 2001.<br />
[11] H. Ma and M. E. Zarki, “Broadcast/Multicast mpeg-2 Video over<br />
Wireless Channels Using Header rRdundancy FEC Strategies,” in Proc.<br />
of SPIE, vol. 3528, Nov. 1998, pp. 69–80.<br />
[12] W. Tan and A. Zakhor, “Error Control for Video Multicast Using<br />
Hierarchical FEC,” in Proc. of 6th Int. Conf. Image Processing (ICIP),<br />
vol. 1, Oct. 1999, pp. 401–405.<br />
[13] P. A. Chou, A. E. Mohr, A. Wang, and S. Mehrotra, “Error Control for<br />
Receiver-Driven Layered Multicast of Audio and Video,” IEEE Trans.<br />
Multimedia, vol. 3, pp. 108–122, Mar. 2001.<br />
[14] Mohr, E. Riskin, and R. Ladner, “Unequal Loss Protection: Graceful<br />
Degradation over Packet Erasure Channels Through Forward Error<br />
Correction,” IEEE J. Sel. Areas Commun., vol. 18, pp. 819–828, Apr.<br />
2000.<br />
[15] W. Tan and A. Zakhor, “Error Control for Video Multicast Using<br />
Hierarchical FEC,” in Proc. of 6th Int. Conf. Image Processing (ICIP),<br />
vol. 1, Oct. 1999, pp. 401–405.<br />
[16] S. Deering, D. Estrin, D. Farinacci, V. Jacobson, C. Liu, and L.Wei, “The<br />
pim architecture for wide-area multicast routing,” IEEE/ACM Trans.<br />
Netw., vol. 4, pp. 153–162, Apr. 1996.<br />
[17] S. Fashandi, S. O. Gharan, and A. Khandani, “Path Diversity over the<br />
Internet: Performance Analysis and Rate Allocation,” IEEE/ACM Trans.<br />
Netw., vol. 18, pp. 1373–1386, Mar. 2010.<br />
[18] G. Apostolopoulos, T. Wong, W. Tan, and S. Wee, “On Multiple<br />
Description Streaming With Content Delivery Networks,” in Proc. od<br />
IEEE INFOCOM, vol. 3, Palo Alto, CA, 2002, pp. 1736–1745.<br />
[19] J. Chakareski and B. Girod, “Rate-Distortion Optimized Packet Scheduling<br />
and Routing for Media Streaming with Path Diversity,” in Proc. of<br />
IEEE Data Compression Conference, Snowbird, Utah, 2003, pp. 203–<br />
212.<br />
[20] H. Han, S. Shakkottai, C. Hollot, R. Srikant, and D. Towsley, “Multi-<br />
Path TCP: A Joint Congestion Control and Routing Scheme to Exploit<br />
Path Diversity in the Internet,” IEEE/ACM Trans. Netw., vol. 14, no. 6,<br />
pp. 1260–1271, 2006.<br />
[21] M. Ghanassi and P. Kabal, “Optimizing Voice-over-IP Speech Quality<br />
Using Path Diversity,” in Proc. of 8th IEEE Workshop on Multimedia<br />
Signal Processing, Victoria, BC, 2006, pp. 155–160.<br />
[22] S. Mao, S. Panwar, and Y. Hou, “On Optimal Partitioning of Realtime<br />
Traffic over Multiple Paths,” in Proc. of INFOCOM 2005, vol. 4, Miami,<br />
Florida, 2005, pp. 2325–2336.<br />
[23] T. Nguyen and A. Zakhor, “Path Diversity with Forward Error Correction<br />
(pdf) System for Packet Switched Networks,” in Proc. of IEEE<br />
INFOCOM, vol. 1, San Francisco, CA, 2003, pp. 663–672.<br />
[24] ——, “Multiple Sender Distributed Video Streaming,” IEEE Trans.<br />
Multimedia, vol. 6, no. 2, pp. 315–326, Mar. 2004.<br />
[25] E. Zitzler and L. Thiele, “Multiobjective Evolutionary Algorithm: A<br />
comparative case study and the Strength Pareto Approach,” IEEE Trans.<br />
Evolutionary Computations, vol. 3, no. 4, Nov. 1999.<br />
[26] Y. G.Woldesenbet, G. G. Yen, and B. G. Tessema, “Constraint Handling<br />
in Multiobjective Evolutionary Optimization,” IEEE Trans. Evolutionary<br />
Computations, vol. 13, no. 3, pp. 514–525, Jun. 2009.<br />
[27] N. J. Radcliffe, “Equivalence Class Analysis of Genetic Algorithms,”<br />
Complex Systems, vol. 5, pp. 183–205, 1991.<br />
[28] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution<br />
Programs. SpringerVerlag, New York, 1992.
70 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
[29] L. F. Fenton, “The Sum of Log-Normal Probability Distributions in<br />
Scatter Transmission Systems,” IRE Trans. Commun. Systems, vol. 8,<br />
pp. 57–67, Mar. 1960.<br />
Mehrdad Alipour is graduated from the Graduate School in Industrial<br />
Engineering, University of Science and Culture (USC), Tehran, Iran. He<br />
received his B.Sc. degree in Industrial Engineering from the Isfahan University<br />
of Technology (IUT), Iran in 2008. His research interests include optimization<br />
in telecommunication networks, meta-heuristic optimization approaches, and<br />
stochastic programming.<br />
Matin Bagherpour is a postdoctoral researcher in centre for Quantifiable<br />
Quality of Service in Communication Systems (Q2S) in the Norwegian<br />
University of Science and Technology (NTNU). She received her PhD<br />
in Industrial Engineering from Tarbiat Modares University (Iran) in May<br />
2008. She has worked as an assistant professor in University of Science<br />
and Culture (Iran) since then. Her research interests include optimization<br />
(mathematical programming, integer programming, meta-heuristic algorithms,<br />
combinatorial optimization), especially optimization of telecommunication<br />
networks (routing, scheduling, resource allocation, pricing, QoS, etc.), and<br />
ICT economics.<br />
Øivind Kure is a Professor at centre for Quantifiable Quality of Service in<br />
Communication Systems (Q2S) in the Norwegian University of Science and<br />
Technology (NTNU). He received his PhD from the University of California,<br />
Berkeley, USA. His current research interests include quality of service in<br />
wired and wireless network, sensor networks, and routing.
ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong> 71<br />
Analytical Model for Virtual Link Provisioning in<br />
Service Overlay Networks<br />
Piotr Krawiec, Andrzej Bęben and Jarosław Śliwiński<br />
Abstract—In this paper, we propose analytical model of Virtual<br />
Link used in the Service Overlay Networks. The Virtual Link<br />
exploits Selective Repeat ARQ scheme with time constrained<br />
number of retransmission and the playout buffer mechanism.<br />
Our model allows deriving equations that express trade-off<br />
between loss and delay characteristics experienced by packets<br />
transferred through VL. The main innovation of our model is<br />
the ability to cope with variable delay experienced by packets<br />
transferred by underlying network. Following the analytical<br />
model, we propose a method for Virtual Link dimensioning. The<br />
accuracy of the proposed model and dimensioning method is<br />
illustrated by simulation results.<br />
I. INTRODUCTION<br />
THE Service Overlay Networks (SON) [1] operate at the<br />
application layer to offer new services in the Internet,<br />
such as QoS [2], reliability [3], multicast [4], privacy [5], etc.<br />
The nodes in SON, so called overlay nodes that are connected<br />
using underlying network, perform service specific functions<br />
related to both packet forwarding and service control. Since<br />
transfer characteristics of underlying network are usually not<br />
adequate for SON requirements, the overlay nodes engage<br />
additional mechanisms to adjust packet transfer characteristics<br />
to specific SON needs. This concept, called Virtual Link (VL),<br />
was introduced in [2] and then it was extended by several<br />
authors, e.g., [6], [7]. These studies show how the ARQ (Automatic<br />
Repeat reQuest) and/or FEC (Forward Error Correction)<br />
mechanisms applied at VL recover lost packets and finally<br />
improve the quality of transferred VoIP or video streams. The<br />
VL concept was enhanced in [8], where authors applied hybrid<br />
ARQ scheme jointly with the playout buffer mechanism to not<br />
only recover lost packets, but also to mitigate packet delay<br />
variation. Such improvement was achieved at the expense of<br />
reduced capacity and increased packet transfer delay.<br />
In this paper, we introduce analytical model for VL with<br />
the Selective Repeat ARQ scheme and the playout buffer<br />
mechanism. Although the anaysis of delay characteristics of<br />
ARQ schemes have been already presented in literature, e.g.,<br />
in [9], [10], [11], [12], [13], [14], they are based on the<br />
assumtion of constant round trip time. The main novelty of<br />
our model are: (1) the ability to cope with variable transfer<br />
delays between sender and receiver (2) time limited number<br />
of retransmissions. These features originate from characteristics<br />
of underlying network, where packet transfer delays<br />
are usually described by parameters of a random variable.<br />
As a consequence, VL behaves similar to a queueing system<br />
Piotr Krawiec, Andrzej Bęben and Jarosław Śliwiński are with Institute of<br />
Telecommunications, Warsaw University of Technology Nowowiejska 15/19,<br />
00-665 Warsaw, Poland Email: {pkrawiec, abeben, jsliwinski}@tele.pw.edu.pl<br />
Fig. 1.<br />
The Virtual Link concept.<br />
with randomly delayed feedback. Such model has not been<br />
solved yet, therefore different approximations are considered,<br />
e.g., [15]. Our model allows to approximate the distribution<br />
of packet transfer time starting from the moment when a<br />
packet arrives to VL at the sender side until the moment<br />
when the packet leaves the receiver side or until it is lost due<br />
to exceeding the threshold of packet transfer delay. On that<br />
basis, we are able to express VL characteristics as a function<br />
of packet transfer characteristics of underlying network and<br />
the assumed transfer time threshold. Using this analysis we<br />
propose a method for dimensioning of the VL.<br />
The paper organisation is the following. In Section II, we<br />
introduce the VL concept. Then in Section III, we present<br />
proposed analytical model of the VL jointly with simulation<br />
results showing its effectiveness. In Section IV, we propose<br />
VL dimensioning method, which takes advantages of proposed<br />
analytical model. Finally, Section V summarises the paper and<br />
gives outline of further works.<br />
II. VIRTUAL LINK<br />
The SON concept assumes that overlay nodes connect each<br />
other by Virtual Connections (VC), which are offered by<br />
underlying network, as presented on Fig. 1. Since usually,<br />
there is no direct relation between SON and underlying network,<br />
the overlay nodes engage additional mechanisms, called<br />
Virtual Link (VL), to adjust packet transfer characteristics<br />
offered by VC to SON needs. As proposed in [2], [8], the<br />
VL engages the selective repeat ARQ mechanism supported<br />
by the playout buffer mechanism. The selective repeat ARQ<br />
mechanism recovers lost packets at the expense of increased<br />
packet transfer delay and reduced link capacity. On the other<br />
hand, the playout buffer mechanism enforces the same delay<br />
for each packet transferred through VL, which emulates the<br />
behaviour of an ordinary synchronous link.
