You are on page 1of 226

Universiteit Gent

Faculteit Ingenieurswetenschappen en Architectuur


Vakgroep Telecommunicatie en Informatieverwerking
Prestatie-analyse van buffers in draadloze netwerken
Performance Evaluation of Buffers in Wireless Networks
Koen De Turck
Proefschrift tot het bekomen van de graad van
Doctor in de Ingenieurswetenschappen:
Computerwetenschappen
Academiejaar 2010-2011
Universiteit Gent
Faculteit Ingenieurswetenschappen en Architectuur
Vakgroep Telecommunicatie en Informatieverwerking
Promotor: Prof. dr. ir. Sabine Wittevrongel
Universiteit Gent
Faculteit Ingenieurswetenschappen en Architectuur
Vakgroep Telecommunicatie en Informatieverwerking
Sint-Pietersnieuwstraat 41, B-9000 Gent, Belgi e
Proefschrift tot het behalen van de graad van
Doctor in de Ingenieurswetenschappen:
Computerwetenschappen
Academiejaar 2010-2011
Dankwoord
De tijd is gekomen om de vele mensen te bedanken die mijn jaren als doctoran-
dus mogelijk, aangenaam of de moeite waard hebben gemaakt. Bij hen heb ik
in de loop der jaren een aanzienlijke schuld opgebouwd, en het treft dus dat er
onder doctorandi en schrijvers in het algemeen sinds lang de opmerkelijke opvat-
ting bestaat dat onze schuldeisers genoopt worden tot een gehele of gedeeltelijke
aossing (of toch tenminste tot een billijke schuldherschikking), indien men hen
in de daartoe voorziene plaats vermeldt.
De lijst luidt aldus:
Mijn promotor Sabine Wittevrongel en vakgroepvoorzitter Herwig Bruneel, voor
de mogelijkheid die zij hebben geboden om dit onderzoek te verrichten, het stim-
ulerende kader en de wetenschappelijke vrijheid; Dieter Fiems, haast een weten-
schappelijke grote broer, om te tonen wat ik met voornoemde vrijheid kon aanvan-
gen, net als ik gedreven door een onverzadigbare nieuwsgierigheid; Stijn De Vuyst
om zijn kwaliteiten als didacticus, gracus, motivator en onderzoeker; Koenraad
Laevens, om zijn diepe inzicht in het onderzoek en nuchtere kijk op de onder-
zoekswereld; Soan De Clercq om zijn aanstekelijk enthousiasme en om (samen
met Dieter) mijn beperkte muzikale talent te dulden; Wouter, lot- en soortgenoot
sinds dag e en; de vele bureaugenoten door de jaren heen: Joost, Pieter, Thomas,
Carla, Frederick, Eline, Luigi, Alessandro, Francesco, Patrick; Thomas Demoor
en Joris Walraevens die samen met Stijn en Dieter mijn adoptiebureau volmaken
waar het steeds gezellig keuvelen en kofe drinken is; elk lid van onze hechte
en gemoedelijke onderzoeksgroep SMACS; het uitstekende technische personeel
van TELIN: Annette, Patrick, Sylvia, Davy en Philippe; vrienden Maarten, Mieke,
Roy, Wendelien, Dieter, Stefanie voor de steeds jne gesprekken; mijn ouders,
broer en zus, om hun jarenlange steun en vertrouwen; Caroline om alles maar ook
om mijn afwezige buien (een neveneffect van aan wiskunde gerelateerd werk) met
gelijk gemoed te verdragen.
Aan hen allen betuig ik oprecht dank, met de belofte dat ze niet noodzakelijk
hoeven te lezen wat ik heb geschreven.
Gent, 2011
Koen De Turck
Table of Contents
Dankwoord i
Nederlandse samenvatting v
English summary ix
1 Introduction 1-1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-1
1.2 A Short Overview of Wireless Networks . . . . . . . . . . . . . . 1-1
1.2.1 Retransmission protocols . . . . . . . . . . . . . . . . . . 1-3
1.2.2 Energy-saving measures . . . . . . . . . . . . . . . . . . 1-6
1.3 Stochastic Modeling . . . . . . . . . . . . . . . . . . . . . . . . 1-8
1.3.1 Recursive characterizations of Markov chains . . . . . . . 1-11
1.3.2 Background Markov chains . . . . . . . . . . . . . . . . 1-12
1.3.3 Censored Markov chains . . . . . . . . . . . . . . . . . . 1-12
1.3.4 Some examples of Markov chains . . . . . . . . . . . . . 1-13
1.3.4.1 Two-state Markov chains . . . . . . . . . . . . 1-13
1.3.4.2 Finite Markov chains . . . . . . . . . . . . . . 1-14
1.3.4.3 Random walks . . . . . . . . . . . . . . . . . . 1-14
1.3.4.4 ARMA models . . . . . . . . . . . . . . . . . . 1-16
1.3.5 Which method to choose? . . . . . . . . . . . . . . . . . 1-17
1.3.6 Large Deviations . . . . . . . . . . . . . . . . . . . . . . 1-18
1.3.7 Series expansions . . . . . . . . . . . . . . . . . . . . . . 1-19
1.4 Dissertation Outline . . . . . . . . . . . . . . . . . . . . . . . . . 1-20
1.5 Main contributions . . . . . . . . . . . . . . . . . . . . . . . . . 1-24
1.6 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-25
1.6.1 Publications in international journals . . . . . . . . . . . 1-25
1.6.2 Book chapters . . . . . . . . . . . . . . . . . . . . . . . . 1-26
1.6.3 Papers in Proceedings of International Conferences . . . . 1-26
1.6.4 Abstracts and Other Presentations . . . . . . . . . . . . . 1-29
1.6.5 Awards . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-30
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-31
iv
2 Delay Analysis of Go-Back-N ARQ 2-1
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-2
2.2 Modeling Assumptions . . . . . . . . . . . . . . . . . . . . . . . 2-3
2.3 Distribution of the Service Time . . . . . . . . . . . . . . . . . . 2-4
2.4 Distribution of the Unnished Work . . . . . . . . . . . . . . . . 2-7
2.5 Distribution of the Packet Delay . . . . . . . . . . . . . . . . . . 2-9
2.6 Computing Moments of the Packet Delay . . . . . . . . . . . . . 2-10
2.7 Channel Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-13
2.7.1 Cyclic Channel Model . . . . . . . . . . . . . . . . . . . 2-13
2.7.2 Gilbert-Elliott Channel Model . . . . . . . . . . . . . . . 2-14
2.8 Numerical results and discussion . . . . . . . . . . . . . . . . . . 2-15
2.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-17
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-21
3 Order-preserving ARQ 3-1
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1
3.2 Model and notations . . . . . . . . . . . . . . . . . . . . . . . . 3-3
3.3 Grammar of an ARQ protocol . . . . . . . . . . . . . . . . . . . 3-3
3.3.1 Stop-and-wait . . . . . . . . . . . . . . . . . . . . . . . . 3-6
3.3.2 Go-back-N . . . . . . . . . . . . . . . . . . . . . . . . . 3-6
3.3.3 Stop-and-wait with m repeated copies . . . . . . . . . . . 3-6
3.3.4 Persistent ARQ . . . . . . . . . . . . . . . . . . . . . . . 3-6
3.3.5 Go-back-N with m repeated copies . . . . . . . . . . . . . 3-7
3.3.6 Protocols with a limited number of transmissions . . . . . 3-7
3.3.7 Protocols with long and short encodings . . . . . . . . . . 3-7
3.3.8 Protocols outside of this class . . . . . . . . . . . . . . . 3-8
3.4 Performance analysis . . . . . . . . . . . . . . . . . . . . . . . . 3-8
3.4.1 Analysis of the throughput . . . . . . . . . . . . . . . . . 3-8
3.4.2 Analysis of the message delay . . . . . . . . . . . . . . . 3-9
3.4.3 Analysis of the buffer content . . . . . . . . . . . . . . . 3-9
3.4.4 Analysis of the modied buffer . . . . . . . . . . . . . . 3-10
3.4.5 Putting it all together . . . . . . . . . . . . . . . . . . . . 3-10
3.4.6 Analysis of the packet delay . . . . . . . . . . . . . . . . 3-11
3.4.6.1 Distribution of the Unnished Work . . . . . . 3-11
3.4.6.2 Distribution of the Packet Delay . . . . . . . . . 3-12
3.5 Asymptotic behavior . . . . . . . . . . . . . . . . . . . . . . . . 3-12
3.6 General Markov models . . . . . . . . . . . . . . . . . . . . . . . 3-13
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-13
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-13
4 Moderate deviations for fading channels 4-1
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-1
4.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-2
4.3 Scalings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-4
4.3.1 Fast-time scaling . . . . . . . . . . . . . . . . . . . . . . 4-5
v
4.3.2 Many-ows scalings . . . . . . . . . . . . . . . . . . . . 4-6
4.3.3 Computing the covariance structure . . . . . . . . . . . . 4-7
4.4 Optimal control . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-7
4.5 Some numerical results . . . . . . . . . . . . . . . . . . . . . . . 4-8
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-9
5 Receiver Buffer Behavior for Selective Repeat 5-1
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-2
5.2 Mathematical Model of the Resequencing Buffer . . . . . . . . . 5-3
5.3 Exact Analysis of the Receiver Buffer Content . . . . . . . . . . . 5-6
5.4 Large-Deviations Analysis . . . . . . . . . . . . . . . . . . . . . 5-8
5.5 A Closer Look at the Large Deviations Limit . . . . . . . . . . . 5-13
5.6 Asymptotic Analysis of the Packet Delay . . . . . . . . . . . . . 5-14
5.7 Approximations for High Error Probabilities . . . . . . . . . . . . 5-15
5.8 Numerical Results and Discussion . . . . . . . . . . . . . . . . . 5-17
5.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-20
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-20
6 Sleep mode for correlated downlink trafc 6-1
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-2
6.2 Queueing Model . . . . . . . . . . . . . . . . . . . . . . . . . . 6-4
6.3 Analysis of the Buffer Content . . . . . . . . . . . . . . . . . . . 6-6
6.3.1 Derivation of V(z) . . . . . . . . . . . . . . . . . . . . . 6-6
6.3.2 Derivation of U
c
(z) . . . . . . . . . . . . . . . . . . . . . 6-7
6.3.3 Derivation of U(z) . . . . . . . . . . . . . . . . . . . . . 6-8
6.4 Analysis of the Delay . . . . . . . . . . . . . . . . . . . . . . . . 6-11
6.4.1 Unnished Work . . . . . . . . . . . . . . . . . . . . . . 6-11
6.4.2 Packet Delay . . . . . . . . . . . . . . . . . . . . . . . . 6-12
6.5 Energy Consumption . . . . . . . . . . . . . . . . . . . . . . . . 6-13
6.6 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . 6-14
6.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-19
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-22
7 Delay versus Energy Consumption for IEEE 802.16e 7-1
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-1
7.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-3
7.3 Queueing Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 7-4
7.4 Mean delay and energy consumption . . . . . . . . . . . . . . . . 7-5
7.5 Examples and discussion . . . . . . . . . . . . . . . . . . . . . . 7-6
7.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-10
vi
8 Sleep mode with bidirectional trafc 8-1
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-2
8.2 Model Description . . . . . . . . . . . . . . . . . . . . . . . . . 8-5
8.3 Analysis of Buffer content . . . . . . . . . . . . . . . . . . . . . 8-7
8.3.1 Packet delay . . . . . . . . . . . . . . . . . . . . . . . . 8-11
8.3.2 Power consumption . . . . . . . . . . . . . . . . . . . . . 8-11
8.4 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . 8-12
8.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-14
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-15
9 Sleep mode analysis for 802.16m 9-1
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-2
9.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-3
9.3 Analysis of Buffer Content and Power Consumption . . . . . . . . 9-6
9.4 Computational method . . . . . . . . . . . . . . . . . . . . . . . 9-9
9.5 Analysis of Packet Delay . . . . . . . . . . . . . . . . . . . . . . 9-12
9.6 Practical Example . . . . . . . . . . . . . . . . . . . . . . . . . . 9-13
9.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-14
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-15
10 Sleep mode under light trafc 10-1
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-2
10.2 Model and Problem Statement . . . . . . . . . . . . . . . . . . . 10-2
10.3 Light-trafc technique . . . . . . . . . . . . . . . . . . . . . . . 10-3
10.4 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-5
10.5 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . 10-7
10.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-8
11 Large Deviations for Sleep Mode 11-1
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-1
11.2 Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-3
11.3 Large Deviations Analysis . . . . . . . . . . . . . . . . . . . . . 11-5
11.3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . 11-5
11.3.2 Limiting process . . . . . . . . . . . . . . . . . . . . . . 11-6
11.3.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-7
11.3.4 Main Result . . . . . . . . . . . . . . . . . . . . . . . . . 11-10
11.4 Numerical method . . . . . . . . . . . . . . . . . . . . . . . . . 11-10
11.5 Practical Application . . . . . . . . . . . . . . . . . . . . . . . . 11-12
11.5.1 Single Mobile Station under S
1
. . . . . . . . . . . . . . 11-12
11.5.2 Single Mobile Station under S
2
. . . . . . . . . . . . . . 11-14
11.5.3 Scenarios with Two Mobile Stations (under mechanism S
2
) 11-14
11.5.4 Energy consumption . . . . . . . . . . . . . . . . . . . . 11-15
11.5.5 Numerical illustrations . . . . . . . . . . . . . . . . . . . 11-15
11.6 General observations and conclusions . . . . . . . . . . . . . . . 11-17
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-18
12 Conclusions 12-1
12.1 Overview of the main results . . . . . . . . . . . . . . . . . . . . 12-1
12.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-2
Nederlandse samenvatting
Dutch Summary
Het onderwerp van dit proefschrift is de stochastische studie van draadloze telecom-
municatienetwerken.
De invloed van draadloze netwerken op ons dagelijks leven hoeft nauwelijks
toegelicht te worden. Dit wordt bijvoorbeeld gellustreerd door de enorme impact
die de GSM op ons sociaal leven heeft, een impact die zo groot is dat het nauwe-
lijks te geloven is dat GSMs nog maar sinds een tiental jaar een wijde verspreiding
kennen.
Wij hebben ons niet verdiept in de sociale component van draadloze netwerken,
maar eerder in de technologische obstakels die een nog wijder gebruik in de weg
staan. Concreet hebben we een grondige stochastische studie uitgevoerd van twee
aspecten van draadloze netwerken, namelijk (1) retransmissieprotocollen over ge-
correleerde kanalen en (2) energiebesparende mechanismen in draadloze proto-
collen. We lichten deze beide luiken hieronder kort toe.
Retransmissieprotocollen zijn populaire mechanismen om een foutloze over-
dracht van informatie te garanderen over een kanaal dat fouten introduceert. Men
vindt ze in vrijwel elke moderne telecommunicatiestandaard, en dit op bijna elke
laag (van datalinklaag tot applicatielaag). Ons uitgangspunt is om niet alleen de
doorvoer van pakketten te onderzoeken, maar ook en vooral het buffergedrag.
Bovendien vertonen draadloze kanalen de complicatie dat het foutproces van het
kanaal in belangrijke mate gecorreleerd is, wat een belangrijke invloed op het
buffergedrag heeft. In het onderzoek werden een aantal nieuwe aspecten be-
licht. Zo geven we een gestroomlijnde analyse voor een deelklasse van de re-
transmissieprotocollen. Ook hadden we aandacht voor een aantal verschillende
kanaalmodellen, Markoviaanse met een eindig aantal toestanden, en fading mo-
dellen met een onderliggend Gaussiaans proces. We stelden ons ten doel prestatie-
maten te vinden die een vollediger beeld schetsen van de prestatie van een protocol
dan enkel de doorvoer (throughput), zonder een volledige bufferanalyse hoeven te
doen en vonden inspiratie bij de heavy-trafc-theorie en de theorie der grote af-
wijkingen (large deviations). We sluiten dit luik af met een asymptotische analyse
van een herordeningsbuffer over een gecorreleerd kanaal, een ander tot hiertoe
onderbelicht aspect.
Het tweede luik omvat energiebesparende maatregelen in draadloze netwerken,
waar momenteel heel wat onderzoek naar gebeurt. Men kan stellen dat batterij-
en de achillespees vormen van draadloze toestellen en gebruikers zijn dan ook
x NEDERLANDSE SAMENVATTING
bereid om een deel van de performantie van draadloze toestellen op te offeren
voor een langere batterijduur. Het hoeft dan ook geen wonder te heten dat zowel
in de opvolger van WiFi, namelijk WiMAX (ofcieel IEEE 802.16) als in de vol-
gende generatie GSM technologie (LTE; long term evolution) een hele reeks en-
ergiebesparende maatregelen opgenomen zijn. E en van de meest prominente wordt
in WiMAX-terminologie sleep mode genoemd en in LTE-kringen DRX (dis-
continuous reception). Beide mechanismen houden in dat een draadloos toestel
met het basisstation kan afspreken om voor een bepaalde tijd naar een energiebe-
sparende modus over te schakelen (het toestel gaat in sleep mode). Pakketten die
gedurende die tijd in het basisstation aankomen en bestemd zijn voor dit draad-
loze toestel, lopen dan uiteraard extra vertraging op. Het is duidelijk dat er een
afweging gemaakt moet worden tussen langere sleep-modeperiodes (en dus meer
energiebesparing) en een verminderde bedieningskwaliteit. Onder de aspecten
die we hebben bestudeerd, vallen onder meer: (1) de invloed van correlatie in
het aankomstproces van pakketten die bestemd zijn voor het draadloze toestel in
kwestie, (2) de invloed van traek in de andere richting, namelijk van het draad-
loze toestel naar het basisstation, (3) optimalisatie van de lengtes van de sleep-
modeperiodes, en (4) de invloed van het sleep-modemechanisme op het optreden
van zeer grote vertragingen. Al deze punten vormden uitbreidingen of vernieuwin-
gen ten opzichte van bestaande literatuur.
In dit proefschrift hebben we naast het eigenlijke studieobject in casu draad-
loze netwerken een even grote aandacht voor het wiskundig gereedschap. Dit is
bij het onderzoek van telecommunicatienetwerken veelal van stochastische aard.
Het belang van stochastiek in het gebied van de telecommunicatie wordt door
haast de hele onderzoeksgemeenschap onderkend. Inderdaad, meer praktisch on-
derzoek wordt onderbouwd door hetzij een statistische analyse van metingen, het-
zij door simulaties. Beide hebben een methodologische basis in de stochastiek.
Bestond het onderwerp van dit proefschrift uit twee aspecten van draadloze
netwerken, het wiskundig gereedschap is op te delen in drie luiken.
Het eerste luik bevat de klassieke technieken van de wachtlijntheorie, met de
traditionele tweedeling tussen probabiliteitsdomein en transformatiedomein, die
eigenlijk een valse tweedeling is, want er bestaat een diepe link tussen beide me-
thoden, zodat de keuze voor deze of gene methode haast eerder een culturele
dan een inhoudelijke beslissing is, waarbij de onderzoeker die zich bij voorkeur in
het probabiliteitsdomein ophoudt eerder in stabiele numerieke oplossingen gen-
teresseerd is, terwijl het transformdomein eerder gebruikt wordt voor wie op zoek
gaat naar uitdrukkingen in gesloten vorm. Dit zijn echter slechts vuistregels, en
hoewel de link tussen probabiliteitsdomein en transformdomein wiskundig goed
bekend is, modellering is niet alleen een wetenschap maar ook een ambacht. In
dit proefschrift staan een aantal voorbeelden om de lezer ervan te overtuigen dat
een kennis van beide methoden de moeite loont. Deze klassieke methoden hebben
hun deugdelijkheid bewezen, maar niet elk model past binnen het vrij specieke
raamwerk die deze methoden vereisen.
Daarom beschouwen we in het tweede luik methoden om kansen op zeldzame
gebeurtenissen te kwanticeren. Als men enkel genteresseerd is in het asympto-
DUTCH SUMMARY xi
tisch gedrag, dan is het mogelijk om breder toepasbare resultaten af te leiden en
wel met de theorie van de grote afwijkingen (large deviations). Een van de doe-
len was de precieze relatie tot de traditionele asymptotische analyse te bepalen,
alsook de voordelen van beide. Methoden gebaseerd op grote afwijkingen hebben
het voordeel zeer algemeen toepasbaar te zijn. Zo konden we bijvoorbeeld Gaus-
siaanse kanaalmodellen onderzoeken in het luik over retransmissieprotocollen, die
buiten het bereik van de klassieke technieken liggen. Ze bieden ook een stevige
basis voor de vuistregel dat een zeer zeldzame gebeurtenis bijna steeds op de meest
waarschijnlijke manier optreedt. Deze vuistregel laat toe om het gedrag van sys-
temen aan te voelen die te complex zijn voor een exacte analyse. Als nadeel kan
genoemd worden dat deze methoden wiskundig vrij technisch zijn.
Het derde luik betreft perturbatietechnieken, welke onder meer nuttig zijn voor
modellen die niet expliciet oplosbaar zijn tenzij een zekere parameter een bepaalde
extreme waarde aanneemt. Het doel is vervolgens om een prestatiekarakteristiek
in de buurt van deze extreme waarde te kunnen schrijven als een machtreeks.
Doorheen het tot stand brengen van dit proefschrift heb ik de voor- en nadelen
van deze diverse technieken beter naar waarde kunnen schatten en het is mijn hoop
dat dit werk ook een bescheiden bijdrage kan leveren tot een grotere appreciatie
voor en verspreiding van deze methoden.
English Summary
The subject of this dissertation is the stochastic study of wireless telecommunica-
tion networks.
The inuence of wireless networks on our daily lives hardly needs any com-
ment. One can think for example of the enormous impact of mobile phones on
our social lifes, which is so great that it is hard to believe that they are only a
phenomenon of the last decade.
However, this is not a thesis on the sociological implications of wireless net-
works. Rather it is a study of the technological obstacles that impede an even more
pervasive use. Specically we undertook a stochastic study of two aspects of wire-
less networks, i.e. (1) retransmission protocols over correlated channels and (2)
energy-saving mechanisms. We give a brief explanation of both themes.
Retransmission protocols are popular mechanisms to provide a reliable trans-
mission of information over an inherently unreliable channel. They are found in
almost every modern telecommunication standard, and on almost every level of
the protocol stack, from datalink layer to application layer. Our vantage point is to
not only investigate the throughput of the packets, but rather the buffer behavior as
a whole. Added complication is that wireless channels exhibit correlated behav-
ior to a considerable extent, which we show to have an important impact on the
buffer behavior. We highlighted new aspects in the course of our investigations.
For example, we give a encompassing analysis for a particular class of retrans-
mission protocols. We considered both Markovian models with a nite number of
states as well as fading models with an underlying Gaussian process. We sought
performance measures that capture the essence of buffer behavior more fully than
just the throughput but do not require a complete buffer analysis, which we found
in the theories of large deviations and heavy trafc. The nal contribution of this
part entails an asymptotic analysis of a reordering buffer over a correlated channel,
which had hitherto received little attention.
The second theme in this dissertation concerns the energy-saving measures in
wireless networks, which is currently a hot topic. One may safely say that batter-
ies are amongst the principal weaknesses of wireless telecommunications. Users
are to some degree prepared to give a part of the performance in order to have
longer battery lives. It stands to reason that both the successor of the popular
WiFi standard, WiMAX, as well as LTE (long term evolution), which is the heir
of the UMTS (Universal Mobile Telecommunications System) have both ample
provisions for energy-saving measures. Among them are sleep mode in WiMAX
protocols and DRX (discontinuous reception) in LTE. Under both mechanisms,
xiv ENGLISH SUMMARY
wireless appliances can negotiate a temporal partial powerdown (i.e. the system
is said to go to sleep mode). Packets destined for this mobile station that arrive
during this time then incur an extra delay. This illustrates the central trade-off
that must be made between longer sleep-mode periods (and thus greater energy
savings) and a diminished quality of service. We investigated different aspects
of sleep mode, among others: (1) the inuence of correlation in the arrival pro-
cess of packets destined for the wireless device, (2) the inuence of uplink trafc
(i.e. trafc that ows from the mobile station to the base station), (3) optimiza-
tion of sleep-mode parameters, (4) the inuence of the sleep-mode mechanism on
the occurrence of very large packet delays. All of these points are extensions or
innovations as compared to existing literature.
As this dissertation is about modelling, and the position occupied by modellers
is in between network engineers and applied mathematicians, it is natural to put an
equally strong focus on the application and on the mathematical tools. These tools
are very often of a stochastic nature.
The importance of stochastic methods is almost universally acknowledged by
the telecommunication community. Indeed, more practically oriented research is
validated by either simulations or statistical studies, which both have stochastical
underpinnings.
Whereas two aspects of wireless networks form the subject of this dissertation,
the mathematical toolset can be divided into three parts.
The rst part consists of the classical techniques of queueing theory, with
the traditional dichotomy between probability domain and transform domain tech-
niques, which is in fact a false dichotomy, as there is a deep link between both
methods, so that choice for the latter or the former is more a matter of taste than
a choice based on content. A researcher that prefers the probability domain is
likely to attach more importance to the stable numerical algorithms, whereas the
transform domain is more likely to be the weapon of choice for nding closed-
form solutions. These are just rules of thumb, and although the link between the
two methods is well-known, modelling is not only a science but also an art. This
dissertation takes a basically bipartisan point of view and provides a few exam-
ples to convince the reader that a knowledge of both methods is recommendable.
These classical methods have shown their worth, but they impose a rather specic
framework.
Therefore, we consider in this dissertation alternative methods, such as the
large-deviations method, which is a method to quantify probabilities of rare events.
If one is only interested in rare events, it is possible to derive very broadly applica-
ble results. We wanted to appreciate the differences between the large-deviations
theory and asymptotics that can be derived from the classical methods. For ex-
ample, in the part on retransmission protocols, we could study Gaussian channel
models which are out of reach of the classical methods. Large-deviations theory
also offers a sound mathematical basis for the rule of thumb that a rare event in
many different situations happens in the most probable way. This rule of thumb
offers intuition in the behavior of systems that are too complex for exact analy-
sis. The considerable mathematical technicality may be cited as somewhat of a
ENGLISH SUMMARY xv
drawback.
The third part considers series expansions, which are useful for models that do
not have an exact solution unless a certain parameter assumes a certain extreme
value. The goal is then to characterize the behavior in the neighborhood of this
extreme value as a power series.
Throughout this doctoral research, I learned the strong and weak points of the
different methods, and it is my hope that this dissertation can also make a modest
contribution to the further dissemination of these methods.
1
Introduction
1.1 Introduction
For many users, wireless communication networks are an attractive alternative to
the traditional wired networks. One may cite advantages such as less clobbering
with cables, connectivity in sparsely populated regions, and so on. There are how-
ever also drawbacks that impede a more widespread use. In this dissertation we
look at two such drawbacks, together with possible solutions, and develop stochas-
tic models for them.
In this introductory chapter, we sketch both the tools and the subjects of this
dissertation, and subsequently elaborate on our own contributions. First, we give a
general overview of wireless networks, with special attention given to retransmis-
sion and energy saving, the two core subjects of this thesis. Next, we give some
thoughts on the nature and utility of stochastic modelling, as well as a short in-
troduction on the basic tools that are used throughout this dissertation. Also, an
overview of the rest of this book is provided, with special attention to the coher-
ence between the different chapters.
1.2 A Short Overview of Wireless Networks
A wireless network consists of a number of base stations and a number of mobile
stations. The base stations (one may think of the radio masts of a GSM network, or
wi hotspots) are on a xed location and are linked to the rest of the internet (usu-
1-2 INTRODUCTION
ally in a wired manner). Mobile stations are hand-held devices (mobile phones,
laptops, ...) that provide internet access to a user. They may be on a xed loca-
tion, but may usually wander around. As the connections are wireless, the medium
is the air, and this is shared by all mobile users. Wireless protocols regulate the
interaction of mobile and base stations. There are many aspects that need to be
stipulated, such as the channel coding, regulating interference of different users,
handling the handover of a mobile user from one base station to another, power
saving measures etc. This partly explains why modern wireless protocol standards
are rather lengthy (see eg. [1, 4], which both count several hundreds of pages).
There are at the time of writing two standards that compete to be the next-
generation wireless broadband telecommunication networks. The rst one is IEEE
802.16e, backed by the IEEE consortium. It is colloquially known as WiMAX and
is the successor of the well known WiFi standard (IEEE 802.11). The other is LTE
(long term evolution). This is a heir to the equally well-known UMTS (Universal
Mobile Telecommunications System) branch. It is unclear whether both protocols
will achieve a breakthrough, and coexist peacefully, or whether only one of them
will achieve success. It is however assuring that the models developed in this
dissertation are largely agnostic to the exact standard employed. For example, in
the rst sleep mode papers we wrote, we had specically WiMAX in mind as LTE
had not fully emerged yet. However we could adapt it with only minor changes to
the new framework.
These two standards also bring their own (rather generous) share of acronyms
and terminology to the table. For example, the IEEE community speaks of mo-
bile subscriber stations (MSS) and base stations (BS), which are the equivalents of
respectively user equipment (UE) and evolved Node-Bs (eNB) in LTE terminol-
ogy. As we try to take as much as possible a technology-neutral stance through
the course of this dissertation, we adopt most of the time the simplied terminol-
ogy of mobile stations, base stations, unless when referring to one technology in
particular.
A universal characteristic of todays wireless networks is the almost pervasive
synchronicity. All communication is organized in a discretized time axis that is
shared by all participating mobile stations and base stations. This validates our
choice of working almost exclusively in discrete time. The basic unit of time is
called a frame in IEEE 802.16 technology, but we opt for the term slot which is
more prevalent in discrete-time queueing theory.
Wireless communications are different from wired ones in a couple of key as-
pects. Firstly, as communication takes place over an inherently shared, changing
medium, transmission errors happen at a more frequent rate as compared to wired
networks. This means that special measures should be taken in order to control er-
rors. A popular scheme is the employment of retransmission protocols to organize
further transmissions of bits of information until the transmission is ensured to be
INTRODUCTION 1-3
A
C
K
N
A
K
N
A
K
transmitter
(wireless) channel
receiver
P
P
waiting time
transmitter delay
transmission time resequencing time
first transmission
of P
removal of from
the transmitter queue
P
P received correctly
arrival of
packet P
Figure 1.1: Basic retransmission set-up.
faultless. This will be discussed in our chapters on retransmission protocols.
In wireless telecommunications, there is not only no access to a wired network,
also there is no continuous access to the power grid. This means that batteries have
to be used as a means of power supply. As a result, the focus on energy efciency
is much larger in research on wireless networks than it is in wired networks. An
intelligent protocol design therefore foresees in adaptable standby or sleep modes,
which switch off the radio interface when no data transmission takes place, at the
while ensuring that the overall performance is not negatively affected. This is the
subject of our research work on sleep mode.
Needless to say, these are not the only two aspects that distinguish a typical
deployment of wireless networks from wired ones. Another factor is the more
dynamic network topology that stems from the mobility of users. Handover issues,
cognitive radio networks and so on are also interesting issues that are specic to
wireless networks, but they will not be addressed in this dissertation.
1.2.1 Retransmission protocols
Consider the situation in which a user wishes to convey information from point A
in the network (the transmitter) to point B (the receiver). This is the whole rai-
son d etre of communication networks, hence this is a very fundamental situation.
The medium between point A and B is called the channel. If the channel is fault-
less and no other users have access to the medium, then the information may be
transmitted in almost any way. However, neither of these assumptions is realistic,
especially not in the case of wireless networks. Much research has been devoted
1-4 INTRODUCTION
to the performance study of communication over faultless channels. The problem
of the channel being a shared medium is dealt with by means of channel access
methods or medium access protocols.
It is however the other problem, the one of how to deal with a faulty channel
that plays a larger role in this dissertation. The general strategy of dealing with
this problem is as simple as it is efcient. The transmitter divides the informa-
tion in equal packets, storing them in a transmitter buffer if necessary, and adds
an error-detecting code. That is, parity bits are added so that the receiver can
check the integrity of the packet by recomputing the parity bits, and checking if
the parity bits still hold true. The receiver then noties the transmitter by means
of acknowledgment messages. Positive acknowledgments are denoted by ACK,
negative ones by NAK. The transmitter can then take action and send a fresh copy
of the corrupted packets. An important characteristic of a retransmission system
is the feedback delay, or roundtrip time, which is the time from the start of trans-
mission of a packet at the senders side until the reception of the corresponding
acknowledgement message. Note that the implementation of every retransmission
mechanism (also frequently called ARQ: Automatic repeat request) needs buffers.
At the senders side, a transmitter buffer is needed in which packets are temporar-
ily stored until their nal acknowledgement is received. Also, at the receivers side
a resequencing buffer can be needed if the order of the packets can be disturbed by
the protocol.
How can we evaluate the performance of retransmission protocols? A basic
and popular way is to consider the throughput of an ARQ protocol. The throughput
of an ARQ protocol is dened as the average number of packets that is successfully
transmitted per time unit, given that there are always packets available to be sent.
However, the throughput does not tell a complete story, as there are various
aspects to the buffer behavior that make the evaluation a little more subtle. As
packets travel through the retransmission buffer, they incur each a different de-
lay. The distribution of the packet delay can be very different for systems with the
same throughput, and has an important impact on the user experience. The packet
delay at the retransmission level constitutes a part of the total latency (together
with latency incurred at higher layers of the protocol stack). This latency is an im-
portant measure of the user experience. This was true already for well-established
applications such as telephony and videoconferencing, but new applications such
as gaming have even stricter latency demands. Also, constant latency is far less
harmful than a uctuating one, as the latter often necessitates buffering capacity in
higher layers as well.
There is a great amount of variety possible in the implementation of the basic
idea. There is the three classic protocols that have increasing degree of complex-
ity in implementation. The simplest protocol in terms of implementation is the
stop-and-wait protocol. After each transmission, the sender waits until it receives
INTRODUCTION 1-5
an acknowledgment from the receiver. The advantage of this scheme is that the
channel is used at any given point by either the sender or the receiver, hence the
channel allocation is simpler (this is known as half-duplex channel). Also, a fur-
ther advantage is that no resequencing buffer is needed at the receivers side. The
main disavantage is that the throughput is suboptimal.
One step up in complexity of implementation is the Go-Back-N protocol. In
this protocol, the sender transmits the packets in a continuous manner, but when
a negative acknowledgement is received, the erroneous packet and all its succes-
sors are to be sent again, regardless of the status of their reception. This protocol
requires a full duplex channel, but as with stop-and-wait, no resequencing mecha-
nisms need to be in place.
The last and most complex basic scheme is the selective repeat protocol. In this
scheme, each packet has an unambiguous identier and only packets for which a
negative acknowledgement is received, are sent again. In this setup, there are
two buffers needed: one at the senders side called the transmit buffer and one
resequencing buffer at the receivers side.
Apart from these basic protocols, there are a lot of other variations. Another
axis in the design space is to replace the error-detecting coding by a partially error-
correcting code. The most radical departure from the basic scheme of ARQ in this
regard is forward error correction (FEC), where instead of an error-detecting code
an error-correcting code is used, which allows to correct a number of errors per
packet (of course the more errors one wants to be able to correct, the larger the
relative length of the error correcting code and the smaller the goodput, the share
of useful information compared to the total amount of information that must be
transmitted over the channel).
When using FEC, acknowledgments and retransmissions are usually done away
with, the rationale being that the probability of a packet error should be mitigated
by error correction only, and that it may be counterproductive to retransmit packets
as they can have become useless after too large a delay.
It is also possible to combine ARQ and FEC, resulting in hybrid ARQ. In this
setup, some errors can be corrected at the receivers side, but if the packet cannot
be recovered completely, then a retransmission is requested anyway.
It is clear that FEC-based and ARQ-based protocols exhibit different behavior:
in periods of bad channel conditions, ARQ leads to a temporary buffer jam and
thus to more variable packet delays. In contrast, FEC leads to more transmission
failures when the channel conditions are very bad. FEC is therefore used in cases
where timely delivery is more important than reliability, or else it must rely on
higher layers of the protocol to provide error-free transmission.
Whereas retransmission protocols have formed a part of numerous protocols
since the beginning of computer networks, the attention for energy saving and
sleep modes (the subject of the next paragraph) is of a more recent date.
1-6 INTRODUCTION
L L
L L L
T
0
T
0 0
2T 0
2T
0
4T
MS
BS
buffer
arriving
downlink
traffic
d
o
w
n
l
i
n
k
sleep
period 1
sleep
period 1
sleep
period 2
sleep
period 2
sleep
period 3
SLEEP SLEEP AWAKE
busy busy idle
AWAKE
Figure 1.2: Basic energy-saving scheme.
1.2.2 Energy-saving measures
Mobile devices in wireless networks are often battery-powered. Users highly value
a long battery autonomy and although battery technology has improved leaps and
bounds, substantial further gains can be made if there are protocols that support
further energy saving measures.
In this dissertation, we refer to sleep mode as the temporary and often partial
switch-off of the radio interface in order to save battery power. During this sleep
period, packets may arrive at the BS that are destined for the sleeping MS, but
these cannot be delivered until the MS has switched on the radio interface again.
Note that usually, the radio interface is not switched-off completely, and certain
types of emergency packets (such as initiation of hand-off procedures etc.) may
still be received even during the sleep mode. A complete switch-off is referred to
as idle mode. The reason why it is possible to save signicant amounts of power
by only selectively receiving messages is that the decoding of a packet is a very
energy-consuming activity, and by decoding only the (simpler and less numerous)
control messages and not the more complex data messages, a signicant fraction
of power can be saved. This arrangement requires that control and data messages
can be easily discerned (for example by having different headers).
The two emerging standards of broadband wireless internet, namely WiMAX
and LTE [4] of the Universal Mobile Telecommunications System (UMTS), each
offer a number of energy-saving measures. These measures are sometimes denoted
by the marketing term green telecommunications which is not really accurate, as
the main motivation of the industry is not environmental but rather to take away
one of the main impediments to a more widespread use of mobile devices, namely
short battery life. A rather more green solution would be if one would not have to
use these heavy-metal-containing batteries at all.
It is worth to note that sleep mode operation has other benets beyond battery
efciency: if a substantial fraction of the mobile stations is switched off, then this
INTRODUCTION 1-7
leads to greater bandwidth efciency, as mobile stations in sleep mode need only
a tiny bit of bandwidth.
Energy-saving mechanisms are referred to as sleep mode and idle mode in
WiMAX [7] and as discontinuous reception (DRX) in LTE [6]. In WiMAX, a
MSS in sleep mode is still registered to a BS and still performs hand-off pro-
cedures. In contrast, idle mode operation (which is optional in current WiMAX
standards) goes further and allows the MSS to be completely switched off and un-
registered with any BS, while still receiving broadcast trafc. In LTE, DRX can
be initiated when the UE is registered with an eNB (RRC CONNECTED: radio
resource control connected state) or when it is not (RRC IDLE: radio resource
control idle state). In the former case, the UE scans the neighboring eNBs when
it detects signal quality degradation with respect to the serving eNB and performs
handover when required by the signal quality. In this paper, we mainly consider
the sleep mode mechanism and DRX in RRC CONNECTED. In either case, the
MSS (or UE) turns itself off for predetermined periods of time which are negoti-
ated with the BS (eNB). In accordance with WiMAX terminology, these periods
are referred to as sleep windows.
Three sleep mode classes are dened by the WiMAX standard. When in Sleep
Mode Class I, the sleep window is progressively doubled in size from a prenegoti-
ated minimum to a prenegotiated maximum value. Having reached the maximum
sleep window length, this sleep window is repeated uninterruptedly, until trafc
arrives. This class is considered to be suitable for best-effort and non-real-time
trafc. Class II features xed-length sleep windows, that is the same sleep win-
dow size is repeated continuously until there is uplink or downlink trafc to be
transmitted. This class is typically employed for UGS (unsolicited grant services)
trafc. Finally, class III negotiates only a one-time sleep period. This is typically
used for management trafc, when the MSS knows when upcoming trafc is to be
expected.
In the case of LTE, the DRX cycle is started if there is neither download nor
upload trafc for a predened amount of time or a predened number of frames.
The choice of the lengths of the sleep windows differs considerably from WiMAX.
First, LTE initiates a closedown period in which the UE is not yet powered down.
Only when no trafc arrives during this period, the UE is powered down for short
periods, i.e. a small sleep window is chosen. If there is still no trafc after a
predened number of these short sleep windows, DRX switches to longer sleep
windows. By opting for a closedown period and a short window size rst, the UE
is allowed to wake up fast in case of unexpected data arrival immediately after the
DRX cycle is enabled.
There is quite a different mechanism that is proposed to be included in the
forthcoming IEEE 802.16m standard. In this scheme, sleep and awake periods are
organized in xed length cycles. When a mobile station goes into sleep mode, it
1-8 INTRODUCTION
does not do so for a predetermined period of time as in the other versions of sleep
mode, but rather for the rest of the current cycle.
All energy saving measures summarized in this section feature the same non-
trivial trade-off: the longer the sleep periods are chosen to be, the lower will be the
power consumption, but also the longer the additional delay.
1.3 Stochastic Modeling
This section aims to convey the general ideas rather than to be completely rigor-
ous. The rationale behind this choice is that technical aspects are elaborated on
in the papers, if and when they appear. Also it would be an overly ambitious and
misguided idea to try to encompass in just a dozen pages the many excellent books
that have appeared on stochastic modelling (and often on just one aspect of it).
We illustrate the concepts of stochastic modelling by means of the theory of
countable Markov chains. We aim to keep the exposition reasonably accessible,
with the idea that technical difculties are handled in the next chapters as they
appear. The choice for the countable Markov framework is motivated by the fact
that it does not bring many of the more technical issues (such as measurability,
topology, etc.) with it, and that much of it can be understood by a solid intuitive
(as opposed to mathematical) understanding of probability theory. We insert some
notes to show the simplications that can occur as we go downstream, to nite
Markov chains, or which extra complications arise when going upstream to more
general state spaces.
A common informal characterization of a Markov chain is that the future is
affected by the past only through the present. Consider a sequence of random
variables X
k
taking values in the set X indexed by a discrete time parameter k.
This sequence is said to satisfy the Markov property if:
P[X
k+1
= i[X
k
= i
k
, , X
0
= i
0
] =P[X
k+1
= i[X
k
= i
k
].
This has as a consequence that a Markov chain is completely characterized by
a transition matrix P with entries:
P
i j
=P[X
k+1
= j[X
k
= i]. (1.1)
It is useful to consider which processes are not Markovian, or cannot be for-
mulated as one. As in Markov chains the future depends only on the present state,
processes that look back a bounded number of instants can trivially be made into
Markov chains by augmenting the state space. Processes that do not forget their
past (i.e. in which new transitions depend on the entire history) therefore do not
t in the Markovian framework (often not even approximately). Still, Markovian
processes form a solid class that can model or approximate a good deal of different
situations.
INTRODUCTION 1-9
Note. On the distinction between Markov chains and Markov processes: in
general, Markov chains are understood to proceed in discrete time, regardless of
the space on which they operate. Markov processes in contrast are continuous-time
stochastic processes. A smaller number of authors makes the distinction between
the two on the basis of the state on which they are operating: if it is nite or
countable, then it is a Markov chain, else it is a Markov process.
End of Note.
States of countable Markov chains are commonly classied in three categories.
A state is called k-periodic if a Markov chain starting in the state only returns at
times that are multiples of k. For k = 1, the state is aperiodic. A state is called
recurrent if the return time is nite with probability one, and transient if it is not.
If one state of a Markov chain is periodic, then the Markov chain is called peri-
odic. Likewise, if one state is recurrent, then the Markov chain is called recurrent.
Recurrent Markov chains hence may have transient states, while transient Markov
chains must have all states transient. The theory of Lyapunov functions (in this
context also denoted as Foster-Lyapunov functions) offers an elegant framework
for studying transience and recurrence.
Lastly, states i and j are said to communicate if there exists a positive prob-
ability of going from i to j and from j to i. The communication relation forms
an equivalence, which divides the system into equivalence classes which in this
context are known as recurrence classes. A Markov chain with a single recurrence
class is called irreducible.
The basic quantitative result that can be obtained from Markov chains is the
distribution of the state under stationarity or steady state. For an irreducible Markov
chain with transition probabilities P
i j
, i, j X, the stationary distribution
i
is the
unique solution of the following equations:

j
=

iX

i
p
i j
and

iX

i
= 1.
The prevalence of this type of result is so great that the expression solving a
Markov chain is taken to mean without ambiguity nding its stationary distribu-
tion. Informally speaking, it tells us how the system behaves after a long period
of time, when the effects of the initial condition have faded away. Although very
popular, there is more to Markov chains than just their stationary distribution, as
evidenced by the fact that Markov chains with very different dynamics may have
the same stationary distribution. Nevertheless, many branches of applied probabil-
ity (such as queueing theory) have focused on stationarity, and so will we for the
most part of this dissertation.
It is useful to write the above equations in matrix form:
= P and 1 = 1, (1.2)
1-10 INTRODUCTION
where 1 denotes a column vector of appropriate size. Row and column vectors
each fulll a specic role in matrix expressions pertaining to Markov chains. Row
vectors represent distributions, i.e. they provide the probabilities of residing in
each of the states of the Markov chain. Also, an expression such as P
k
represents
the distribution after k steps, given that the initial distribution is denoted by .
Column vectors represent functions on the state space. Hence, P
k
f is a column
vector whose ith component represents the value of this function after the Markov
chain has run for k steps, given that we start from state i. Important examples of
such functions include [f]
i
= i
m
, so that f signies the kth moment over the state
space, and also [f]
i
=z
i
, which leads to the generating-functions approach and thus
to the rather successful marriage of linear algebra and complex algebra (also see
further on).
Often it is useful to combine these two equations into a single one. To this
purpose, let be any probability vector, then
(I P+1) = , (1.3)
is equivalent to the above. This reformulation hints at the important topic of the
Poissons equation in Markov chains.
(I P)f = g1g. (1.4)
In most cases the function g is given and then f is called the solution of the Pois-
sons equation. If a solution exists, then it is equal to f = Zg, where Z is called the
deviation matrix (or sometimes group inverse) and is given by:
Z =

n=0
[P
n
1] = (I P+1)
1
1, (1.5)
where the last equality is only guaranteed to make sense when the state space is
nite. Observe that the inverse of the matrix at the lefthand side of (1.3) appears
in the formula of Z, provided = . This matrix Z (and Poissons equation) is
important in surprisingly many contexts: in the theory of controlled Markov chains
(also known as Markov decision processes (MDPs)); in the derivation of average
rst hitting times; it appears in the calculation of moments in Chapter 2 under the
name of group inverse. Lastly, it appears in the light-trafc approximations of
Chapter 10.
The spectrum of the transition matrix P provides plenty of information on the
behavior of the corresponding Markov chain. First, note that 1 always belongs
to the spectrum of P, and according to the Perron- Frobenius theory, there are no
larger eigenvalues (in absolute value). The multiplicity of eigenvalue 1 gives the
number of recurrent classes. Also, if there is more than one eigenvalue on the unit
circle, then this indicates that the chain is periodic, the period being given by the
INTRODUCTION 1-11
number of eigenvalues. The spectral gap, i.e. the gap between the magnitudes
of the largest (in casu 1) and second largest eigenvalue gives an indication of the
speed with which a chain evolves towards the stationary distribution. Consider an
aperiodic and irreducible Markov chain. If the state space is nite, then so is the
number of eigenvalues and thus the spectral gap is strictly positive. This line of
reasoning does not hold anymore when the number of states is innite. For one
thing, the spectrum of innite matrices is only dened with respect to a norm (and
also depends on the specic norm that we choose). Also, as the cardinality of the
spectrum is innite, the second largest eigenvalue is no longer guaranteed to be at a
positive distance from the unit circle. The existence of a spectral gap is guaranteed
by a stronger form of ergodicity, called V-uniform ergodicity (see [14]).
1.3.1 Recursive characterizations of Markov chains
There is an alternative characterization of Markov chains. Consider an auxiliary
sequence of independent and identically distributed Y -valued random variables
Y
k
with common law
Y
. The sequence X
k
is a Markov chain if there exists a
function f : X Y X such that the following stochastic recursion holds:
X
k+1
= f (X
k
,Y
k
).
The recursive description is often both easier to construct and more intuitive
when investigating a specic problem. Also, a Monte-Carlo simulation of the
Markov chain is very straightforward: implement the function f and provide a
function that generates samples according to the law
Y
, then the following crude
Monte-Carlo method provides estimates of the expectation of g(X

) (provided that
the chain is ergodic):
E[g(X

)]
1
K
K

k=1
g(X
k
).
The advantage of organizing simulations in this way is that the state and the ran-
dom input are made explicit, which makes it easier to reason about and which
straties the use of variance-reduction techniques such as importance sampling.
The formulation in terms of transition matrices on the other hand seems more
useful when investigating Markov chains in general. Part of the skill in analyzing
Markov chains is knowing which framework is the most suitable one in the given
situation. Note that pairs ( f ,
Y
) satisfying a certain Markov chain model are not
unique.
Note. We can easily convert a recursive description of a Markov chain into a
transition matrix representation:
P
i j
=P[ f (i,V
k
) = j] =
V
(v V[ f (i, v) = j).
1-12 INTRODUCTION
The expression between braces is the pre-image of the function f (i, .). Hence the
conversion is, at least in theory, straightforward. If X and Y are nite then there
is a simple algorithm with [Y [[X[ runtime. This means that the computation of
the stationary distribution (with time complexity O([X[
3
) unless the Markov chain
has extra structure), far outweighs the determination of the transition probabilities,
hence, in cases where one has to resort to numerical computation of the stationary
probabilities, the recursive form is just as effective as the form based on probabil-
ities, and often more elucidating.
1.3.2 Background Markov chains
The recursive description of a Markov chain leads to an elegant characterization
of a background Markov chain. Suppose there exist functions f
1
and f
2
such that
X
k+1
= f
1
(X
k
,Y
k
); Y
k+1
= f
2
(Y
k
, Z
k
),
then the rst equation alone does not characterize a Markov chain, as the random
variables are not independent and identically distributed (iid). However, the two
equations together do form a Markov chain, with state space X Y , iid driving
sequence Z
k
, and recursive function f ((x, y), z) = ( f
1
(x, y), f
2
(y, z)).
The concept of background Markov chains allows to construct complex Markov
chains from simple ones in a modular fashion and it will frequently be used in this
dissertation to incorporate models for correlated arrival processes, correlated error
processes etc. into buffer models.
1.3.3 Censored Markov chains
For any subset A of X , we can consider a censored chain. Let X
A
n
be the stochas-
tic process whose nth transition is the nth time of the Markov chain X
k
being
in the set A. Informally speaking, the Markov chain evolves as before, but we
count only those time instants on which the Markov chain visits a state in A. Note
the difference with so-called quenched processes, in which we consider processes
conditioned on the fact that they stay in a certain subset A of the state space.
The rst fact about X
A
n
is that it is also a Markov chain, and has the transition
probability matrix P
A
P
A
= P
AA
+P
AA
c

P
A
c
A
c P
A
c
A
, (1.6)
where P
AB
(i, j) = P
i j
1
iA, jB
and

P
BB
=

k=0
P
k
BB
. Also, A
c
denotes the comple-
ment of set A, and 1

denotes the indicator function.
The power of this technique is perhaps not readily appreciated. The technique
of the embedded points can be considered as a rst important special case: some-
times the structure of the Markov chain is simpler when looking at certain event
times (called embedded points), which we let correspond to the states in set A (it
INTRODUCTION 1-13
may be necessary to introduce additional states). The following formula relates
the stationary distribution
A
at event times with the general distribution :
=
A
(I +P
AA
c

P
A
c
A
c ). (1.7)
The other (and in a way opposite) technique consists of augmenting the Markov
chain by adding supplementary variables, in the hope that this chain will be easier
to solve. The process of interest is then a censored Markov chain with respect to
the augmented chain.
These two tricks form the basic tools in a Markovian context with which
one can manipulate the model at hand into a favorable structure. They appear
throughout this dissertation in various forms.
Note. The method of the embedded points as well as the method of supplemen-
tary variables have a long history. In the earliest examples, the embedded points
formed a renewal process (i.e. iid interevent times). The renewal theorem is then
used to transform the distribution at embedded points to arbitrary points. Note that
the renewal theorem can be derived as a special case of formula (1.7) when applied
to the chain of forward recurrence times: for renewal lengths given by iid random
variables B
k
, we associate a Markov chain with state space 0, 1N, driving se-
quence Y
k
= B
k
and function f ( x, x, y) = (1
x=1
, (x1)1
x>1
+y1
x=1
). The set
A is given by (1, x), x N.
The censored formula in fact goes well beyond renewal processes, allowing
for a Markov dependence between event times. It can in turn be seen as a special
case of the (discrete-time) Palm inversion formula [10]. Indeed, the latter formula
allows to write distributions at marked time instants into distributions at event
times and hence it should not come as a surprise that (1.7) is a translation of the
inversion formula from the Palm framework to the Markov framework.
End of Note.
1.3.4 Some examples of Markov chains
We show some examples of Markov chains, of course paying extra attention to the
ones that play a role in this dissertation.
1.3.4.1 Two-state Markov chains
This is arguably the simplest possible Markov chain. The transition matrix P is
given by
P =
_
p 1p
1q q
_
. (1.8)
Almost every characteristic of Markov chains has a simple solution for this par-
ticular Markov chain (provided that it makes sense for two states): the stationary
distribution, the transient behavior, and so on.
1-14 INTRODUCTION
In this dissertation, it is used as a simple example of a background chain for
either the error process or the arrival process.
1.3.4.2 Finite Markov chains
Finite Markov chains form an obvious generalization of two-state Markov chains.
They also appear throughout this dissertation as a popular source of background
chains, modelling correlated error channels or correlated arrival streams.
It is an often overlooked fact that the stationary distribution of any nite Markov
chain has a closed-form expression. Indeed, consider for example Cramers rule,
which writes the solution of a linear system of equations as a division of two deter-
minants. This complies to the denition of a closed form formula, i.e. a formula
that is obtained by a nite number of elementary operations. This puts the exalted
status of closed-form expressions a bit into perspective: not only nite Markov
chains have routinely stationary distributions, it also (to some degree) devaluates
the value of a closed-form expression for the stationary distribution of an innite
Markov chain, if one keeps in mind that they are in many cases sold as theoretical
approximations (for example, buffer sizes are always nite in practice)!
The moral of this story of course is that one cannot go without numerical
considerations even for closed-form solutions: The number of operations should
be taken into account (which for the solution with Cramers rule increases very
quickly as a function of the number of states), as well as the numerical stabil-
ity: in the computation of determinants, many subtractions appear, so unless one
can substantially rewrite the expressions, a closed-form formula may have worse
numerical properties than a purely numerical algorithm. On the other hand, com-
puter algebra systems have changed the rules of the game a bit to the advantage of
closed-form expressions.
It should also be noted that other characteristics, such as the spectrum of P,
which is important for the transient behavior, can only in a few special cases be
solved in closed form.
1.3.4.3 Random walks
Random walks, and especially reected random walks, are the bread and butter of
queueing theory (and many other branches of applied probability as well). They
are characterized by the following simple updating formula:
X
k+1
= X
k
+V
k
. (1.9)
Random walks may evolve on one-dimensional as well as on multidimensional
spaces. The random walks that are most frequently seen in queueing theory are of
the reected type:
X
k+1
= (X
k
+V
k
)
+
. (1.10)
INTRODUCTION 1-15
where the notation (.)
+
denotes max(0, .). This is the most basic queueing model
in discrete time: X
k
models the evolution of the queue content from one time
slot to the next, where V
k
represents the net change in number of customers: the
number of arriving customers minus the number of possibly departing customers.
The transition probability matrix of such Markov chains is almost of a Toeplitz
form. Recall that Toeplitz matrices have equal diagonals. In the one dimensional
case, the transition probability matrix has the following form:
P =
_
p
0
p
q T
_
, (1.11)
where p
0
is a scalar, p a row vector, q a column vector, and T a Toeplitz ma-
trix. The suitability of the generating-functions method in this kind of structure is
largely due to the following property:
(T)z = (z)T(z), (1.12)
where z is a column vector (with zero-based indices) such that [z]
i
= z
i
and T(z)
is the generating function associated with the (innite) Toeplitz matrix: T(z) =

+
i=
T
k,i+k
z
i
, and (z) =
i

i
z
i
. This property allows to perform calculations
in either the generating functions domain (using the tools of complex analysis) or
with matrix computations (using the tools of linear algebra). From the point of
view of linear algebra, it boils down to the inversion of an innite Toeplitz matrix,
which corresponds to Wiener-Hopf factorization in the complex plane. We refer to
Latouche [9] and Asmussen [15] for further details. Generally, solution methods
are of an iterative nature, be it by identifying zeros of a power series (the complex-
analysis approach) or solving a linear system of equations.
We mention an important special case and a generalization. When the random
walk has the special property of being skip-free to the left, which means that V
k
is
equal to or greater than 1, then inversion of the problem is possible with direct
means.
The generalization is when the random variables V
k
are not iid, but rather
modulated by a nite background chain. The transition probability matrix is in this
case as in (1.11), but every scalar is substituted by a block of size NN, with N the
cardinality of the background chain. Combining this with the skip-free property,
then we get the popular matrix-analytic method. It is most useful to divide the state
space into two components: the state corresponding to the background process,
traditionally called the phase, and the level, which is the state corresponding to the
random walk. The transition probability matrix has the following block structure:
P =
_
_
_
_
_
_
_
B
0
B
1
B
2

A
0
A
1
A
2

A
0
A
1
A
2

A
0
A
1
A
2

.
.
.
.
.
.
.
.
.
_
_
_
_
_
_
_
. (1.13)
1-16 INTRODUCTION
Numerical solutions of this type of Markov chain revolve since the pioneering
work of Neuts [8] around the computation of the fundamental matrix G, which is
the smallest nonnegative solution of the matrix equation
G =

k=0
A
k
G
k
. (1.14)
The stationary probability vector = [
0
,
1
, ] that satises = P, can be
found by the equations