72 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
packets<br />
arriving from<br />
ON1 input ports<br />
output port of ON1<br />
packets are<br />
served with<br />
CVL rate<br />
delay control<br />
mechanism<br />
time<br />
stamp<br />
VL sender side<br />
ARQ RTX<br />
buffer<br />
VC shaper<br />
VC in underlying<br />
network of CVC capacity<br />
(CVL < CVC )<br />
ARQ feedback<br />
input port of ON2<br />
VL receiver side<br />
playout buffer packets going<br />
to ON2 output<br />
ports with<br />
CVL rate<br />
packet transfer<br />
delay control<br />
ARQ decision<br />
point<br />
delay experienced on VC and D max denotes the time limit<br />
for successful packet delivery through VL. This time limit<br />
determines number of packet retransmissions. Equation (2)<br />
defines worst case limit of IP LR V L as a function of packet<br />
loss ratio on VC, defined by IP LR V C and the minimum<br />
number of packet retransmissions<br />
Fig. 2.<br />
The mechanisms of Virtual Link.<br />
Fig. 2 presents exemplary VL, which is established between<br />
neighbouring overlay nodes ON 1 and ON 2 . The packet arriving<br />
to the output buffer in ON 1 enters the queue, which<br />
is served with the rate of VL, named C V L . The VL begins<br />
packet service by the ARQ mechanism, where the copies of<br />
particular packets are stored in ARQ retransmission buffer<br />
(ARQ RTX). To each packet, we add time stamp with its<br />
arrival time to VL. We used this time stamp at the receiver side<br />
to recover traffic profile in the playout buffer. The receiving<br />
side controls the sequence of the incoming packets and it sends<br />
acknowledgements for received packets as well as requests<br />
for retransmissions for lost packets. If a packet is lost, then<br />
the sending side retransmits it upon receiving retransmission<br />
request or expiration of the time-out. The number of retransmissions<br />
is limited by value D max , which defines the<br />
time limit for successful packet delivery. At the receiver side,<br />
we put received packets into the playout buffer. The playout<br />
buffer delays the packet departure to allow for retransmissions<br />
of previously lost packets and mitigate the variable packet<br />
transfer time in VL. Moreover, playout buffer recovers the<br />
sequence and the inter-packet gaps of transferred packets,<br />
based on their timestamps. When a packet arrives too late,<br />
the playout buffer simply drops it. More detailed description<br />
of VL mechanisms is presented in [8].<br />
The starting point in the VL analysis are characteristics<br />
of VC. They correspond to available capacity, denoted as<br />
C V C , and the packet transfer characteristics expressed in<br />
terms of QoS metrics [16] such as: 1) minimum IP Packet<br />
Transfer Delay, minIP T D V C , 2) IP Packet Delay Variation,<br />
IP DV V C , jointly with random variable describing random<br />
part of IP Packet Transfer Delay, as well as, 3) IP Packet Loss<br />
Ratio, IP LR V C . These data may come from contracts agreed<br />
with Internet Service Provider or from the measurements<br />
performed by overlay nodes. Anyway, in our analysis, we left<br />
the problem of gathering VC characteristics for further studies,<br />
assuming that VC characteristics are known a priori.<br />
Summarising, the VL mechanisms give a trade-off between<br />
packet transfer delay and packet transfer loss characteristics<br />
provided by VL. In principle, greater value of IP T D V L<br />
allows VL mechanisms for more retransmissions what improve<br />
packet loss characteristics but on the other hand might not<br />
be acceptable for delay sensitive traffic. This trade-off may<br />
be expressed by generic equations (1), (2) and (3). Equation<br />
(1) defines the value of IP T D V L experienced by packets<br />
transferred through VL<br />
IP T D V L = minIP T D V C + D max = const, (1)<br />
where minIP T D V C is the minimum value of packet transfer<br />
IP LR V L = IP LR V C<br />
1+⌊ Dmax<br />
RT Tmax ⌋ , (2)<br />
where RT T max is the maximum value of round trip time<br />
experienced by packets transferred through VC and turned<br />
around to the source by reverse VC.<br />
Note that every retransmission reduces the effective value<br />
of C V L . Therefore, we can approximate the allowed capacity<br />
of VL by the value corresponding to the infinite number of<br />
retransmissions<br />
C V L ≤ C V C ∗ (1 − IP LR V C ). (3)<br />
Taking into account that implementation of the VL requires<br />
introducing some overhead in the form of a VL header, we<br />
obtain the final value of allowed capacity for the VL by<br />
multiplying C V L from equation (3) by expression L d/(L d +L V L )<br />
, where L d denotes size of packet arriving to the VL, and L V L<br />
is the VL header length.<br />
Remark that presented above equations are valid for the VL,<br />
which uses selective repeat ARQ scheme and playout buffer<br />
mechanism.<br />
III. PROPOSED MODEL<br />
The Virtual Link features a retransmission scheme combined<br />
with delay based decision process. Consequently, the<br />
model for performance analysis has to take into account the<br />
correlation between retransmissions and the packet transfer<br />
characteristics between sender and receiver (in both directions).<br />
Our analysis aims to derive the packet transfer characteristics<br />
of the Virtual Link expressed by packet transfer<br />
delay IP T D V L and packet loss ratio IP LR V L with regard<br />
to its assumed capacity C V L and retransmission delay limit<br />
D max .<br />
A. Definition of model<br />
We assume that Virtual Connection has the capacity C V C<br />
and has different propagation delays for data and acknowledgement<br />
packets, which are t pd and t pa , respectively. Furthermore,<br />
we assume that data (acknowledgement) packets are of<br />
constant size L d (L a ) with transmission delay t Xd = L d/C V C<br />
(t Xa = La /C V C ). Sum of propagation delay t p and transmission<br />
delay t X equals minIP T D V C metric. Moreover, the<br />
random variable X i (Z i ) defines the variable part of packet<br />
transfer delay in the data (acknowledgement) direction. We<br />
presume that capacities of VCs are significantly lower than<br />
capacities avaiable in underlying network, and hence there is<br />
no correlation between handling, in underlying nodes, of two<br />
consecutive packets transferred through given VC. It allowes<br />
us to neglect correlation between losses and delays of consequtive<br />
packets transferred through VC. Therefore, the packet<br />
losses may follow independent model with loss probability
KRAWIEC et al.: ANALYTICAL MODEL FOR VIRTUAL LINK PROVISIONING IN SERVICE OVERLAY NETWORKS 73<br />
VL<br />
input<br />
Fig. 3.<br />
time<br />
stamp<br />
Dq<br />
Input<br />
buffer<br />
RTO<br />
start<br />
RTX<br />
buffer<br />
CVC<br />
RTX<br />
start<br />
Drtx<br />
The Virtual Link model.<br />
Dt<br />
tXd + tpd + Xi ; pd<br />
tXa + tpa + Zi ; pa<br />
p d (p a ) in data (acknowledgement) direction. Additionally,<br />
the above assumptions implies that random variables defining<br />
packet transfer delay (X i and Z i ) and packet losses are all<br />
independent of each other.<br />
The packets arriving to the Virtual Link have constant bit<br />
rate profile with bit rate C V L (constant packet inter-arrivals).<br />
We assume, that the sequence number space ARQ mechanism<br />
utilises, is large enough to not stop the packet sending process.<br />
The overview of the Virtual Link model is presented in<br />
Fig. 3. The variables used in the model are the following:<br />
• Dq – random variable which describes the waiting time<br />
in the input buffer for “fresh” packets. Dq duration is<br />
defined by the time instant when a packet enters the input<br />
buffer and by the time instant when its service begins, i.e.,<br />
the first transmission in Virtual Connection.<br />
• Dt – random variable which describes the packet transfer<br />
delay from the sender to the receiver (in Virtual Connection).<br />
It is defined as time interval from the start of first<br />
transmission of the “fresh” packet in Virtual Connection,<br />
until the moment of arrival of the first packet copy to the<br />
receiver. Virtual Link drops packets having no chance<br />
for reception before the deadline D max ; therefore, we<br />
assume that for dropped (i.e., lost) packets Dt takes<br />
infinite value.<br />
• Dp – random variable which describes the duration of<br />
packet’s stay in the playout buffer.<br />
• Drtx – random variable which describes a queueing<br />
delay of retransmitted packets. We assume that retransmitted<br />
packets access the Virtual Connection with higher<br />
priority than “fresh” packets.<br />
According to the Virtual Link concept, assuming the packet<br />
is not lost, its total transfer delay in VL is constant and it is<br />
given by (see eq. 1)<br />
IP T D V L = t Xd + t pd + Dmax = Dq + Dt + Dp (4)<br />
Random variables Dq and Dt are constrained by IP T D V L<br />
pd<br />
Yi<br />
1-pd<br />
Dp<br />
Playout<br />
buffer<br />
CVL<br />
VL<br />
output<br />
Dq + Dt ≤ IP T D V L . (5)<br />
Duration Dp, which complements the sum of Dq and Dt to<br />
constant time IP T D V L , varies from 0 to Dmax<br />
Dp = IP T D V L − Dq − Dt, (6)<br />
Dp ∈ [0; Dmax] .<br />
Let P s defines the probability of the event that packet is<br />
successfully transferred by Virtual Link in time t i.e., the<br />
packet’s copy arrives to the receiver before t<br />
P s = P r{Dq + Dt ≤ t}. (7)<br />
In the Virtual Link, the rate of incoming packets is limited by<br />
capacity C V L . Next, the packets are taken for service from<br />
input queue with rate C V C > C V L or they wait in the input<br />
queue until service of retransmitted packets from RTX buffer<br />
is finished. Because packet losses higher than a few percent<br />
are rather rare in normal operation of network ([17]), for<br />
further analysis we assume that queueing time Dq is negligible<br />
compared to Dt. Consequently, we can state that<br />
P s = P r{Dt ≤ t}. (8)<br />
Thanks to the assumed independence between packet delay<br />
and packet losses characteristics, we can write the following<br />
formula which defines the conditional probability of packet’s<br />
successful delivery in time not longer than t, assuming that<br />
there were i th transmission attempts<br />
P r{Dt ≤ t} =<br />
=<br />
∞∑<br />
P r{Dt ≤ t ∧ T r = i} (9)<br />
i=1<br />
∞∑<br />
P r{Dt ≤ t | T r = i}P r{T r = i},<br />
i=1<br />
where T r is the random variable describing number of<br />
packet transmissions (including retransmissions) until successful<br />
packet delivery to the receiver. Assuming independent<br />
packet loss model, T r has a geometric distribution<br />
P r{T r = i} = (1 − p d )p (i−1)<br />
d<br />
. (10)<br />
Next, we analyse duration between time instant when the<br />
first transmission of a packet starts, until the successful reception<br />
after i th transmission. Probability that packet transfer<br />
time is not greater than t, assuming success on the first attempt<br />
(T r = 1), is given by<br />
P r{Dt ≤ t | T r = 1} = (11)<br />
= P r{t Xd + t pd + X i ≤ t ∧ T r = 1}<br />
P r{T r = 1}<br />
= P r{t Xd + t pd + X i ≤ t}P r{T r = 1}<br />
P r{T r = 1}<br />
= P r{t Xd + t pd + X i ≤ t}<br />
In this case the time Dt has two components: 1) data packet<br />
transmission time on VC link (t Xd ), and 2) variable propagation<br />
time through VC link (t pd + X i ).<br />
Probability that packet transfer time is not greater than t,<br />
assuming success on the second attempt (T r = 2), equals<br />
P r{Dt ≤ t | T r = 2} = (12)<br />
= P r{min[RT O; Y i + t Xd + t pd + X i<br />
+t Xa + t pa + Z i ] + Drtx +<br />
+t Xd + t pd + X i ≤ t}
74 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
Ld/CVL<br />
S<br />
P1<br />
R<br />
S<br />
P1<br />
R<br />
Yi<br />
Tf = 4<br />
RTO<br />
P1<br />
NACK1<br />
NACK1<br />
RTX<br />
P1<br />
Fig. 4.<br />
An Y i random variable.<br />
RTX<br />
P1<br />
In this case time Dt consists of the four components: 1)<br />
time required to detect lost packet (ARQ decision delay; it is<br />
defined as a minimum of a retransmission timeout RTO and<br />
time, after which NACK for lost packet arrives to a sender),<br />
2) waiting time in the retransmission buffer (Drtx), 3) data<br />
packet transmission time on VC link (t Xd ), and 4) variable<br />
propagation time of retransmitted data packet through VC link<br />
(t pd + X i ).<br />
Time, after which NACK arrives to a sender, is a sum of<br />
following values (see Fig. 4): 1) time Y i , 2) data packet transfer<br />
time (t Xd + t pd + X i ), and 3) transfer time (t Xa + t pa + Z i )<br />
of acknowledgement for data packet, which carries NACK for<br />
lost packet.<br />
Random variable Y i denotes time required for the sender<br />
to transmit consecutive data packets, which are essential in<br />
order to receive an acknowledgement indicating the packet<br />
loss. Because packets arrive to VL with rate C V L/L d , r.v. Y i<br />
can take the following discrete values (see Fig. 4):<br />
Y i = j · L d<br />
C V L<br />
, j ∈ {1, 2, . . .} (13)<br />
Let T f be a random variable which describes number of<br />
consecutive packets, which must be sent by the sender to<br />
receive information about packet loss. Fig. 4 illustrates the<br />
case for T f = 4. The T f depends on the data and the ACK<br />
packets loss probabilities (p d and p a , respectively) and has<br />
geometric distribution (see Appendix A):none<br />
P r{T f = j} = (1 − p d )(1 − p a )[p d + (1 − p d )p a ] (j−1) (14)<br />
Using the law of total probability, we can rewrite the<br />
formula (12) as:<br />
P r{Dt ≤ t | T r = 2} = (15)<br />
∞∑<br />
{ [<br />
= P r min RT O; j · L d<br />
+<br />
C V L<br />
j=1<br />
]<br />
+t Xd + t pd + X i + t Xa + t pa + Z i +<br />
}<br />
+Drtx + t Xd + t pd + X i ≤ t ·<br />
·P r {T f = j}<br />
The SR ARQ scheme used in Virtual Link assumes, that<br />
only the first packet retransmission is controlled by NACK,<br />
Fig. 5.<br />
RTX<br />
ACK1<br />
The Virtual Link retransmission schema.<br />
while the second, third, etc., occurs after retransmission interval<br />
RTX, as it is shown on Fig. 5. Such approach allows<br />
us to maintain high responsiveness for the first retransmission<br />
and, at the same time, to impose the limit on the additional<br />
traffic generated due to retransmissions. Taking this feature<br />
into account, we can express the probability that packet<br />
transfer time is not greater than t, assuming that i − 1 packet<br />
transmission attempts fail and there is a success in the i th<br />
attempt (i ≥ 2,) by<br />
P r{Dt ≤ t | T r = i} = (16)<br />
∞∑<br />
{ [<br />
= P r min RT O; j · L d<br />
+<br />
C V L<br />
j=1<br />
+t Xd + t pd + X i + t Xa + t pa + Z i<br />
]<br />
+<br />
+(i − 2)RT X + Drtx + t Xd + t pd +<br />
}<br />
+X i ≤ t · P r {T f = j}<br />
Finally, according to the formula (none9) the probability,<br />
that Dt (successful packet transfer from the sender to the<br />
receiver) is not greater than t, assuming unlimited number of<br />
retransmission, has the following form<br />
P r{Dt ≤ t} = (17)<br />
= P r {t Xd + t pd + X i ≤ t} (1 − p d )p 0 d +<br />
[<br />
∞∑ ∑ ∞ { [<br />
+ P r min RT O; j · L d<br />
+<br />
C V L<br />
i=2<br />
j=1<br />
+t Xd + t pd + X i + t Xa + t pa + Z i<br />
]<br />
+<br />
+(i − 2)RT X + Drtx + t Xd + t pd +<br />
}<br />
]<br />
+X i ≤ t · P r {T f = j} (1 − p d )p (i−1)<br />
d<br />
where P r {T f = j} is given by formula (14).<br />
B. Model evaluation<br />
We evaluate the analytical results, which are obtained using<br />
the proposed model, with results of simulations. The following
KRAWIEC et al.: ANALYTICAL MODEL FOR VIRTUAL LINK PROVISIONING IN SERVICE OVERLAY NETWORKS 75<br />
assumptions are taken:<br />
• since RTO timer is usually set to relatively high value,<br />
we consider RTO timeout expiration as exceptional, rare<br />
event, therefore the sender obtains information about<br />
packet loss mainly thanks to receiving negative acknowledgements;<br />
• taking into account, that retransmitted packets have<br />
higher priority and packet losses at VC are usually a few<br />
percent at most, we consider queueing delay Drtx of<br />
retransmitted packets as a negligible part of total packet<br />
transfer time Dt.<br />
Therefore, we can write formula (17) as:<br />
P r{Dt ≤ t} = (18)<br />
= P r {t Xd + t pd + X i ≤ t} (1 − p d )p 0 d +<br />
[<br />
∞∑ ∑ ∞ { j · Ld<br />
+ P r + t Xd + t pd + X i +<br />
C V L<br />
i=2<br />
j=1<br />
+t Xa + t pa + Z i + (i − 2)RT X + t Xd + t pd +<br />
}<br />
]<br />
+X i ≤ t · P r {T f = j} (1 − p d )p (i−1)<br />
d<br />
Let consider the expression under the summation:<br />
j · L d<br />
C V L<br />
+ t Xd + t pd + X i + t Xa + t pa + (19)<br />
+Z i + (i − 2)RT X + t Xd + t pd + X i<br />
For fixed i and j we can write it as a sum of independent<br />
random variables and constants:<br />
X i + Z i + X i + a + j · L d<br />
C V L<br />
+ (i − 2)RT X (20)<br />
where a is a constant value, which equals to 2(t Xd + t pd ) +<br />
t Xa + t pa .<br />
Probability density function of Dt can be obtained as a<br />
convolution of pdf s for X i and Z i random variables and<br />
the constants, by exploitation of the Laplace transform (δ()<br />
denotes the Dirac delta function):<br />
f Dt (t) = L −1{ L { } { ( )} }· X i · L δ t − tXd − t pd (21)<br />
⎡<br />
∞∑ ∞∑<br />
·P r{T r = 1} + ⎣ L<br />
{L −1 { X i<br />
}·<br />
i=2<br />
j=1<br />
L { } { } { ( j · L d<br />
Z i · L Xi · L δ t − a − −<br />
C V<br />
⎤ L<br />
−(i − 2)RT X )}} · P r{T f = j} ⎦ · P r{T r = i}<br />
We calculate distribution of Dt using (21) for the case,<br />
when random variables which describe variable part of packet<br />
transfer delay through VC link are exponentially distributed<br />
with mean equals none 1 /m X for direction sender-to-receiver<br />
and 1 /m Z for direction receiver-to-sender, respectively: X i ∼<br />
Exp(m X ), Z i ∼ Exp(m Z ). We assume that VC links for<br />
CCDF<br />
10 0<br />
10 -1<br />
10 -2<br />
10 -3<br />
10 -4<br />
10 -5<br />
pd=pa=5% - analysis<br />
pd=pa=5% - simulation<br />
10 -6<br />
pd=pa=0.5% - analysis<br />
pd=pa=0.5% - simulation<br />
10 -7<br />
0 20 40 60 80 100 120 140 160 180<br />
Dt [ms]<br />
Fig. 6. Complementary CDF of time Dt for case#1: VC with low packet<br />
transfer delay variation.<br />
CCDF<br />
10 0<br />
10 -1<br />
10 -2<br />
10 -3<br />
10 -4<br />
10 -5<br />
10 -6<br />
10 -7<br />
pd=pa=5% - analysis<br />
10 -8<br />
pd=pa=5% - simulation<br />
10 -9 pd=pa=0.5% - analysis<br />
pd=pa=0.5% - simulation<br />
10 -10<br />
0 20 40 60 80 100 120 140 160 180<br />
Dt [ms]<br />
Fig. 7. Complementary CDF of time Dt for case#2: VC with high packet<br />
transfer delay variation.<br />
both directions are the same: m x = m z , t pd = t pa , p d = p a ,<br />
with C V C = 1 Mbps and two values of packet loss ratio: 5%<br />
and 0.5%. The VL bit rate C V L = 650 Kbps and packet size<br />
L d = 200 B. We consider two cases:<br />
• case#1: VC is characterised by low packet delay variation<br />
(t pd = t pa = 25 ms, m x = m z = 1 ms, IP DV V C =<br />
7 ms);<br />
• case#2: VC is characterised by relatively high packet<br />
delay variation (t pd = t pa = 5 ms, m x = m z = 5 ms,<br />
IP DV V C = 34.5 ms).<br />
During simulation experiments we generate at least 4·10 8<br />
packets for each considered case.<br />
Fig. 6 and Fig. 7 depicts complementary cumulative distribution<br />
function of time Dt, calculated on the basis of<br />
formula (21), as well as obtained experimentally (solid line<br />
and dashed line, respectively), for both cases. We observe<br />
that the proposed model closely approximates the simulation<br />
results. The reason of small shift between the appropriate
76 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
CCDF<br />
10 0<br />
10 -1<br />
10 -2<br />
10 -3<br />
10 -4<br />
10 -5<br />
10 -6<br />
10 -7<br />
10 -8<br />
pd=pa=5% - analysis<br />
10 -9 pd=pa=5% - simulation<br />
pd=pa=0.5% - analysis<br />
10 -10<br />
pd=pa=0.5% - simulation<br />
10 -11<br />
0 20 40 60 80 100 120 140 160 180<br />
Dt [ms]<br />
Fig. 8. Complementary CDF of time Dt, in case when X i and Z i random<br />
variables have uniform distribution.<br />
curves, is that the approximation takes the worst case from<br />
the point of view of packet transmission time. We assume,<br />
that each packet is transmitted with rate limited by the VC<br />
shaper to C V C (see Fig. 2). During simulations, such limitation<br />
occurred rarely, and transmission time for most packets was<br />
much shorter, because they were sent with physical link bit<br />
rate, a hundred times greater than VC bit rate. Consequently,<br />
the proposed model can be treated as an upper bound for Dt<br />
time distribution.<br />
Notice, that in some intervals the simulated curves are above<br />
the calculated ones for greater values of Dt in case#2 (see<br />
Fig. 7). This is due to packet stream integrity condition, which<br />
we assumed during simulations. To avoid packet reordering,<br />
which is possible in high packet delay variation case, we allow<br />
only such values of single packet delay, generated according to<br />
exponential distribution, which keep packet order. In this way,<br />
distribution of X i and Z i random variables used in simulations<br />
differ slightly comparing to calculation.<br />
In Fig. 8 we present complementary cumulative distribution<br />
function of time Dt for case#2, i.e. when VC is characterized<br />
by high packet delay variation, but when the random variables<br />
X i and Z i have a uniform distribution over interval [0 ms,<br />
35 ms]. In this case, a distribution of packet transfer delay<br />
through VC link used for calculation was identical with the<br />
distribution applied in simulation experiments. Consequently,<br />
the analytical curve is above the simulated one, and can be<br />
treated as an upper bound for numerical results, as in case#1<br />
(see Fig. 6).<br />
IV. APPLICATION FOR VIRTUAL LINK DIMENSIONIG<br />
The Virtual Link concept assumes, that packet transfer is<br />
successful, if its copy arrives to the receiver in time not greater<br />
than IP T D V L :<br />
P s = P r{Dt ≤ IP T D V L } (22)<br />
Therefore, the packet loss probability on VL link can be<br />
defined as:<br />
TABLE I<br />
IP LR V L FOR VC WITH LOW PACKET TRANSFER DELAY VARIATION<br />
(CASE#1).<br />
pd (= pa) IP T D V L IP LR V L<br />
analytical model simulation<br />
40 ms 5·10 −2 5·10 −2 ± 4·10 −5<br />
5% 70 ms 5·10 −2 5·10 −2 ± 3 · 10 −5<br />
110 ms 2.5·10 −3 2.5·10 −3 ± 5 · 10 −6<br />
40 ms 5·10 −3 5·10 −3 ± 6·10 −6<br />
0.5% 70 ms 5·10 −3 5·10 −3 ± 9·10 −6<br />
110 ms 2.5·10 −5 2.5·10 −5 ± 3·10 −7<br />
IP LR V L = 1 − P s = 1 − P r{Dt ≤ IP T D V L } (23)<br />
However, exact computation of P r {Dt ≤ IP T D V L } is<br />
not trival, since in the VL number of possible transmission<br />
attempts for each packet is limited by time. Sender transmit<br />
packet only if it has chance to be received by receiver in the<br />
assumed time interval limit IP T D V L . Othervise, packet is<br />
droped. Therefore, we propose to approximate an IP LR V L<br />
metric by calculating distribution P r {Dt ≤ t} with assumption<br />
of unlimited number of retransmissions, as defined by<br />
equation (17). Next, we determine packet loss probability on<br />
the VL as a fraction of packets, for which transfer time Dt<br />
was greater than IP T D V L :<br />
IP LR V L ≈ P r{Dt > IP T D V L } (24)<br />
In tables I and II we present values of IP LR V L metric<br />
obtained from simulations and calculated analytically according<br />
to rule presented above. Simulations were performed for<br />
the same scenarios and the values of parameters as in section<br />
III-B. Table I refers to case#1, with low delay variation,<br />
whereas table II refers to case#2, with greater delay variation.<br />
Parameter Dmax had appropriate values to obtain packet<br />
transfer delay in the VL equals to 40, 70 and 110 ms. The<br />
results were obtained by repeating the simulation tests 10 times<br />
and calculating the mean values with the corresponding 95%<br />
confidence intervals. For each iteration, we simulated at least<br />
50·10 6 packets.<br />
Results presented in table I show, that proposed method<br />
allows to determine packet loss ratio in the Virtual Link<br />
with high accuracy in the case, when variation of delay is<br />
relatively low comparing to constant part of packet transfer<br />
delay through underlying VC. In the case#2, for tests with<br />
higher values of IP T D V L and packet losses in VC equal<br />
to 0.5%, the IP LR V L measured in simulations is above the<br />
calculated values (see table II). This mismatch is similar to the<br />
values observed for analytical and simulated curves depicted<br />
in Fig. ??. However, analytical results still constitute a good<br />
approximation for the IP LR V L obtained by simulations.<br />
Summarizing, presented analytical method helps us to approximate<br />
packet loss ratio provided by the Virtual Link,<br />
taking as input data the packet transfer characteristics of<br />
VC, such as: bit rate C V C , packet loss ratio p d and p a ,<br />
constant (t pd , t pa ) and variable (X i , Z i ) transfer delay, as
KRAWIEC et al.: ANALYTICAL MODEL FOR VIRTUAL LINK PROVISIONING IN SERVICE OVERLAY NETWORKS 77<br />
TABLE II<br />
IP LR V L FOR VC WITH HIGH PACKET TRANSFER DELAY VARIATION<br />
(CASE#2).<br />
pd (= pa) IP T D V L IP LR V L<br />
analytical model simulation<br />
40 ms 1.9·10 −2 1.4·10 −2 ± 3 · 10 −5<br />
5% 70 ms 9.1·10 −4 1.1·10 −3 ± 4·10 −6<br />
110 ms 1.3·10 −5 1.7·10 −5 ± 1 · 10 −7<br />
40 ms 2.8·10 −3 1.7·10 −3 ± 4 · 10 −7<br />
0.5% 70 ms 2.8·10 −5 5.4·10 −5 ± 7 · 10 −7<br />
110 ms 4.0·10 −8 6.6·10 −8 ± 1.5 · 10 −8<br />
well as assumed parameters of the VL: bit rate C V L and delay<br />
IP T D V L . Jointly with formula (3), which defines maximum<br />
allowed capacity of VL, it can be used for dimensioning of<br />
the Virtual Link. A procedure of the VL dimensioning we<br />
present below. Notice, that the VL admits a controlled tradeoff<br />
between packet loss level, offered capacity and constant<br />
delay introduced by the VL.<br />
A. Virtual Link dimensioning algorithm<br />
Using the formula (24) to determine upper bound of packet<br />
loss ratio in the VL, together with the formula (3) for approximation<br />
of the allowed VL capacity, we can dimension the<br />
Virtual Link according to the following algorithm:<br />
• Step 1: determine values of parameters for Virtual Connection,<br />
which is used for establishing the VL (bit rate<br />
of VC, packet loss ratio, minimum packet transfer delay<br />
on VC, distribution of the variable part of packet transfer<br />
delay on VC).<br />
• Step 2: for given value of the VL capacity C V L , which<br />
satisfies the condition described by formula (3), calculate<br />
distribution of packet transfer time between sending and<br />
receiving overlay node P r = {Dt ≤ t} (according to<br />
formula (18) ).<br />
• Step 3: according to formula (24), find such value T V L<br />
in obtained Dt time distribution, for which probability<br />
P r = {Dt > T V L } equals required value of<br />
IP LR V L . In case when the input parameter is value of<br />
IP T D V L , from obtained Dt time distribution find value<br />
of IP LR V L as probability P r = {Dt > IP T D V L }.<br />
• Step 4: calculate value of D max parameter for the VL as<br />
the difference between time T V L and minimum packet<br />
transfer delay on VC: D max = T V L – minIP T D V C .<br />
In case when the input parameter is value of IP T D V L ,<br />
then D max = IP T D V L – minIP T D V C .<br />
Using the following steps, at the top of given VC we can<br />
establish the VL with bit rate C V L , constant packet transfer<br />
time IP T D V L (= T V L ) and packet loss ratio not greater than<br />
IP LR V L .<br />
V. SUMMARY<br />
In this paper we considered the performance analysis of<br />
Virtual Link. The Virtual Link is established between nodes of<br />
the Service Overlay Network in order to improve packet transfer<br />
characteristics of the underlying network. The presented<br />
analysis focuses in the packet transfer delay distribution as<br />
observed during handling in the VL when the retransmission<br />
delay limit is infinite. Comparing to other analytical methods<br />
for ARQ systems, our model features variable transfer delays<br />
between sender and receiver, and moreover, it uses delaybound<br />
scheme for number of retransmissions. The accuracy<br />
of the proposed model was verified by means of simulation<br />
in exemplary scenario with exponentially distributed delay<br />
characteristics of the underlying network. The analytical and<br />
numerical results differ slightly due to worst case assumptions<br />
and due to the method of packet transfer delay emulation<br />
that maintains the order of packets in the underlying network<br />
(for the case with greater value of delay variation). Finally,<br />
we proposed a method for dimensioning of VL with finite<br />
retransmission delay limit that allows for controlled use of<br />
trade-off between packet transfer delay and packet losses.<br />
One of the remaining problems, which is not directly<br />
related to the VL analysis, is the reliable characterisation of<br />
packet transfer characteristics in the Virtual Connection. In the<br />
situation when operator of underlying network do not provide<br />
required parameters of the VC, for example, in the form of<br />
an SLA (Service Level Agreement) contract, the SON owner<br />
can obtain them using measurements. Those measurements<br />
can be performed by external tools, such as OWAMP (One-<br />
Way Active Measurement Protocol) tool [18], or by internal<br />
measurement module integrated with the VL. In the latter case,<br />
the packet transfer characteristics of the VC can be measured<br />
by means of a passive measurement method. For this purpose,<br />
we can use timestamps and sequence numbers, which are<br />
carried in each VL packet header, to determine packet transfer<br />
delay and loss characteristics.<br />
In the further work we will extend the proposed model to<br />
include the FEC mechanism.<br />
REFERENCES<br />
[1] Z. Duan, Z.-L. Zhang, and Y. T. Hou, “Service overlay networks: SLAs,<br />
QoS, and bandwidth provisioning,” IEEE/ACM Trans. Netw., vol. 11,<br />
no. 6, pp. 870–883, 2003.<br />
[2] L. Subramanian, I. Stoica, H. Balakrishnan, and R. H. Katz, “OverQoS:<br />
an overlay based architecture for enhancing internet QoS,” in NSDI’04:<br />
Proceedings of the 1st conference on Symposium on Networked Systems<br />
Design and Implementation. Berkeley, CA, USA: USENIX Association,<br />
2004, pp. 6–6.<br />
[3] D. Andersen, H. Balakrishnan, F. Kaashoek, and R. Morris, “Resilient<br />
overlay networks,” SIGOPS Oper. Syst. Rev., vol. 35, no. 5, pp. 131–145,<br />
2001.<br />
[4] M. Castro, P. Druschel, A.-M. Kermarrec, and A. Rowstron, “Scribe: A<br />
large-scale and decentralized application-level multicast infrastructure,”<br />
IEEE Journal on Selected Areas in Communicastions, vol. 20, no. 8, pp.<br />
1489–1499, October 2002.<br />
[5] R. Dingledine, N. Mathewson, and P. Syverson, “Tor: The secondgeneration<br />
onion router,” in Proceedings of the 13th USENIX Security<br />
Symposium. USENIX Association, August 2004.<br />
[6] Y. Amir, C. Danilov, S. Goose, D. Hedqvist, and A. Terzis, “An<br />
overlay architecture for high-quality voip streams,” IEEE Transactions<br />
on Multimedia, vol. 8, no. 6, 2006.<br />
[7] Z. Cen, M. W. Mutka, D. Zhu, and N. Xi, “Supermedia Transport for<br />
Teleoperations over Overlay Networks,” in Proceedings of NETWORK-<br />
ING 2005: 4th International IFIP-TC6 Networking Conference, ser.<br />
LNCS, vol. 3462. Springer, May 2005, pp. 1409–1412.<br />
[8] W. Burakowski, J. Śliwiński, A. Bęben, and P. Krawiec, “Constant Bit<br />
Rate Virtual Links in IP Networks,” in Proceedings of the 16th Polish<br />
Teletraffic Symposium. Łódź, Poland: Technical University of Łódź,<br />
2009, pp. 23–30.