0
=
0
B

0
and
n
=
_

0
B

n
+
n1

i=1

i
A

ni+1
_
(I A

1
)
1
, for n > 0, (1.15)
where
A

n
=

i=n
A
i
G
in
, B

n
=

i=n
B
i
G
in
. (1.16)
There exist a lot of algorithms that aid in computing the desired quantities of
interest. Even a straightforward translation of formula (1.14) constitutes a solid
and numerically stable computational method, especially when the Markov chain
possesses a strong negative drift. There exist more sophisticated methods as well,
for an overview see e.g. [9].
Generating-functions methods are also possible, this time by considering the
function (i, j) z
i
e
j
, where i denotes the level and j denotes the phase, and e
j
denotes a column vector with M entries, with every entry equal to 0 except for the
jth, which is equal to 1. After some manipulations, we nd that = P has its
equivalent in the transform domain with the following functional equation:
(z) = ((z) (0))
1
z
A(z) +(0)B(z), (1.17)
where (z) =
i

i
z
i
, A(z) = A
i
z
i
and B(z) = B
i
z
i
.
1.3.4.4 ARMA models
In this section we make a short excursion into the world of the Markov chains with
a general (uncountable) state space. Apart from using a different language (row
vectors get replaced by measures; column vectors by measurable functions; transi-
tion probability matrices by kernels or operators), and a greater level of technicity
(for example irreducibility gets replaced by such concepts as -irreducibility and
-irreducibility), it is remarkable that almost every concept has its analogue in
this more general setting, see [14]. We focus on one particular example, called
ARMA, autoregressive moving average model, given by the following recursive
formulation:
INTRODUCTION 1-17
X
k+1
= AX
k
+V
k
, (1.18)
where X
k
,V
k
are complex-valued vectors of dimension N, and A is an N N ma-
trix. A frequent assumption is that V
k
, the noise vector, has a (complex-valued)
Gaussian distribution.
We mention this because it is a popular model for the channel gain over a
fading channel. The error probability P
k
is a function of this channel gain (for
more details see Chapter 4).
Skip-free reected random walks that are modulated by a Markov chain of this
general type, are known since Tweedie [14] to have an operator-analytic solution.
As compared to the matrix-analytic case however, the numerical task is daunting,
and only a few researchers have ventured into this area. In fact, just as with large
nite Markov chains which in principle have a closed-form stationary solution,
but which may better be solved using numerical methods, one may argue that the
numerical method reaches its breaking point for the operator-analytic models, and
are best tackled with (possibly adapted) Monte-Carlo methods.
In this dissertation, we investigated them with large and moderate deviations
instead (see Section ??).
1.3.5 Which method to choose?
We encountered three different ways of nding the stationary solution of Markov
chains: (1) the analytic method that aims at closed-formexpressions; (2) the linear-
algebra method; and (3) the Monte-Carlo method. Without a doubt, the last one
is the most generally applicable, but it also has some disadvantages. As not the
exact values are obtained, but rather a series of random samples, statistical analysis
is needed to ensure that no improper conclusions are drawn. The sensitivity of
input parameters on the output, or the probability that certain rare events occur
are hard to obtain using naive Monte-Carlo techniques, and may only be obtained
through the implementation of specialized algorithms. Optimization is also a hard
task. The analytic method scores best on these points, with the numerical method
holding middle ground.
If we look at very large nite Markov chains, analytic methods are undesir-
able (safe for some special kind of structure). Whereas Monte-Carlo simulations
show good scalability with respect to the complexity of the model (as measured
by the number of states) (typically logarithmic dependence is observed), and slow
convergence (typically with rate N
1/2
, where N is the number of samples), the
linear algebra techniques suffer from the reverse. For direct methods, the rate of
convergence is not applicable but the complexity is approximately K
3
where K
is the number of states. When the Markov model has the desirable property that
its probability matrix is sparse, and thus vector-matrix computations may be per-
1-18 INTRODUCTION
formed efciently (linearly in the most advantageous case), then one can apply
iterative methods that exhibit a worse scalability but a better convergence rate.
This reects conventional wisdom that says that linear algebra methods are great
for intermediate problems whereas simulation-based methods are better for very
large problems.
1.3.6 Large Deviations
Large deviations theory is the study of exponential decay in rare events. We refer
the reader to a.o. [12] and [13] for references on this topic.
What is a large deviations result? Given a sequence of probability measures

n
, large deviations say when

n
(A) K(n)exp(s
n
inf
xA
I(x)),
is a good approximation for large values of n. In this formula, I is known as the rate
function, and s
n
is the scaling factor or speed. Often s
n
=n. K(n) is a function that
decays slower than exp(s
n
) about which large-deviations theory usually gives no
information.
A number of key features of large-deviations theory can readily be deduced
from even this informal denition. Firstly, large deviations is about exponential
decay in one key parameter of the model. It gives exponential asymptotics instead
of complete asymptotics. Secondly, large-deviations transforms a probabilistic
problem into an optimization problem (in order to extract information, one has to
compute the inmum of a function over a set). Thirdly, the setup with the sequence
of probability measures is exible and permits the study of many different limits,
including ones that go beyond what is usually understood as tail asymptotics (also
see below).
The habitual formulation of large deviations results is more formal. A se-
quence of random variables X
L
in a Hausdorff space X with Borel sigma alge-
bra B is said to satisfy a large deviations principle with good rate function I if for
any B B,
inf
xB
o
I(x) liminf
n
1
s
n
log
n
(B)
limsup
n
1
s
n
log
n
(B) inf
x

B
I(x)
where I : X R has compact level sets, and B
o
and

B denote the interior
and closure of the set B respectively.
There are three levels in large deviations theory (as introduced by the great
pioneers Varadhan and Stroock):
INTRODUCTION 1-19
Level I large deviations concerns the study of deviations from the sample
mean. Cram ers Theorem forms the foundation of this branch of the theory.
Level II large deviations considers the probability of empirical distributions
deviating from the true distribution. Sanovs theorem is central here.
Finally, level III deviations is the most technical and considers the large
deviations on processes: more specically, it is concerned with the proba-
bilities that a process deviates from its most likely path (as predicted by uid
limits or law of large numbers arguments), and follows another path. It is in
this branch (and mostly only in this branch) that the full complexity of large
deviations theory and the connection with topology is needed.
Let us compare the large-deviations asymptotic with the more traditional tail
asymptotics. Tail asymptotics in a queueing context are usually understood to
be of the following form. Consider the probability that the buffer content U in
stationarity exceeds a certain level n, then the expression f (n) is the (exact) tail
asymptotic of the buffer content if
lim
n
P[U > n]
f (n)
= 1. (1.19)
Large-deviations theory can (partially) provide the same answers by considering
the sequence of measures
n
(A)
.
= P[
U
n
A]. We write partially since we obtain
exponential asymptotics instead of exact asymptotics. This is called the large-
buffer or fast-time limit.
It is important to note that this is not the only possibility. A particularly in-
teresting and often applied alternative is the many-sources limit. In this setup, we
consider a sequence of arrival processes, where the nth arrival process is equal to
n iid superpositions of the original arrival process. The nth measure
n
considers
a system where (1) the arrival process is replaced by n iid superpositions of the
original arrival process; (2) the service rate is n times faster as well. This limit
captures the time scales better than the fast-time scaling (see [13]).
1.3.7 Series expansions
In this section we briey introduce the third methodology that is applied in this
dissertation. Our experience with this methodology is still increasing, with one
chapter applying this technique having been included.
Series expansions are a way of obtaining results for models that are hard to
solve explicitly. For example, consider the situation where a model depends on
a parameter such that for = 0 we can nd an explicit solution, but not for
,= 0. It is a natural idea to consider expansions around = 0, so that an expres-
sion of the form c

= c
(0)
+c
(1)
+
2
c
(2)
+ provides a good approximation
1-20 INTRODUCTION
of the performance measure c

for sufciently small values of . This idea has


a broad applicability in various branches of applied mathematics and engineering.
Also in the specic context of stochastic modelling it has been applied in many
forms and avors, and it is known under various names as perturbations, light-
trafc expansions, Taylor-series expansions and so on. As with large-deviations
theory, this technique is in principle not conned to the Markovian framework
(see e.g. [16], which uses the Palm theory), although many interesting examples
indeed fall within the Markovian framework. There is a distinction between regu-
lar perturbations, which means that the Markov chain is irreducible in = 0, and
singular perturbations, which means that the Markov chain is reducible in = 0
(but typically irreducible in the neighborhood of = 0). In this dissertation, we
focus on regular perturbations.
There are roughly three methods to establish series expansions of stochastic
models. The rst makes use of sample-path arguments. Consider for example the
case that denotes a particular event rate (for example, for light-trafc approxima-
tions, denotes the arrival rate). One may apply the non-rigorous reasoning that a
sample path containing n such events has a probability of order
n
, and hence, for
an nth order expansion, only sample paths with n or less events must be consid-
ered. For many specic examples, this is indeed the case. However due to the fact
that the number of sample paths is uncountable and thus the probability of every
individual path is zero, making this rigorous is non-trivial.
The second approach involves formulating the model in the transform domain,
which leads to a functional equation for the generating function of the stationary
distribution. Symbolic manipulations may then allow to identify a power series
expression that solves the functional equation.
The third approach is via the following updating formula, which has been es-
tablished in very general Markovian settings, see eg. [17]:

=
0

k=0
[(P

P
0
)Z]
k
where P

denotes the transition probability matrix parametrized in and Z denotes


the deviation matrix of P
0
, illustrating another application of this concept. In this
case, the derivation revolves around nding this deviation matrix (or deviation
kernel in case of general Markovian state spaces).
Personal taste and the specic model determine the preferred approach.
1.4 Dissertation Outline
In this section we lay out how the chapters of this dissertation t together. As
mentioned in the previous section, we look at two main problems arising in the
context of wireless networks through the lens of three methodologies.
INTRODUCTION 1-21
Chapter 1:
Introduction
Chapter 2: Delay
of Go-Back-N
Chapter 3: Uni-
ed ARQ analysis
Chapter 4:
Moderate
deviations over
fading channels
Chapter 5:
Large deviations
of SR rese-
quencing buer
Chapter 6:
Sleep-mode
analysis with
Markov arrivals
Chapter 7:
802.16e sleep
mode op-
timization
Chapter 8:
Sleep mode with
interruptions
Chapter 9: Sleep
mode for 802.16m
Chapter 10:
Sleep mode
under light trac
Chapter 11:
Large-deviations
analysis of
sleep mode
Retransmission
protocols
goal: fewer
packet errors
Sleep mode
goal: better en-
ergy eciency
Figure 1.3: Overview of this dissertation.
1-22 INTRODUCTION
We start with the part on retransmission protocols.
The earliest research done as part of this dissertation was the delay analysis
of the go-back-N protocol (Chapter 2) over a nite-state Markovian error channel.
We built on existing studies that assume an independent error process. We found
great utility in building the analysis around the pgm of the service time.
In Chapter 3, we tried to stretch this concept of the service time pgmfurther and
analyse thus an entire class of protocols. We work according to a tried and tested
recipe of combinatorics. We dene a simple formal language along with rules to
transform this language into an mgf of the length of words of such languages.
The last two chapters on retransmission protocols make use of large-deviations
theory. In Chapter 4, we look beyond the nite-state Markovian models that we
hitherto considered. Specically, we look at the fading models as described in
the previous section, which are popular among communications theorists. The
strategy taken thus far when applying queueing theory to systems involving fading
channels has always been to reduce the occurring stochastic processes to nite-
state Markov chains, either by proposing a certain geometry and then estimate
the transition probabilities or else by discretizing non-Markovian processes or un-
countable Markov chains. In either way the problem of these approaches is that
there is a veil between the parameters of the fading process and the performance
measures resulting from the queueing analysis, that can only be lifted through
quite heavy numerical work. Also, it is not clear to determine whether a given
nite-state Markov chain is a good approximation of the original process. Also,
it is an unresolved question whether the performance measure we are interested
in affects the optimal approximating Markov chain, and if it does, to what extent.
One can imagine for example, that the optimal chain is different when the mean
delay is sought as compared to when tail asymptotics are wanted.
Chapter 5, the nal chapter of the part on retransmission protocols stands apart
for several reasons. Firstly, it considers the resequencing buffer at the receivers
side instead of the transmitter buffer. Secondly, this type of buffer cannot be anal-
ysed by a random-walk type Markov chain.
The part on sleep mode starts off in Chapter 6 with an analysis of sleep mode
as included in the IEEE 802.16e standard. Its use of the method of the embedded
points is not followed through to the end, unlike in further chapters. The results of
the analysis include expressions for the distribution of the buffer content and the
packet delay, as well as expressions for the mean power consumption.
Having performed the analysis of the two main elements of the performance
tradeoff, i.e. the mean delay and the mean power consumption, we explicitly in-
vestigated the question of optimization in Chapter 7.
Next, in Chapter 8 we look at the problem of interrupted sleep periods: if
there is trafc in the uplink direction as well, then it makes no sense to continue
the ongoing sleep period. The resulting interruptions have a quite severe impact
INTRODUCTION 1-23
on the performance. In this chapter, we take a more extreme vantage point with
respect to the method of the embedded points than we did in previous sleep mode
analyses.
Chapter 9 is devoted to the study of a different sleep-mode mechanism as pro-
posed by the companies Samsung and Nokia. The new mechanism is different as
it organizes sleep and awake modes in cycles of xed length. This leads to some
extra problems, both in terms of analysis as in terms of computation. Indeed, the
occurrence of listening and closedown intervals leads to a large number of corner
cases so that the analysis should be conducted very carefully. Secondly, as com-
pared to the analyses of the previous chapters, the cardinality of the phase set is a
lot larger.
Lastly, we look at two alternative methodologies to assess the performance
of the sleep mode protocol. Although the classical method as elaborated in the
rst four chapters on sleep mode leads to quite efcient results, there is room for
improvement on two accounts. Firstly, it would be interesting to have a direct
(i.e., non-numerical) assessment of the inuence of arrival correlation (in terms
of the autocorrelation function for example). Secondly, scenarios with multiple
mobile stations, although not quite out of reach of these methods, require so much
additional calculations that the utility of the analysis as compared to simulations
or numeric inversion seems limited.
Notice that many of the customary approximation techniques lter away the
effects of sleep mode and are thus unsuited. As we remarked in the introductory
paragraphs on approximation, many of these techniques work by neglecting the
boundary, that is by neglecting what happens when the system gets empty, and
this boundary is exactly where sleep mode differs from standard models.
For example, heavy trafc considers the limit of the load of the system going
to 1. Of course, this also means that the fraction of the time that the queue is empty
gets also vanishingly small. Hence, the occurrence of sleep mode becomes a rare
event whose impact can be neglected. (Indeed, this is also obvious by looking
at plots of mean delay versus load of sleep-mode models). Also tail asymptotics
do not show the inuence of sleep mode: the probability of reaching a very large
delay for example will be dominated by the queueing delay, not by the sleep delay.
Therefore, we consider in Chapter 10 light-trafc approximations. This makes
perfect sense as the sleep-mode mechanism is most commonly applied if the load
is quite small. Moreover, the biggest effect is seen under exactly this condition. In
the chapter, we follow the sample-path approach to light-trafc approximations,
but in the notes we give some comments on the relative advantages of each of the
methods.
We conclude the part on the sleep-mode mechanism just as we did the part on
retransmission protocols, namely with a chapter on large-deviations (Chapter 11).
The aim of the chapter is to give an assessment of rare overow events, especially
1-24 INTRODUCTION
in scenarios with multiple mobile stations. The large-deviations analysis moves the
problem from the stochastic domain to a deterministic optimization problem. This
problem can be solved for systems with one mobile station and also for multiple
mobile stations (although the solution gets increasingly messy). For a higher num-
ber of mobile stations, we propose a novel numerical method to nd the overow
decay rates and the most likely path to overow.
1.5 Main contributions
In this section, we list some of what we consider to be the main contributions of
the doctoral research.
We performed a unifying analysis of a number of ARQ protocols by identi-
fying the central role played by service length pgms and showed that these
pgms are central in both an exact queueing analysis, as well as in the through-
put and in the asymptotic variance and the rate function, which play a role
in resp. heavy-trafc approximations and large-deviations theory.
We performed an asymptotic buffer analysis for a retransmission system
with a Gaussian fading (non-Markovian) error model.
We performed a large-deviations analysis of an ARQ resequencing buffer,
and laid bare a hitherto unnoticed link with the famous coupon-collector
problem.
To the best of our knowledge, we were the rst to study the sleep mode
mechanism with a correlated arrival process. However, it is intuitively clear
that correlated arrivals have an inuence on the best way to organize sleep
mode. This is also evidenced by our numerical examples (Chapter 6).
We looked at the important question of how we should optimally choose the
sleep mode parameters (Chapter 7).
We studied also another aspect that was previously neglected, namely the
inuence of uplink trafc on the sleep mode mechanism (Chapter 8).
During our investigations, we encountered a proposal for a new sleep mode
mechanism backed by some well-known companies and we performed an
analysis for this new proposal as well (Chapter 9).
We found a scaling that effectively captures the sleep mode mechanism and
performed a large-deviations analysis on it.
INTRODUCTION 1-25
We investigated a numerical method for solving the optimization problem
that arises from the large-deviations analysis. Numerical methods are (a
bit surprisingly) rarely studied, especially in a telecommunications context.
The approach we chose was to discretize the problem space and perform
Dijkstras algorithm on the resulting grid.
1.6 Publications
The research performed during the doctoral research led to a number of publica-
tions in recognised international research fora journals and conferences.
1.6.1 Publications in international journals
1. K. DE TURCK, S. WITTEVRONGEL, Delay analysis of the go-back-N
ARQ protocol over a time-varying channel, Lecture Notes in Computer Sci-
ence, 2005, vol. 3670, pp. 124-138.
2. K. DE TURCK, S. DE VUYST, D. FIEMS, S. WITTEVRONGEL, Perfor-
mance analysis of the IEEE 802.16e sleep mode for correlated downlink
trafc, Telecommunication Systems, 2008, vol. 39, no. 2, pp. 145-156.
3. S. DE VUYST, K. DE TURCK, D. FIEMS, S. WITTEVRONGEL, H. BRU-
NEEL, Delay versus energy consumption of the IEEE 802.16e sleep-mode
mechanism, IEEE Transactions on Wireless Communications, 2009, vol. 8,
no. 11, pp. 5383-5387.
4. K. DE TURCK, S. WITTEVRONGEL, Receiver buffer behavior for the se-
lective repeat protocol over a wireless channel: an exact and large-deviations
analysis, Journal of Industrial and Management Optimization, 2010, vol. 6,
no. 3, pp. 603-619.
5. K. DE TURCK, S. WITTEVRONGEL, Delay and stability aspects of the
go-back-N ARQ scheme for correlated error channels, Wireless Personal
Communications, accepted for publication.
6. D. FIEMS, B. PRABHU, K. DE TURCK, Analytic approximations of queues
with lightly- and heavily-correlated autoregressive service times, Annals of
Operations research, accepted for publication.
7. S. DE CLERCQ, K. DE TURCK, B. STEYAERT, H. BRUNEEL, Frame-
bound priority scheduling in discrete-time queueing systems, Journal of In-
dustrial and Management Optimizations, 2011, vol. 7, no.3.
1-26 INTRODUCTION
8. E. MOROZOV, W. ROGIEST, K. DE TURCK, D. FIEMS, H. BRUNEEL,
Stability of Multi-Wavelength Optical Buffers with Delay-Oriented Schedul-
ing, submitted to European Transactions on Telecommunications.
9. K. DE TURCK, S. DE VUYST, D. FIEMS, S. WITTEVRONGEL, H. BRU-
NEEL, Performance Analysis of Sleep Mode Mechanisms in the Presence
of Bidirectional Trafc, submitted to IEEE Transactions on Wireless Com-
munications.
10. K. DE TURCK, S. DE VUYST, D. FIEMS, H. BRUNEEL, S. WITTEVRON-
GEL, Efcient performance analysis of newly proposed sleep-mode mecha-
nisms for IEEE 802.16m in case of correlated downlink trafc, submitted to
Performance Evaluation.
1.6.2 Book chapters
1. K. DE TURCK, S. WITTEVRONGEL, Delay analysis of Go-Back-N ARQ
for correlated error channels, chapter 11 in: Mobility Management and
Quality-of-Service for Heterogeneous Networks, River Publishers, Aalborg,
2009 (ISBN: 978-87-92329-20-2), pp. 245-267.
1.6.3 Papers in Proceedings of International Conferences
1. K. DE TURCK, S. WITTEVRONGEL, Delay and stability aspects of the
go-back-N ARQ scheme for correlated error channels, Proceedings of the
Third International Working Conference on Performance Modelling and Eval-
uation of Heterogeneous Networks, HET-NETs 05 (Ilkley, 18-20 July 2005),
Technical Proceedings, pp. P47/1-P47/10.
2. K. DE TURCK, S. WITTEVRONGEL, A unied analysis of ARQ proto-
cols over a correlated channel, Proceedings of the Third Euro-NGI Work-
shop on New Trends in Modelling, Quantitative Methods and Measurements
(Torino, 22-23 June 2006), paper 15, pp. 1-4.
3. K. DE TURCK, S. WITTEVRONGEL, Amatrix-geometric approach to dis-
tributional forms of Littles theorem, Proceedings of the Fourth International
Working Conference on Performance Modelling and Evaluation of Hetero-
geneous Networks, HET-NETs 06 (Ilkley, 11-13 September 2006), Part I:
Research Papers, pp. P33/1-P33/9.
4. K. DE TURCK, S. WITTEVRONGEL, Buffer analysis for a number of
ARQ protocols with correlated errors, Proceedings of the Nineteenth Be-
larusian Winter Workshop on Queueing Theory, BWWQT-2007 (Grodno,
29 January - 1 February 2007), pp. 35-41.
INTRODUCTION 1-27
5. K. DE TURCK, S. DE VUYST, D. FIEMS, S. WITTEVRONGEL, Analy-
sis of IEEE 802.16e sleep mode with correlated trafc, Proceedings of the
Euro-FGI Workshop on New Trends in Modelling, Quantitative Methods
and Measurements (Ghent, 31 May 1 June 2007), pp. 75-78.
6. K. DE TURCK, S. DE VUYST, S. WITTEVRONGEL, Performance of a
class of retransmission protocols in case of variable feedback delays, Pro-
ceedings of the Fourteenth International Conference on Analytical and Stochas-
tic Modelling Techniques and Applications, ASMTA2007 (Prague, 4-6 June
2007), pp. 50-55.
7. K. DE TURCK, S. DE VUYST, D. FIEMS, S. WITTEVRONGEL, An an-
alytic model of IEEE 802.16e sleep mode operation with correlated trafc,
Proceedings of the 7th International Conference on Next Generation Tele-
trafc and Wired/Wireless Advanced Networking, NEW2AN 2007 (Saint-
Petersburg, 10-14 September 2007), Lecture Notes in Computer Science,
2007, vol. 4712, pp. 109-120.
8. K. DE TURCK, S. WITTEVRONGEL, Performance of the selective repeat
protocol over a wireless link: analysis and computation, Proceedings of the
International Science and Training Conference Modern Computer and In-
formation Technologies, mcIT-2008 (Grodno, 21-24 April 2008), Part 2, pp.
7-10.
9. K. DE TURCK, D. FIEMS, S. WITTEVRONGEL, H. BRUNEEL, Perfor-
mance of sleep-mode mechanisms under light-trafc conditions, Proceed-
ings of the Twentieth Belarusian Winter Workshop on Queueing Theory,
BWWQT-2009 (Minsk, 26-29 January 2009), pp. 30-35.
10. K. DE TURCK, S. ANDREEV, S. DE VUYST, D. FIEMS, S. WITTEVRON-
GEL, H. BRUNEEL, Performance of the IEEE 802.16e sleep mode mech-
anism in the presence of bidirectional trafc, Proceedings of the First Inter-
national Workshop on Green Communications, GreenComm 09, Workshop
of the IEEE International Conference on Communications, ICC 2009 (Dres-
den, 14-18 June 2009), pp. 428-432.
11. K. DE TURCK, S. WITTEVRONGEL, Exact and large-deviations analy-
sis of the receiver buffer of selective repeat retransmission protocols over
correlated channels, Proceedings of the Fourth International Conference on
Queueing Theory and Network Applications, QTNA 2009 (Singapore, 29-
31 July 2009), paper 20, pp. 1-7.
12. K. DE TURCK, S. DE VUYST, D. FIEMS, S. WITTEVRONGEL, H. BRU-
NEEL, Performance of the sleep-mode mechanismof the newIEEE 802.16m
1-28 INTRODUCTION
proposal for correlated downlink trafc, Proceedings of the Third Euro-
NF International Conference on Network Control and Optimization, NET-
COOP 2009 (Eindhoven, 23-25 November 2009), Lecture Notes in Com-
puter Science, 2009, vol. 5894, pp. 152-165.
13. K. DE TURCK, M. MOENECLAEY, S. WITTEVRONGEL, Moderate de-
viations of retransmission buffers over a wireless fading channel, Proceed-
ings of the Sixth International Working Conference on Performance Mod-
elling and Evaluation of Heterogeneous Networks, HET-NETs 2010 (Za-
kopane, 14-16 January 2010), pp. 367-378.
14. W. ROGIEST, K. DE TURCK, K. LAEVENS, S. WITTEVRONGEL, H.
BRUNEEL, Contention resolution for optical switching: tuning the channel
and delay selection algorithm, Proceedings of the Ninth International Con-
ference on Networks, ICN 2010 (Menuires, 11-16 April 2010), pp. 10-15.
15. S. DE CLERCQ, K. DE TURCK, B. STEYAERT, H. BRUNEEL, ADiscrete-
Time Queueing System under Frame-bound Priority, Proceedings of the 5th
International Conference on Queueing Theory and Network Applications
(QTNA2010), Beijing, July 24-26, 2010, p.112-119.
16. W. ROGIEST, K. DE TURCK, K. LAEVENS, D. FIEMS, S. WITTEVRON-
GEL, H. BRUNEEL, Performance analysis of FDL buffers: a heuristic ap-
proach with impatience and quantization, Proceedings of the 5th Interna-
tional Conference on Queueing Theory and Network Applications, QTNA
2010 (Beijing, 24-26 July 2010), pp. 148-155.
17. K. DE TURCK, D. FIEMS, S. DE VUYST, H. BRUNEEL, S. WITTEVRON-
GEL, Large-deviations analysis for energy-saving mechanisms in wireless
networks, Proceedings of the 5th International ICST Conference on Perfor-
mance Evaluation Methodologies and Tools, Valuetools 2011 (Ecole Nor-
male Suprieure de Cachan, 16-20 May 2011).
18. K. DE TURCK, D. FIEMS, S. WITTEVRONGEL, H. BRUNEEL, Taylor-
series expansions for queueing systems with train arrivals, Proceedings of
the 5th International ICST Conference on Performance Evaluation Method-
ologies and Tools, Valuetools 2011 (Ecole Normale Suprieure de Cachan,
16-20 May 2011).
19. W. ROGIEST, K. DE TURCK, K. LAEVENS, D. FIEMS, H. BRUNEEL, S.
WITTEVRONGEL, Optimized channel and delay selection for contention
resolution in optical networks, Proceedings of the IEEE International Con-
ference on Communications, ICC 2011 (Kyoto, 5-9 June 2011).
INTRODUCTION 1-29
20. V. PLA, S. DE VUYST, K. DE TURCK, E. BERNAL-MOR, J. MARTINEZ-
BAUSET, S. WITTEVRONGEL, Saturation throughput in a heterogeneous
multi-channel cognitive radio network, Proceedings of the IEEE Interna-
tional Conference on Communications, ICC 2011 (Kyoto, 5-9 June 2011).
1.6.4 Abstracts and Other Presentations
1. K. DE TURCK, S. WITTEVRONGEL, Effect of error correlation on the
performance of Go-Back-N automatic repeat request protocol, Book of Ab-
stracts of the Twentieth Conference on Quantitative Methods for Decision
Making, ORBEL 20 (Ghent, 19-20 January 2006), pp. 31-32.
2. K. DE TURCK, S. DE VUYST, S. WITTEVRONGEL, Buffer analysis of a
class of retransmission protocols with variable feedback delays, Book of Ab-
stracts of the Euro-NGI Workshop on Transfer Control with Delayed Feed-
back (Eindhoven, 11 November 2006), p. 3.
3. K. DE TURCK, S. WITTEVRONGEL, Performance analysis of retransmis-
sion protocols in a wireless environment, Book of Abstracts of the 7th FirW
PhD Symposium (Ghent University, 29 November 2006), p. 91.
4. K. DE TURCK, S. WITTEVRONGEL, System content versus system time
in discrete-time queueing systems, Book of Abstracts of the Second Benelux
Workshop on Performance Analysis of Communication Systems (Antwerp,
5-6 February 2007), p. 13.
5. K. DE TURCK, S. DE VUYST, D. FIEMS, S. WITTEVRONGEL, H. BRU-
NEEL, Optimal sleep mode strategies for IEEE 802.16e wireless links, Book
of Abstracts of the 22nd Conference on Quantitative Methods for Decision
Making, ORBEL 22 (Brussels, 16-18 January 2008), pp. 57-58.
6. K. DE TURCK, S. WITTEVRONGEL, Analysis of discrete-time resequenc-
ing buffers, Book of Abstracts of the Seventh International Petrozavodsk
Conference on Probability Methods in Discrete Mathematics (Petrozavodsk,
2-6 June 2008), pp. 766-767.
7. S. DE VUYST, K. DE TURCK, D. FIEMS, S. WITTEVRONGEL, H. BRU-
NEEL, QoS versus energy efciency of the IEEE 802.16e sleep-mode mech-
anism, Presentation at the Second Euro-NF Plenary Meeting (Kaiserslautern,
3-5 June 2009).
8. K. DE TURCK, W. ROGIEST, D. FIEMS, H. BRUNEEL, S. WITTEVRON-
GEL, Fluid limits and maximal load in multi-wavelength optical buffer sys-
tems, Book of Abstracts of the Third Madrid Conference on Queueing The-
ory, MCQT 10 (Toledo, 28 June 1 July 2010), p. 30-31.
1-30 INTRODUCTION
9. K. DE TURCK, D. FIEMS, S. WITTEVRONGEL, H. BRUNEEL, Light-
trafc analysis of queueing systems with train arrivals, Book of Abstracts of
the 5th International Workshop on Applied Probability, IWAP 2010 (Madrid,
5-8 July 2010), pp. 129-130.
10. D. FIEMS, K. DE TURCK, S. WITTEVRONGEL, H. BRUNEEL, Taylor-
series expansions for queueing systems with train arrivals, Poster at the IFIP
WG 7.3 Conference on Computer Performance, Modeling, Measurements
and Evaluation, PERFORMANCE 2010 (Namur, 17-19 November 2010),
pp. 13-14.
1.6.5 Awards
1. Best Paper Award at the Fourth International Conference on Queueing The-
ory and Network Applications, QTNA 2009, for the paper: Exact and large-
deviations analysis of the receiver buffer of selective repeat retransmission
protocols over correlated channels (with S. Wittevrongel).
INTRODUCTION 1-31
References
[1] IEEE 802.16e-2005, Part 16: Air interface for xed and mobile broadband
wireless access systems Amendment 2: physical and medium access con-
trol layers for combined xed and mobile operation in licensed bands
Corrigendum 1, February 2006.
[2] The Draft IEEE 802.16m System Description Document, IEEE 802.16m-
08/003r4, July 2008.
[3] B. Li, Y. Qin, C.P. Low, C.L. Gwee, A survey on Mobile WiMAX, IEEE
Communications Magazine, Vol. 45, No. 12, pp. 7075, 2007.
[4] 3GPP TS 36.300 E-UTRAN Overall description Stage 2, v. 10.1.0,
September, 2010.
[5] Y. Xiao, Energy saving mechanism in the IEEE 802.16e wireless MAN,
IEEE Communications Letters, Vol. 9, No. 7, 2005, pp. 595597.
[6] C. Bontu and E. Illidge, DRX mechanism for power saving in LTE, IEEE
Communications Magazine, Vol. 47, No. 6, pp. 4855, 2009.
[7] Y. Xiao, Energy saving mechanism in the IEEE 802.16e wireless MAN,
IEEE Communications Letters, Vol. 9, No. 7, 2005, pp. 595597.
[8] M.F. Neuts, Structured stochastic matrices of M/G/1 type and their appli-
cations, volume 5 of Probability: Pure and Applied. Marcel Dekker Inc.,
New York, 1989.
[9] D.A. Bini, G. Latouche, B. Meini, Numerical methods for structured
Markov chains, Oxford University Press, 2005.
[10] F. Baccelli, P. Br emaud, Elements of Queueing Theory, Springer Berlin,
second edition. 2003.
[11] B. Meini. An improved FFT-based version of Ramaswamis formula.
Stochastic Models, Vol. 13, No. 2 (1997) pp. 223238.
[12] A. Dembo and O. Zeitouni. Large deviations techniques and applications,
volume 38 of Applications of Mathematics (NewYork). Springer-Verlag, New
York, second edition, 1998.
[13] A. Ganesh, N. OConnell, and D. Wischik. Big Queues. Springer, 2004.
[14] S. Meyn, R. Tweedie, Markov chains and stochastic stability, 2nd Edition
(2009) Springer-Verlag, London.
1-32 INTRODUCTION
[15] S. Asmussen, Applied Probability and Queues, second edition Springer,
Berlin, 2003.
[16] B. Baszczyszyn, Factorial-moment expansion for stochastic systems.
Stoch. Proc. Appl. 56, 321-335 (1995).
[17] B. Heidergott and A. Hordijk. Taylor Series Expansions for Stationary
Markov Chains. Advances in Applied Probability, 35(4):10461070, 2003.
2
Delay Analysis of Go-Back-N ARQ
for Correlated Error Channels
Koen De Turck and Sabine Wittevrongel
published as chapter 11 in: Mobility Management and Quality-of-Service for
Heterogeneous Networks, River Publishers, 2009 (ISBN: 978-87-92329-20-2).
Abstract. We investigate the performance of the Go-Back-N ARQ (Automatic
Repeat reQuest) protocol over a wireless channel. Data packets are sent from
transmitter to receiver over the wireless transmission channel. When a packet is
received, the receiver checks whether it has been received correctly or not, and
sends a feedback message to notify the transmitter of the condition of that packet.
When the transmitter is notied of a transmission error, the incorrectly received
packet is sent again, as well as every following packet.
Our modeling assumptions are based on two convictions. On the one hand, a
good view of the performance of an ARQ protocol not only requires an analysis
of the throughput, but also of the buffer behavior. Therefore, we offer a complete
queueing analysis of the transmitter buffer, in addition to a throughput analysis.
Secondly, due to the highly variable nature of the error process in wireless net-
works, we have to take error correlation explicitly into account.
Hence, we model the channel by means of a general Markov chain with M
states and a xed error probability in every state. The transmitter buffer is mod-
2-2 DELAY ANALYSIS OF GO-BACK-N ARQ
eled as a discrete-time queue with innite storage capacity and independent and
identically distributed packet arrivals from slot to slot.
We nd concise expressions for the probability generating functions of the un-
nished work and the packet delay of the transmitter buffer. Furthermore, we show
explicit expressions for the mean and the variance of both system characteristics
and we derive some heavy-load approximations. Finally, we provide some numer-
ical examples.
2.1 Introduction
A popular way of protecting against transmission errors is the so-called Automatic
Repeat reQuest (ARQ). For this mechanism to work, the transmitter adds a simple
error checking code to each packet so the receiver can detect the most common
transmission errors. When a packet is received, the receiver checks whether it has
been received correctly or not, and sends a feedback message to notify the trans-
mitter of the condition of that packet. That is, a positive acknowledgement (ACK)
is sent in case of a correct transmission; a negative acknowledgement (NAK) is
sent in case of a transmission error. This organization requires a bi-directional
channel between the sender and the receiver. Since not all arriving packets can be
sent immediately and moreover the transmitter has to keep a copy of each packet
until it is correctly transmitted, a buffer to store packets is required at the transmit-
ter side.
Various types of ARQ protocols have been proposed in the literature [1]. They
differ in the way the transmissions and retransmissions of packets are organized.
In this paper, we focus on the Go-Back-N ARQ protocol (GBN-ARQ). Its opera-
tion is illustrated in Fig. 1. In case of GBN-ARQ, the transmitter keeps on sending
packets to the receiver without interruptions until a NAK is received. Upon re-
ception of a NAK, the incorrectly received packet is sent again, as well as every
following packet. We introduce the term feedback delay, which consists of (1) the
time during which a packet is travelling through the channel, (2) the processing
time of that packet in the receiver, and (3) the time during which the feedback
message is travelling back.
The performance of the GBN-ARQ protocol has been investigated before.
However, in many existing studies it is assumed that transmission errors occur in-
dependently, which leads to a considerably simpler analysis see e.g. [2] [4]. Such
an assumption is not realistic in case of wireless transmission channels subject to
e.g. fading effects, see [5] for details. In particular, the time-varying nature of the
channel will result in correlation in the occurrence of transmission errors. In this
paper, we therefore model the channel by means of a Markov chain with M states
and with a xed error probability in every state. A special case of our model, where
M =2, is known as the Gilbert-Elliott model [6, 7]. Previous work on the behavior
CHAPTER 2 2-3
B
B
B
B
B
B
B
B
B
B
B
B
transmitter
receiver
channel
...
...
...
...
error error
- -
ignored ignored
-
s
f
f
f x