78 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
[9] L. Badia, M. Rossi, and M. Zorzi, “Queueing and delivery analysis<br />
of SR ARQ on Markov channels with non-instantaneous feedback,”<br />
in Proceedings of the IEEE Global Telecommunications Conference<br />
GLOBECOM’05. IEEE Communications Society, 2005, pp. 3717–<br />
3721.<br />
[10] J. G. Kim and M. Krunz, “Delay analysis of selective repeat ARQ for<br />
a Markovian source over a wireless channel,” IEEE Transactions on<br />
Vehicular Technology, vol. 49, no. 5, pp. 1968–1981, September 2000.<br />
[11] M. Rossi, L. Badia, and M. Zorzi, “SR ARQ delay statistics on N-state<br />
Markov channels with non-instantaneous feedback,” IEEE Transactions<br />
on Wireless Communications, vol. 5, no. 6, pp. 1526–1536, June 2006.<br />
[12] W. Luo, K. Balachandran, S. Nanda, and K. Chang, “Delay analysis<br />
of selective-repeat ARQ with applications to link adaptation in wireless<br />
packet data systems,” IEEE Transactions on Wireless Communications,<br />
vol. 4, no. 3, pp. 1017–1028, May 2005.<br />
[13] N. Cardwell, S. Savage, and T. Anderson, “Modeling TCP latency,” in<br />
Proceedings of the Nineteenth Annual Joint Conference of the IEEE<br />
Computer and Communications Societies INFOCOM 2000. IEEE<br />
Communications Society, March 2000, pp. 1742–1751.<br />
[14] B. Sikdar, S. Kalyanaraman, and K. S. Vastola, “Analytic models for the<br />
latency and steady-state throughput of TCP Tahoe, Reno, and SACK,”<br />
IEEE/ACM Transactions on Networking, vol. 14, no. 6, pp. 959–971,<br />
December 2003.<br />
[15] E. A. Pekoz and N. Joglekar, “Poisson Traffic Flow in a General<br />
Feedback Queue,” Journal of Applied Probability, vol. 39, no. 3, pp.<br />
630–636, September 2002.<br />
[16] ITU-T Recommendation Y.1541, “Network performance objectives for<br />
IP-based services,” May 2002.<br />
[17] V. Paxson, “End-to-End Internet Packet Dynamics,” IEEE/ACM Transactions<br />
on Networking, vol. 7, no. 3, pp. 277–292, June 1999.<br />
[18] S. Shalunov, B. Teitelbaum, A. Karp, J. Boote, and M. Zekauskas,<br />
“A One-way Active Measurement Protocol (OWAMP),” RFC 4656,<br />
September 2006.<br />
Tf = 1<br />
S<br />
R<br />
P1<br />
NACK1<br />
a) Example, when r.v. Tf is equal 1<br />
S<br />
R<br />
P1<br />
Tf = 2<br />
Tf = 2<br />
NACK1<br />
b) Example, when r.v. Tf is equal 2<br />
S<br />
R<br />
P1<br />
Tf = 3<br />
Tf = 3<br />
NACK1<br />
S<br />
R<br />
P1<br />
S<br />
S<br />
S<br />
P1<br />
P1<br />
P1<br />
NACK1<br />
NACK1<br />
R<br />
R<br />
R<br />
APPENDIX A: TF R.V. DISTRIBUTION<br />
Random variable T f describes number of consecutive packets,<br />
which must be sent by the sender to receive information<br />
about lost packet.<br />
R.v. T f takes value 1, if the first packet sent after lost<br />
packet, reachs a receiver, and sender receives acknowledgement<br />
for that packet (which contains NACK for lost packet) -<br />
see Fig. 9.<br />
Probability, that T f is equal 1, is:<br />
Fig. 9.<br />
Tf = 3<br />
Tf = 3<br />
NACK1<br />
c) Example, when r.v. Tf is equal 3<br />
A T f random variable.<br />
NACK1<br />
P r{T f = 1} = (1 − p d )(1 − p a ) (25)<br />
Probability, that T f is equal 2, is (see Fig. 9):<br />
P r{T f = 2} = p d (1 − p d )(1 − p a ) + (26)<br />
+ (1 − p a )p a (1 − p d )(1 − p a )<br />
= (1 − p d )(1 − p a )[p d + (1 − p d )p a ]<br />
Probability, that T f is equal 3, is (see Fig. 9):<br />
P r{T f = 3} = p 2 d(1 − p d )(1 − p a ) + (27)<br />
+ p d (1 − p a )p a (1 − p d )(1 − p a )<br />
+ (1 − p d )p a p d (1 − p d )(1 − p a )<br />
+ (1 − p d )p a (1 − p d )p a (1 − p d )(1 − p a )<br />
= (1 − p d )(1 − p a )[p d + (1 − p d )p a ] 2<br />
Finally, random variable T f has the geometric distribution:<br />
P r{T f = j} = (1 − p d )(1 − p a )[p d + (1 − p d )p a ] (j−1) (28)<br />
where p d and p a denotes packet loss probability for direction<br />
sender-to-receiver and receiver-to-sender, respectively.<br />
Piotr Krawiec received M.Sc. and Ph.D. degrees in telecommunications from<br />
the Warsaw University of Technology, P oland, in 2005 and <strong>2011</strong>, respectively.<br />
Now he works as assistant at the Institute of Telecommunications, Warsaw<br />
University of Technology. His research interests focus on quality of service<br />
in IP networks, NGN architecture and new networks techniques.<br />
Andrzej Bęben received M.Sc. and Ph.D. degrees in telecommunications<br />
from Warsaw University of Technology (WUT), Poland, in 1998 and 2001,<br />
respectively. Since 2001 he has been assistant professor with the Institute<br />
of Telecommunications at Warsaw University of Technology, where he is<br />
a member of the Telecommunication Network Technologies research group.<br />
His research areas include IP networks (fixed and wireless), content aware<br />
networks, traffic engineering, simulation techniques, measurement methods,<br />
and testbeds.<br />
Jarosław Śliwiński was born in Toruń, Poland, in 1979. He received M.Sc.<br />
and Ph.D. degrees from Warsaw University of Technology in 2003 and 2008,<br />
respectively. His research interests cover traffic control, systems’ design and<br />
implementation methodology.
ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong> 79<br />
Decisive Factors for Quality of Experience of<br />
OpenID Authentication Using EAP<br />
Charlott Lorentzen, Markus Fiedler, and Peter Lorentzen<br />
Abstract—When using the web, long response times are bones<br />
of contention for users, i.e. they damp the Quality of Experience<br />
(QoE). Though, if one knows the cause of the long response<br />
time one may examine what could be done to eliminate the<br />
obstacle. In this paper, we determine the weak point of the<br />
Extensible Authentication Protocol Method for GSM Subscriber<br />
Identity Modules (EAP-SIM) with the OpenID service with<br />
regards to excessive authentication times, which determine the<br />
response times. In order to provoke controlled increases of the<br />
latter, we emulate bad network performance by introducing<br />
bi-directional delay between the supplicant (client) and the<br />
authentication server. The same procedure is also applied to<br />
several other EAP methods. Based on a recent, exponential<br />
relationship between QoE and response time, we then identify,<br />
quantify and compare the decisive factors for QoE reduction<br />
as functions of the components of the authentication times. The<br />
results we obtain clearly show that one task of the EAP-SIM<br />
authentication contributes significantly more to the total response<br />
times than the other tasks, which points out the direction for<br />
future optimisation of user perception of authentication times.<br />
I. INTRODUCTION<br />
OPENID [1] is a system that allows for automatic<br />
confirmation of a user’s identity when visiting<br />
other authentication-enabled sites or communities supporting<br />
OpenID. In other words, a user is always authenticated, but<br />
only the first authentication will have to be initiated by the<br />
user. OpenID is promising for uses in a seamless environment,<br />
where the goal is to remain seamlessly authenticated while<br />
switching network, device or even application.<br />
An authentication procedure typically produces a chain of<br />
messages before completing. And the more messages the chain<br />
consists of, the greater the risk of unacceptable response times<br />
(RT) becomes. These kinds of chains of messages, and/or<br />
chains of requests to different servers and databases form<br />
service chains (SC) [2] that can be quite large and complex.<br />
We previously discovered significant RTs for the authentication<br />
to the OpenID server when using networks with<br />
low bandwidth, such as mobile networks. Such waiting times<br />
challenge user patience [3] and increase the risk of users trying<br />
to bypass or turn off security features. Once the Quality of<br />
Experience (QoE) [4] is really bad, users might even abandon<br />
the service [5]. The concept of QoE refers to the totality of end<br />
user experience of the delivered service [6], and in this case<br />
of the RT of the service. Typically different parts, or steps,<br />
of a service contribute to a certain RT and there might be a<br />
specific part that contributes more than others to the total RT.<br />
Charlott Lorentzen, Markus Fiedler, and Peter Lorentzen are with School<br />
of Computing, Blekinge Institute of Technology, Karlskrona, Sweden, Email:<br />
@bth.se<br />
Given the background above, this paper will identify and<br />
quantify the decisive factors for QoE of Extensible Authentication<br />
Protocol Method for GSM Subscriber Identity Modules<br />
(EAP-SIM) with the OpenID authentication service as functions<br />
of network impairments in form of additional delay. The<br />
study will show what parts of the EAP-SIM authentication<br />
make the greatest contribution to RT, when authenticating via<br />
OpenID. Furthermore, the study will find the decisive factors<br />
of the following authentication methods: EAP Message-Digest<br />
algorithm 5 Challenge (EAP-MD5), EAP Tunneled Transport<br />
Layer Security (EAP-TTLS) with MD5, EAP-TTLS with<br />
Password Authentication Protocol (PAP), EAP-TTLS with<br />
Challenge Handshake Authentication Protocol (CHAP), EAP-<br />
TTLS with Microsoft CHAP version 2 (MSCHAPv2), and<br />
Protected EAP (PEAP) with MSCHAPv2. The decisive factors<br />
for each authentication method will then be compared.<br />
There are several studies dealing with the evaluation and<br />
optimisation of different EAP authentication methods in different<br />
network environments and scenarios, such as studies<br />
on performance evaluation of EAP for roaming [7] and handover<br />
[8] in WLAN environments. However, to the best of<br />
our knowledge, the combination of EAP-SIM authentication<br />
method with OpenID for web authentication has not been yet<br />
investigated, nor have there been studies done on decisive<br />
factors of different EAP methods.<br />
The organization of this paper is as follows: Section II<br />
gives an overview of the authentication method and services<br />
used for the OpenID EAP-SIM authentication. Section III<br />
discusses the impact of different network parameters and<br />
describes the methodology of the study, and Section IV<br />
describes the OpenID EAP-SIM authentication experiments<br />
with corresponding setup, procedure and RT measurements.<br />
The results of the latter study and the following analysis<br />
of the results are presented in Section V, followed by a<br />
discussion of the results in aspects of QoE in Section VI.<br />
Section VII describes an additional study of decisive factors<br />
for several authentication protocols, and presents and discusses<br />
the results. Finally, Section VIII provides a conclusion of the<br />
paper and points out future work.<br />
A. OpenID<br />
II. TECHNICAL BACKGROUND<br />
OpenID is a service that handles one’s authentications. If<br />
users are logged in to their OpenID server, they are automatically<br />
logged in at visited web pages that have previously been<br />
enabled with OpenID for their particular user account.<br />
An OpenID identity is a unique URL which contains the<br />
trusted provider and the username. The provider is the host
80 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
User<br />
PC<br />
Internet<br />
Service provider<br />
AAA/<br />
MAP GW<br />
SIM<br />
Supplicant<br />
Authenticator<br />
Fig. 1.<br />
Setup of the Ubisafe OpenID authentication, with BTH client side.<br />
of the URL, in our case Ubisafe AS, in Norway, and the<br />
username/URL will be openid.ubisafe.no/, but<br />
the provider can also be for example Yahoo, or any other<br />
site that provides OpenID as authentication service. With<br />
OpenID users need one password or authentication credential<br />
and username to be able to authenticate oneself to all enabled<br />
sites, and the password or authentication credential only needs<br />
to be used when logging in at the OpenID server.<br />
When a user account have OpenID enabled, for example on<br />
a community, the user logs in at the OpenID server and then<br />
it is possible to visit the community and provide the OpenID<br />
username/URL, and the authentication to the community will<br />
be completed without an additional password. This applies to<br />
all OpenID-enabled pages during one web browsing session.<br />
OpenID was chosen according to the requirements for<br />
seamless network and service access, to be provided by the<br />
IMS platform of the Mobicome project [9], [10].<br />
B. EAP-SIM Authentication<br />
In the EAP-SIM authentication method the actors are the<br />
SIM, the supplicant, the authenticator and the authentication<br />
server. The authenticator and the authentication server can be<br />
the same actor or situated in the same physical device (see<br />
Fig. 1). The supplicant is the user client, and the SIM-card is<br />
connected to the supplicant. The user just plugs in the SIMcard<br />
or makes sure the supplicant has access to it, and the<br />
supplicant is then the entity that communicates with the SIMcard.<br />
The authentication procedure is as follows.<br />
• A connection (physical and virtual) is established between<br />
the supplicant and authenticator.<br />
• The authenticator requests the identity (ID) from the<br />
supplicant (EAP-Request/Identity).<br />
• The supplicant produces a response and sends its ID to<br />
the authenticator (EAP-Response/Identity).<br />
• The authenticator challenges the supplicant in one or<br />
several steps to verify the ID of the supplicant (EAP-<br />
Request/SIM/Start and EAP-Request/SIM/Challenge).<br />
• The supplicant handles and responds to the challenge(s)<br />
from the authenticator (EAP-Response/SIM/Start and<br />
EAP-Response/SIM/Challenge).<br />
• If everything is correct with the challenge response the<br />
supplicant is authenticated and a success notification is<br />
sent from the authenticator (EAP-Success).<br />
C. Authentication service chain<br />
When a user requests to login to the Ubisafe OpenID server<br />
on the web page [11], a chain of messages to supply the<br />
Fig. 2.<br />
SC for EAP-SIM authentication with OpenID.<br />
service, i.e. a SC, is started. The SC for this authentication<br />
method is visualised in Fig. 2. In the sequel, let T Ik denote<br />
internal durations within the supplicant, while T Nk refer to<br />
durations involving network communications. The user request<br />
enters the supplicant which starts the setup of a connection to<br />
the authenticator, after making sure the SIM-card authentication<br />
is a valid method for both supplicant and server (duration<br />
T I1 ). Once the connection is established, the authenticator<br />
sends back a request (end of time T N1 ) for the ID of the<br />
supplicant, as described in the list in the previous section<br />
(Sect. II-B). Please, note that T N1 includes the time for setting<br />
up a connection between the authenticator and the supplicant,<br />
before sending the ID response to the supplicant. The setup<br />
of a connection is made visible with the dotted line in Fig. 2<br />
and the vertical dots below it, since these messages are not<br />
EAP-SIM messages.<br />
When the supplicant receives requests from the authenticator,<br />
the SIM-card is needed to produce a response to the<br />
requests (time T I2 ) since the SIM-card has the ID, and the<br />
keys, before sending the response to the authenticator. The<br />
SIM-card is also needed to produce a response (time T I3 ) for<br />
the SIM-challenge.<br />
Two further durations shown in Fig. 2, namely ɛ (between<br />
user click and initiation of the authentication) and δ (between<br />
completion of the authentication and displaying the results to
LORENTZEN et al.: DECISIVE FACTORS FOR QUALITY OF EXPERIENCE OF OPENID AUTHENTICATION USING EAP 81<br />
the user) have shown to be of minor importance in the context<br />
of this study. We will therefore assume that the RT is well<br />
approximated by the authentication time, i.e. the sum of its<br />
internal and network components.<br />
III. NETWORK IMPACT<br />
In the course of our work, we seek to determine the RTs<br />
for the different part of the authentication method in question.<br />
The user perception is of course based on a whole RT, but if<br />
the greatest contribution to the RT was found, then it might<br />
also be minimised or made more scalable for large network<br />
delays with the goal to preserve a good QoE.<br />
Timestamps were recorded for each task within the authentication,<br />
and RTs were calculated from the start timestamp and<br />
end timestamp for each task.<br />
The objective of calculating RTs is to see whether there<br />
are any parts of the authentication that contribute most to<br />
the total RT or whether some parts are particularly sensitive<br />
to degradations in network performance. For this reason, we<br />
measure and compare the parts of the RT, in particular to see<br />
differences between RT for different parts of the authentication<br />
process, as well as their impact.<br />
On network level, in most cases, adding delay or imposing<br />
bandwidth constraints gives a similar effect, namely higher RT<br />
values. In this study, bad network performance is emulated<br />
with a traffic shaper situated in the supplicant that shapes bidirectionally<br />
on the network interface of the latter.<br />
Bandwidth can in some shapers be difficult to use for<br />
provoking the results we are looking for. For a single packet<br />
message the bandwidth constraint might never give any effect<br />
as the shaper might use a previous packet arrival to determine<br />
the next one. The latter was visible in the trials with bandwidth<br />
that were done early on in this study, where the RT for<br />
some parts did not change when decreasing the bandwidth<br />
and finally a timeout was received without any change in the<br />
RT for those parts.<br />
Loss will result in higher RTs because of necessary retransmissions,<br />
but if one packet is lost in each part of the<br />
authentication, it will have the same impact on the RT for each<br />
part. To compare the effect of loss for each part would be quite<br />
difficult because of the encrypted traffic for the authentication.<br />
Therefore, loss has not yet been used as network performance<br />
degradation parameter in this study.<br />
Delay is added to every packet that passes the traffic shaper,<br />
and the delay can be constant or variable. Variable delay has<br />
been tested in a previous project for map services [2]. Even<br />
though a constant delay might not be the most realistic case<br />
for emulating delay, it has shown a crucial enabler for the<br />
quantitative results presented in this study.<br />
IV. EXPERIMENTS<br />
The experiment setup consisted of a client computer with a<br />
SIM dongle, a traffic shaper for adding delay on the network<br />
interface of the client, and a server situated in Oslo, Norway, as<br />
shown in Fig. 1. All trials on the client computer were carried<br />
out on campus, during the same period of the day to withhold<br />
consistency, namely during evenings when most personnel<br />
were not at work. The delays that were added by the shaper<br />
were 0 ms, 250 ms, 500 ms, 750 ms, and 1 s in both directions.<br />
The timestamps were recorded via a JavaScript. Even though<br />
JavaScript logging of timestamps have proven to be only fairly<br />
accurate [12], the accuracy is sufficient for this experiment.<br />
Although the shaper adds constant delays on network level,<br />
the corresponding RT values are varying slightly, as there are<br />
many random impacts affecting the way between user and<br />
authentication service. Nevertheless, the chosen delays allowed<br />
to change the order of magnitudes of the RT such that trends<br />
regarding QoE could be clearly seen [3].<br />
The experiment considered the login procedure on the<br />
Ubisafe OpenID server web page. On the web page “USB-<br />
SIM Dongle” was chosen in the Java applet handling the login.<br />
After clicking “Login”, the Java applet logged timestamps for<br />
starting and ending all parts of the EAP-SIM authentication<br />
(see Fig. 2). When the login was completed and the new page<br />
was loaded, a logout was done and then the procedure was<br />
repeated.<br />
For each delay the experiment was done 45 times, and the<br />
log file was saved for later analysis. Although caching was<br />
disabled and cookies were not saved, the first five trials for<br />
each delay were discarded in order to avoid any potential bias<br />
of the measurements. The results were averaged, and 95 %<br />
confidence intervals were calculated; the latter have however<br />
shown to be too small to be visible in the plots of Figure 3.<br />
V. RESULTS AND ANALYSIS<br />
The authentication procedure consists of 16 steps, from initiation<br />
to success, including both internal processing time (T I )<br />
and communication time spent in the network (T N ). Though, it<br />
can be abstracted down to seven steps, of which three (indexed<br />
by I1 to I3) are internal durations and four (indexed by N1 to<br />
N4) are external communication, i.e. network communication<br />
outside the supplicant (cf. Fig. 2). These steps are formalised<br />
as<br />
T R − δ − ɛ =<br />
4∑<br />
T Nk +<br />
k=1<br />
3∑<br />
T Ik = T N + T I , (1)<br />
k=1<br />
where T R is the total RT, T N is the total time for the<br />
network communication steps, and T I is the total time for the<br />
internal processing steps, including communication between<br />
the supplicant and the SIM-card. As indicated before, we<br />
assume δ → 0 and ɛ → 0.<br />
When comparing T N and T I , the main contributions to<br />
the RT change with the increase of the RT. In case of no<br />
or low delays, RT is dominated by the processing time, T I .<br />
For high delays, the RT is instead dominated by the network<br />
communication time, T N . For a delay of 1 s the RT of about<br />
24 s consist of more than 90 % of network communication<br />
time, while for a transparent shaper, the relation is almost the<br />
opposite, as the processing time takes up about 70 % of the<br />
total RT.<br />
The steps including network communication are affected by<br />
the delay, whilst the internal communication, e.g. communication<br />
between supplicant and SIM-dongle, is not affected. The<br />
parts of the RT that are interesting in the results are therefore
16505<br />
16509<br />
16513<br />
16517<br />
16521<br />
16525<br />
16529<br />
16533<br />
16537<br />
16541<br />
16545<br />
16549<br />
16553<br />
16557<br />
16561<br />
16565<br />
16569<br />
16573<br />
16577<br />
16581<br />
16585<br />
More<br />
Response time [s]<br />
Frequency<br />
82 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
25<br />
20<br />
8<br />
7<br />
6<br />
15<br />
10<br />
5<br />
0<br />
0 200 400 600 800 1000<br />
Delay added in shaper [ms]<br />
TN1<br />
TN2<br />
TN3<br />
TN4<br />
TR<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
Response time [ms]<br />
Fig. 3.<br />
RT components versus bi-directional delay added in the shaper.<br />
Fig. 4.<br />
Frequency of T N1 , for a bi-directional delay d =1 s, bin width 2 ms.<br />
T N1 , T N2 , T N3 and T N4 , whereas the rest of the tasks or steps<br />
include only internal processing times that will not change<br />
with regard to increasing delay.<br />
T N1 provides the largest contribution to the total RT when<br />
there is no delay added, as can be seen in Fig. 3. It can<br />
also be seen that T N1 is most sensitive to additional delay of<br />
the four network communication steps. When the additional<br />
delay is 1 s, both ways, T N1 is already about 16 s long.<br />
For the remaining three tasks the RT grows equally fast, but<br />
substantially slower than for task N1, and they do start at a<br />
lower RT values in the case of no additional delay introduced<br />
by the shaper. Comparing the four tasks behind T N , it is also<br />
for T N1 that the most roundtrips in communication can be<br />
seen, due to the setup of a connection.<br />
For T N1 the linear growth in T R , with respect to changes<br />
of the network delay added in the shaper d, is eight times as<br />
large as compared to T N2 , T N3 and T N4 :<br />
while<br />
T N1 ≈ 16d = 8 × 2d (2)<br />
T N2,N3,N4 ≈ 2d (3)<br />
Thus, for the total network time, we get the approximation<br />
T N ≈ 16d + (3 × 2d) = 22d. (4)<br />
The fact that the tasks with one round trip get a RT of<br />
double the delay (cf. Equation 3), as the delay is added in<br />
both directions, might indicate a relation between the number<br />
of packets sent back and forth and the factor of growth. If<br />
two messages, counting both ways, get two times the delay,<br />
then 16 times the delay should indicate 16 messages when<br />
counting both ways, and thus eight round trips. Eight is also<br />
the factor between T N1 (Equation 2) and the other components<br />
(Equation 3).<br />
When looking at the distribution of the RT values, they are<br />
a bit different from trial to trial, and from delay to delay.<br />
Most of the distributions are similar to a normal distribution,<br />
but in some cases with a (rather short) tail on the right-hand<br />
side. In some cases there are a few values that are bigger than<br />
the average, the median and the 90 % percentile. Such values<br />
belong to the so-called tail and can be quite bad when it comes<br />
to (perceived) network performance. However, for a RT in the<br />
order of magnitude of around 16 s, a parts of a second is not<br />
that large of a difference. In Fig. 4 the tail value is about 70 ms<br />
from the center of the distribution and the RTs are all larger<br />
than 16 s. Short tails, in this order of magnitude, do not have<br />
to be considered.<br />
VI. QOE ASPECTS FOR EAP-SIM<br />
Considering the user perception of the RT of the authentication,<br />
or QoE, and the changes of the latter with regard to the<br />
changes in RT, a previously researched user model [3] is used.<br />
Equation 5 was developed in the previous study for the same<br />
system and in the same environment and represents basically<br />
a Mean Opinion Score (MOS) [13], which enables us to use<br />
the equation in this study in a straightforward manner:<br />
QoE ≈ 4.7e −0.1TR/s . (5)<br />
Obviously, each additional second factor of the network part<br />
of the RT yields a relative damping of the QoE by factor 0.9.<br />
From the measurements, it has been observed that processing<br />
time is approximately constant, which can be formulated as<br />
3∑<br />
T Ik ≈ 1.8 s, (6)<br />
k=1<br />
Thus, equation 5 can be rewritten as<br />
which yields<br />
QoE ≈ 4.7e −0.18 e −0.1TN/s<br />
≈ 3.9e −0.1TN/s (7)<br />
QoE ≈ 3.9e −2.2d/s . (8)<br />
Obviously, for the fixed line connection used in this experiment,<br />
it can be seen in Equation 8 that the QoE cannot exceed<br />
3.9.<br />
In Figure 5 one may see that, because of the exponential<br />
slope, already at 150∼200 ms of delay, the MOS has gone<br />
below the rating “Fair”, and at 250 ms of delay the MOS is<br />
closing in on the rating “Poor”. The QoE reaches the MOS<br />
value 1, or “Bad”, at about 650 ms of added delay. Since the
Response Time [s]<br />
MOS<br />
LORENTZEN et al.: DECISIVE FACTORS FOR QUALITY OF EXPERIENCE OF OPENID AUTHENTICATION USING EAP 83<br />
5<br />
4<br />
3<br />
User<br />
PC<br />
Supplicant<br />
Switch<br />
Authentication<br />
Server<br />
Server<br />
Authentication<br />
Server<br />
2<br />
1<br />
0<br />
0 250 500 750 1000<br />
Delay added in shaper [ms]<br />
Fig. 6. Setup of the EAP authentication experiments.<br />
1,2<br />
Tn2<br />
Tn4<br />
1<br />
Tn1<br />
0,8<br />
Fig. 5. QoE in terms of MOS versus added delays, using Equation 8.<br />
MOS scale goes from 1 to 5 when users are rating, values<br />
below 1 can be transformed to 1 as shown in Equation 2 in<br />
[14].<br />
Dividing Equation 8 into the parts that grow equally gives<br />
QoE ≈ 3.9e −1.6d/s (e −0.2d/s ) 3 , (9)<br />
where the first e-term shows the impact of T N1 and the second<br />
e-term shows the joint impact of the remaining times, namely<br />
T N2 , T N3 and T N4 on QoE. One may see that the first part<br />
has a significantly higher impact:<br />
γ = e−1.6d/s<br />
e −0.6d/s = e−d/s < 1. (10)<br />
Equation 10 describes the QoE damping factor γ between<br />
the impact of the connection setup time, T N1 , and the impact of<br />
the remaining network times, T N2 , T N3 and T N4 , as function<br />
of the one-way delay d introduced by the shaper. It can be<br />
seen that, as d is growing, the damping impact of the connection<br />
setup supersedes the one of all the remaining network<br />
communication times. Even for low delays d, the connection<br />
setup time has a greater impact than all the remaining times,<br />
though in a lower order of magnitude.<br />
Assume that one could reduce the number of messages<br />
during the connection setup by 50 %, and thereby also reduce<br />
the connection setup time T N1 , one would observe a much<br />
less critical impact of the delay on the QoE, namely a factor<br />
of e −0.2d/s .<br />
As far as we can tell from this study, it is the setup of a<br />
connection and secure tunnel between the supplicant and the<br />
authenticator in OpenID with EAP-SIM that does not scale<br />
nicely with an increasing delay.<br />
VII. PERFORMANCE OF OTHER EAP METHODS<br />
In this study the authentication methods EAP-MD5, PEAP-<br />
MSCHAPv2, EAP-TTLS-PAP, EAP-TTLS-MD5, EAP-TTLS-<br />
CHAP, and EAP-TTLS-MSCHAPv2 were studied in a laboratory<br />
environment.<br />
The experiments were done in a similar, but laboratory,<br />
environment, which excluded the Internet but included all<br />
other parties as in the previous study (see Fig 6). Though, the<br />
0,6<br />
0,4<br />
0,2<br />
0<br />
0 20 40 60 80 100<br />
Delay added in shaper [ms]<br />
Fig. 7. RT components for EAP-MD5 versus bi-directional delay added in<br />
the shaper.<br />
authenticator and the authentication server are two separate<br />
devices in this setup. The traffic shaper is placed between<br />
the authenticator and server, in this experiment setup. The<br />
new setup has the network traffic shaper on the server side<br />
of the authenticator, since the delay is introduced on the<br />
network layer (IP layer), and the traffic between supplicant<br />
and authenticator is not sent using IP addresses. This setup<br />
gives the same result in addition of network delay as in the<br />
previous study for all tasks, except for the first task including<br />
the two initial messages, which are only sent to and from the<br />
authenticator.<br />
Time stamps were logged during all experiments; both at the<br />
supplicant side and the server side, and RTs were calculated<br />
for each run in each experiment. The experiments were run 40<br />
times for each delay setting. The network delays added in the<br />
network shaper in this experiment were 0 ms, 20 ms, 40 ms,<br />
60 ms, 80 ms, and 100 ms, in both directions.<br />
The EAP methods were analyzed to find the decisive factors.<br />
The RTs in these additional experiments are considerably<br />
lower, but the linear behavior in RT increase with increased<br />
delay shows the same “per packet” proportionality as in the<br />
above study.<br />
A. Authentication method overview<br />
EAP is the underlying protocol for authentication procedure<br />
and TTLS is used to setup a tunnel for the authentication.<br />
The tunnel is established between the supplicant and the<br />
authenticator for secure data, and between supplicant and<br />
authentication server for secure password authentication.<br />
PAP, MD5, CHAP, and MSCHAPv2 are the authentication<br />
algorithms used in the authentication procedure. The security
Response Time [s]<br />
Response Time [s]<br />
Response Time [s]<br />
84 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
1,2<br />
1<br />
0,8<br />
Tn3<br />
Tn2<br />
Tn4<br />
Tn1<br />
1,2<br />
1<br />
0,8<br />
Tn4<br />
Tn3<br />
Tn2<br />
Tn1<br />
0,6<br />
0,6<br />
0,4<br />
0,4<br />
0,2<br />
0,2<br />
0<br />
0 20 40 60 80 100<br />
Delay added in shaper [ms]<br />
0<br />
0 20 40 60 80 100<br />
Delay added in shaper [ms]<br />
Fig. 8. RT components for EAP-TTLS-PAP/EAP-TTLS-CHAP versus bidirectional<br />
delay added in the shaper.<br />
Fig. 10. RT components for PEAP-MSCHAPv2 versus bi-directional delay<br />
added in the shaper.<br />
1,2<br />
1<br />
0,8<br />
0,6<br />
0,4<br />
0,2<br />
0<br />
Tn3<br />
Tn4<br />
Tn2<br />
Tn1<br />
0 20 40 60 80 100<br />
Delay added in shaper [ms]<br />