1 2 3 4 2 3 4 5 6 4
1 2 3 4 2 3 4 5 6 4 5
ACK NAK ACK ACK NAK
Figure 2.1: Operation of the Go-Back-N protocol (feedback delay s = 2).
of GBN-ARQ for correlated errors is reported in [8], where the throughput and the
transmitter buffer content are studied. In our paper, on the other hand, we present
a full analysis of another important performance metric, namely the packet delay.
Other related work is given in [5] where the performance of Stop-and-Wait ARQ
over a dynamic two-state channel is investigated. The selective-repeat protocol
with correlated errors is analyzed in [9] under the assumption of so-called ideal
ARQ, where the feedback delay is neglected and acknowledgement messages are
received instantly.
The paper is organized as follows. The modeling assumptions are given in Sec-
tion 2.2. In Section 2.3, we introduce the probability generating matrix (pgm) of
the service time, which plays a crucial role in the further analysis. Next, we derive
the distribution of the unnished work at an arbitrary slot boundary in Section 2.4,
as a preparatory step for the analysis of the probability generating function (pgf)
of the delay in Section 2.5. The actual computation of the moments of the packet
delay is elaborated on in Section 2.6. We also derive some heavy-load approxima-
tions. We introduce a generalization of the popular Gilbert-Elliott channel model,
which we dub the cyclic channel model in Section 2.7. In Section 2.8, we provide
numerical examples to show the soundness of our approximations and the impact
of the error correlation. Finally, conclusions are drawn in Section 2.9.
2.2 Modeling Assumptions
We set out to analyze the behavior of the transmitter buffer for the GBN-ARQ
protocol. Throughout the analysis, we assume that the data to be transmitted are
divided into xed-length packets. The time axis is likewise assumed to be divided
into xed-length slots, where a slot corresponds to the time needed to transmit one
packet. Synchronous transmission is used, i.e., transmission always starts at the
beginning of a slot. Packets are transmitted according to a rst-come-rst-served
2-4 DELAY ANALYSIS OF GO-BACK-N ARQ
discipline. After a constant period of s slots, an acknowledgement message from
the receiver indicating whether or not the packet was correctly received, arrives at
the transmitter (see Fig. 2.1). This interval of s slots is referred to as the feedback
delay of the channel. It is assumed that no errors occur in the acknowledgement
messages. This is a reasonable assumption as the information content of an ac-
knowledgement message is low and can thus be heavily protected. Moreover, if we
do not make this assumption, then we cannot ensure completely reliable communi-
cation, but only minimize the risk, see the so-called Two Generals Problem [10].
Let the random variable a
k
denote the number of packets arriving at the trans-
mitter during slot k. The a
k
s are assumed to be independent and identically dis-
tributed (iid) variables with mean E[A], variance Var[A] and pgf A(z). The trans-
mitter buffer is assumed to have an innite storage capacity and packets are stored
in this buffer until they (and their predecessors) are successfully transmitted over
the channel. Note that due to the synchronous transmission mode, a packet arriving
in an empty buffer is not transmitted until the next slot.
The error process in the channel is modeled in this paper by means of a general
Markov chain with M states. Let the random value c
k
denote the channel state at
slot boundary k (i.e. at the beginning of slot k), then the entries of the transition
probability matrix Q associated with the Markov chain are given by
q
i j
.
=P[c
k+1
= j[c
k
= i]. (2.1)
State i has an error probability of e
i
. When there is an error, the packet sent dur-
ing that slot will be incorrectly received, and the transmitter will receive a NAK
message s slots later. We dene the matrix E as the diagonal matrix with elements
e
1
, . . . , e
M
. We also introduce the notation x
.
= 1 x. Likewise,

E is a diagonal
matrix with elements e
1
, . . . , e
M
.
In the rest of this paper, we will often rely on a trick that was introduced by
Towsley and Wolf [2]. The trick is to consider a slightly modied system in which
a packet immediately leaves the transmitter buffer after being successfully trans-
mitted, instead of staying another s slots until the positive acknowledgement is
received. Note that the delay in the modied system of every packet is exactly s
slots less than in the original system, which makes it easy to convert results for the
delay obtained for the modied system into results for the original system, as will
be elaborated on later.
2.3 Distribution of the Service Time
By putting the evolution of the service time into a convenient form, we set the
stage for the queueing analysis proper. As we will see, the results in this section
will greatly simplify the derivations in the subsequent sections.
CHAPTER 2 2-5
state i state j
service time of packet P
time
Figure 2.2: Service time starting in state i and followed by state j.
The service time of a packet is dened as the time interval (expressed in slots)
that starts at the slot boundary where the packet is transmitted for the rst time
and ends with the slot boundary where the packet leaves the system. We consider
the modied system, so packets leave the buffer at the end of the slot where they
are correctly transmitted. Remark that this also means that in the modied system,
service times do not overlap. Service times in the real system are s slots longer,
and may overlap.
In view of the above modeling assumptions, the exact distribution of the ser-
vice time of a packet will depend on the channel state in the slot during which the
service of the packet starts. Therefore, consecutive service times in the considered
model are not iid, unlike in the case of uncorrelated transmission errors. In order
to study the service time, let us introduce a pgm S(z) of dimension MM. The
element [S(z)]
i j
is the partial conditional pgf of the service time that is followed
by a slot with channel state j, given that the service time starts in channel state i
(see Fig. 2.2), that is
[S(z)]
i j
= E[z
S
1(c
k
= j)[c
k
= i], (2.2)
where the random variable S denotes a service time, k and k

denote the slot bound-


ary where the service starts and ends respectively, and 1(.) is the indicator function
of the event between brackets. Note that a service time has some kind of general-
ized geometric property: when there is no error, the service nishes after exactly
one slot; when there is an error, a new transmission starts after s +1 slots, and this
can be seen as the start of a completely new service time, but this time starting in
state c
k+s+1
. Indeed, when there is an error, the total service lasts s +1 slots plus a
remaining service time. Given the state c
k+s+1
at the start of this remaining service
time, the mechanism that determines the length of the remaining service time is
completely equivalent to the one that determines the length of a full service time
and therefore this remaining service time also has pgm S(z), i.e. it has the pgm of
2-6 DELAY ANALYSIS OF GO-BACK-N ARQ
a completely new service time. From these observations, we can see that
[S(z)]
i j
= zP[no error at k, c
k
= j[c
k
= i] (2.3)
+z
s+1
M

=1
P[error at k, c
k+s+1
= i

[c
k
= i] E[z
S
1(c
k
= j)[c
k+s+1
= i

]
(2.4)
= z(1e
i
)[Q]
i j
+z
s+1
M

=1
e
i
[Q
s+1
]
ii
[S(z)]
i

j
, (2.5)
where we have used the fact that the (s +1)-step transition probabilities for the
Markov chain of the channel are given by the transition matrix Q
s+1
. Exploiting
the matrix notation, we can write pgm S(z) succinctly as
S(z) =

EQz +EQ
s+1
z
s+1
S(z). (2.6)
From the above relation, we can derive S(z) as
S(z) = (I EQ
s+1
z
s+1
)
1

EQz, (2.7)
where I denotes the MM identity matrix.
The advantage of capturing the distribution of the service time in the pgm S(z)
is that we can express the pgm of the length of n subsequent service times simply
as S(z)
n
. Equation (2.7) also allows us to compute the throughput of the protocol.
The throughput is dened as the inverse of the mean service time under the
assumption that there are always packets available. Notice that the matrix S(1) is
the transition matrix that records the channel state transition between the start and
the end of a service. The steady-state probability vector is given by
= S(1) and 1 = 1, (2.8)
where 1 is an N1 column vector of ones. The vector records the probabilities of
nding the channel in a particular state at the beginning of a service, if newservices
start uninterruptedly. Now, [S

(1)]
i j
equals E[S 1(c

k
= j)[c
k
= i], i.e. [S

(1)]
i j
denotes the mean length of a service time that is followed by state j, given that it
starts in state i. Therefore, we have that
S

(1)1 =
M

i=1
M

j=1
P[c
k
= i] E[S 1(c

k
= j)[c
k
= i]
= E[S]. (2.9)
That is, S

(1)1 is the expected length of a service time, given that there are no
gaps between services. This is equal to the reciproque of the throughput :
=
1
S

(1)1
. (2.10)
CHAPTER 2 2-7
k k + 1

state g
k
state g
k+1
unnished work w
k
at start of slot k
k
unnished work w
k+1
at start of slot k + 1

k
time
Figure 2.3: Denitions of w
k
,
k
and g
k
.
2.4 Distribution of the Unnished Work
In this section we derive an expression for the pgf of the unnished work w
k
at
the beginning of an arbitrary slot k. This is the time in slots needed to serve the
packets that are present in the buffer at the beginning of slot k. We again consider
the modied system here.
Let
k
denote the amount of work (in slots) that enters the buffer during slot
k, i.e. the sum of the service times of all packets arriving in slot k. Then the
unnished work evolves according to the following system equation:
w
k+1
= (w
k
1)
+
+
k
, (2.11)
where ( )
+
=max(0, ). The randomvariables w
k
and
k
on the right-hand side
of (2.11) are not independent, because both depend on the state of the channel.
This means that the set w
k
does not form a Markov chain. We can make it
a Markov chain by adding an extra variable. The most convenient choice is the
variable g
k
, which indicates the state of the channel in the earliest slot where work
arriving during slot k could start. So, when w
k
= 0, g
k
corresponds to the channel
state during slot k +1. The case where w
k
> 0 is illustrated in Fig. 2.3. We have
that
g
k+1
=
_
g
(1)
k
when w
k
=
k
= 0,
g
(
k
)
k
otherwise,
(2.12)
where the notation g
(n)
k
stands for the state of the channel n slots after the slot with
state g
k
. From (2.11) and (2.12), it is easily seen that the set of random variables
w
k
, g
k
indeed forms a Markov chain.
Let us now dene the partial pgfs
W
j,k
(z) =

n=0
P[w
k
= n, g
k
= j]z
n
,
and introduce the vector notation W
k
(z) = (W
1,k
(z), ,W
M,k
(z)). Then it is pos-
sible to rewrite the above system equations in the z-domain as:
W
k+1
(z) =
1
z
(W
k
(z) W
k
(0))A(S(z)) +W
k
(0)[A(S(z)) A(0)I +A(0)Q] .
2-8 DELAY ANALYSIS OF GO-BACK-N ARQ
Here we have introduced the convenient shorthand notation A(S(z)) to denote a
matrix that is a power series expansion in the pgm S(z) with the same coefcients
as the power series expansion in z of A(z). Similar notations will also be used later
in this paper. Note that
A(S(1)) =

k=0
P[a = k]S(1)
k
= , (2.13)
which shows that transition matrices S(1) and A(S(1)) have the same steady-state
probability vector. The system will reach an equilibrium if the average amount of
work per slot (commonly called the load ) is strictly smaller than 1. The load
satises
= E[S] E[A] = A

(1)S

(1)1. (2.14)
In the steady state both W
k
(z) and W
k+1
(z) will converge to a common lim-
iting vector W(z) = (W
1
(z), ,W
M
(z)). Taking limits for k and solving the
resulting equation for W(z), we obtain
W(z)[zI A(S(z))] = W(0)[(z 1)A(S(z)) +zA(0)(QI)] . (2.15)
Note that this formula still contains the unknown vector W(0), whose computation
is the subject of a huge number of studies [11], [12]. The proposed algorithms fall
apart into two broad categories. The rst class are the so-called spectral methods,
which are based on the property that there are a number of values z
i
inside the unit
circle, such that z
i
IA(S(z
i
)) is a singular matrix, and such that the corresponding
right null spaces R
i
have together a total rank of M1 [11]. Then we can form a
set of M1 equations
0 = W(0)[(z
i
1)A(S(z
i
)) +z
i
A(0)(QI)] R
i
. (2.16)
The last equation for W(0) is supplied by the well-known property of single-server
queues that utilization equals the load. The utilization equals 1W(0)1, while the
load is given by Eq. (2.14), so that we complete the set of equations by stating that
W(0)1 = 1. (2.17)
Some authors question the numerical stability of this method, especially if there
are z
i
which have a corresponding null space of a high dimension, and they give
preference to the second method.
This other method is based on the weak canonical Wiener-Hopf factorization
of the Laurent series I
A(S(z))
z
, which says that this expression can be factorized
under mild conditions into
I
A(S(z))
z
= F(z)(I
1
z
G). (2.18)
CHAPTER 2 2-9
k k + 1

state s(P) state r(P)
w
k service times
delay d(P) of packet P
packet P arrives
time
Figure 2.4: Illustration of the various components of the delay.
Here power series F(z) has no roots inside the unit circle, whereas the roots of
expression I
1
z
Glie inside or on the unit circle. It can be shown [12] that matrix G
is the well-known fundamental matrix as it occurs in M/G/1-type Markov chains,
for which many algorithms have been developed. If we substitute the Wiener-
Hopf factorization into Eq. (10.1), and identify the terms in z
1
, we nd after
some manipulations that
W(0) = W(0)(G+A(0)(I Q)). (2.19)
Hence the computation of the boundary probabilities is reduced to a linear system
of equations, which must again be supplemented by Eq. (2.17).
2.5 Distribution of the Packet Delay
Now that we have found an expression for the pgf of the unnished work, it is
rather straightforward to analyze the delay of an arbitrary packet P arriving in slot
k. Let us rst focus again on the modied system. We denote the channel state
during the arrival slot of P as s(P). The channel state in the rst slot after the
service of P is denoted by r(P).
The delay d(P) of packet P consists of two components, as illustrated in
Fig. 5.8: (a) the remaining number of slots from slot k +1 onwards needed to
execute the unnished work present in the system at the beginning of slot k, i.e.,
(w
k
1)
+
, and (b) the service times of the packets arriving in the buffer in
slot k that will be served no later than (but including) P. The distribution of the
unnished work has been derived in the previous section. Owing to the fact that P
is an arbitrary packet, the pgf L(z) of can be derived as (see e.g. [13]):
L(z) =
z(1A(z))
E[A](1z)
. (2.20)
Now let D
j
(z) denote the partial pgf of the packet delay, provided that r(P) =
j. Then we can write, again using vector notation,
D(z) =
1
z
(W(z) +(z 1)W(0))L(S(z)). (2.21)
2-10 DELAY ANALYSIS OF GO-BACK-N ARQ
The unconditional pgf

D(z) of the packet delay in the modied system is given by

D(z) = D(z)1. From (2.21), it then follows that


z

D(z) = (W(z) +(z 1)W(0))L(S(z))1. (2.22)
From this, we can nally obtain the pgf of the delay D(z) in the real system:
D(z) = z
s

D(z). (2.23)
2.6 Computing Moments of the Packet Delay
In this section, we explain in detail how to compute the mean and higher-order
central moments of the packet delay. Once we have the boundary probabilities in
the form of vector W(0), only a moderate amount of numerical computation is
needed to get the rst few moments of the delay. We illustrate this by computing
the rst two central moments, as higher-order moments are obtained by similar
means.
Notice that the mean and the variance of the delay in the real system can be
trivially obtained in function of the same quantities of the modied system:
E[D] = E[

D] +s and Var[D] = Var[

D]. (2.24)
Let us now compute the rst two derivatives of

D(z) as given in Eq. (2.22),
and evaluate them for z = 1:

(1) +1 = (W

(1) +W(0))1+W(1)

(1); (2.25)

(1) +2

D

(1) = W

(1)1+2(W

(1) +W(0))

(1) +W(1)

(1), (2.26)
where we introduced the column vector (z) = L(S(z))1. Note that (1) = 1. In
view of the fact that Var[D] =

D

(1) +

D

(1)

D

(1)
2
, we nd after some further
manipulations that
E[D] = E[W] +W(1)

(1) +s; (2.27)


Var[D] = Var[W] +2(W

(1) +W(0))

(1) +W(1)

(1)
(1+W(1)

(1))(2E[W] +W(1)

(1)), (2.28)
where we introduced E[W] =W

(1)1 and Var[W] =W

(1)1+W

(1)1(W

(1)1)
2
.
So far, we have reduced the problem of nding the moments of the delay into the
problem of nding the row vectors W(1), W

(1) and W

(1) associated with the


unnished work, and the column vectors

(1) and

(1), which are connected to


the amount of work that arrives during an arrival slot.
CHAPTER 2 2-11
Let us rst focus our attention on the unknowns W(1), W

(1) and W

(1).
When we evaluate Eq. (10.1) for z = 1 and do the same for its rst two derivatives
in z, we obtain
W(1)(I A(1)) = W(0)A(0)(QI);
(2.29)
W

(1)(I A(1)) +W(1)(I A

(1)) = W(0)(A(1) +A(0)(QI));


(2.30)
W

(1)(I A(1)) +2W

(1)(I A

(1)) W(1)A

(1) = 2W(0)A

(1), (2.31)
where we have introduced A(z) = A(S(z)) for reasons of convenience. We see
that W(1), W

(1) and generally W


(n)
(1) satisfy a linear equation of the form x(I
A(1)) = b. However, as A(1) is a stochastic matrix, the system of equations is
singular, and hence we need an extra equation for each W
(n)
(1). Note that the
solution of an inhomogeneous system of equations can generally be written as
the sum of one particular solution of the inhomogeneous system, plus a solution
of the homogeneous system, which in this case has the form c, for an arbitrary
constant c. We dene the particular solution of the problem with the help of the
group inverse [14], which is the subject of the Appendix. The useful property of
the group inverse for our purposes is that the group inverse (I A(1))
#
gives us
the solution for x for the equation x(I A(1)) = b, such that x1 = 0. Let

W(1),

(1) and

W

(1) denote the particular solutions of (2.29)(2.31) found by means


of the group inverse, i.e.

W(1) = W(0)A(0)(QI)(I A(1))


#
; (2.32)

(1) = (W(0)(A(1) +A(0)(QI)) W(1)(I A

(1)))(I A(1))
#
; (2.33)

(1) = (2W(0)A

(1) 2W

(1)(I A

(1)) +W(1)A

(1))(I A(1))
#
. (2.34)
Then we can write
W(1) =

W(1) +W(1)1; (2.35)
W

(1) =

W

(1) +W

(1)1; (2.36)
W

(1) =

W

(1) +W

(1)1, (2.37)
where we still have to determine the scalars W(1)1, W

(1)1 and W

(1)1. Note
that the normalization condition learns us that W(1)1 = 1. Furthermore, when we
substitute Eq. (2.36) into Eq. (2.31), and postmultiply by the column vector 1, we
nd
2(

W

(1) +W

(1)1)(I A

(1))1W(1)A

(1)1 = 2W(0)A

(1)1. (2.38)
Since (I A

(1))1 = 1, this leads to


2(1)W

(1)1 = 2W(0)A

(1)1+W(1)A

(1)12

W

(1)(I A

(1))1. (2.39)
2-12 DELAY ANALYSIS OF GO-BACK-N ARQ
Similarly, from the evaluation of the third derivative of Eq. (10.1) for z = 1 we
can infer the value for W

(1)1.
Having expounded on the computation of the derivatives of W(z) in z = 1, we
now set up a similar scheme for the computation of

(1) and

(1). Note that


E[A](S(z) I)(z) = S(z)(A(z) I)1. (2.40)
The rst two derivations of this expression with respect to z, evaluated in z = 1,
are equal to
E[A](S

(1)1+(S(1) I)

(1)) = S(1)A

(1)1; (2.41)
E[A](S

(1)1+2S

(1)

(1) +(S(1) I)

(1)) = 2S

(1)A

(1)1+S(1)A

(1)1.
(2.42)
Employing again the technique of the group inverse, we establish that

(1) =

(1) +1

(1), (2.43)
where

(1) =
1
E[A]
(I S(1))
#
(E[A]S

(1)1S(1)A

(1)1). (2.44)
We nd the scalar

(1) by plugging the previous equation into Eq. (2.42), and


premultiplying by :
2

(1) =2S

(1)A

(1)1+A

(1)1E[A]S

(1)12E[A]S

(1)

(1). (2.45)
We omit the entirely analogous derivation of

(1).
The computational complexity of our method compares favorably to alterna-
tives. After we have computed 2 group inverses of matrices of order MM, we
only have to compute matrix-vector multiplications and vector additions, which
have complexities O(M
2
) and O(M) respectively. In practice, the real cost is the
computation of the boundary probability vector W(0), which is almost always per-
force an iterative procedure. We therefore elaborate on an approximative method
for heavy loads (i.e. 1) which permits to skip the computation of W(0) alto-
gether.
The idea behind this approximation is that as the load approaches one, W(0)
becomes almost equal to the zero vector, so that the terms in which W(0) occurs,
vanish for loads close to one. Let us mark the system characteristics as they hold
for heavy-load conditions by a subscript h, e.g. E[D
h
], W
h
(1), W

h
(1), . First
note that the vector W
h
(1) in that case approaches the vector . This can also be
intuited from the fact that for heavy loads, empty periods get scarce, and hence
services follow each other without gaps, which is exactly the denition of the
steady-state vector . Similar simplications in Eq. (2.39) lead to
E[W
h
] =
1
2(1)
(A