Fig. 9. RT components for EAP-TTLS-MD5/EAP-TTLS-MSCHAPv2 versus<br />
bi-directional delay added in the shaper<br />
level is a bit different for these algorithms. In PAP the user<br />
name and password are sent unencrypted. Without a tunnel<br />
PAP would, in other words, be insecure and therefore it is<br />
not supported by EAP itself, but by TTLS. In TTLS the<br />
tunnel hides the password, so sending it unencrypted is not<br />
a security issue. Though, the password is stored as a hash at<br />
the authentication server side and can therefore not be stolen.<br />
CHAP was standardised for PPP from the beginning and<br />
is only supported by TTLS, since it is not an EAP method.<br />
The authentication server challenges the supplicant, which<br />
responds to the challenge by proving the possession of a shared<br />
secret, i.e. a password. The password is sent as a hash over<br />
the network, but must be available in plain text at both the<br />
supplicant side (by typing it) and the authentication server<br />
side (in a stored file, e.g. /etc/passwd) in order to confirm that<br />
the hash in the challenge response is valid.<br />
EAP-MD5 was standardized with EAP from the beginning.<br />
MD5 is supported by EAP and can be used with EAP, PEAP<br />
or TTLS. Though, the password needs to be available on<br />
both supplicant side and authentication server side, as in<br />
CHAP. When MD5 is used as EAP-MD5, the ID is sent from<br />
the beginning, but when it is used in EAP-TTLS-MD5, the<br />
supplicant is anonymous in the beginning and the ID is instead<br />
sent in the authentication phase, which is reflected in the<br />
response times for those phases, as there is one more roundtrip<br />
from supplicant to server for this task.<br />
MSCHAPv2, like MD5 and CHAP, sends a hash of the<br />
password, but in this case it does not have to be stored in<br />
plain text on both sides. A particular one-way hash of the<br />
password, which will then serve as the password, is stored at<br />
the authentication server side. The supplicant, who knows the<br />
hash algorithm, will be able to produce a matching “password”<br />
from the original password, which can be used in the response<br />
of a challenge from the authentication server. MSCHAPv2 also<br />
provides mutual authentication. MSCHAPv2 can be used with<br />
EAP, PEAP and TTLS.<br />
B. Results and Analysis<br />
For each method the messages have been separated into<br />
four tasks, and what is included in the RT for each task is are<br />
presented in a list below. The RTs are all have the supplicant as<br />
base. All methods include all the four tasks, except for EAP-<br />
MD5 which does not setup a tunnel for the communication,<br />
and thus lacks the response time increase from T n3 .<br />
• T n1 : Initiation of the connection with the authenticator.<br />
• T n2 : Negotiation of authentication protocol with the authentication<br />
server.<br />
• T n3 : Setup of a secure tunnel to the authenticator and<br />
authentication server.<br />
• T n4 : Authentication with challenge(s) and response(s).<br />
One of the authentication methods, namely EAP-MD5, does<br />
not have a setup of a tunnel. In this method the authentication<br />
challenge has the longest RT, T n4 (see Fig. 7). The increase in<br />
RT of the decisive task, T n4 , is four times the added delay, and<br />
since the delay is added in both directions four times means<br />
two round trips between the supplicant and the authentication<br />
server.<br />
For EAP-TTLS-PAP and EAP-TTLS-CHAP the task with<br />
the RT of largest impact is the setup of the TTLS tunnel (see<br />
Fig. 8). The RT of this task, T n3 , increases with six times the<br />
added delay. The increase in RT for T n3 is just two times the
Complexity: Response time [s]<br />
Response Time [s]<br />
LORENTZEN et al.: DECISIVE FACTORS FOR QUALITY OF EXPERIENCE OF OPENID AUTHENTICATION USING EAP 85<br />
1,2<br />
1<br />
0,8<br />
0,6<br />
0,4<br />
0,2<br />
PEAP-MSCHAPv2 : Tn4<br />
EAP-TTLS-PAP/CHAP : Tn3<br />
EAP-TTLS-MD5/MSCHAPv2 : Tn3<br />
EAP-MD5 : Tn2<br />
TABLE I<br />
DECISIVE FACTORS AND SECURITY LEVEL FOR EACH METHOD.<br />
EAP-MD5 e −0.4d/s 1<br />
EAP-TTLS-CHAP e −0.6d/s 2<br />
EAP-TTLS-PAP e −0.6d/s 3<br />
EAP-TTLS-MD5 e −0.6d/s 4<br />
EAP-TTLS-MSCHAPv2 e −0.6d/s 5<br />
PEAP-MSCHAPv2 e −1.0d/s 6<br />
EAP-TTLS-CHAP e −1.6d/s -<br />
Fig. 11.<br />
0<br />
0 20 40 60 80 100<br />
Delay added in shaper [ms]<br />
The decisive factors for each authentication method.<br />
The decisive factors for each of these methods are derived<br />
from the RT of the task with the largest impact on the total<br />
RT. The T n for each method is translated into a dependence<br />
on delay based in Equation 5. The decisive factors for each<br />
method are presented in Table I.<br />
2,5<br />
2<br />
1,5<br />
1<br />
0,5<br />
0<br />
1 2 3 4 5 6<br />
Security level<br />
Fig. 12. Security vs. Complexity for several authentication methods, with<br />
100 ms added delay.<br />
added delay larger than for the task with the next largest RT,<br />
T n2 , which is the time for the negotiation of authentication<br />
protocol. The increase in response time for the authentication<br />
challenge is only two times the added delay.<br />
EAP-TTLS-MD5 and EAP-TTLS-MSCHAPv2 also have<br />
the largest impact on the total RT from the setup of the TTLS<br />
tunnel, with the same increase in RT. The RT of this task, T n3 ,<br />
increases with six times the added delay (see Fig. 9), whereas<br />
the tasks that has the RTs with the next largest impact on the<br />
total RT, namely T n2 and T n4 , are four times the added delay.<br />
The EAP method that has the task which distinguishes itself<br />
the most, in terms of largest impact on the total RT, is PEAP-<br />
MSCHAPv2. Though the next largest RT, T n3 , increases with<br />
six times the added delay, the RT of T n4 increases with ten<br />
times the added delay (see Fig. 10).<br />
Fig. 11 shows the RTs with the largest impact for each<br />
of the EAP methods in this additional study. As seen in the<br />
figure, PEAP-MSCHAPv2 is the method that has the highest<br />
RT with T n4 . EAP-MD5 has the lowest RT with T n2 , which in<br />
fact is lower than the next largest RT for PEAP-MSCHAPv2,<br />
and equal to the next largest RT for EAP-TTLS-MD5, and<br />
EAP-TTLS-MSCHAPv2.<br />
C. Simplicity versus Security<br />
In the previous study [15] the compromising relationship<br />
between simplicity and security is discussed. If one party<br />
is increased the other party will sooner or later suffer from<br />
degradation. In this section we look at the relationship between<br />
security and simplicity with regard to the six EAP methods,<br />
described and analysed above.<br />
The EAP methods have been arranged from highest security<br />
level (6) to lowest security level (1) in a simple manner (see<br />
Table I). For example, a plain text password is, within this<br />
comparison example, considered to provide a lower security<br />
level than a hashed password, and a tunnel is considered to<br />
add security. Response time will, in this example, be compliant<br />
with complexity, which is the inverse of simplicity. The higher<br />
the RT, the lower the simplicity. The RT order of magnitude<br />
is due to the number of messages sent back and forth for each<br />
task.<br />
In Fig. 12 it can be seen that the simplicity is increasing with<br />
the decrease in security. The EAP method that is considered to<br />
have the lowest security level, EAP-MD5, is in fact the simplest<br />
method, and the PEAP-MSCHAPv2, which is considered<br />
to have the highest security together with EAP-TTLS-MD5<br />
and EAP-TTLS-MSCHAPv2 has the least simplicity. This is of<br />
course a generalisation and the model is not exactly compliant<br />
with every existing authentication method. Though, it gives<br />
a rough picture of the options that exist when choosing an<br />
authentication method. If the situation or solution requires a<br />
high level of security, then it can be justified to add complexity<br />
to the system, even though it might give higher RTs.<br />
VIII. CONCLUSION AND FUTURE WORK<br />
This paper has described the study of finding the most<br />
vulnerable part of EAP-SIM authentication method using<br />
the OpenID authentication service. After initial tests of the<br />
methods and the network connection, the experiments were<br />
performed with constant delay added in a shaper on the<br />
network interface of the supplicant, or client machine. A<br />
constant delay, which was increased for each trial, was added<br />
to provoke a change in RT for the different parts of the<br />
authentication.
86 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, <strong>SEPTEMBER</strong> <strong>2011</strong><br />
When analysing the results for OpenID authentication with<br />
EAP-SIM, it could be clearly seen that one of the network<br />
times is growing faster than the others, namely the one that<br />
refers to the initiation and setup of a secure connection<br />
between the supplicant and the authenticator, followed by an<br />
ID request from the authenticator. The fact that this initial<br />
task has the largest contribution to the total RT and therefore<br />
also had the greatest impact on QoE was shown from several<br />
angles. The initial is also the task with most packets sent back<br />
and forth.<br />
The results for OpenID using EAP-SIM were then connected<br />
to the QoE user model that was developed in the<br />
previous study of the same system. From this mapping it<br />
was shown that the increase in RT for the initial secure<br />
connection setup task would result in a degradation of QoE to<br />
the extent that it reaches a rating of 1 (lowest grade, meaning<br />
“bad”) at about 650 ms of added delay in both directions. To<br />
achieve a factor of the impact on the RT that is more tolerable<br />
and scalable with large delays, one would need to reduce<br />
the number of messages during the connection setup. The<br />
reduction of the latter by factor two would entail a significant<br />
gain in scalability, seen from a smaller damping factor.<br />
The other EAP based authentication methods were found to<br />
have the lower decisive factors than the OpenID solution with<br />
EAP-SIM. The largest contribution to the total RT for PEAP-<br />
MSCHAPv2 is given by T n4 , which is the authentication phase<br />
including the challenge. The method that had the decisive<br />
factor with the lowest impact was EAP-MD5. In EAP-MD5<br />
the initiation of a connection between the supplicant and the<br />
server gave the largest impact on RT, with T n2 . The rest of<br />
the authentication methods had almost equal decisive factors,<br />
with the largest impact in RT from T n3 , i.e. the setup of the<br />
secure tunnel.<br />
After comparing the simplicity and security levels of the<br />
EAP based authentication methods, the compromise between<br />
security and simplicity was shown in a quantitative manner.<br />
Adding a security level will compromise the simplicity in<br />
most cases, depending on how the increase in security level is<br />
defined.<br />
The accurate impact of variable delay, and also loss, could<br />
be evaluated in the future. Since bandwidth has already been<br />
tried without giving a realistic result, a suitable traffic shaper<br />
has to be found before it can be further evaluated. Variable<br />
delay will perhaps result in a bit lower RT than constant<br />
delay, but that needs to be proved, or counter-proved. Also,<br />
loss would be interesting to evaluate as network performance<br />
parameter.<br />
The setup of a connection between the supplicant and<br />
the authenticator for OpenID using EAP-SIM needs to be<br />
examined closer, to see if there is a possibility to optimize it.<br />
Since the supplicant and the authenticator shares information<br />
from the SIM-card, there might be other possibilities to setup a<br />
connection. Then it might also be a good option if the OpenID<br />
authentication service is provided by the operator issuing the<br />
SIM-card. These possibilities will be closer examined with<br />
regards to trustworthiness and functionality in future work.<br />
ACKNOWLEDGEMENTS<br />
The authors would like to thank Dr. Ivar Jørstad at Ubisafe<br />
AS in Olso, Norway, for providing and supporting the authentication<br />
server setup, and Johan Lindh for his assistance in<br />
gathering parts of the underlying data.<br />
REFERENCES<br />
[1] “The OpenID Foundation website,” http://openid.net, [online], cited<br />
<strong>2011</strong>-01-15.<br />
[2] M. Fiedler, C. Eliasson, S. Chevul, and S. Eriksén, “Quality of Experience<br />
and Quality of Service in a Service Supply Chain,” in EuroFGI<br />
IA.7.6 Workshop on Socio-Economic Issues of Future Generation Internet,<br />
Santander, Spain, Jun. 2007.<br />
[3] C. Lorentzen, M. Fiedler, H. Johnson, J. Shaikh, and I. J. rstad, “On User<br />
Perception of Web Login - A Study on QoE in the Context of Security,”<br />
in Proc. of Australasian Telecommunication Networks and Applications<br />
Conference (ATNAC 2010), Auckland, New Zealand, Nov. 2010.<br />
[4] Vocabulary for performance and quality of service. Amendment 2:<br />
New definitions for inclusion in Recommendation P.10/G.100, ITU-T<br />
Recommendation P.10/G.100 (2006)/Amendment 2 (07/2008).<br />
[5] M. Fiedler, T. H. feld, and P. Tran-Gia, “A Generic Quantitative<br />
Relationship between Quality of Experience and Quality of Service,”<br />
IEEE NETWORK, Special Issue on Improving QoE for Network Service,<br />
vol. 24, no. 2, pp. 36–41, Mar. 2010.<br />
[6] P. Reichl, “From Quality-of-Service and Quality-of-Design to Qualityof-Experience:<br />
A Holistic View on Future Interactive Telecommunication<br />
Services,” in Proc. of Software Telecommunications and Computer<br />
Networks, Split, Croatia, Sep. 2007.<br />
[7] J. Cordasco, S. Wetzel, and U. Meyer, “Implementation and Performance<br />
Evaluation of EAP-TLS-KS,” in Proc. of Security and Privacy<br />
in Communication Networks (SecureComm’06), Baltimore, MD, USA,<br />
Aug. 2006.<br />
[8] S. F. Hasan, N. H. Siddique, and S. Chakraborty, “On Evaluating the<br />
Latency in Handing Over to EAP-enabled WLAN APs from Outdoors,”<br />
in Prof. of IEEE, IET International Symposium on Communication Systems,<br />
Networks and Digital Signal Processing (CSNDSP’10), Newcastle,<br />
UK, Jul. 2010, pp. 278–282.<br />
[9] “The Mobicome website,” http://www.mobicome.org [online], [Cited<br />
<strong>2011</strong>-01-15.].<br />
[10] I. J. rstad, D. V. Thuan, T. J. nvik, and D. V. Thanh, “Utilising Emerging<br />
Identity Management Frameworks in IMS,” in Proc. of the 12th International<br />
Conference on Intelligence in service delivery Networks (ICIN),<br />
Bourdeaux, France, Oct. 2008.<br />
[11] “The Ubisafe AS OpenID server website,” https://openid.ubisafe.no/<br />
[online], [Cited <strong>2011</strong>-01-15.].<br />
[12] K. Wac, M. Fiedler, R. Bults, and H. Hermens, “Estimations of Additional<br />
Delays for Mobile Application Data from Comparative Output-<br />
Input Throughput Analysis,” in Proc. of NOMS 2010, Osaka, Japan, Apr.<br />
2010.<br />
[13] Methods for subjective determination of transmission quality, ITU-T<br />
Recommendation P.800, 1996.<br />
[14] M. Fiedler and T. Hossfeld, “Quality of Experience-related differential<br />
equations and provisioning-delivery hysteresis,” in Proc. of 21st Specialist<br />
Seminar on Multimedia Applications & Traffic, Performance and<br />
QoE, Miyazaki, Japan, Mar. 2010.<br />
[15] C. Eliasson, M. Fiedler, and I. J. rstad, “A Criteria-Based Evaluation<br />
Framework for Authentication Schemes in IMS,” in Proc. of the 4th International<br />
Conference on Availability, Reliability and Security (AReS),<br />
Fukuoka, Japan, Mar. 2009, pp. 865–869.<br />
Charlott Lorentzen received her master degree in electrical engineering from<br />
Blekinge Institute of Technology, Karlskrona, Sweden, in 2006. After working<br />
in some research projects she proceeded with doctoral studies in telecommunication,<br />
with emphasis on network security and network performance, at<br />
Blekinge Institute of Technology. She just received her licentiate degree, in<br />
May <strong>2011</strong>, with the licentiate dissertation “User Perception and Performance<br />
of Authentication Procedures”.