(1)1+2A

(1)(I A(1))
#
A

(1)1). (2.46)
CHAPTER 2 2-13
The extra terms in the expression (7.6) for the mean delay E[D] disappear in the
heavy trafc limit and hence E[D
h
] = E[W
h
]. Note that E[W
h
] quanties the delay
incurred by the buffer, whereas the other terms quantify the delay incurred by
other packets arriving during the same slot, and the delay incurred by the feedback
channel. It is clear that the delay due to buffering will dwarf the other terms in
case of high loads. Note that the computational complexity of the heavy-load
mean delay is not so much greater than the complexity of the throughput, yet it
offers a lot of additional insight in the performance.
2.7 Channel Models
A vast amount of papers have been written on the modeling of communication
channels in general [6, 7] and wireless communication channels in particular [15
17]. Many of the proposed models t into the discrete Markovian framework we
have used in this paper.
It is intuitively clear that a larger number of background states can more faith-
fully model the behavior of real communication channels. However, there are also
disadvantages. A model with a large number of background states needs in general
a large number of parameters (order M
2
). If these have to be estimated from obser-
vations, one needs a lot of data to reliably estimate all these parameters. Moreover,
it gets difcult to assign an interpretation to the parameters.
2.7.1 Cyclic Channel Model
We propose a channel model that is very versatile yet simple enough to be reasoned
about: we allow only cyclic transitions of the channel state, and the sojourn time
in each state i, 1 i M is (shifted) geometrically distributed with parameter q
i
.
This chain is illustrated in Fig. 2.5. We have the following transition matrix:
q =
_
_
_
_
_
_
_
q
1
q
1
q
2
q
2
.
.
.
.
.
.
q
M1
q
M1
q
M
q
M
_
_
_
_
_
_
_
. (2.47)
State i has error probability e
i
. Hence, 2M parameters dene the channel model
(M error probabilities and M transition probabilities). As the mean sojourn time in
state i equals q
1
i
, the mean length of an entire cycle (a visit to all states 1 to M, in
order) equals
L
c
=
M

j=1
1
1q
j
. (2.48)
2-14 DELAY ANALYSIS OF GO-BACK-N ARQ
1 2 M 1 M
. . .
1 q
1
1 q
M1
1 q
2
1 q
M2
1 q
M
q
1
q
2
q
M1
q
M
error error error error
probability probability probability probability
e
1
e
2
e
M1
e
M
Figure 2.5: The Markov chain of a cyclic channel model
This is an important characteristic of the burstiness of the channel: the higher the
mean cycle length, the burstier the channel. The steady-state probabilities
i
of
the channel states are easily obtained as

i
=
q
1
i
L
c
. (2.49)
Instead of using the transition probabilities to describe the channel model, we can
also use the steady-state probability vector and the mean cycle length, which may
give a somewhat more intuitive description. However, not all cycle lengths are
possible for a given probability vector. Considering that q
i
1, for all i, 1 i M,
we see from Eq. (2.49) that L
c

1
i
for all i, 1 i M. Hence the minimal cycle
length L
min
c
equals
L
min
c
=
1
min
i

i
. (2.50)
In the numerical examples, we characterize a cyclic channel model by the cycle
length L
c
, and two vectors and E
c
, which record respectively the steady-state
probabilities
i
and the error probabilities e
i
.
2.7.2 Gilbert-Elliott Channel Model
For M = 2, the cyclic channel model reduces to a Gilbert-Elliott channel model
[6, 7], which is the staple model in wireless communication modelling.
The two states of the Gilbert-Elliott model are usually labelled 1 and 2, or
GOOD and BAD (see Fig. 2.6). The parameters e
1
and e
2
are the error proba-
bilities of the channel in resp. state 0 and 1. Of course, the designations GOOD and
BAD make only sense when e
1
< e
2
, but this is not a requirement for the analysis.
CHAPTER 2 2-15
1 q
1
1 q
2
q
1
q
2
error error
probability probability
e
1
e
2
GOOD BAD
1 2
Figure 2.6: The Gilbert-Elliott channel model
The transitions in the Gilbert-Elliott model are completely dened by the pa-
rameters q
1
and q
2
, where
q
i
=P[state in next slot is i[ state in current slot is i].
Rather than using q
1
and q
2
, we dene the parameters
=
1q
1
2q
1
q
2
and K =
1
2q
1
q
2
(2.51)
to be understood as follows: is the fraction of the time that the system is in state
2, while the parameter K can be seen as a measure for the mean lengths of 1- and 2-
periods. Specically, the mean length of a 2-period is K/, of a 1-period it is K/ .
Therefore, the factor K can be seen as a measure for the absolute lengths of the 1-
periods and 2-periods, while characterizes their relative lengths. The parameter
K thus characterizes the degree of correlation in the channel state and is therefore
referred to further on as the correlation factor, the value K = 1 corresponds to an
uncorrelated channel state from slot to slot.
Note that in this model, the sojourn times in both states are geometrically dis-
tributed, while some measurements have indicated [17] that even though good pe-
riods can be modelled more or less faithfully with geometric lengths, bad periods
cannot. A cyclic channel with more than one state designated to model the bad
periods, can offer a solution to this problem.
2.8 Numerical results and discussion
In this section, we provide some numerical examples. First, we show some prob-
ability mass functions of the delay in Figs. 2.7(a) and 2.7(b), for a Gilbert-Elliott
channel model and geometrically distributed batch arrivals with mean E[A] = 0.3.
The gures show the inuence of the correlation factor K, and the feedback delay
s respectively. These curves have been obtained from the pgf D(z) by means of
an inversion method explained in [18]. We see that the packet delay over a more
2-16 DELAY ANALYSIS OF GO-BACK-N ARQ
K = 50
K = 100
K = 20
1e1
1e2
1e3
1e4
0 20 40 60 80 100
n
P
r
[
D
=
n
]
(a) for s = 4, = 0.2, e
1
= 0.05, e
2
= 0.5
and K = 20, 50, 100.
s = 0 s = 4
s = 5
1e1
1e2
1e3
1e4
0 20 40 60 80 100
n
P
r
[
D
=
n
]
(b) for = 0.2, e
1
= 0.05, e
2
= 0.5, K = 20
and s = 0, 4, 5.
Figure 2.7: Logarithmic plots of P[

D = n]
correlated channel has a heavier tail, which shows that correlation of the channel
has an important inuence on the performance of the system. A larger feedback
delay also has a negative effect on the delay performance: we see a heavier tail for
larger values of s.
The huge impact of the correlation is also observed in Fig. 2.8(a), where
we plot the mean delay E[D] against the cycle length L
c
. We used a Poisson ar-
rival process, with pgf (A(z) = exp((z 1))), with =
1
4
, feedback delay s = 6,
and a four-state cyclic channel model with steady-state probability vector =
(0.4, 0.2, 0.2, 0.2), and xed error probabilities for the rst three states (respec-
tively e
1
= 0, e
2
= 0.1, e
3
= 0.1). We let the error probability e
4
take on different
values, as indicated in the plot. We see an almost linear increase of the mean delay
in function of the cycle length. Hence bursty channels deteriorate the performance
of the channel tremendously, although the throughput is actually an increasing
function for the very same parameters (see Fig. 2.8(b)). This makes a strong case
for the fact that the throughput is not sufcient to evaluate the performance of a
retransmission protocol.
Next, we show that bad periods with non-geometric lengths have indeed a dif-
ferent performance than bad periods with geometric lengths in Fig. 2.9. We com-
pare channel model A with parameters E
c
= (0.0, 0.3), and = (0.2, 0.8) with
channel model Bwith parameters E
c
=(0.0, 0.3 0.3) and =(0.2, 0.1, , 0.1).
That is, we compare a Gilbert-Elliott model where the bad period length is geo-
metrically distributed, with a model where the bad period consists of 8 consecutive
geometrically distributed subperiods. Note that the error probability in a bad slot
in both cases equals 0.3, and moreover, the average length of a bad period is equal
to 0.8L
c
in both cases. In the second model however, the length of a bad period has
a signicantly lower variance, and (hence) shows a markedly better performance
than the Gilbert-Elliott model. The difference is particularly large for more corre-
CHAPTER 2 2-17
e
4
= 0.4
e
4
= 0.3
e
4
= 0.2
e
4
= 0.1
0 1000 2000 3000 4000
20
40
60
80
L
c
E
[
D
]
(a)
e
4
= 0.1
e
4
= 0.2
e
4
= 0.3
e
4
= 0.4
0 1000 2000 3000 4000
0.55
0.6
0.65
0.7
0.75
L
c
t
h
r
o
u
g
h
p
u
t

(b)
Figure 2.8: Mean delay (left) and throughput (right) versus cycle length L
c
, with Poisson
arrivals ( =
1
4
), s = 6, = (0.40.20.20.2) and E
c
= (0.0, 0.1, 0.1, e
4
), for different values
of e
4
.
lated channels (longer L
c
). This gure shows that more complicated models may
indeed be necessary in order to faithfully capture the performance of a protocol
over a wireless link.
Finally, we demonstrate the soundness of our heavy-load approximation in
Fig. 2.10. The computationally inexpensive approximations prove to be accurate
for loads higher than 0.7 and sometimes even well before that point.
2.9 Conclusions
We have studied the transmitter buffer behavior for a Go-Back-N ARQ protocol
over a time-varying channel. In particular, we have analyzed and found expres-
sions for the steady-state distributions of the unnished work and of the packet de-
lay, as well as some very efcient heavy-load approximations. Finally, by means
of some examples we have discussed the inuence of the model parameters on the
delay performance.
Appendix: Group Inverse of a Matrix
In this appendix, we elaborate on the technique of the group inverse, which con-
stitutes an alternative for spectral decomposition, although some results [19] hint
at a close relationship between the two techniques.
The group inverse of a matrix M is the unique matrix M
#
which satises
MM
#
M = M, M
#
MM
#
= M
#
, and MM
#
= M
#
M. We refer to [14] for more de-
2-18 DELAY ANALYSIS OF GO-BACK-N ARQ
A, L
c
=1e4
B, L
c
=1e4
A, L
c
=1e3
B, L
c
=1e3
0 0.2 0.4 0.6 0.8 1
0
500
1000
1500
2000

E
[
D
]
Figure 2.9: A plot of the mean delay versus the load, with feedback delay s = 6 for a
two-state channel model A (E
c
= (0.0, 0.3), and = (0.2, 0.8)) and a nine-state channel
model B (E
c
= (0.0, 0.3, , 0.3) and = (0.2, 0.1, , 0.1)). The load is varied by
varying parameter of the Poisson arrival process.
CHAPTER 2 2-19
L
c
=5e3
L
c
=1e3
0 0.2 0.4 0.6 0.8 1
0
500
1000
1500
2000

E
[
D
]
Figure 2.10: A plot of the mean delay versus the load, with the heavy-load approximation
indicated with dashed lines, for a four state model, with = (0.4, 0.3, 0.2, 0.1) and
E
c
= (0.0, 0.3, 0.3, 0.3) for s = 6 and L
c
= 1000, 5000. The load is varied by varying
parameter of the Poisson arrival process.
2-20 DELAY ANALYSIS OF GO-BACK-N ARQ
tails and recall here only that for a transition matrix A(1) of an irreducible Markov
chain the group inverse of I A(1) equals
(I A(1))
#
= (I A(1) +1)
1
1, (2.52)
where is the stationary probability vector of A(1). It can readily be seen that
(I A(1))
#
1 = 0 and (I A(1))
#
= 0. In general, the solution of x(I A(1))b is
given by x = b(I A(1))
#
+x. Note that the group inverse does not necessarily
have to be computed via Eq. (52). Indeed, an alternative algorithm can be found
in [20], where one does not have to compute rst, but rather is obtained as a
simple by-product of the group-inverse computation. Note that the group inverse
is also known as the deviation matrix or fundamental matrix (the latter term is not
to be confused with the fundamental matrix G of the matrix analytic method).
Postscriptum
This chapter captures some of the oldest work done as part of this PhD, as an ex-
tension of Stijns work on the Stop-and-Wait protocol over a Gilbert-Elliott chan-
nel [5]. Spectral decomposition was the preferred tool in that article, i.e. one
would write S(z) as
S(z) =
M

m=1

m
(z)r
m
(z)
m
(z),
where
m
(z) and
m
(z), r
m
(z) are the eigenvalues and the corresponding left and
right eigenvectors. The matrix expression for D(z) can thus be reduced to a scalar
one. This allows to follow the habitual procedure of scalar probability generating
functions.
This becomes tricky when the number of states is higher than 2, as we can
have not only numerical instabilities but also conceptual difculties. We avoided
the use of spectral decomposition, relying instead on the matrix-analytic method
to nd the zeros in the unit circle and on the group inverse to nd expressions for
the moments.
An advantage of the group inverse is that it needs fewer assumptions than spec-
tral decomposition. It is enough to assume that the matrix A(1) has a simple eigen-
value in 1 (ergodicity of the corresponding Markov chain is a sufcient condition)
to ensure that the group inverse is given by Eq. (52), whereas spectral decompo-
sition techniques often need additional assumptions, such as the assumption that
A(1) be diagonizable. Another tricky aspect that the use of the group inverse man-
ages to circumvent is the normalization of the eigenvectors.
CHAPTER 2 2-21
References
[1] Bhunia, C.T.: ARQ Review and modications. IETE Technical Review 18
(2001) 381401
[2] Towsley, D., Wolf, J.K.: On the statistical analysis of queue lengths and
waiting times for statistical multiplexers with ARQ retransmission schemes.
IEEE Transactions on Communications 25 (1979) 693703
[3] Konheim, A.G.: A queueing analysis of two ARQ protocols. IEEE Transac-
tions on Communications 28 (1980) 10041014
[4] De Munnynck, M., Wittevrongel, S., Lootens, A., Bruneel, H.: Queueing
analysis of some continuous ARQ strategies with repeated transmissions.
Electronics Letters 38 (2002) 1295 1297
[5] De Vuyst, S., Wittevrongel, S., Bruneel, H.: Queueing delay of Stop-and-
Wait ARQ over a wireless Markovian channel. Heterogeneous Networks,
Vol. III, River Publishers, 2008.
[6] Gilbert, E.N.: Capacity of a burst-noise channel. The Bell System Technical
Journal 39 (1960) 12531265
[7] Elliott, E.O.: Estimates of error rate for codes on burst-noise channels. Bell
System Technical Journal 42 (1963) 19771997.
[8] Towsley, D.: A statistical analysis of ARQ protocols operating in a non-
independent error environment. IEEE Transactions on Communications 27
(1981) 971981
[9] Kim, J.G., Krunz, M.: Delay analysis of selective repeat ARQ for trans-
porting Markovian sources over a wireless channel. IEEE Transactions on
Vehicular Technology 49 (2000) 19681981
[10] E. A. Akkoyunlu , K. Ekanadham , R. V. Huber, Some constraints and
tradeoffs in the design of network communications. Proceedings of the fth
ACM symposium on Operating systems principles, p.67-74, November 19-
21, 1975, Austin, Texas, United States
[11] Gail, H. R., Hantler, S. L., Taylor, B. A.: Spectral analysis of M/G/1 and
G/M/1 type Markov chains. Adv. Appl. Prob. 28 (1996) 114165
[12] Bini, D. A., Latouche, G., Meini, B.: Numerical methods for structured
Markov chains. Oxford University Press (2005)
[13] Bruneel, H.: Buffers with stochastic output interruptions. Electronics Letters
19 (1983) 735737
2-22 DELAY ANALYSIS OF GO-BACK-N ARQ
[14] Campbell, S.L., Meyer, C. D.: Generalized Inverses of Linear Transforma-
tions. Dover Publications (1991)
[15] Kim, Y.Y., Li, S.Q.: Capturing important statistics of a fading/shadowing
channel for network performance analysis. IEEE Journal on Selected Areas
in Communications, 17, (1999) 888901
[16] Kumwilaisak, W.: Fading channel modeling via variable-length Markov
chain technique. IEEE Transactions on Vehicular Technology 57 (2008)
13381358
[17] McDougall, J., Miller, S.: Sensitivity of wireless network simulations to a
two-state Markov model channel approximation. Proceedings of GLOBE-
COM 2003, (1-5 December 2003, San Francisco, USA)
[18] Abate, J., Whitt, W.: Numerical inversion of probability generating func-
tions. Operations Research Letters 12 (1992) 245251
[19] Meyer, C. D., Stewart G. W.: Derivatives and Perturbations of Eigenvectors.
SIAM Journal on Numerical Analysis, 25 (1988) 679691
[20] Meyer, C. D.: The role of the group generalized inverse in the theory of nite
Markov chains. SIAM Review, 17 (1975) 443464
3
Buffer analysis of order-preserving
ARQ protocols over correlated
channels
Koen De Turck and Sabine Wittevrongel
Manuscript in preparation.
Abstract.
We analyze the transmitter buffer behavior for a number of ARQ protocols
operating over channels with correlated errors. The buffer is modeled as a discrete-
time queue with an innite size. We model the channel as a nite-state Markov
chain, where each state has a xed error probability. We represent ARQ protocols
as regular languages, and subsequently use these formal languages to derive an
expression of the service time as a probability generating matrix. This allows us to
analyze the queue content and the packet delay with abstraction of the exact ARQ
protocol. Explicit expressions are obtained for the probability generating functions
of the queue content and the delay.
3.1 Introduction
Telecommunication channels are by nature error-prone. Automatic Repeat reQuest
(ARQ) is an often used mechanism ( [1]) to ensure reliable transmission over noisy
3-2 ORDER-PRESERVING ARQ
channels. With this technique, if the receiver detects an error in a received packet,
it sends back a negative acknowledgement (NAK) to the transmitter in order to
request the retransmission of the packet; otherwise it sends back a positive ac-
knowledgement (ACK) meaning that the transmitter can safely remove the packet
from its buffer. ARQ forms an important part of modern wireless digital com-
munication. Over wireless channels, the error process very often has a heavily
correlated nature. We want to investigate the inuence of error correlation on the
system performance. Numerous studies have analyzed the impact of correlation
on the throughput; in this paper we will investigate the behavior of the transmitter
buffer.
There are many different types of ARQ protocols, but the principle is always
the same. The sender divides the data into packets, adds an error detecting code
and sends them to the receiver. The latter checks the incoming packets, and in-
forms the sender of the state of each packet by sending positive and negative ac-
knowledgement messages (ACK and NAK respectively).
In this paper we analyze the transmitter buffer under an order-preserving re-
transmission protocol. Roughly speaking, order-preserving means that there is no
resequencing buffer needed at the receivers side. We dene this term more for-
mally later on.
There has been a lot of research on the queueing analysis of ARQ protocols.
However, the majority of research has been under the assumption of independent
transmission errors, see e.g. [2]- [4]. The model we propose in this paper is dif-
ferent from these studies in that we allow the errors occurring in the channel to be
correlated in time. This assumption is necessary when dynamical, time-varying
channels are studied (such as wireless channels). Specically, the channel state
process is modeled by means of a Markov chain with M states with a different
error probability in every state.
Some previous work on the buffer analysis of ARQ protocols over a time-
varying channel exists. For example, in [5] the GBN-ARQ protocol has been
considered, but only the throughput and the buffer content have been analyzed,
whereas in this paper, we also provide a full analysis of the packet delay. In [6],
the queueing performance of stop-and-wait ARQ over a dynamic channel has been
studied for the case of M = 2. The selective-repeat protocol with correlated errors
has been investigated in [7], but the model is basically an ideal ARQ system. This
means that there is no feedback delay: acknowledgements are received instantly,
which simplies the analysis. In this paper we present a full queueing analysis of
both buffer content and delay for a broad class of protocols for nonindependent
channel errors.
The outline of this chapter is as follows. In section 3.2 we describe the mathe-
matical model and elaborate on the class of protocols that are studied in this paper.
Section 3.3 elaborates on the grammar of ARQ protocols and section 3.4 contains
CHAPTER 3 3-3
the analysis. Next, in section 3.5 we discuss asymptotic behavior and in section
3.6 we discuss some generalizations.
3.2 Model and notations
We model the transmitter buffer of an ARQ system as an discrete-time queue of
innite capacity. The time axis is divided into slots which correspond to the time
to transmit one (xed-size) packet. Information on the success of a transmission is
received after a xed time of s slots after transmission in the form of an ACK (pos-
itive acknowledgement) or NACK (negative acknowledgement) message. Pack-
ets arrive at the transmitters side according to an independent and indentically
distributed batch arrival process. The probability of having i packets in a ran-
dom slot is equal to a(i). The probability generating function (pgf) is given by
A(z) =
i
a(i)z
i
. We assume late arrivals, i.e. arriving packets cannot be transmit-
ted in the slot in which they arrive, but at the earliest in the next slot.
The probability that a packet is transmitted erroneously during a slot is modu-
lated by a nite-state Markov chain with state set 1, , M and transition prob-
ability matrix Q of dimension MM. The transmission of a packet is erroneous
with probability p(i) when the channel state during the transmission slot is i. We
dene the diagonal matrices E and

E with entries p(i) and p(i) respectively. The
steady-state probability vector of the Markov chain is . We assume throughout
the chapter that acknowledgement messages arrive at the sender a constant time of
s slots after the transmission slot. Note that the correctness of a transmission only
depends on the channel state at the beginning of the transmission, and not on the
channel states during the rest of the transmission.
3.3 Grammar of an ARQ protocol
In this section, we carve out a fairly broad class of retransmission protocols, such
that we can derive a unied analysis of the buffer performance for this class.
As has been shown in the literature (e.g. see [3]), an exact buffer performance
of selective-repeat ARQ and related protocols leads to a very large state space that
increases exponentially as a function of the roundtrip time, and makes numerical
evaluation thus very costly. We would like to dene a class such that the state
space needed to perform a buffer performance remains reasonably small, and nd
the order-preserving protocols to be a particularly useful restriction. A protocol
is order-preserving if copies of packets must arrive at the receivers side in the
original order. If a copy arrives out of order, the receiver discards it and requests a
new copy. From an application point of view, this means that no resequencing is
needed at the receivers side, which makes an implementation much simpler. From
3-4 ORDER-PRESERVING ARQ
the analysis point of view, this restriction has as a consequence that we need keep
track of a single packet only, thus keeping the state space small.
We keep the analysis concise by adopting a recipe that has proven its worth in
the eld of combinatorics, see e.g. [8]. It consists of three steps: rst we dene a
grammar representing the objects of interest, and then we transform this grammar
into a generating function of the length of the objects. Developing a (small) set
of rules makes this step almost automatic. From the generating functions, one can
subsequently extract useful information on different performance characteristics.
The difcult part of this recipe is usually selecting the suitable kind of objects.
We choose as objects the string of the subsequent successes, failures and discards,
which we denote as the transmission string. We associate with each time slot a
symbol from the set a, n, o. We denote by symbol a the occurence of a successful
transmission (i.e. a transmission that the receiver answers with an ACK message),
by symbol n the occurence of an unsuccessful transmission. Symbol o denotes that
either there is no transmission during that slot, or the packet is to be discarded at
the receivers side anyway. One of the key points of this paper is that we consider
the thus obtained sequence of symbols as a formal language.
In order to keep the exposition reasonably concrete, let us rst consider two
simple examples. We start with the most basic protocol, stop-and-wait. Under this
protocol, the transmitter sends a packet and then waits until the acknowledgement
message returns before sending either the next copy of the same packet (in case
of a NACK), or, in case of an ACK message, gets on with the transmission of a
new packet. For this protocol, the transmission string has the following recursive
production rule T:
T a o
s
[ n o
s
T, (3.1)
where o
s
is a shorthand for s occurrences of terminal o, and [ is a notation for
the choice operator. The formula in words reads as follows, the transmission of a
packet under stop-and-wait consists of either: (1) a slot in which a successful trans-
mission attempt takes place, followed by s slots without transmission attempts, or
(2) an unsuccessful transmission attempt, followed by rst s slots without transmis-
sion attempts, and then a period that can be seen as a complete new transmission
attempt.
For the second example, we look at a variant of stop-and-wait where the packet
is transmitted twice at every attempt:
T a o
s
[ n a o
s
[ n n o
s
T. (3.2)
Indeed, either the rst transmission is correct followed by a period of s slots in
which the sender awaits an acknowledgment message, or the rst one is incor-
rect and the second one is correct, or both are incorrect, in which case the same
procedure is started again.
CHAPTER 3 3-5
The protocols of two examples are of what we dub the stop-and-wait type.
With a few small adaptations, we can also allow for protocols of the go-back-n
type. For protocols of the rst type, the sender waits until the transmission of a
packet is certainly terminated before starting a new transmission. For the second
type, transmissions of new packets may take place although additional transmis-
sions of the old packet may be necessary. This second type is more complicated
and forces us to associate two transmission strings with each packet. The complete
transmission string of a packet is the substring that starts with the rst transmis-
sion of that packet and ends with the departure of that packet from the transmitter
buffer. We also dene partial transmission strings. This is a substring that starts
with the rst transmission of a packet and ends at the time instant that the rst
(non-ignored) transmission a new packet can start (assuming that there is a new
packet). For protocols of the stop-and-wait type, these two strings trivially coin-
cide.
By way of example, we derive the partial and complete transmission strings
(T
p
and T
c
respectively) for the standard go-back-n protocol.
T
c
a o
s
[ n o
s
T
c
T
p
a o [ n o
s
T
p
. (3.3)
Indeed, the production rule for a complete go-back-n transmission is the same as
for the stop-and-wait transmission. A partial transmission is s slots shorter.
These production rules hold regardless of the channel model, indeed without
any probabilistic framework at all. For the further analysis however, we need to
connect these abstract strings with the probabilistic channel model. As will be-
come clear in the following, all information we need can be obtained from the
matrix generating functions (mgf) of the length of the partial and complete trans-
mission strings. These are matrices of partial pgfs of dimension MM, with
entries dened as follows:
[F
T
(z)]
i j
= E[z
L
T
1( = j)[ = i], (3.4)
where L
T
, and are random variables denoting respectively the length of non-
terminal T, the background state at the rst time instant of T, and the background
state after the last time instant of T.
The following four laws sufce to derive the mgf of all grammars in this paper.
1. F
/ 0
(z) = I.
2. F
o
(z) = Qz.
3. F
S
1
S
2
(z) =F
S
1
(z)F
S
2
(z).
4. F
aS
1
[nS
2
(z) =

EQzF
S
1
(z) +EQzF
S
2
(z).
3-6 ORDER-PRESERVING ARQ
The proof of these properties is trivial.
The derivation of the mgfs with given production rules is now straightforward.
We now make an inventory of protocols we can analyse within this framework,
state the production rules for T
c
and T
p
and derive the corresponding mgfs S
c
(z)
.
=
F
T
c
(z), and S
p
(z)
.
=F
T
p
(z).
3.3.1 Stop-and-wait
From the production rule (3.1), we straightforwardly nd the mgf of the service
time. As it is the rst such derivation, we write it out in full:
S
c
(z) =F
a o
s
[ n o
s
T
(z)
=

EQzF
o
s (z) +EQzF
n o
s
T
(z)
=

EQ
s+1
z
s+1
+EQ
s+1
z
s+1
F
T
(z)
= (I EQ
s+1
z
s+1
)
1

EQ
s+1
z
s+1
. (3.5)
Also, we have that S
p
(z) = S
c
(z).
3.3.2 Go-back-N
For go-back-n, the complete transmission mgf S
c
(z) is exactly the same as for
stop-and-wait. The partial transmission mgf S
p
(z) is equal to
S
p
(z) = (I EQ
s+1
z
s+1
)
1

EQz. (3.6)
3.3.3 Stop-and-wait with m repeated copies
Next, we describe a stop-and-wait protocol with m repeated copies, hence a gener-
alization of our second example (3.2). This protocol necessitates the inclusion of
an auxiliary rule:
T
c
T
m
T
(i)
a o
s
[ n T
(i1)
T
(0)
n o
s
T
(m)
. (3.7)
This leads to
S
p
(z) = (I (EQ)
m
Q
s
z
s+m
)
1
(I (

EQ)
m
z
m
)(I

EQz)
1
EQ
s+1
z
s+1
. (3.8)
3.3.4 Persistent ARQ
This variant takes the principle of sending multiple copies to an extreme and keeps
sending the same packt until a positive acknowledgment is received.
T a o
s
[ n T. (3.9)
CHAPTER 3 3-7
For the mgf we have:
S
p
(z) = (I EQz)
1

EQ
s+1
z
s+1
. (3.10)
3.3.5 Go-back-N with m repeated copies
We can translate the principle of multiple repeated copies also to the go-back-N
family. A complete transmission is identical to a stop-and-wait transmission with
m copies. One should be a little careful when deriving the rules for a partial trans-
mission string: after a successful transmission of say the (mi)th transmission,
the sender sends i copies (or less if an ACK is received earlier) before it goes on
to the next packet, so it takes min(i, s) slots extra.
T
p
T
(m)
T
(i)
a o
min(i,s)
[ n T
(i1)
T
(0)
o
s
T
(m)
. (3.11)
3.3.6 Protocols with a limited number of transmissions
Some ARQ protocols limit the transmission of each packet to a certain number,
discarding the packet if the last try fails. The reasoning is that large delays render
packets useless, either because the packets form a part of a stream with real-time
requirements, or else because the ARQ layer is in place only to correct minor
disturbances of the channel, while failing early in case of major disturbances (for
example so that higher layers can take appropriate action). We discern a stop-and-
wait variant:
T T
m
T
i
a o
s
[ n o
s
T
i1
T
0
/ 0, (3.12)
and there also exists a go-back-N variant, for which the T
c
is equal to the stop-and-
wait variant, and T
p
is equal to:
T
p
T
m
T
i
a [ n o
s
T
i1
T
0
/ 0. (3.13)
3.3.7 Protocols with long and short encodings
On the fringes of the class of protocols that we consider in this chapter, one may
nd protocols that can choose between encodings of different lengths, so that we
3-8 ORDER-PRESERVING ARQ
can study the trade-off between longer transmission times and bigger packet error
rates. To this purpose, we introduce the symbols a
i
and n
i
, for i = 0, 1, with a new
translation rule:
F
a
i
S
1
[n
i
S
2
(z) =

P
i
z
n
i
F
S
1
(z) +P
i
z
n
i
F
S
2
(z),
where n
i
denotes the transmission length for encoding i and E
i
,

E
i
denote the
partial transition matrices after a successful and unsuccessful transmission respec-
tively. They are thus the equivalents of

EQ and EQ respectively.
A possible application of this generalized scheme is a protocol that for each
packet rst sends a shorter, less protected version and for every subsequent trans-
mission of the same packet, a better protected packet is chosen.
3.3.8 Protocols outside of this class
The prototypical example of a protocol that is not order-preserving is the selective
repeat protocol. This protocol (and its variants) can be modeled by including the
status of the last s transmissions (as demonstrated for example in [3]). This means
that per level, there are M2
s
different states which means that the state space in-
creases very quickly as a function of s, so that even moderate values of s stay out
of reach of an exact analysis. We note however that both in heavy trafc as well
as when tail behavior is concerned, the selective-repeat protocol coincides in be-
havior with the stop-and-wait protocol with s = 0 (also known as ideal ARQ). We
therefore spend in this dissertation no special attention to the transmitter side of
the selective repeat protocol. We do however spend a chapter on the resequencing
buffer of the selective repeat protocol (Chapter 5).
3.4 Performance analysis
3.4.1 Analysis of the throughput
The rst important measure that can be used to compare the various protocols
under different channel conditions is the throughput. The throughput of an ARQ
protocol is dened as the average number of transmitted packets per slot, provided
that new packets are always available. This is equivalent to the reciproque of the
average partial transmission time of a random packet.
If S
p
(z) denotes the mgf for a partial transmission, then the throughput is
CHAPTER 3 3-9
given by
= E[L
T
]
1
= (
M

i=1
M

j=1
P[ = i] E[L
T
1( = j)[ = i])
1
= (S

p
(1)1)
1
, (3.14)
where vector captures the steady-state probabilities that an partial transmission
starts in a particular state. As new packets are always available, the end of an
partial transmission always marks a the beginning of another, and hence is given
as the steady-state probability vector of transition matrix S
p
(1).
3.4.2 Analysis of the message delay
Some papers derive the distribution of a metric message delay an ARQ system
which is only a part of the the complete packet delay of the queueing system which
we will derive in section 3.4.6. The denition is as follows. Assume that a message
consisting of n packets arrives at the transmitter while the transmitter queue is
empty. Then the message delay is a random variable denoting the number of slots
to transmit successfully the entire message consisting of n packets. This is given
by the pgf M
n
(z),
M
n
(z) = S
p
(z)
n1
S
c
(z)1, (3.15)
as the entire delay is given by the adapted service times of the rst n 1 pack-
ets, followed by a complete service time of the last packet. The moments of the
message delay can be computed by the moment-generating function of generating
functions.
3.4.3 Analysis of the buffer content
In this section, we obtain the vector generating function (vgf) U(z) of the buffer
content at a random time instant. Entries of this vgf are dened as follows:
[U(z)]
i
= E[z
u
1( = i)], (3.16)
where u and are random variables denoting respectively the buffer content and
the background state at a random time instant.
As the overlapping transmissions of the packets make the head-on analysis
rather complicated, we introduce the notion of the modied buffer, which works
just the same as the original buffer, except that packets leave the buffer imme-
diately after their partial transmission is nished. By denition, no overlapping
service times occur in the adapted buffer model. (In queueing theory, it is more
common to speak about service times of packets than about transmission times).
3-10 ORDER-PRESERVING ARQ
The content of the original and modied buffer have a simple relationship, as we
will see. A similar trick was used by Towsley and Wolf [2] (specically for go-
back-n), but in this paper, we extend this to a more general class of protocols (the
difference in length between a partial and a complete service is not necessarily s
slots).
3.4.4 Analysis of the modied buffer
The evolution of the modied buffer content is of M/G/1-type, and we derive the
exact transitions in this section with the transform method. In the modied buffer,
the packets undergo partial services, given by mgf S
p
(z).
We mark the time instants at the beginning of an empty period in the modied
buffer with a symbol E, and all time instants for which the adapted buffer is empty
with symbol e. Probability vectors u
E
and u
e
record for each background state the
probability that a random time instant is marked with E and e respectively.
u

e
= u

E
+u

e
QA(0). (3.17)
Let marks B, b and c denote respectively those time instants at the beginning
of a busy period, at the beginning of a service and at a service completion.
A busy period starts after an empty slot in which arrivals occur; the buffer
content is equal to the number of arrivals in that slot.
U

B
(z) = u

e
Q(A(z) A(0)). (3.18)
At a service completion, the buffer content equals the buffer content at a service
beginning minus one (the packet itself) plus the packets arriving during the service.
U

c
(z) = U

b
(z)
S
p
(A(z))
z
. (3.19)
A beginning of a service is either the beginning of a busy period as well, or
else it is a service completion that is not empty.
U

b
(z) = U

B
(z) +U

c
(z) u

E
. (3.20)
Equations (3.17)(3.20) give us 4 equations in 5 unknowns. Hence, we can
solve u

e
, U

B
(z), U

c
(z) and U

b
(z) with only one unknown vector, u

E
remaining.
This vector is the boundary vector of the associated M/G/1-type queue and can be
found with a variety of methods see Chapter 1 for a discussion.
3.4.5 Putting it all together
Now we make the step from the modied buffer to the original buffer. Note that
beginnings of services occur at exactly the same time instants in both systems.
CHAPTER 3 3-11
This does not mean that at the beginning of a service the content of the adapted
buffer is the same as the content of the original buffer, rather, in the original buffer
are a number of packets whose adapted service has completed, yet their full service
has not.
As service completions in the original buffer occur in the order of arrival, we
can say that at the service completion of a packet in the original buffer, following
packets are in the queue: the packets that were in the modied buffer at the start
of the service of the packet, minus the packet itself, plus the packets that arrive
during the complete service of the packet. Hence the vgf U
c
(z) of the content of
the original buffer is given by
U
c
(z) = U

b
(z)
S
c
(A(z))
z
. (3.21)
The desired result, the vgf of the buffer content at random time instants, U(z)
is now only one step away:
U(z)(z 1) = U
c
(z)(A(z) 1). (3.22)
3.4.6 Analysis of the packet delay
In this section, we derive the stationary distribution of the packet delay, which is
dened as the delay of a random packet arriving in the transmitter buffer, in which
packets arrive in iid batches from slot to slot, with pgf A(z). As the rst step in this
analysis, we derive the distribution of the unnished work at the beginning of a
random slot. After this, analysing the distribution of the delay of a random packet
is an easy task.
3.4.6.1 Distribution of the Unnished Work
The unnished work at the beginning of a slot is dened as the time it takes to
serve all packets that are present in the system at that time instant, provided that
no new packets arrive. When the unnished work during a random slot is denoted
by random variable w
k
, the system equation is given by
w
k+1
= (w
k
1)
+
+g
k
, (3.23)
where g
k
is the number of slots of work that arrive during slot k. As g
k
depends on
the length of the service times, and the service times depend on the channel state
during which they start, the sequence of random variables w
k
does not form a
Markov chain. To form a Markov chain, we include an extra random variable of
the channel state during the slot immediately after the last service slot.
3-12 ORDER-PRESERVING ARQ
Let W(z) denote the vector generating function (vgf) of the unnished work
during steady state.
W(z) =
1
z
(W(z) W(0))A(S
p
(z)). (3.24)
3.4.6.2 Distribution of the Packet Delay
The delay of a random but tagged packet is equal to the sum of three terms: (1) the
unnished work during the arrival slot of that packet; (2) the partial service times
of the packets that arrive during the same slot but earlier, (3) the complete service
of the tagged packet. Hence, we see that
D(z) = W(z)A

(S
p
(z))S
c
(z)1. (3.25)
where
A

(z) =
A(z) 1
A

(1)(z 1)
. (3.26)
3.5 Asymptotic behavior
In this section we summarize the information that is contained in the Perron-
Frobenius eigenvalue of the service mgf S
p
(z). Recall that (z), (z) is an eigen-
value and corresponding left eigenvector of S
p
(z) if
(z)S
p
(z) = (z)(z).
The Perron-Frobenius Theorem says that for every irreducible nonnegative matrix,
there exists a unique largest real eigenvalue and the corresponding left and right
eigenvectors have positive entries. This guarantees the existence of a PF eigen-
value of S
p
(z) for at least z ]0, 1]. Also note that if the entries of a matrix depend
analytically on a parameter, then so do the eigenvalues and eigenvectors. We can
thus dene a analytic continuation of the PF eigenvalue also outside this region. In
particular, if [S
p
(z)]
i j
< for all z [1, z
c
[, then we can extend the PF eigenvalue
at least to ]0, z
c
[.
One may ask why the eigenvalue of S
p
(z) and not S
c
(z) ? An answer that
appeals to intuition is that the information contained in the PF eigenvalue is about
the long term, large buffer behavior, and the pgm of the complete buffer content
captures essentially a fringe effect.
Three things can be derived from this PF eigenvalue. The rst thing is the
throughput, noting that

(1) = S

p
(1)1. Secondly, the second derivative offers
insight into the asymptotic variance and hence the heavy-trafc behavior. Thirdly,
dominant poles can be derived as the smallest root z
D
> 1 of the equation z =
A((z)) for the delay and as the smallest root z
U
> 1 of the equation z = (A(z)).
Note that z
U
= (z
D
) and z
D
= A(z
U
).
CHAPTER 3 3-13
3.6 General Markov models
In this section, we explore briey the consequences of leaving the nite-state
Markov-chain framework and venturing into the Markov chains with general state
spaces. Basically, a lot of the analysis stays valid, with kernel generating functions
taking the place of matrix generating functions, dened as follows:
F
T
(z, x, A) = E[z
L
T
1( A)[ = x], (3.27)
with L
T
, and dened as before. As mentioned in Chapter 1, the full buffer anal-
ysis proceeds with operator-analytic tools, which are quite well understood from
a theoretical point of view, but on which the numerical experience is limited. The
asymptotics as mentioned in the previous section do admit easier generalizations,
as explored in the next chapter (for the ideal ARQ protocol).
3.7 Conclusion
We identied a class of retransmission protocols and proposed a streamlined anal-
ysis for them. We follow a strategy that has been applied many times in combina-
torics: we dene a suitable language and subsequently derive generating functions.
Having derived matrix generating functions, we nd expressions for the distri-
butions of the packet delay and the buffer content, as well as asymptotics on these
performance characteristics.
References
[1] Bhunia, C.T.: ARQ - Review and modications. IETE Technical Review 18
(2001) 381401
[2] Towsley, D., Wolf, J.K.: On the statistical analysis of queue lengths and
waiting times for statistical multiplexers with ARQ retransmission schemes.
IEEE Transactions on Communications 25 (1979) 693703
[3] Konheim, A.G.: A queueing analysis of two ARQ protocols. IEEE Transac-
tions on Communications 28 (1980) 10041014
[4] De Munnynck, M., Wittevrongel, S., Lootens, A., Bruneel, H.: Queueing
analysis of some continuous ARQ strategies with repeated transmissions.
Electronics Letters 38 (2002) 1295 1297
[5] Towsley, D.: A statistical analysis of ARQ protocols operating in a non-
independent error environment. IEEE Transactions on Communications 27
(1981) 971981
3-14 ORDER-PRESERVING ARQ
[6] De Vuyst, S., Wittevrongel, S., Bruneel, H.: Delay analysis of the Stop-and-
Wait ARQ scheme under correlated errors. Proceedings of HET-NETs 2004,
Performance Modelling and Evaluation of Heterogenous Networks (26-28
July 2004, Ilkley, West Yorkshire, UK), p. 21/121/11
[7] Kim, J.G., Krunz, M.: Delay analysis of selective repeat ARQ for trans-
porting Markovian sources over a wireless channel. IEEE Transactions on
Vehicular Technology 49 (2000) 19681981
[8] P. Flajolet and R. Sedgewick: Analytic combinatorics. Cambridge University
Press, Cambridge, 2009.
[9] Gilbert, E.N.: Capacity of a burst noise channel. The Bell System Technical
Journal 39 (1960) 12531265
[10] Gail, H. R., Hantler, S. L., Taylor, B. A.: Spectral analysis of M/G/1 and
G/M/1 type Markov chains. Adv. Appl. Prob. 28 (1996) 114165
[11] Gantmacher, F. R.: The Theory of Matrices, Volume 1. AMS Chelsea Pub-
lishing, Providence, Rhode Island (1959)
[12] Bruneel, H.: Buffers with stochastic output interruptions. Electronics Letters
19 (1983) 735737
[13] Kleinrock, L.: Queueing Systems, Volume I: Theory. Wiley, New York
(1975)
[14] Bruneel, H., Kim, B.G.: Discrete-time models for communication systems
including ATM. Kluwer Academic Publishers, Boston (1993) (ISBN: 0-
7923-9292-2)
[15] Abate, J., Whitt, W.: Numerical inversion of probability generating func-
tions. Operations Research Letters 12 (1992) 245251
4
Moderate deviations of retransmission
buffers over wireless fading channels
Koen De Turck, Marc Moeneclaey, Sabine Wittevrongel
presented at the HETNETS conference January 2010, Zakopane, Poland
Abstract. We study the buffer performance of retransmission protocols over
a wireless link. In contrast to the previous chapters, the channel gain is modeled
as a discrete-time complex-valued Gaussian model. The advantage of this channel
model over simpler alternatives (such as nite-state Markov models) is that the
correspondence with the physical situation is more transparent. In order to keep
the performance analysis of the buffer content tractable, we opt for heavy-trafc
and moderate deviations scalings.
4.1 Introduction
The popularity of wireless telecommunications is increasing rapidly. However,
apart from the obvious advantages over wired networks, such as increased user
mobility and easier deployment, wireless communications also have a number of
drawbacks. For example, the dependence on batteries requires a more careful
energy management. Secondly, the presence of fading and interference is also
particular to wireless links and may cause a severe degradation of performance.
4-2 MODERATE DEVIATIONS FOR FADING CHANNELS
In this chapter, we look at the performance loss due to fading. This loss mani-
fests itself in a reduced throughput of the link, but other performance metrics may
be severely affected as well, such as the mean packet delay, the overow proba-
bility of the buffer at the transmitters side or the delay jitter. This is by no means
a new topic; we refer to [1, 3, 4] as a sample of how this problem has been tack-
led, some of the papers focus on throughput only, others on the complete buffer
performance. The novelty of this analysis resides in the fact that we combine a
couple of elements in a way that has not been done before, which manages to cir-
cumvent a rather huge computation burden. Firstly, we make use of a complex
Gaussian process as a model of the channel gain. This type of model is more
popular in communications theory circles, than for queueing analyses, because a
direct computation of the buffer content distribution is too resource-intensive to be
useful. For this reason, researchers interested in the buffer content have hitherto
focused on nite-state Markov models. The undeniable advantage of Gaussian
models however is that the correspondence with physical parameters is transpar-
ent: metrics such as the SNR (signal-to-noise-ratio) and coherence time feature
directly.
During recent years, scaling has become a respected and full-edged analysis
tool. When confronted with a problem for which a direct computation is very
costly or plainly impossible, the probabilist might opt to scale the original problem
in such manner that a much simpler model arises, one in which the salient features
of the original model are retained, but other stochastic uctuations get ltered
away. We look at two scaling methods in particular, namely heavy-trafc and
moderate deviations. Heavy-trafc analysis is easily the oldest scaling method
known in queueing theory. Kingman [5] in the sixties was the rst to exploit
the deep link between queueing systems operating under the limit load 1
and diffusion processes. Moderate deviations do not have such a long history.
Its promise is to combine the strong points of large deviations and heavy trafc
methods. Essentially, it is a rigourous way of looking at tails of asymptotically
Gaussian processes. We are indebted to the scaling approach taken in [6, 7].
The structure of the chapter is as follows. In section 4.2, we detail the channel
and the buffer model; in section 4.3, we review the moderate deviations and heavy
trafc scalings and apply them to the model at hand. Next, we add some notes on
optimal control in section 4.4. We look at some numerical examples in section 4.5
and nally, we draw conclusions in section 4.6.
4.2 Model
Consider a wireless station (the transmitter) delivering data packets to another
wireless station (the receiver). Time is considered to be slotted, where the du-
ration of a slot corresponds to the transmission time of a data packet of length
CHAPTER 4 4-3
L bits. The transmission buffer has room for B data packets. The channel over
which the information is sent is subject to fading, which we model as follows. The
channel gain h
t
during slot t N forms a discrete-time complex Gaussian process.
We assume wide-sense stationarity, and moreover Eh
t
= 0. The process is thus
characterized completely by an autocorrelation function r
t
:
r
t
.
= E(h

s
h
t+s
) = E(h

0
h
t
),
where z

denotes the complex conjugate of z.


The process h
t
can also be characterized as ltered white noise. Indeed, con-
sider a sequence
t
of independent and identically distributed (iid) complex normal
variables with zero mean and unit variance, and a lter bank with parameters g
t
such that:
h
t
=

s
g
s

ts
.
The two representations are in fact equivalent. A popular choice in this case is the
Butterworth lter. In this paper, however, we do not further elaborate on the lter
representation of the channel process.
Two choices of r
t
are particularly popular. Jakes model [9] is perhaps the most
well-known choice. It was derived from theoretical considerations, and expresses
r
t
in terms of a Bessel function of the rst kind:
r
t
= J
0
(2f
d
t),
where f
d
is the Doppler frequency. There is a simple relation between the the
Doppler frequency f
d
, the carrier frequency f
c
and the velocity of the receiver:
f
d
=
v
c
f
c
,
where c denotes the speed of light. This shows the strong inuence of the car-
rier frequency and the velocity on the nature of the fading process. The Doppler
frequency also corresponds to the cut-off frequency of the lter.
The other popular form for the autocorrelation function is Gaussian:
r
t
= exp
_

t
2
2
2
_
, (4.1)
where is a form factor regulating the width of the autocorrelation function. We
can relate this to the Doppler frequency by determining the cut-off frequency in
the frequency domain. A Gaussian function with form factor in the time domain
is mapped unto a Gaussian function with form factor
1
in the frequency domain.
Some manipulations yield the following formula for the n-dB cut-off frequency:
f
d
=
_
n
5
log10
1
. (4.2)
4-4 MODERATE DEVIATIONS FOR FADING CHANNELS
Our overview of the channel model is completed by the link between the chan-
nel gain and the transmission error process. The bit error probability is a function
of the channel gain h as follows:
p
b
(h) = Q
_
_
2E
b
N
0
[h[
2
_
,
where
E
b
N
0
denotes the SNR and Q(x) denotes the error function; it is equal to the
probability that a normal random variable with zero mean and unit variance is
larger than x. The packet error probability p(h) is the probability that at least one
bit of the packet is incorrect:
p(h) = 1(1p
b
(h))
L
.
Let c
t
denote the transmission process: c
t
is equal to 1 when the transmission
during slot t is successful and 0 otherwise. We have that
c
t
= i
1p(h
t
)
,
where i
q
denotes a Bernoulli random variable with success probability q.
Let a
t
be the stationary stochastic process of the number of packet arrivals
during slot t. We will impose more specic assumptions for each particular scaling
we are considering. Let
.
= Ea
0
;
.
= Ec
0
. The load of the system is dened as
= /.
In this paper, we look at the transmitter buffer performance, with the so-called
ideal ARQ (ARQ stands for automatic repeat request) protocol: packets are re-
transmitted until they are received correctly, (until c
t
= 1). The adjective ideal
refers to the assumption that there is no feedback delay. That is, the transmis-
sion status of a packet is directly known. The scalings that we consider in this
paper involve letting the load approach 1, and under such conditions ARQ with
non-zero feedback delay converges to ideal ARQ. The queue content process u
t

is formulated in terms of the arrival and transmission processes, by means of the


well-known Lindley recursion:
u
t+1
= [u
t
+a
t
c
t
]
B
0
.
where [x]
B
0
.
= max(0, min(B, x)).
4.3 Scalings
We obtain asymptotic results on the queue content distribution by appropriately
scaling the arrival and transmission streams. In this context, it is customary to
dene the net-input process w
t
.
= a
t
c
t
. Even within the class of scaling methods
(which are by themselves already approximative) we have to be careful as to which
methods offer good approximations for a reasonably low computational effort.
CHAPTER 4 4-5
4.3.1 Fast-time scaling
We consider a set of scalings that involve speeding up the net-input process. Let
w
L
denote the net-input process sped up by a factor L:
w
L
t
.
=
(t+1)L1

s=tL
w
s
.
Let us look at a family of scalings of the form:
w = L
(1)/2
(L
1
w
L
()1). (4.3)
where [0, 1] and 1 denotes a constant process and equal to 1. For = 0, we get
the so-called central limit scaling, whereas = 1 is known as the large deviations
scaling, (which is essentially the same as the scaling used for the law of large
numbers).
Let us rst have a look at the central limit scaling = 0. Under the condition
that the rst two moments exist and the net-input process is mixing, the scaled
process converges to a (discrete sample of) Brownian motion with zero drift and
diffusion parameter V
w
:
V
w
= lim
t
1
t
Var[
t

s=1
w
s
].
The queue content process under this scaling converges to a reected Brownian
motion with drift , diffusion parameter V
w
, and boundaries at 0 and B. This
leads to a couple of simple performance formulae [8]:
E[u]
1

+
B
1exp(B)
. (4.4)
and
P[u b]
e
B
e
B
1
. (4.5)
where
.
= 2()/V
w
.
Large and moderate deviations scalings under this scaling. They involve com-
puting a rate function, which for the transmission process at hand is either compu-
tationally complex (in the large deviations case), or leads only to the asymptotic
formula (4.5) of the central limit result (in the moderate deviations case). Indeed,
for the large deviations case, computations center around the scaled cumulant gen-
erating function (scgf) ():
() = lim
T
1
T
logE[exp(w
T
0
)],
from which the rate function can be obtained by a Legendre-Fenchel transform.
The computation of () requires the evaluation of a high-dimensional integral
4-6 MODERATE DEVIATIONS FOR FADING CHANNELS
(with the dimension tending to innity). This can only be solved by costly Monte-
Carlo techniques, and as we set out to nd easy to compute performance formulae,
we will not further pursue this path. We note that the moderate deviations limit in
fact corresponds to a second-order approximation of the rate function [7].
4.3.2 Many-ows scalings
We now look at a scaling that preserves the time-covariance structure of the orig-
inal net-input process: instead of speeding up this process, we denote by w
L
an
aggregate of L independent copies of the same net-input process. The family of
scalings has now the following form:
w = L
(1)/2
(L
1
w
L
()1), (4.6)
again for [0, 1]. In the central limit scale = 0, the scaled process now con-
verges to a Gaussian process (not necessarily Brownian motion) with the same
drift and covariance structure as the original net-input process w. Although the
queue-content process also converges to a Gaussian process, it is generally dif-
cult to derive closed-form performance metrics for it. Only in special cases (such
as single-server Brownian queues), closed form results are available ( for an ac-
count, see [11]). This is why we resort to moderate deviations in this case. Under
mild conditions [6, 7], the scaled process satises a moderate deviations principle
(MDP) for (0, 1) with rate function I
t
:
lim
L
L

logP[ w

S] inf
t>0
inf
x

S
I
t
(x), (4.7)
where I
t
(x) is equal to
I
t
(x) = sup
R
t

T
x
1
2

T
C
t
, (4.8)
with C
t
the covariance matrix of the net-input process (with dimension t t):
[C
t
]
i j
=
[ij[
= Cov[w
i
, w
j
]. (4.9)
The tail asymptotics of the queue content process are given by [7]:
logP[u b] I, (4.10)
where
I = inf
t0
(b+()t)
2
2V
t
. (4.11)
We detail in the next subsection how to compute the variance function V
t
, which is
dened as follows: V
t
=
i, j
[C
t
]
i j
. One can also use the rened asymptotics of
the Bahadur-Rao type [10]:
P[q b]
1

2V
t

e
I
, (4.12)
where t

is the t that minimizes (11.5), and

= (b+()t)/2V
t
.
CHAPTER 4 4-7
4.3.3 Computing the covariance structure
In this section, we show how to compute the function V
t
that appears in the asymp-
totic performance measures of the previous section. First, note that the net-input
process is the sum of two independent processes: the arrival process and the trans-
mission process, which means that V
t
can be split up likewise:
V
t
=V
a
t
+V
c
t
. (4.13)
For the arrival process, we opt in this paper for the parsimonious fractional
Brownian process, which has three parameters: a drift , a diffusion parameter

2
and a Hurst parameter H, where H (0, 1). For H =
1
2
, we have the standard
Brownian motion with independent increments, In case of H <
1
2
the increments
are negatively correlated, and positively correlated for H >
1
2
. We have that V
a
t
=

2
t
2H
.
The function V
c
t
of the transmission process can be found via the auxiliary
sequence
t
:

t
.
= E[(c
0
)(c
t
)]
= E[(1p(h
0
) )(1p(h
t
) )]. (4.14)
where the last transition is due to the denition of i(.). The computation of this
sequence is best done numerically. The computational complexity is relatively
minor, however: for each t we must evaluate a four-dimensional integral (recall
that h
t
are complex-valued random variables, thus yielding two dimensions each).
Sequences V
c
t
and
t
are related as follows:
V
c
t
=
t1

i=0
t1

j=0

[ij[
. (4.15)
The asymptotic variance, which plays a central role in fast-time scalings, is
equal to the limit V = lim
t
V
t
/t. Note that this limit may not exist, for example
for fractional Brownian with Hurst parameter H >
1
2
.
4.4 Optimal control
The transmission over wireless channels poses a challenging control problem to
the designer of wireless networks: which level of the SNR represents the optimal
trade-off between quality-of-service and energy consumption ? The parsimonious
performance formulae presented in the previous section offer a feasible path to the
static optimization. Indeed, assume given as a QoS constraint that the overow
probability must be smaller than P. The SNR inuences the transmission rate
4-8 MODERATE DEVIATIONS FOR FADING CHANNELS
and the variability of the transmission process V
c
t
. The buffer size b inuences the
overow probability. The control problem is thus as follows:
Find the minimal buffer size b and SNR such that:
logP > inf
t0
(b+((SNR) )t)
2
2(V
a
t
+V
c
t
(SNR))
.
One can also adapt the SNR dynamically according to the perceived current
channel and trafc state. This dynamic optimization problem is a lot harder, and
is the subject of future research.
4.5 Some numerical results
Consider a scenario in which a transmitter sends data packets of L = 10000 bits
to a receiver over a wireless channel subject to fading. The duration of a packet
transmission is 5 ms. The carrier frequency f
c
of the transmission is 1Gz, and the
receiver moves relative to the sender with a velocity of v. In the rst pair of gures,
we show autocorrelation function of the channel gain for different velocities and
for the Gaussian and Jakes model respectively. Note that the manner in which
the two models decay is completely different. We also show the corresponding
autocorrelation function
t
of the transmission process c
t
for the same scenarios in
gure 2. Note that again, for a Jakes model there are a lot of small bumps after
the main bump, whereas for the Gaussian model there is only one bump. Although
the bumps appear small they have an considerable inuence on the function V
t
.
Next, we turn our attention to the buffer performance proper. We plot the tail
probabilities of the buffer occupancy for different velocities in the left subplot of
gure 3. We see that the velocity has a huge effect on the buffer performance. In
the right subplot, we look at the log-probability of the buffer exceeding a certain
level (b = 80) versus the velocity v. We see that above some speed the inuence
gets minimal.
In the last gure, we show that when the arrival source is really bursty (Hurst
parameter H = 0.7, signifying a large positive correlation), the performance of the
buffer deteriorates to the extent that the inuence of the fading channel is hardly
seen.
4.6 Conclusion
We studied the moderate deviations asymptotics of a retransmission buffer over
a wireless fading channel. We found easy to evaluate performance formulae that
link important physical parameters such as signal-to-noise ratios, coherence time
and so on. The most important conclusions are: (1) the throughput alone does
not sufce to characterize the buffer performance, (2) the lower the velocity of
CHAPTER 4 4-9
v = 10
v = 1
0 100 200 300
0.5
0
0.5
1
t
r
t
Figure 4.1: Channel autocorrelation functions r
t
with Gaussian form (dashed lines) and
Besselian form (Jakes model; full lines) for different values of v(in km/h).
v = 10
v = 1
0 50 100 150 200
0
0.05
0.1
0.15
t

t
Figure 4.2: Transmission autocorrelation function
t
for different values of v.
Gaussian-form models are in dashed lines; Jakes models are in full lines. Values of the
other parameters are: f
c
= 1 GHz; T
p
= 5ms;L = 10000; SNR=14dB.
the receiver the worse the buffer performs (3) the effects of the fading channel
might be swamped by really bursty (or Hursty) arrival sources, especially when
the packet error probability is reasonably low, and (4) Gaussian and Jakes fading
models give different tail behavior.
References
[1] Konheim, A.G.: A queueing analysis of two ARQ protocols. IEEE Transac-
tions on Communications 28 (1980) 10041014
[2] De Vuyst, S., Wittevrongel, S., Bruneel, H.: Delay analysis of the Stop-and-
Wait ARQ scheme under correlated errors. Proceedings of HET-NETs 2004,
Performance Modelling and Evaluation of Heterogenous Networks (26-28
July 2004, Ilkley, West Yorkshire, UK), p. 21/121/11
4-10 MODERATE DEVIATIONS FOR FADING CHANNELS
v = 1
v = 10
0 20 40 60 80 100
12
10
8
6
4
2
0
b
l
o
g
P
(
u

b
)
SNR=10
SNR=14
0 20 40 60 80
8
6
4
2
v
l
o
g
P
(
u

8
0
)
Figure 4.3: Top: Log probabilities of the buffer content u for different values of v;
Gaussian-form models are in dashed lines; Jakes models are in full lines. Bottom:
logP(u 80) versus the velocitity v of the receiver, for a Gaussian-form model. Values of
other parameters are: SNR=14dB;H = 0.5;V
a
= 0.1.
[3] R. Fantacci. Queueing analysis of the Selective Repeat automatic repeat re-
quest protocol wireless packet networks. IEEE Trans. Veh. Technol, 45:258
264, 1996.
[4] Kim, J.G., Krunz, M.: Delay analysis of selective repeat ARQ for trans-
porting Markovian sources over a wireless channel. IEEE Transactions on
Vehicular Technology 49 (2000) 19681981
[5] J.F.C. Kingman: On queues in heavy-trafc. J. Royal Stat. Soc. ser. B, 24,
383392 (1962).
CHAPTER 4 4-11
H = 0.5
H = 0.7
0 20 40 60
4
3
2
1
0
1
b
l
o
g
P
(
u