LORENTZEN et al.: DECISIVE FACTORS FOR QUALITY OF EXPERIENCE OF OPENID AUTHENTICATION USING EAP 87<br />
Markus Fiedler received his doctoral degree in electrical engineering/ICT<br />
from Universitt des Saarlandes, Saarbrücken, Germany, in 1998. Since then,<br />
he has been with Blekinge Institute of Technology, Karlskrona, Sweden. He<br />
has been holding the Docent degree in telecommunication systems since 2006.<br />
Being an Associate Professor within the School of Computing (COM) at<br />
BTH, he performs and supervises research on Quality of Experience; seamless<br />
communications; network virtualization; service chains; and networks of the<br />
future (NF). He is leading and participating in several national and European<br />
projects. He is serving on the Steering Board of the European Network of<br />
Excellence Euro-NF, and is coordinating Euro-NF’s research activities.<br />
Peter Lorentzen has recently finished his master thesis work, “Evaluation<br />
of EAP-methods”. He will get his master degree in software engineering,<br />
with emphasis on telecommunication, from Blekinge Institute of Technology,<br />
Karlskrona, Sweden, in the second half of <strong>2011</strong>.
88 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, SEPTEMPBER <strong>2011</strong><br />
Agent based VoIP Application with Reputation<br />
Mechanisms<br />
Grzegorz Oryńczak and Zbigniew Kotulski<br />
Abstract—In this paper we introduce our new VoIP model the<br />
aim of which is to meet the challenges of modern telephony.<br />
We present project concepts, details of implementation and our<br />
testing environment which was designed for testing many aspects<br />
of VoIP based systems. Our system combines mechanisms for<br />
ensuring best possible connection quality (QoS), load balance<br />
of servers in infrastructure, providing security mechanisms and<br />
giving control over the packet routing decisions. The system is<br />
based on Peer-to-Peer (P2P) model and data between users are<br />
routed over an overlay network, consisting of all participating<br />
peers as network nodes. In the logging process, each user is<br />
assigned to a specific node (based on his/her geographic location<br />
and nodes load). Every node also has a built-in mechanism<br />
allowing to mediate between the user and the main server (e.g.<br />
in logging process). Besides that, because nodes are participating<br />
in data transmission, we have control over the data flow route.<br />
It is possible to specify the desired route, so, regardless of the<br />
external routing protocol, we can avoid paths that are susceptible<br />
to eavesdropping. Another feature of presented system is usage<br />
of agents. Each agent acts with a single node. Its main task<br />
is to constantly control the quality of transmission. It analyzes<br />
such parameters like link bandwidth use, number of lost packets,<br />
time interval between each packets etc. The information collected<br />
by the agents from all nodes allows to built a dynamic routing<br />
table. Every node uses Dijkstra’s algorithm to find the best at<br />
the moment route to all other nodes. The routes are constantly<br />
modified as a consequence of changes found by agents or<br />
updates sent by other nodes. To ensure greater security and<br />
high reliability of the system, we have provided a reputation<br />
mechanism. It is used during updating of the information about<br />
possible routes and their quality, given by other nodes. Owing to<br />
this solution nodes and routes which are more reliable get higher<br />
priority.<br />
Index Terms—voice over IP, IP telephony security, speech<br />
quality control, software agents<br />
I. INTRODUCTION<br />
VOICE sending is an incredibly useful and worth developing<br />
feature among many possibilities given by the<br />
Internet. One of the first attempts of creating a protocol for<br />
transferring human speech over computer network was the<br />
Network Voice Protocol [1] (NVP) made by Danny Cohen of<br />
the InformationSciences Institute fromUniversityof Southern<br />
California in 1973. NVP was used to send speech between<br />
distributed sites on the ARPANET. Since that time telephony<br />
based on Internet Protocols has become more and more popular.<br />
Nowadays, it becomes a serious competitor to standard<br />
Grzegorz Oryńczak is a PhD student at Jagellonian University, Department<br />
of Physics, Astronomy and Applied Computer Science, Cracow, Poland<br />
(corresponding author, email grzegorz.orynczak@uj.edu.pl)<br />
Zbigniew Kotulski a professor at Institute of Fundamental Technological<br />
Research of the Polish Academy of Sciences and professor at Department of<br />
Electronics and Information Technology of Warsaw University of Technology,<br />
Poland (email: zkotulsk@ippt.gov.p)<br />
telephony. Many advantages of this form of communication<br />
like cheap (or free) calls, wide range of additional features<br />
(video calls, conference calls, etc.) made it popular among<br />
companies and ordinary homes. Taking into account the continuous<br />
increase of Voice over IP (VoIP) users, it is safe to<br />
say that internet telephony will be one of the main forms of<br />
communication. However, there are still some challenges that<br />
it has to face, like providing a mechanism to ensure proper<br />
quality of service (QoS) and good security for data transfer<br />
and signaling.<br />
Because VoIP is a real-time application it has specific<br />
requirements from the lower layers. The most important of<br />
them are related to delay, jitter and packet loss. In telephony,<br />
the callers usually notice roundtrip voice delays of 250 ms<br />
or more, sensitive people are able to detect about 200 ms<br />
latencies. If that threshold is passed, communication starts<br />
to be annoying. ITU-T G.114 [2] recommends maximum of<br />
150 ms one-way latency. And because it includes the entire<br />
voicepath,thenetworktransmitlatencyshouldbesignificantly<br />
smaller than 150 ms. Unfortunately, for real-time applications<br />
we cannot use standard internet transport protocols such as<br />
TCP and UDP because they are not designed for this specific<br />
use, so they do not give us control over delay and jitter.<br />
BecauseTCPisaconnectionorientedprotocolitisslowerthen<br />
UDP and built-in retransmission mechanism is often useless<br />
forreal-timetransmission–retransmittedpacketsareoutdated.<br />
For multimedia data, reliability is not as important as timely<br />
delivery,so UDP is a preferable choice to base on for building<br />
real-time protocols. Although UDP has its benefits when it<br />
comes to speed, protocols based on it have to deal with lack<br />
of some important mechanisms. First of them is a congestion<br />
control mechanism which is not present in UDP and if the<br />
sender exceeds transmission rate that can be handled by the<br />
networkit leadsto congestionproblemsandnetworkoverload.<br />
The protocol should also implement mechanisms for timestamping<br />
packets to allow synchronization and minimize jitter<br />
problems. RTP/RTCP defined in RFC 1889 [3] is currently<br />
most widely used transport protocol for real-time services. It<br />
can estimate and control actual data transmission rate but QoS<br />
is still not guaranteed.<br />
In this paper we introduce our new VoIP model the aim<br />
of which is to meet the challenges of modern telephony. As<br />
opposedtostandardclient/serverarchitectureusedforexample<br />
in SIP [4] or H.323 [5], we chose to base our system on<br />
Peer-to-Peer (P2P) model. During last years P2P systems have<br />
become popular not only in domains like file sharing but also<br />
proven to be successful for voice and video communication<br />
(e.g. Skype). There are many benefits of using this network
ORYŃCZAK AND KOTULSKI: AGENT BASED VOIP APPLICATION WITH REPUTATION MECHANISMS 89<br />
Fig. 1. Overlay network model.<br />
Fig. 2. Infrastructure components.<br />
model; they are described in the next section. Another choice<br />
that we made was using agents for analyzing infrastructure.<br />
In many tasks agent-based solutions appeared to be more<br />
efficient [6], this paper shows that they are also useful in<br />
VoIP applications. Our goal was to design a secure system,<br />
which will ensure the best possible connectionquality(QoS) –<br />
which we achieved by path switching technique based on our<br />
own routing protocols assisted with reputation mechanisms.<br />
Security mechanismsthat we providedand biggercontrolover<br />
the packet routing decisions (owing to P2P model) make this<br />
system a good choice even in the environment where a high<br />
security level is required. To make the system more efficient<br />
and reliable, mechanisms for node load balance were also<br />
provided.<br />
The paper is organized as follows. In the Section 2 the<br />
system architecture is described with node model and general<br />
communication flow. The Section 3 is devoted to security and<br />
QoS mechanisms. Finally, the Section 4 concludes our work.<br />
II. SYSTEM ARCHITECTURE<br />
VoIP application presented in this paper is based on a P2P<br />
network. As opposed to traditional client/server architecture<br />
nodes of the P2P system (peers) act both as a client and a<br />
server and sometimes as relay for other peers in the network.<br />
When using a P2P model it is possible to implement an<br />
abstract overlay network that is build at Application Layer<br />
and consists of all participating peers as network nodes. This<br />
abstract network allows to build system independent from the<br />
physical network topology, because data transport service is<br />
provided by Transport Layer considered as part of underlying<br />
network.<br />
We chose the P2P model for our application because of<br />
several important reasons. First of all, owing to overlay<br />
network it is possible to make our own routing decisions<br />
and be more independent from external routing protocols.<br />
It is important because widely used routing protocols are<br />
often not real-time traffic friendly [7], but now, by monitoring<br />
path quality between peers, it is possible to choose the best<br />
route for packets transmission and quickly respond to any<br />
quality changes. Another P2P feature, that has proven to<br />
be useful in described VoIP application, is self-organization,<br />
which implies that any peer can enter or leave network at any<br />
time without a risk of overall system stability degradation.<br />
Owing to self-organization,system can be easily extendedand<br />
is more reliable and less vulnerable to failures and attacks.<br />
Additionally, nodes load-balancing mechanisms implemented<br />
in this application increase overall performance and stability.<br />
Another benefit of this system is an automatic elimination<br />
of problems with clients that are behind NATs. In normal<br />
circumstances, when both clients are behind NATs they are<br />
unable to establish direct connection. Although there are<br />
techniques like Session Traversal Utilities for NAT (STUN)<br />
[8] that can detect the presence of a network address translator<br />
and obtain the port number that the NAT has allocated for<br />
the applications UDP connection, they are often ineffective<br />
[9]. In this case additional server for traversal transmission<br />
between clients is required. In most cases that additional relay<br />
server is not on optimal path between those two clients, so it<br />
imposes additional delay in real-time communication. On the<br />
other hand, in P2P network peers are used for data routing, so<br />
no additional server is required, and because we chose peers<br />
that form the optimal path between clients transmission delay<br />
is minimized. Finally, owing to overlay network architecture<br />
and own routing protocol, we were also able to implement<br />
additional security mechanisms; they are described in Section<br />
III.B.<br />
A. Application components<br />
Our VoIP applications consists of three main elements:<br />
Login Server<br />
As the name suggests, the main task of the server is to<br />
provide services for authentication and authorization. Each<br />
user who wants to connect to the VoIP system must be<br />
previously logged into this server. Registration of new users<br />
and account management is also supported by the server. It<br />
can be used for charge calculation as well. Also every node<br />
must previously bypass the authorization check before it can<br />
be attached to the infrastructure. In the user logging process<br />
each user is assigned to a specific node, selection is based<br />
on a geographic location and load information. The server<br />
contains up-to-date information about every user state, its
90 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, SEPTEMPBER <strong>2011</strong><br />
current IP address, port number and assigned node ID, so it is<br />
used to determine current user location by other VoIP users.<br />
Because the server has full knowledge about current state of<br />
nodes and path qualities it plays the most important function<br />
in informing other nodes about any changes, so nodes can<br />
quickly recalculate their routing tables. As one may notice,<br />
the presence of the server is critical for this VoIP application,<br />
so it is important to provide appropriated hardware resources<br />
for its operation. For big systems, it is also possible to split<br />
these services between several servers.<br />
Nodes<br />
Nodes are essential part of our application and every P2P<br />
network. Their main task is to handle end-user support. As<br />
it was written before, during the logging process each user is<br />
assigned to a specific node and that node is used to route VoIP<br />
signaling and real-time data between other users and nodes.<br />
Theyalso haveabuilt-inmechanismallowing themto mediate<br />
between the user and the Logging Server (e.g. in logging<br />
process),owingtoitinfrastructureismoreresistanttoblocking<br />
(e.g. by the Internet Service Provider). Every node in our<br />
infrastructure has knowledge about other nodes and qualities<br />
of paths between them - this knowledge is used for building<br />
route tables. Nodes cooperate with the Logging Server by<br />
exchanging information about paths. They report status of the<br />
path between neighborsand receive information about the rest<br />
of the infrastructure. Because in our system nodes have to<br />
perform many tasks, we decided to use agents that will take<br />
over some of them. That allowed us to decompose the code,<br />
so it becomes more transparent. The system gains flexibility<br />
- nodes can be easily upgraded just by changing agents<br />
(even remotely). Additionally, different kinds of agents can<br />
be used, e.g. intelligent agents with ability to learn and adapt<br />
to different network conditions and by communicating with<br />
each other they can share knowledge and make routing more<br />
efficient. Each agent acts within the single node. Its main task<br />
is to constantly control the quality of transmissions relayed<br />
by this node. It analyzes such parameters as link bandwidth<br />
usage, number of lost packets, time interval between packets<br />
etc. Agents also test state of the other temporarily not being<br />
used links for detecting any changes. More about agents and<br />
routing mechanism is written in the next section.<br />
From the security perspective, we distinguish two types of<br />
nodes: standard and trusted. Every machine that has required<br />
resources can join our network and become a standard node.<br />
Nodes that had been previously verified as trusted can be used<br />
for routing data that requires a higher level of security. Apart<br />