b
)
Figure 4.4: A plot of logP(u b) versus b for an SNR of 10dB (dashed lines) and of 14dB
(full lines) for two values of the Hurst parameter. Values of the other parameters are:
V
a
= 1; v = 5km/h.
[6] Ayalvadi Ganesh, Neil OConnell, Damon Wischik: Big Queues. Springer
Verlag Berlin (2004)
[7] Damon Wischik: Moderate deviations in queueing theory. (submitted for
publication)
[8] Ward Whitt: Stochastic-process Limits. Springer Verlag, New York (2002).
[9] William C. Jakes: Microwave Mobile Communications. Wiley and Sons,
New York, 1975.
[10] Montgomery, M., DeVeciana, G.: On the relevance of time scales in perfor-
mance oriented trafc characterizations. Proceedings of INFOCOM 1996.
[11] Michel Mandjes. Large deviations for Gaussian queues: modelling commu-
nication networks. Wiley and Sons, New York, 2007.
5
Receiver Buffer Behavior for the
Selective Repeat Protocol over a
Wireless Channel: an Exact and
Large-Deviations Analysis
Koen De Turck and Sabine Wittevrongel
published in Journal of Industrial and Management Optimization, 2010, vol.
6, no. 3, pp. 603-619.
Abstract. In this chapter, we formulate and analyze a model of the resequenc-
ing buffer at the receivers side for the Selective Repeat protocol over a general
class of transmission channels. Thanks to its efciency, Selective Repeat is a ubiq-
uitous error control mechanism in many different settings, in particular in wireless
protocols such as WiMax and WiFi. In contrast to the asymptotic behavior of the
buffer at the senders side, the buffer at the receivers side was until now a less
researched subject.
In view of the correlated nature of transmission errors over wireless channels,
the receiver buffer model considers a general Markovian error process. We provide
both an exact mathematical analysis of the receiver buffer behavior as well as a
computationally efcient large-deviations result. An asymptotic analysis of the
delay is also given. Numerical examples show that the correlation of the error
5-2 RECEIVER BUFFER BEHAVIOR FOR SELECTIVE REPEAT
process has an important inuence on the performance of the receiver buffer.
5.1 Introduction
Retransmission protocols are a crucial part of almost any means of telecommunica-
tions, in almost any layer of the protocol stack. They play an especially important
role in protocols over wireless links such as WiFi and WiMAX, as these links often
exhibit an error process that is both bursty and has high error rates. The principle
is simple: the sender transfers data (in the form of a numbered packet stream) to a
receiving unit over an unreliable channel. The receiver acknowledges the status of
each of the received packets, to which the sender responds by resending the incor-
rectly received packets. Many retransmission protocols have been proposed, each
distinguished by the way in which the retransmissions and the acknowledgements
are organized.
Of the basic types, the Selective Repeat protocol is the most efcient protocol
in terms of throughput, and as such it is widely implemented in numerous appli-
ances. However this efciency comes at a price, as it is one of the most complex
retransmission protocols to implement (and to analyze).
As packets might arrive out of order, Selective Repeat requires in contrast to
other protocols a resequencing buffer at the receivers side, in which a correctly
received packet can be stored until the original packet order can be restored.
We provide in this chapter both an exact and a large-deviations analysis of the
receiver buffer behavior. The latter results in computationally simple and insightful
formulas of the essential performance characteristics. It is our belief that these
may be more useful to the network analyst than an exact and elaborate analysis
that is tedious to compute and to interpret. The analysis in this chapter might
prove to be instructive beyond the present application, as it shows a way to obtain
efciently computable performance measures for a model that is particularly struck
by the curse of dimensionality. That is, the state space explodes exponentially as
a function of the roundtrip time, which means that traditional solution techniques
become intractable for even moderate values of that parameter. We show that
we can nevertheless extract with only minor efforts the decay rate of the buffer
content distribution, a performance measure that is often argued to be among the
most useful for teletrafc engineers.
The goal of the chapter is rst of all to study the performance characteristics
of the SR protocol and especially to clarify the behavior of the reordering buffer
at the receiver side and the true effect of error correlation. These issues have not
yet been properly claried before. A further goal of the chapter is to communi-
cate another successful application of large-deviation theory to the performance
modeling community.
Previous work has been done on the queueing performance of the Selective
CHAPTER 5 5-3
Repeat protocol, most papers focusing on the buffer at the senders side. The
classic paper by Konheim [17] provides an exact solution, that however suffers
from the above-mentioned curse of dimensionality. Various approximations with
diverse levels of accuracy and mathematical justication have been proposed in
[1, 11, 16]. More recent papers also study the impact of correlation on the protocol
performance (see e.g. [4, 7, 8, 13]). A key asset of the present work is that we
provide mathematical proofs of the proposed approximations.
Papers which focus on the queue at the receivers side have been less numerous.
In the papers [5, 19], independent and identically distributed (IID) error processes
were considered, whereas in the present paper we extend this to general Markovian
error processes. More recent efforts are reported in [6] and [22]. In [6], as in the
present work, Markovian instead of IID processes were considered, however in
the arrival process instead of the error process. The analysis of [22] shares with
the present work a strong emphasis on tractable formulas, but it is applied to a
situation where packet delays are variable and packet errors are non-existent, quite
different from the lower-layer Selective Repeat setting that is the subject of this
chapter.
The outline of this chapter is as follows. In section 5.2, we describe the math-
ematical model of the resequencing buffer. Section 5.3 contains a method to com-
pute the exact solution. The large-deviations analysis is performed in section 5.4,
and in section 5.5 we look at how exactly large buffer occupancies occur. We study
the asymptotic distribution of the packet delay in section 5.6. Thereafter, in sec-
tion 5.7, we explore the similarity with a well-known combinatorial problem, the
coupon-collector problem. Next, we show some numerical results in section 5.8.
We conclude in section 5.9.
5.2 Mathematical Model of the Resequencing Buffer
We extend the mathematical model of [5] and [19] to more general error processes.
The sender is assumed to operate under saturated-buffer conditions, which means
that there are always packets available to be sent. This assumption frees us from
specifying any arrival process, and can moreover be considered as a worst case.
Indeed, the fact that sometimes there are no new packets to be sent, can only have
a positive effect on the performance of the system. We assume that the time is di-
vided into slices of equal length, called slots, and the transmission of one packet is
assumed to take exactly one slot. The receiver sends an acknowledgement message
of each packet. If a packet is incorrectly received, the transmitter sends a new copy,
that arrives at the receiver exactly L slots later. This duration of L slots is called
the roundtrip time or the feedback delay of the channel, and has a tremendous ef-
fect on the performance of the system, as we will see. If the transmitter receives
a positive acknowledgement message, it just takes a fresh packet and attempts to
5-4 RECEIVER BUFFER BEHAVIOR FOR SELECTIVE REPEAT
send it.
When looking at a sample path as in Figure 5.1, it is handy to t it into a
rectangular strip of width L. Time proceeds as in a Western text, left to right,
top to bottom. As in [5] we call a vertical line in such a plot a group and note
that every transmission attempt of the same packet takes place during the same
group. How can we determine how many packets are waiting in the resequencing
buffer at an arbitrary slot boundary t ? It is simply the number of all the correctly
received packets that were rst transmitted after the eldest erroneously received
packet (EEP). This EEP can be detected as the longest uninterrupted gray region
looking from bottom to top. The buffer content at time t is hence simply the
number of white squares, counted from time t backwards until we meet the EEP
(the dashed squares do not count as they represent packets that are elder than the
EEP). This is basically the reasoning in [5] and in [19].
Figure 5.1: First method to determine the resequencing buffer content at a given time
instant (L = 6). Erroneous transmissions are indicated in gray. Dashed boxes represent
transmissions of packets that have already left the buffer. We identify the EEP (here:
packet 10) and then count the number of packets with higher sequence numbers that are
already successfully transmitted. We thus nd 10 packets in the resequencing buffer.
In the present paper, we take a related, and arguably a somewhat more elegant
approach (and in case of correlated errors, perhaps the only feasible one). We de-
ne a process that runs backwards in time from a random slot boundary onwards,
and we count the number of correct packets that we encounter, as illustrated in Fig-
ure 5.2. We stop the process when there has been a correct packet in every group.
It appears that the buffer content at that random slot boundary is the total number
of correct packets minus L. This means that the buffer content at stationarity has
the same distribution as the random variable U
T
, which is dened in terms of the
discrete-time process (U
t
, V
t
) and the stopping time T, which have the following
CHAPTER 5 5-5
denition:
U
0
=L; V
0
=
L
..
(0, 0, , 0);
U
t+1
=U
t
+P
t
; V
t+1
= rotate(V
t
) (P
t
, 0, , 0), (5.1)
and
T = mint : V
t
= 1. (5.2)
A few words on the notation. The random vector V
t
is a Boolean vector of length
L (i.e., it consists only of zeros and ones). We denote the set of such vectors by
B
L
. The function rotate(.) rotates a vector as follows:
rotate(v
1
, v
2
, , v
L
) = (v
2
, , v
L
, v
1
), (5.3)
and the operator performs an elementwise logical or. Finally, P
t
is the stationary
ergodic error process, with P
t
=0 if an error occurs at time t, and P
t
=1 otherwise,
where it must be understood that P
t
runs backwards in time. The random variable
U of the stationary buffer content hence corresponds in distribution to the random
variable U
T
.
It is easily veried that T is indeed a stopping time with respect to the process
(U
t
, V
t
), i.e., it is adapted to the natural ltration of the process. The above process
can be viewed as a variant of the classic coupon-collector problem [12], which
goes as follows. Let n objects be picked repeatedly according to a certain random
process. Find the earliest time at which all n objects have been picked at least
once. The difference between the problem at hand and the classic problem is that
in the former, one attempts to get a specic object at any given time, in a cyclic
manner, whereas in the latter one typically receives a random coupon at every
attempt. In section 5.7 we show that in high-error environments, we obtain in
the limit the familiar coupon-collector solution. Note that the same solution was
obtained in [5], although without mentioning the relation with the coupon-collector
problem.
With respect to the packet error process, we mainly consider two classes: (1)
the classical independent and identically distributed (IID) error process, in which
packet errors occur with a xed error probability p; and (2) nite state-space
Markovian error processes, which have become the model of choice in many per-
formance studies over wireless links [3, 4, 9, 20]. The large-deviations analysis can
be extended to more general error processes as well, a fact we briey elaborate on
in section 5.4.
The Markovian error processes we consider in this paper consist of a discrete-
time Markov chain (the background Markov chain) with a nite number of states
M, which we require in the present work to be irreducible. Furthermore, each
state has a distinct packet error probability. Let matrix Q denote the transition
5-6 RECEIVER BUFFER BEHAVIOR FOR SELECTIVE REPEAT
Figure 5.2: The alternative method to determine the resequencing buffer content, which
accents the kinship with the classical coupon-collector problem. Erroneous transmissions
are again indicated in gray. We travel backwards and count the number of successful
transmissions until every group has at least one. Here, the dashed boxes represent
successful transmissions that are not counted (that is, after time T). The buffer content is
then found by subtracting the length of a roundtrip time. Hence, we nd 166 = 10
packets in the resequencing buffer.
probability matrix of the background Markov chain, and let vectors p
0
and p
1
record respectively the error and success probabilities during each state. The vector
denotes the steady-state probability vector of the Markov chain. That is, is
the unique normalized vector which satises the equation = Q. As is well-
known, the transition matrix Q
(r)
of a time-reversed Markov process is related to
the transition matrix Q of the original process by the following formula:
Q
(r)
= diag()
1
Q
T
diag(), (5.4)
where diag(v) denotes the diagonal matrix constructed from vector v and Q
T
is the
transpose of Q. In order to minimize notational clutter, we denote in the sequel by
Q the transition matrix of the Markov process that travels backwards in time, as
this will prove to be by far the most prevalent. We also introduce the matrices P
0
and P
1
, dened as follows:
P
0
= diag(p
0
)Q and P
1
= diag(p
1
)Q. (5.5)
5.3 Exact Analysis of the Receiver Buffer Content
In this section we present a method for the exact analysis of the buffer content
distribution. That is, we want to nd the distribution of U
t
when V
t
hits the state
(1, , 1), (i.e. at stopping time T). This is related to so-called exit problems.
CHAPTER 5 5-7
While exit problems are about the distribution of the time until a certain set of
states are hit, here we are interested in the value of an auxiliary process U
T
at the
hitting time.
We dene a vector generating function U
v
(y, z) that keeps track of the time, of
U
t
, V
t
, and of the background state x
t
of the error process:
[U
v
(y, z)]
i
=

t=0
y
t
E[z
U
t
1
V
t
=v,x
t
=i
], (5.6)
where notation 1
A
denotes the indicator function, i.e., it is equal to 1 when event
A is fullled and zero otherwise. Note that U
v
(y, z) is a transient generating
function, that is, it explicitly keeps track of the time t through the argument y.
After fairly straightforward manipulations in the transform domain, we can
derive from the denition of the process (U
t
, V
t
) that the following set of equations
holds:
U
0
(y, z) = z
L
+yU
0
(y, z)P
0
; (5.7)
U
1v
(y, z) = yzU
v0
(y, z)P
1
+yU
v1
(y, z)(P
0
+zP
1
); (5.8)
U
0w
(y, z) = yU
w0
(y, z)P
0
; (5.9)
U
1
(y, z) = yzU
10
(y, z)P
1
, (5.10)
for all v B
L1
1 and for all w B
L1
0.
Looking at the denition of the process (U
t
, V
t
), we can deduce that the gen-
erating function U(z) of the buffer content under steady-state conditions is equal
to
U(z) = U
1
(1, z)1. (5.11)
In principle, the above derivations supply us with a system of 2
L
equations
from which the generating function U(z) can be obtained. By way of example, we
explicitly derive the buffer content distribution for the case L = 2. We have that
U
00
(1, z) = z
2
(I P
0
)
1
;
U
10
(1, z) = zU
00
(1, z)P
1
+U
01
(1, z)(P
0
+zP
1
);
U
01
(1, z) = U
10
(1, z)P
0
;
U
11
(1, z) = zU
10
(1, z)P
1
.
From this set of equations, we see that the buffer content distribution has the fol-
lowing generating function:
U(z) = (I P
0
)
1
P
1
(I P
2
0
zP
0
P
1
)
1
P
1
1. (5.12)
Although for L =2 the expression is simple enough, for larger L it simply becomes
intractable to solve this system of 2
L
equations. As we already stated in the intro-
duction, this problem is severely affected by the curse of dimensionality. Note
5-8 RECEIVER BUFFER BEHAVIOR FOR SELECTIVE REPEAT
that, as in [5], we can circumvent this curse for IID error processes, as in that case
we can perform some tricks to obtain an explicit expression for U(z) for general L.
The non-commutativity of matrices however prevents us from using these tricks in
case of Markovian error processes.
As a numerical procedure, we nd that this kind of analytical solution falls
a bit short. In fact, for large L, it is very well possible that even a Monte-Carlo
simulation of the stochastic process (U
t
, V
t
) can produce meaningful numerical
results faster and with less memory usage than the solution sketched above, espe-
cially if that simulation program makes optimal use of importance sampling, or a
similar variance reduction technique. In the next section we therefore show some
computationally efcient analytical approximations.
5.4 Large-Deviations Analysis
Since the rst steps were made in the works of Cram er in the 1930s, and especially
after Donsker and Varadhan put it on a rm mathematical basis, large-deviations
(LD) theory has always been a useful tool for practitioners in such diverse ar-
eas as risk management, statistical physics and teletrafc engineering. Roughly
speaking, large-deviations theory concerns the asymptotic behavior of stochastic
processes, that is of situations that are large deviations from the mean. Typically, it
can only tell something about the exponential decay rate of the probability of rare
events, and as such it is not the most detailed of analytical tools. However, the de-
cay rate of certain rare events is often the most crucial performance measure, and
furthermore, the power of large-deviations theory resides in the fact that it can be
generalized rather easily, and that it can, with relative simplicity tell us something
about systems for which other methods including simulations may fail.
In the following, we set out to compute the exponential decay rate for the buffer
content distribution. That is, it is often observed (from simulations etc.) that for
large k,
P[u > k] Ce
k
, (5.13)
and we would like to compute the decay rate efciently, (which is in practice of
a far greater importance than C). Formally, is dened as the limit
= lim
k
1
k
logP[u > k], (5.14)
provided that this limit exists.
We start out fromthe denition of the stochastic process (U
t
, V
t
). If U
T
ends up
with a value larger than k, it is because U
t
reaches a value of k while the stopping
time has not yet occurred, i.e., there exists a time instant t for which U
t
becomes k
CHAPTER 5 5-9
and V
t
,= 1. This leads to:
P[U
T
> k] =

t
1
t
P[U
t1
= k 1, V
t
,= 1, P
t1
= 1]. (5.15)
If the stopping time has not yet been reached, it means that no successful trans-
mission has occurred in (at least) one of the groups. Let us dene the modied
process Y
s
t
where parameter s indicates that group s is entirely erroneous:
Y
s
t
=U
t
1
P
t
=1
k0,kL+s<t
1
P
kL+s
=0
. (5.16)
We derive a large-deviations principle for process Y
s
t
by means of the G artner-
Ellis Theorem [10]. In order to do so, we dene the scaled cumulant generating
function
s,t
():

s,t
()
.
=
1
t
logE[e
Y
s
t
], (5.17)
and the corresponding limit t :

s
() = lim
t

s,t
(). (5.18)
For Markovian error processes,
s,t
() is equal to

s,t
() = logP()
s1
R()
k
P()
r
P
1
e

1, (5.19)
where k and r are uniquely dened by relations t = s +kL+r +1 and 0 r < L.
Also, we dened P() =P
0
+e

P
1
, and R() =P
0
P()
L1
. Note that when t goes
to innity, so does k, with k L
1
t whereas s and r remain bounded. This enables
us to invoke the Perron-Frobenius theory as follows:

s
() = lim
k
1
kL
logv()R()
k
w() (5.20)
=
1
L
log(R()), (5.21)
where the notation (.) denotes the spectral radius of a matrix. In the rst step,
we performed a change of variables in the limit, and assembled the asymptotically
vanishing parts into the row vector v() and the column vector w(). In the second
step, we made use of a property that is proved in [10], 3.1, and that holds when-
ever matrix R() is an irreducible matrix. Note that the spectral radius in that case
is identical to the largest eigenvalue, and hence an intuitive explanation resides in
the fact that in the limit, the contribution of the largest eigenvalue dominates all
the other contributions. Also note that
s
() is independent of s, and hence we
drop the s when no ambiguity can arise and simply write ().
5-10 RECEIVER BUFFER BEHAVIOR FOR SELECTIVE REPEAT
The G artner-Ellis Theoremstates that the rate function is equal to the Legendre-
Fenchel transform of ():

(x) = sup

(x ()), (5.22)
under certain regularity conditions on () (that is, it must be nite in a neighbor-
hood of = 0, essentially smooth and lower-semicontinuous). These conditions
are fullled in this case, see eg. [14], p. 43.
The rate function

(x), in turn, allows us to nd the exponential decay of


probabilities associated with process Y
s
t
: for sufciently nice sets A (such as
intervals) the following powerful result holds:
P[t
1
Y
s
t
A] = exp(t inf
xA

(x) +o(t)), (5.23)


where o(t) denotes a function that decays faster than any linear function. Also
note that the application of the G artner-Ellis Theorem is not restricted to the nite
Markov error processes we are considering here. As long as the process Y
s
t
admits
a scaled cumulant function, for which the limit
s
() exists and fullls the above
mentioned regularity conditions. Its computation can be performed either in closed
form, numerically, or even by sampling (i.e., by a Monte-Carlo based technique).
Now, we are ready to derive the large-deviations result for the buffer content
distribution proper.
Theorem 5.4.1. The buffer content distribution satises
lim
x
1
x
logP[U
T
> x] =,
where = inf
x>0

(x)
x
or equivalently,
= sup > 0 : () < 0. (5.24)
Proof. We start out by showing that the two denitions of the decay rate are
equivalent. In fact, this is proved in many articles and books on large deviations,
but for the sake of completeness, we repeat the proof here. Note that
< = inf
x>0

(x)
x
iff x

(x) < 0 for x > 0.


By the duality of the Legendre-Fenchel transform under convexity, we get
() = sup
x>0
(x

(x)), (5.25)
and hence, < , if and only if () < 0, which leads to the desired
= sup > 0 : () < 0. (5.26)
CHAPTER 5 5-11
It is this last relation that proves to be far more useful. It amounts to nding a
root of the cumulant generating function, without any need for Legendre-Fenchel
transforms and so on.
Next, we derive a lower bound for the exponential decay rate. Observe that for
any given s, t and x,
P[Y
s
t
> x] P[U
T
> x], (5.27)
as the event Y
s
t
> x implies U
T
> x. Hence, for any c > 0, and x > 0, and s : 0
s < L,
1
x
logP[U
T
> x]
1
c
x
c

logP[
Y
s

x
c

x
c

> c], (5.28)


where the notation a denotes the smallest integer not smaller than a. Choose
t =
x
c
, then we get
liminf
x
1
x
logP[U
T
> x]

1
c
liminf
t
1
t
logP[
Y
s
t
t
> c] =

(c)
c
. (5.29)
Since c is chosen arbitrarily (as long as c > 0), we have indeed proved the lower
bound
liminf
x
1
x
logP[U
T
> x] . (5.30)
For the upper bound, we start from the inequality
P[U
T
> x]

t1,0s<L
P[Y
s
t
x], (5.31)
The reasoning behind this inequality is that if the event U
T
> x occurs, then
there exists s, t for which Y
s
t
x. Hence, summing over all possible s and t, we
get the desired inequality.
Using Chernoffs bound, we get:
P[Y
s
t
x] exp(x)E[exp(Y
s
t
)]
exp(x +t
s
t
()), (5.32)
where in the second step we applied the denition of
s
t
(). Pick some > 0
such that () < 0 (we know that there exists such since R(0) is a substochastic
matrix, and hence (0) < 0 and () is differentiable in the neighborhood of
= 0). Choose > 0 such that () +2 < 0. Since
s
t
() (), there exists
t
0
such that for t >t
0

s
t
() < () +, (5.33)
and hence
P[U
T
> x] e
x
_

1tt
0
,0s<L
e
t
s
t
()
+

t>t
0
,0s<L
e
t
_
. (5.34)
5-12 RECEIVER BUFFER BEHAVIOR FOR SELECTIVE REPEAT
The rst sum is nite because each of the terms is nite, and the second sum is
clearly nite. Hence
limsup
x
1
x
logP[U
T
> x] , (5.35)
As we can choose as long as () < 0 or equivalently 0 < < , this concludes
the proof.
Note that thanks to a careful choice of the processes (U
t
, V
t
) and Y
t
, we were
able to derive the decay rate along similar lines as for the standard G/G/1 queue,
see for example the proofs in [14, 15, 18], which is somewhat remarkable as the
workings of the resequencing buffer are very different. For example, it is hard to
attribute a meaning to workload processes in resequencing buffers, which are the
central tool in the LD derivation of standard queues.
For IID error processes (with error probability p), we are able to come up with
an exact solution. Note that the cumulant generating function of Y
s
t
is in this case
equal to
() =
1
L
log(p(p+(1p)e

)
L1
), (5.36)
so that
= log
p

1
L1
p
1p
. (5.37)
One of the appeals of large-deviations techniques is that one can often make an
educated guess about the rate function and subsequently substantiate that guess
with rigourous derivations. Here, the intuition would be something along the lines
of: in Selective Repeat resequencing buffers, large buffer contents occur only when
there are many erroneous transmissions for one particular packet.
Note that the value of stays invariant under the reversion of the Markov chain.
Indeed, let () be an eigenvalue of R(), and let v() be the corresponding left
eigenvalue. When we construct the corresponding matrix expression R
(r)
() for
the reversed background Markov process Q
(r)
, we get
R
(r)
() = diag()
(1)
diag(p
0
)Q
T
_
(diag(p
0
) +e

diag(p
1
))Q
T
_
L1
diag()
= diag()
(1)
diag(p
0
)(P()
L1
)
T
Q
T
diag(). (5.38)
We nd that w() =diag()
(1)
diag(p
0
)v()
T
is a right eigenvector of R
(r)
(),
corresponding to eigenvalue (). Indeed,
R
(r)
()w() = ()diag()
(1)
diag(p
0
)v()
T
. (5.39)
As this holds for every eigenvalue (), this means that the spectra of R() and
R
(r)
() are identical, so that naturally the spectral radii are the same as well.
Summarizing the results of this section, in order to nd the decay rate ,
we must nd the smallest real root of an expression involving a spectral radius.
CHAPTER 5 5-13
By combining a good root-nding algorithm (such as the secant method or the
Newton-Raphson method) with an efcient algorithm to compute the spectral ra-
dius (see e.g. [21]), this is a computationally feasible task, even for larger matrix
dimensions.
5.5 A Closer Look at the Large Deviations Limit
In the previous section, we looked at the probability that an exceptionally large
buffer content occurs. In this section, we have a closer look at how this rare event
exactly occurs. Indeed, we know that one group will stay erroneous during a large
time, but what is the expected number of successful transmissions in the other
groups? Is it larger or smaller than during typical runs of the process? Is it the
same for all groups? These questions are not merely of theoretical interest, but
they are also important for the design of efcient simulation algorithms.
For an IID error processes, the error rate during a build-up to the rare event is
of course the same for every group except the erroneous one. In our nite Markov-
chain setup, the situation is not as simple. An equal error rate in each of the
groups is not assured. We introduce the L 1-dimensional functions () and

(x), which are respectively the limiting cumulant generating function and the
rate function. They are dened as follows:
() = log(P
0
P(
1
) . . . P(
L1
)) and

(x) = sup

( x()). (5.40)
Intuitively, the rate function

(x
1
, x
2
, ) gives the decay rate when the success
rate during the rst group is x
1
, during the second, x
2
, and so on. We thus bring
the large-deviations principle to a more detailed space.
We bring the denition (5.24) of also into this ner space as follows:
= sup > 0 : (, , ) < 0. (5.41)
As eigenvalues are differentiable in a parameter whenever the original matrix is
differentiable in that parameter, we have that () is differentiable in the relevant
domain. Hence, the Legendre-Fenchel transform of () reduces to the simpler
Legendre transform, and in particular,

(x) = (, , ) x(, , ), where x = ()[


=(, ,)
. (5.42)
It is exactly this vector x in which we are interested. With the help of some
matrix manipulations we get simple formulas for x. Indeed, its ith component is
equal to
x
i
=

i
[
=(, ,)
=
1
(R())

i
(P
0
P(
1
) . . . P(
L1
))[
=(, ,)
. (5.43)
5-14 RECEIVER BUFFER BEHAVIOR FOR SELECTIVE REPEAT
Due to the denition of , the fraction is equal to 1, whereas the remaining factor
can be rewritten by the formula for the derivative of an eigenvalue. Indeed, let
and r denote the left and right eigenvectors corresponding to the Perron-Frobenius
eigenvalue of R(), then we have that
x
i
=

i
(P
0
P(
1
) . . . P(
L1
))r
= P
0
P()
i1
P
1
P()
Li1
r. (5.44)
Hence, we found that in sample paths leading to large buffer contents, the
success rate in each of the groups is not necessarily equal but given by equation
(5.44). After some trivial manipulations, we nd that for IID error processes, the
success rate is the same in each of the groups, and equal to 1 p
L
L1
, which is
(slightly) higher than during a typical run.
The relationship between large-deviations results and exponentially tilted im-
portance sampling is well-known, see e.g. [2]. From the fact that the optimal decay
rate is reached in (, , ) (with all arguments the same), we see that the opti-
mal change of measure tilts the Markov chain equally with parameter in all of the
groups, except for the rst group, where it is tilted with parameter . We have
put this knowledge to use to improve the efciency of our simulation program.
5.6 Asymptotic Analysis of the Packet Delay
We now tackle the asymptotics of the packet delay. The heuristic reasoning is as
follows. An exceptionally long delay of a packet is only possible when there is a
long burst of errors in one of the groups. Let D be the random variable denoting
the delay of an arbitrary packet. Then we postulate the following on the decay rate
of the packet delay:
Theorem 5.6.1. The packet delay distribution has an exponential asymptotic with
decay rate :
lim
k
1
k
logP[D > k] =,
where
=
1
L
log(Q
L1
P
0
). (5.45)
Proof. We nd lower and upper bounds of the probability for the event D > k,
and show that they have the same exponential asymptotic.
We start with an upper bound. If we have a delay larger than k, then necessarily,
there is a group with minimally (k1)/L consecutive errors. Let i be the channel
state of the slot during which series of errors starts. For the upper bound we are
CHAPTER 5 5-15
interested in the channel state that maximizes the probability. Thus we have has
an upper bound:
P[D > k] sup
i
P[# errors in a group =(k +1)/L]
= sup
i
e
i
(Q
L1
P
0
)
(k+1)/L
1, (5.46)
where e
i
is a rowvector for which the ith component is 1, and the other components
are zero. From the Perron-Frobenius theorem it is hence easily seen that
lim
k
1
k
logP[D > k] . (5.47)
Now we look at the lower bound. We construct a superset E of D > k in the
event space, so that when E holds, then necessarily also D > k. The probability
of E is then smaller than P[D > k], giving us the desired lower bound. Informally
speaking, we need two ingredients so that a sample path produces a packet delay
larger than k: (1) there must be a packet that is erroneously transmitted for a large
number of times, say n, where n (k+1)/L+1, and (2) there must be an arrival
in the reordering buffer of a packet that is younger than said packet. This means
we have a situation as in gure 5.3. But since we track the delay of an arbitrary
packet, we condition on the event that the arrival indeed occurs. Also note that for
the lower bound, we look for the combination of channel state i and gap j that
minimizes the probability. Hence we have:
P[D > k] inf
i, j,m
P[n errors[packet arrives with certain i, j, m]
= inf
i, j,m
e
i
P
1
Q
j1
P
0
(Q
Lj1
P
0
Q
j1
P
0
)
m1
Q
Lj1
P
1
Q
j1
P
0