from that, to ensure greater security and high reliability of the<br />
presented VoIP application, we provide it with a reputation<br />
mechanism. Every node has its own reputation index assigned<br />
by Logging Server, based on node reliability and long time<br />
behavior. Those reputation indexes are used for supporting<br />
mechanisms for building routing tables.<br />
End terminals<br />
End terminals are applications installed on computers (or<br />
smartphones) that are used for making and receiving calls.<br />
Because application uses only two ports: TCP for signaling<br />
and UDP for real-time data transfer, it is easy to configure<br />
firewall to workwith it. After loggingin to the LoggingServer<br />
Fig. 3. RTP/RTCP transmission schema.<br />
(directly or if direct connection is unenviable/blocked by any<br />
ofnodesbelongingto theinfrastructure– it hasalist oftrusted<br />
nodes in memory) application connects to a node assigned by<br />
Server and it is ready for use.<br />
When user is logged in, or has just changed his/her status,<br />
other users that have such a user on their contact lists are<br />
informed about this change. This mechanism works due to<br />
bilateral relation between users stored in the Login Server<br />
database. After being added to someone’s contact list user is<br />
asked for permission for checking his status. If permission is<br />
granted, relation between those users is stored on server so<br />
they can be immediately informed about status changes.<br />
Other elements<br />
Apart from mentioned elements, the presented application<br />
can be easily extended with additional specialized nodes, like<br />
public switched telephone network (PSTN) gateway or other<br />
IP telephony standards (e.g. SIP or Skype) gateways.<br />
III. DESIGN AND IMPLEMENTATION DETAILS<br />
In this section we give an overview of design issues based<br />
on our implementation.<br />
A. Quality of Service<br />
As it was said before, the standard best effort Internet is<br />
not real time traffic friendly.Although there are techniquesfor<br />
providing QoS like Intserv [10] or Diffserv [11] that reserve<br />
certain network resources for handling real-time traffic; they<br />
still depend on service providers policies and are often unable<br />
to ensure requiredend-to-endquality. Method proposedin this<br />
paper can be used to complement those existing mechanisms.<br />
Our application combines traffic flow adjustment method and<br />
path switching technique to ensure best possible connection<br />
quality. Traffic flow adjustment method is popular and widely<br />
usedincontrollingreal-timetraffic.Theessenceofthismethod<br />
is to adjust codec configuration parameters (output rate, voice<br />
frame size, etc.) and play buffer size to adapt to current<br />
network state. To make proper adjustments it is necessary to<br />
determine actual quality so feedback information is needed.<br />
It can be provided by using additional feedback channel (like<br />
in RTCP) or added into real-time traffic flow: into audio(e.g.<br />
using watermarking techniques) or into packet header [12].<br />
To make is simple, in this application we chose to add<br />
feedback information about quality into real-time traffic packets<br />
header, so if quality falls below desired level end user
ORYŃCZAK AND KOTULSKI: AGENT BASED VOIP APPLICATION WITH REPUTATION MECHANISMS 91<br />
terminal will modify audio parameters. Also every node on<br />
the path is informing its neighbor about the quality of links<br />
between them. It is done by inserting additional information,<br />
like number of sent and received packets, average delay and<br />
jitter, into header. By analyzing that data, the agent within the<br />
node has knowledge about the current quality of the link, and<br />
if it detects any changes, it may decide to re-route traffic by<br />
choosing another nodes to relay data. Logging Server is also<br />
informed about these changes and it passes this information<br />
to other nodes. Additionally, frequent changes in link quality<br />
affect on this link reputation by decreasing it. Temporary not<br />
being used links are also regularly tested by agents, they<br />
are sending (with desired time interval) series of test packets<br />
to simulate real-time traffic and analyze the responses. For<br />
performingroutingdecisions every node is building graph that<br />
represent current network state, then Dijkstra’s shortest path<br />
algorithm[13]is applied,butinstead ofshortestpathcounting,<br />
paths with best end-to-end quality are chosen. It is done by<br />
assigning to each edge in the graph its cost index, which is<br />
calculated by multiplying the correspondinglink quality index<br />
by its reputation. If any link state has changed, graph needs<br />
to be updated and Dijkstra’s algorithm applied again.<br />
B. Security<br />
In case of designing security mechanism for real-time<br />
traffic, it is very important to select appropriate security level.<br />
It must be chosen so that it ensures safety of transmission but<br />
also is not too demanding for resources (additional bandwidth<br />
and CPU power).If toomanysecuritymechanismsare applied<br />
it can affects on QoS, so call quality may be degraded. It<br />
is also possible that VoIP users may choose to disable these<br />
mechanisms to get better call quality. In this application we<br />
used following security schema:<br />
User logging – for logging process TLS connection is<br />
established. The user verifies the authenticity of the Logging<br />
Server using his CA certificate (with was previously delivered<br />
with client program, or downloaded from WWW page). Next,<br />
Digest method is used for user authentication. Afterward,<br />
serverchoosesnodethatwillhandlethisclient,andwithserver<br />
assist (server-node connection is also secure) node and client<br />
exchange their public keys, client updates information about<br />
his contact list, connection ends.<br />
User to node connection – secure TLS connection is<br />
established, for two-way authentication previously exchanged<br />
with Login Sever assist keys are used.<br />
Signaling and real-time traffic. Nodes are used to assist<br />
in signaling between end-users. Main reason of choosing<br />
this signaling method is willingness to provide mechanism<br />
for maintaining secrecy of end-users location, so IP address<br />
needs to be hidden from other users. In order to establish<br />
phone-to-phonecall, only user names and indexes of nodes to<br />
with they are attached are needed (indexes are not necessary<br />
- they can be retrieved for the Login Server, but in this<br />
schema server load and time needed to establish connection is<br />
reduced). Node, to which calling user is connected establishes<br />
TLS connection with destination user node, then they forward<br />
signaling data between users. Diffie-Hellman key exchange<br />
Fig. 4. User logging schema.<br />
protocol is used to establish encrypting key for real-time<br />
transmission. To avoid problems related with maintaining the<br />
Public Key Infrastructure (PKI) users do not use certificates<br />
for authentication. But in case additional security is needed,<br />
we providedmechanism for to-way user authentication: Login<br />
Server as a trusted intermediary is used.<br />
For real-time transfer TLS cannot be used because it is<br />
based on TCP, so it can cause additional delays. In this<br />
application we used AES in integer counter mode [14] (with<br />
the key agreed within signaling process) as a stream cipher.<br />
Bits from cipherare XORed with sounddata, and SHA-1 hash<br />
function is used to ensure packet integrity.<br />
Apart from that, because nodes are participating in data<br />
transmission, we have greater control over the data flow route.<br />
If higher security level is required, it is possible to specify the<br />
desired route, so regardlessof the external routingprotocolwe<br />
can avoid paths that are vulnerable to eavesdropping.<br />
C. Implementation and testing<br />
Our system was written in C# and uses DirectSound to<br />
access the sound device. Also, we created a simple agent<br />
platform for our needs: agents can be run as separate threads<br />
and communicate with each other using sockets.<br />
For testing our infrastructure in many different network<br />
configurations additional simulation software was written.<br />
This simulator allows to graphically create desired network<br />
infrastructureby adding nodes and connectionsbetween them,<br />
real-time data flow is created by streaming audio files, then<br />
links parameters and nodes behaviors can be changed in order<br />
to simulate different cases. For simplicity, simulator is using<br />
the same software that is running on nodes: it is running<br />
them as threads, and configuring by assigning different port<br />
numbers on the same IP. Node state, link delay, jitter, and<br />
packet dropping percentage can be set.<br />
IV. CONCLUSIONS AND FUTURE WORK<br />
In this paper we presented new, based on Peer-to-Peer<br />
network model, IP telephony system. The system model,<br />
infrastructure elements and some implementation details were
92 ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS, VOL. 2, NO. 3, SEPTEMPBER <strong>2011</strong><br />
be controlledby agentsand will work betweeneach two nodes<br />
that participate in data routing and are directly connected in<br />
the overlay network.<br />
Fig. 5. Infrastructure simulator.<br />
described. Also benefits of using P2P networks for building<br />
real-time infrastructures have been mentioned. Additionally,<br />
by placing an agent on each infrastructure node and indicating<br />
the advantages of this approach, we showed that agent<br />
based programming could be a valuable tool for designing<br />
this kind of systems. In our application we implemented<br />
mechanisms for ensuring the highest possible call quality,<br />
and security. In particular, a mechanism for improving QoS<br />
throughcontinuousmeasurementofsoundqualityateachnode<br />
in the overlay network, building dynamic routing tables and<br />
path switching technique. System security is guaranteed by<br />
using secure connection with authentication for login process,<br />
AES encryption of real-time data and SHA-1 hash function<br />
for packets integrality. Additionally, owing to P2P model,<br />
maintaining the secrecy of users location is possible, and by<br />
using internal routing protocols we can avoid unsafe paths.<br />
So far,we havebuilt a workingimplementationofpresented<br />
VoIP system, but it is still in its early testing phase. Many<br />
changes and improvements are still being made, so many<br />
elements, like i.e. reputation mechanism behavior or quality<br />
drops tolerance before path switch occurs, still needs to be<br />
tweaked and validated by simulations. Also, besides software<br />
simulations, we are planning to build real infrastructure and<br />
test system behavior in real environment.<br />
In parallel, we are testing new features, like for example<br />
fast retransmission mechanism for real-time packets, that will<br />
REFERENCES<br />
[1] D. Cohen, “A Protocol for Packet-Switching Voice Communication,”<br />
Computer Networks, vol. 2, no. 4–5, 1976.<br />
[2] One Way Transmission Time, ITU-T, G.114, 2003.<br />
[3] H. Schulzrinne, S. Casner, R. Frederick, and V. Jacobson, “RTP: A<br />
Transport Protocol for Real-Time Applications,” IETF, RFC 3550, Tech.<br />
Rep., Jul. 2003.<br />
[4] H.323. Packet-Based Multimedia Communication Systems, ITU-T, Jul.<br />
2003.<br />
[5] J. Rosenberg, H. Schulzrinne, G. Camarillo, A. Johnston, J. Peterson,<br />
R. Sparks, M. Handley, and E. Schooler, “SIP: Session Initiation<br />
Protocol,” IETF, RFC 3261, Tech. Rep., Jun. 2002.<br />
[6] M. Wooldridge and N. R. Jennings, “Intelligent Agents: Theory and<br />
Practice,” The Knowledge Engineering Review, 1995.<br />
[7] X. Che and L. J. Cobley, “VoIP Performance over Different Interior<br />
Gateway Protocols,” in IJCNIS, Apr. 2009.<br />
[8] J. Rosenberg, R. Mahy, P. Matthews, and D. Wing, “Session Traversal<br />
Utilities for NAT,” RFC 5389, Tech. Rep., Oct. 2008.<br />
[9] Z. Hu, “NAT Traversal Techniques and Peer-to-Peer Applications,” in<br />
HUT T-110.551 Seminar on Internetworking, Apr. 2005.<br />
[10] R. Baden, D. Clark, and S. Shenker, “Integrated Services in the Internet<br />
Architecture: An Overview,” IETF RFC 1633, Tech. Rep., Jun. 1994.<br />
[11] S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, and W. Weiss, “An<br />
Architecture for Differentiated Services,” IETF RFC. 2475, Tech. Rep.,<br />
Dec. 1998.<br />
[12] W. Mazurczyk and Z. Kotulski, “Adaptive VoIP with Audio Watermarking<br />
for Improved Call Quality and Security,” Journal of Information<br />
Assurance and Security, vol. 2, no. 3, pp. 226–234, 2007.<br />
[13] M. Pioro and D. Medhi, “Routing, Flow, and Capacity Design in Communication<br />
and Computer Networks,” The Morgan Kaufmann Series in<br />
Networking, 2004.<br />
[14] M. Dworkin, “Recommendation for Block Cipher Modes of Operation,”<br />
NIST Special Publication 800-38A, NIST, 2001.<br />
Zbigniew Kotulski received his M.Sc. in applied mathematics from Warsaw<br />
University of Technology and Ph.D. and D.Sc. Degrees from Institute of<br />
Fundamental Technological Research of the Polish Academy of Sciences.<br />
He is currently professor at IFTR PAS and professor and head of Security<br />
Research Group at Department of Electronics and Information Technology of<br />
Warsaw University of Technology, Poland.<br />
Grzegorz Oryńczak received his M.Scin computer science from Jagiellonian<br />
University. He is currently a Ph.D. student in computer science at the<br />
Jagiellonian University and Institute of Fundamental Technological Research<br />
of the Polish Academy of Sciences. He also works as a senior specialist at<br />
the National Center for Nuclear Research, Świerk.
About the journal<br />
ADVANCES IN ELECTRONICS AND TELECOMMUNICATIONS is a peer-reviewed journal published by Poznan University of<br />
Technology. It publishes scientific papers devoted to several problems in the area of contemporary electronics and telecommunications.<br />
Its scope is focused on, but not limited to the following issues:<br />
• electronic circuits and systems,<br />
• microwave devices and systems,<br />
• DSP structures and algorithms for wireless and wireline communication systems,<br />
• digital modulations,<br />
• data transmission techniques,<br />
• multiple access techniques and MAC issues,<br />
• information and channel coding theory and its applications,<br />
• software defined radio and cognitive radio technologies,<br />
• wireless local area networks (WLANs),<br />
• satellite communication,<br />
• navigation and localization,<br />
• synchronization subsystems,<br />
• time and timing,<br />
• modeling techniques of package & on-chip interconnects,<br />
• radiation & interference, electromagnetic compatibility,<br />
• propagation aspects in wireless communication,<br />
• UWB channel modeling,<br />
• measurements and wireless sensor networks,<br />
• web technologies,<br />
• e-learning,<br />
• multimedia communication,<br />
• audio and speech processing,<br />
• image and video processing,<br />
• software and hardware system implementation,<br />
• advanced A/D and D/A conversion techniques and their applications,<br />
• SDI - Software Defined Instruments,<br />
• effective measurement, estimation and computation of signal parameters,<br />
• consumer electronics.<br />
Detailed information about the journal can be found at: www.advances.et.put.poznan.pl.<br />
The Editorial Board invites paper submissions on the above topics for Open Call.