(Q
L1
P
0
)
n1
1
(e
i
P
1
Q
j1
P
0
(Q
Lj1
P
0
Q
j1
P
0
)
m1
Q
Lj1
P
1
1)
1
. (5.48)
For large k, the contribution of the matrix power on the middle line dominates and
hence we can apply the Perron-Frobenius theorem as before. Even if m is large,
it cannot dominate as the terms in numerator and denominator cancel each other.
Hence we have the lower bound
lim
k
1
k
logP[D > k] , (5.49)
and this concludes the proof.
5.7 Approximations for High Error Probabilities
When the error probabilities approach one, the time between two error-free trans-
missions will be sufciently high, so that (1) the groups of successive error-free
5-16 RECEIVER BUFFER BEHAVIOR FOR SELECTIVE REPEAT
Figure 5.3: Necessary event for the occurence of a large packet delay. As before, white
blocks denote correct transmissions, gray blocks denote incorrect ones and dashed blocks
denote transmissions which may be either. The variables j, m and n are as dened in the
text.
transmissions are almost independent, and (2) the transient effects of the Markov
chain of the channel have died out. Under such conditions, we can directly apply
the results of the coupon-collector problem. The value of interest in the original
problem is the time (which in our problem translates to the number of successful
transmissions) until we have collected all coupons. Under the above stated condi-
tions, the buffer content during regime U
T
can be approximated as U
T
=L. We
mention the mean, variance and tail probabilities (the derivations can be found in
most textbooks on combinatorics, e.g. in [12]):
E[U
T
] = L(H
L
1) LlogL+L+
1
2
Var[U
T
] = L
2
H
(2)
L1
LH
L1


2
6
L
2
P[U
T
> k] e
k
L
L1
. (5.50)
In these equations, H
L
denotes the Lth harmonic number, and H
(2)
L
the Lth sec-
ond order harmonic number. By we denote the Euler-Mascheroni constant:
0.5772157 .
Note that the mathematical elegance of this high-error approximation surpasses
its practical usability, as consensus seems to be that the usage of retransmission
CHAPTER 5 5-17
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
d
e
c
a
y
r
a
t
e

p
L=4
L=8
L=16
Figure 5.4: Plot showing the inuence of the error probability p on the decay rate of the
buffer content distribution, for different values of the feedback delay L.
protocols over channels with an extremely high packet error rate leads to poor
performance.
5.8 Numerical Results and Discussion
In this section, we show the validity of our large-deviations technique by means of
some specic numerical examples. First, we consider some cases where the error
process is IID. In Figure 5.4, the decay rate of the buffer content distribution is
plotted against the packet error probability p, for different values of the feedback
delay L. We notice that the decay rate is higher for small error probabilities, which
is consistent with the intuition that in that case, higher buffer contents are very
unlikely. Also as intuitively expected, the decay rate decreases with increasing
values of the feedback delay. For p 1, the decay rate approaches the value of
log
L
L1
that is predicted by the high error rate analysis.
In Figure 5.5, we show that our model indeed predicts the decay rates right.
We repeat that large deviations in general predict only the decay rate and not the
offset of the tail, and therefore, we have chosen an offset of zero (corresponding to
C = 1 in the formCexp(n). We obtain accurate simulation results with the help
of importance sampling.
In the next pair of plots, we show some results in case of Markovian error
processes. We take a three-state error model that has been derived in [20] by means
of the Hidden Markov estimation techniques. They found for a fading margin
F = 20dB and a Doppler frequency normalized to the slot length of f
D
T = 0.01,
5-18 RECEIVER BUFFER BEHAVIOR FOR SELECTIVE REPEAT
L = 4
L = 8
L = 16
-20
-15
-10
-5
0
0 10 20 30 40 50 60
l
n
P
r
[
U
>
k
]
k
Figure 5.5: Plot showing the exact distribution of the buffer content, obtained by
simulations (with importance sampling; full lines) and also the tail approximations
obtained by large deviations (dashed lines), for p = 0.3 and different values of the
feedback delay L.
the following matrices:
P
0
=
_
_
0 0.002708 0.000213
0 0.646523 0
0 0 0.878338
_
_
,
P
1
=
_
_
0.997079 0 0
0.353477 0 0
0.121662 0 0
_
_
. (5.51)
In Figure 5.6, we show the inuence of the correlation on the decay rate. That
is, we compare the decay rate of the Markovian channel with the decay rate of
the corresponding IID channel (i.e. with the same time-average error probability,
which in this case is given by p = 0.009324). We see that especially for smaller
feedback delays, there is a considerable difference between the correlated and un-
correlated error process. In Figure 5.7, we compare the obtained asymptotics with
simulations.
Finally, in gure 5.8, we show the decay rates of buffer content and the packet
delay, for both the Markov process and the corresponding IID process. A fairly
remarkable fact is that although for smaller roundtrip times there are important
differences in decay rates, for very round trip times the four curves converge to the
same asymptote.
CHAPTER 5 5-19
0
1
2
3
4
5
6
7
2 3 4 5 6 7 8 9
d
e
c
a
y
r
a
t
e

L
Markovian
IID
Figure 5.6: Plot showing the inuence of error correlation on the decay rate of the buffer
content distribution.
-60
-50
-40
-30
-20
-10
0
0 5 10 15 20 25 30
l
n
P
r
[
U
>
k
]
k
HMM
HMM, LD
IID
IID, LD
Figure 5.7: Plot showing the exact distribution of the buffer content (obtained by
simulations) versus the LD approximations for the 3 state Markov model (HMM) and the
corresponding IID error process. The feedback delay L equals 5.
5-20 RECEIVER BUFFER BEHAVIOR FOR SELECTIVE REPEAT
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
10 20 30 40 50 60 70 80 90 100
d
e
c
a
y
r
a
t
e

L
Buer content; Markov
Buer content; IID
Packet delay; Markov
Packet delay; IID
Figure 5.8: Plot showing the decay rates of the packet delay and the buffer content, each
for IID and Markovian error processes.
5.9 Conclusions
We have analyzed the performance of the resequencing buffer for the selective-
repeat protocol over correlated channels. We found an exact solution for the buffer
content distribution over Markovian error channels, as well as computationally
efcient large-deviations results. We touched upon the similarity with the coupon-
collector problem, to which it reduces in the limit as packet error probabilities
increase. Comparisons with simulations conrm the validity of our approach.
References
[1] M. Anagnostou and E. Protonarios. Performance analysis of the selective
repeat protocol. IEEE Transactrions on communications, 34(2):127135,
FEB 1986.
[2] S. Asmussen and P. W. Glynn. Stochastic simulation: algorithms and anal-
ysis, volume 57 of Stochastic Modelling and Applied Probability. Springer,
New York, 2007.
[3] L. Badia. On the Impact of Correlated Arrivals and Errors on ARQ Delay
Terms. IEEE Transactions on communications, 57(2):334338, feb 2009.
[4] L. Badia, M. Rossi, and M. Zorzi. SR ARQ packet delay statistics on Markov
channels in the presence of variable arrival rate. IEEE Transactions on Wire-
less Communications, 5(7):16391644, JUL 2006.
CHAPTER 5 5-21
[5] H. Bruneel, J. De Vriendt, and C. Ysebaert. Receiver buffer behavior for
the selective-repeat arq protocol. Comput. Netw. ISDN Syst., 19(2):129142,
1990.
[6] S. R. Chakravarthy and S. Chukova. A nite capacity resequencing model
with Markovian arrivals. Asia-Pac. J. Oper. Res., 22(3):409443, 2005.
[7] S. De Turck, Koen; Wittevrongel. Delay analysis of Go-Back-N ARQ for
correlated error channels, pages 245267. River Publishers, Aalborg, 2009.
[8] S. De Vuyst, K. Tworus, S. Wittevrongel, and H. Bruneel. Analysis of stop-
and-wait ARQ for a wireless channel. 4OR, 7(1):6178, 2009.
[9] S. De Vuyst, S. Wittevrongel, and H. Bruneel. Queueing delay of stop-
and-wait arq over a wireless markovian channel. Mobility Management and
Quality-of-Service for Heterogeneous Networks, pages 131, 2009.
[10] A. Dembo and O. Zeitouni. Large deviations techniques and applications,
volume 38 of Applications of Mathematics (NewYork). Springer-Verlag, New
York, second edition, 1998.
[11] R. Fantacci. Queueing analysis of the selective repeat automatic repeat re-
quest protocol wireless packet networks. IEEE Trans. Veh. Technol, 45:258
264, 1996.
[12] P. Flajolet and R. Sedgewick. Analytic combinatorics. Cambridge University
Press, Cambridge, 2009.
[13] S. Fujii and Y. Hayashida. Delay performance of a selective-repeat ARQ
scheme with a Markovian error channel. Electronics and communications in
Japan Part I - Communications, 81(6):3141, June 1998.
[14] A. Ganesh, N. OConnell, and D. Wischik. Big Queues. Springer, 2004.
[15] P. W. Glynn and W. Whitt. Logarithmic asymptotics for steady-state tail
probabilities in a single-server queue. Studies in Applied Probability, 1994.
[16] J. G. Kim and M. Krunz. Delay analysis of selective repeat arq for a marko-
vian source over a wireless channel. IEEE Trans. Veh. Technol, 49:1968
1981, 2000.
[17] A. Konheim. A queueing analysis of two ARQ protocols. IEEE Trans. Com-
mun., 10:10041014, July 1992.
[18] E. Morozov. Communications Systems: Rare Events Simulation and Effective
Bandwidths. Public University of Navarre, 2004.
5-22 RECEIVER BUFFER BEHAVIOR FOR SELECTIVE REPEAT
[19] Z. Rosberg and N. Shacham. Resequencing delay and buffer occupancy un-
der the selective-repeat arq. IEEE Trans. Inform. Theory, 35(1), 1989.
[20] W. Turin and M. Zorzi. Performance analysis of delay-constrained commu-
nications over slow rayleigh fading channels. IEEE Trans. Wireless Comm.,
1(4):801807, 2002.
[21] R. J. Wood and M. J. ONeill. Finding the spectral radius of a large sparse
non-negative matrix. ANZIAM J., 48((C)):C330C345 (2008), 2007.
[22] Y. Xia and D. Tse. Analysis on packet resequencing for reliable network
protocols. Perform. Eval., 61(4):299328, 2005.
6
Performance Analysis of the IEEE
802.16e sleep mode for correlated
downlink trafc
Koen De Turck, Stijn De Vuyst, Dieter Fiems and Sabine Wit-
tevrongel
published in Telecommunication systems.
Abstract. We evaluate the performance of the IEEE 802.16e sleep mode mech-
anism in wireless access networks. This mechanism reduces the energy consump-
tion of a mobile station (MS) by allowing it to turn off its radio interface (sleep
mode) when there is no trafc present at its serving base station (BS). After a sleep
period expires, the MS briey checks the BS for data packets and switches off for
the duration of another sleep period if none are available. Specically for IEEE
802.16e, each additional sleep period doubles in length, up to a certain maximum.
Clearly, the sleep mode mechanism can extend the battery life of the MS consid-
erably, but also increases the delay at the BS buffer. For the performance analysis,
we use a discrete-time queueing model with general service times and multiple
server vacations. The vacations represent the sleep periods and have a length de-
pending on the number of preceding vacations. Unlike previous studies, we take
the (short-range) trafc correlation into account by assuming a D-BMAP arrival
process, i.e. the distribution of the number of packet arrivals per slot is modulated
6-2 SLEEP MODE FOR CORRELATED DOWNLINK TRAFFIC
by the transitions in a Markov chain with N background states. As results, we
obtain the distribution of the number of packets in the queue at various sets of
time epochs, the distribution of the packet delay and the antenna activity rate. We
apply these results to the IEEE 802.16e sleep mode mechanism with correlated
downlink trafc. By means of some examples, we show the inuence of both the
conguration parameters and the trafc correlation on the delay and the energy
consumption.
6.1 Introduction
The new IEEE 802.16e standard [1] [2] for Broadband Wireless Access networks
holds extensions to the original 802.16 standard for support of subscriber mobil-
ity. Among these extensions are an efcient handover procedure and an optional
energy saving mechanism, the latter of which is the topic of this chapter.
The idea behind the energy saving mechanism is to allow a Mobile Station
(MS) to enter sleep mode, which means that its radio interface is switched off
during certain intervals of time. While in such a sleep period, the MS cannot be
reached by its serving Base Station (BS) and any arriving trafc must be buffered
there until the MSs sleep period ends. At the end of each sleep period, the MS
switches its radio back on for a short time (listening interval) to check whether
data packets are available at the BS. If not, the MSs sleep mode continues with
another sleep period. However, if any trafc arrived during the last sleep period,
the MS remains powered and enters awake mode. This allows the BS to transmit
the packets in its buffer, which it does in an exhaustive manner, i.e. the BS keeps
transmitting packets until its buffer is completely empty. At this point, the MS
switches back to sleep mode. Thus, the MS alternates between sleep and awake
mode, as is illustrated in Fig. 7.1.
Specically for the IEEE 802.16e standard, an exponential increase strategy is
used for updating the lengths of the sleep periods. The rst sleep period after en-
tering sleep mode has length T
0
. Each subsequent sleep period however, has twice
the length of the previous sleep period, i.e. 2T
0
, 4T
0
, 8T
0
, . . . , until a maximum
length T
max
is reached. After that, all additional periods have length T
max
.
If during a sleep period, packets need to be sent from the MS in the uplink
direction, the sleep mode is interrupted immediately and the MS stays in awake
mode until all packets in both directions have been transmitted. In this chapter, we
consider uplink trafc only, but we perform a thorough study of the
The sleep mode mechanism has the inherent drawback that it leads to a degra-
dation of QoS (Quality of Service). As the transmission of packets arriving at the
BS must be postponed until the current sleep period of the MS is nished, it is
clear that the overall packet delay will suffer. Hence, a trade-off needs to be made
with respect to the sleep period lengths. Short sleep periods result in too many un-
CHAPTER 6 6-3
L L
L L L
T
0
T
0 0
2T 0
2T
0
4T
MS
BS
buffer
arriving
downlink
traffic
d
o
w
n
l
i
n
k
sleep
period 1
sleep
period 1
sleep
period 2
sleep
period 2
sleep
period 3
SLEEP SLEEP AWAKE
busy busy idle
AWAKE
Figure 6.1: Indication of busy, idle and sleep periods at the BS and sleep mode vs. awake
mode at the MS. An idle period consists of a number of consecutive sleep periods and ends
when arrivals have occurred during its last sleep period. During busy periods, the packets
in the buffer are served and transmitted to the MS until the buffer is empty, after which a
new idle period starts. The listening interval at the end of each sleep period is indicated by
L.
needed activations of the MS radio interface which is less energy efcient, while
very long sleep periods result in excessive packet delays. Therefore, it is important
to be able to predict the inuence of the sleep mode parameters on both MS energy
consumption and packet delay. According to the IEEE 802.16e standard, the sleep
mode parameters are (a.o.) the initial sleep period T
0
and the maximal sleep pe-
riod T
max
which are pre-negotiated each time the MS enters sleep mode. However,
we assume here that these parameters remain xed throughout the operation of the
system.
Quite a few authors have shown interest in the performance of the sleep mode
operation, either in case of IEEE 802.16 or other technologies. In [3], the average
energy consumption of the MS is obtained in case of downlink trafc only, as well
as an approximate expression for the mean packet delay. The energy consump-
tion of the MS in case of both downlink and uplink trafc is considered in [4].
Both [3] and [4] model the incoming (and outgoing) trafc as a Poisson process.
An accurate assessment of the delay experienced at the BS buffer however, re-
quires a queueing model. For IEEE 802.16e, in [5] the BS buffer is modelled as
a continuous-time nite-capacity queue with a Poisson arrival process and deter-
ministic service times. A semi-Markov chain analysis leads to expressions for the
mean packet delay and the mean energy consumption by the MS. The analysis
in [6] is based on an M/G/1/K queueing model with multiple vacations and ex-
haustive service, where the vacations represent the sleep periods. Similar work
can also be found in [7], where the length of a vacation is assumed to depend on
the previous vacation length. In [8], the sleep mode operation in Cellular Digital
Packet Data (CDPD) services is evaluated. The difference with IEEE 802.16e is
6-4 SLEEP MODE FOR CORRELATED DOWNLINK TRAFFIC
that the subsequent sleep periods do not increase in length. The system can thus
be modelled as a queueing system with multiple vacations and exceptional rst
vacation. The loss probability in both [6] and [8] is obtained as well. A simula-
tion study of CDPD sleep mode performance is found in [9]. An alternative to the
exponential increase of the sleep period lengths is evaluated by simulation in [10].
All of the above models however, specically assume an uncorrelated arrival
process (i.e. a Poisson process) which is known to be unrealistic for most types of
data trafc. If the trafc correlation is not taken into account, the buffer content
and the delay a packet experiences in a buffer may be seriously underestimated (see
e.g. [11], [12]). The main contribution of this paper is that we analyse the sleep
mode operation for trafc modelled by a D-BMAP (Discrete Batch Markovian
Arrival Process) [13]. This is a very general model that can capture most trafc
characteristics up to any desired precision. Although in theory, self-similar or
long-range dependent (LRD) trafc is precluded from Markov modelling, many
studies show that this need not be restrictive, see e.g. [14], in realistic scenarios.
As we demonstrate, the presence of correlation in the carried trafc stream has a
major impact on the behaviour of the BS buffer, and hence also on the efciency
of the sleep mode operation.
The model we propose is also generic with respect to the lengths of the sub-
sequent sleep periods. Instead of restricting ourselves to the exponential increase
strategy of the IEEE 802.16e standard, we assume the deterministic lengths of the
rst, second, third, . . . sleep period to be free parameters of the model. This way,
the model can capture any deterministic sleep period updating strategy (not only
exponential) and also account for the presence of the listening intervals and an
additional negotiating interval when the MS enters sleep mode.
The remainder of this paper is organised as follows. In section 6.2, we de-
scribe the queueing model under study. Section 6.3 presents the analysis of the
buffer content. In section 6.4, we derive the distribution of the delay. The energy
consumption is studied in section 6.5. In section 6.6, some numerical examples
are discussed and nally, the paper is concluded in section 6.7.
6.2 Queueing Model
We consider a discrete-time model of an innite-capacity single-server buffer.
Time is divided into xed-length time slots and packets are served (transmitted)
according to a rst-come-rst-served (FCFS) discipline. The service times of
the packets are assumed to be independent and identically distributed (i.i.d.). Let
the random variable S denote a service time, its probability mass function is de-
noted by s(n) = P[S = n slots], n 1, and its probability generating function by
S(z) =

n=1
s(n)z
n
.
The main peculiarity of our model is that whenever the server has nished
CHAPTER 6 6-5
time
. . . . . .
t
1
t
2
t
3

0
= 0

1

2

3
busy period busy period idle period
Figure 6.2: The anatomy of an idle period.
all the work in the buffer, it goes to sleep; an internal timer is then started and
the server awakes to check the buffer content after a sleep period of t
1
time slots.
When upon awaking the server nds that there are still no packets, it goes to sleep
again, this time for a sleep period of t
2
slots. The next sleep period would be of
length t
3
, and so on. In general, t
n
, n 1, denotes the length (expressed in slots) of
the nth sleep period after the buffer has become empty. If after a sleep period, the
server nds a non-empty buffer, it serves all packets present at that point and also
all new packets that arrive while the server is working, until the buffer becomes
empty again and the whole procedure is repeated. As indicated in Fig. 7.1, we will
refer to the entire time interval between two subsequent busy periods (i.e., where
the server is busy doing work) as an idle period, which consists of a number of
sleep periods. For convenience, we also dene
n
.
=
n
m=1
t
m
, n 1, and
0
= 0.
Note that in case the origin of the time axis is set to coincide with the start of the
idle period, each
n
marks the end of a short listening interval where the server
checks for packets in the buffer. The denition of the t
n
s and the
n
s is further
illustrated in Fig. 6.2.
We will assume that packets arrive in the buffer according to a Markov modu-
lated batch arrival process. We assume a nite number N of background states.
The arrival process is completely dened by the values a(k, j[i); k 0; i, j
1, , N, denoting the probability that if the background state is i during a slot,
there are k arrivals during this slot and the background state during the next slot is
j. We put these probabilities in a matrix generating function A(z) with dimension
NN, whose entries are dened as follows:
[A(z)]
i j
.
=

k=0
a(k, j[i)z
k
. (6.1)
This arrival process, whereby the distribution of the number of arrivals in a slot
depends on the transition made by a Markov chain in that slot, is called the Discrete
Batch Markovian Arrival Process (D-BMAP) [13]. Delayed access is assumed, i.e.
arriving packets do not enter the buffer until the end of their arrival slot.
Let denote a probability vector (of dimension 1 N), with []
i
the steady-
state probability of being in background state i during a slot. The vector satises
A(1) = and 1 = 1, (6.2)
6-6 SLEEP MODE FOR CORRELATED DOWNLINK TRAFFIC
time
. . . . . .
idle period
k packets
state i
state j
Figure 6.3: The denition of V(z).
with 1 a column vector of ones of appropriate size. The probability generating
function of the number of arrivals while being in background state i is denoted by
A
i
(z), which can be found as the ith entry of the vector A(z)1. The average number
of arrivals during a slot in background state i is E[a[i] = A

i
(1); the average number
of arrivals during a random slot is E[a] = A

(1)1.
6.3 Analysis of the Buffer Content
In this section, we derive the steady-state distribution of the buffer content at the
beginning of an arbitrary slot. We have opted for a step-by-step approach: we rst
investigate the buffer content at special instants, and then we derive the distribution
of the buffer content at random slots. The advantage of this approach is that the ex-
pressions remain reasonably concise, and each expression has a clear probabilistic
interpretation.
Specically, the rst step is to derive the matrix generating function V(z), with
the following entries:
[V(z)]
i j
.
=

k=1
v(k, j[i)z
k
, (6.3)
where v(k, j[i) is the probability that an idle period ends with k packets in the buffer
and with background state j, provided that the background state was i during the
rst slot of the idle period, as indicated in Fig. 6.3. Below we derive V(z) in terms
of A(z), the
n
s and the t
n
s.
The second step is to analyse the buffer content at service completion instants,
i.e. at the beginning of a slot following a packet departure. This results in the 1N
vector generating function U
c
(z), whose ith entry equals the partial generating
function of the buffer content at a service completion instant when the background
state during the following slot is i. The last step is to study the buffer content at
the beginning of an arbitrary slot, which has vector generating function U(z).
6.3.1 Derivation of V(z)
Consider an idle period, consisting of n sleep periods. The background state during
the rst slot of the idle period is i; the background state during the rst slot after
CHAPTER 6 6-7
the idle period is j; k packets arrive during the nth sleep period. Then we can
write:
[V(z)]
i j
=

n=1

k=1
v
n
(k, j[i)z
k
, (6.4)
where v
n
(k, j[i) is the probability that an idle period consists of n sleep periods and
ends with a buffer content of k in background state j, given that it started in state i.
We know that an idle period consists of n sleep periods if and only if there are no
arrivals during the rst
n1
slots and there are a positive number, say k, arrivals in
the last t
n
slots. Let i

be the background state during the rst slot of the nth sleep
period. Then we can write:
[V(z)]
i j
=

n=1
N

=1
P[0 arrivals in
n1
slots, i

[i]

k=1
P[k arrivals in t
n
slots, j[i

]z
k
.
(6.5)
In view of the denition of A(z), we can rewrite the above expression as
[V(z)]
i j
=

n=1
N

=1
[A(0)

n1
]
ii
[A(z)
t
n
A(0)
t
n
]
i

j
. (6.6)
The inner sum in (6.6) is in fact nothing more than a matrix multiplication, hence
we nally nd the following expression for V(z):
V(z) =

n=1
A(0)

n1
_
A(z)
t
n
A(0)
t
n
_
. (6.7)
In the special case that the t
n
s for n above a certain threshold J are equal, i.e. t
n
=t
J
for n J, we can rewrite (6.7) as a nite sum:
V(z) =
J

n=1
A(0)

n1
_
A(z)
t
n
A(0)
t
n
_
+A(0)

J
(I A(0)
t
J
)
1
_
A(z)
t
J
A(0)
t
J
_
,
(6.8)
where I denotes the identity matrix.
6.3.2 Derivation of U
c
(z)
In order to derive an expression for the vector generating function U
c
(z) of the
buffer content at service completion instants, we note the following. When the
buffer content is non-zero, say u, after a certain service completion, the buffer
content after the next service completion will be u 1 +, where is the num-
ber of packets that arrive during the service time. When the buffer is empty after
a service completion, the server will go into a series of sleep periods, and even-
tually awake with v packets in the buffer, hence there will be v 1 + packets
6-8 SLEEP MODE FOR CORRELATED DOWNLINK TRAFFIC
after the next service completion. Note that V(z) is the matrix generating function
corresponding to the buffer content of v packets at the end of the idle period. The
matrix generating function that corresponds to the number of packets is given by
the expression S(A(z)) =

k=1
A(z)
k
s(k). These observations lead to the following
equation for U
c
(z):
U
c
(z) = (U
c
(z) U
c
(0))
S(A(z))
z
+U
c
(0)V(z)
S(A(z))
z
. (6.9)
From this equation, U
c
(z) is obtained as
U
c
(z) = U
c
(0)(V(z) I)S(A(z))(zI S(A(z)))
1
= U
c
(0)(V(z) I)(zI S(A(z)))
1
S(A(z)). (6.10)
The last step deserves perhaps some explanation: as zI S(A(z)) is commuta-
tive with S(A(z)), its inverse (zI S(A(z)))
1
is also commutative with S(A(z)).
We can also prove this by using the more general property of the Appendix:
(zI S(A(z)))
1
and S(A(z)) are commutative because both can be written as a
power series of S(A(z)) (or of A(z)).
The computation of the only unknown in this formula, i.e. the vector U
c
(0), is
a standard problem in queueing theory and the subject of many papers. We refer
the interested reader to the seminal paper of Gail et al. [15].
6.3.3 Derivation of U(z)
To nd the distribution of the buffer content at the beginning of an arbitrary slot,
we rst derive the distribution at the beginning of an arbitrary busy slot, and the
distribution at the beginning of an arbitrary idle slot. Then we combine these two.
Note that the following relationship exists between the vector generating func-
tion U
b
(z) of the buffer content at the beginning of a service and the vector gener-
ating function U
c
(z) of the buffer content at the end of a service:
U
b
(z)
S(A(z))
z
= U
c
(z). (6.11)
Indeed, the matrix generating function S(A(z)) represents the number of packets
that arrive during a service time, and the factor z
1
is due to the fact that one packet
leaves the buffer at the end of the service. Combination of (7.1) and (6.11) then
leads to
U
b
(z) = U
c
(0)(V(z) I)zzI S(A(z))
1
. (6.12)
The vector generating function U
B
(z) of the buffer content at the beginning of
CHAPTER 6 6-9
a random service slot (i.e. a random busy slot), satises the following relation [16]:
U
B
(z) =
E
S
[
S1
j=0
U
b
(z)A(z)
j
]
E[S]
=
U
b
(z)
E[S]
(I S(A(z)))(I A(z))
1
, (6.13)
where E[S] is the mean length of a service time. Note that this relation is based on
the fact that a randomly selected service slot belongs to a service time of length n
with probability
ns(n)
E[S]
(i.e. a manifestation of the typical renewal-theory paradox,
see e.g. [17]).
In order to study the buffer content at the beginning of an idle slot, let u
0
denote the probability that the buffer content is zero at the end of a service. Note
that u
0
= U
c
(0)1. Also, we introduce the vectors v
n
. The ith entry of v
n
, [v
n
]
i
,
is the probability that during an idle period, there are at least n sleep periods, and
the nth sleep period starts in background state i. The vectors v
n
, n 1, can be
expressed as
v
n
=
1
u
0
U
c
(0)A(0)

n1
. (6.14)
The mean length of an idle period E[length idle] can be obtained as
E[length idle] =

n=1
t
n
v
n
1. (6.15)
If we randomly select a slot from an idle period, the probability that it is a slot
from the nth sleep period is recorded in the vector v

n
. Specically, [v

n
]
i
denotes
the probability that a random idle slot belongs to the nth sleep period of an idle
period (consisting of at least n sleep periods) and this nth sleep period starts in
background state i. The following relation holds between the vectors v
n
and v

n
:
v

n
=
t
n
v
n
E[length idle]
, (6.16)
which is again a manifestation of the renewal-theory paradox.
The (partial) vector generating function U
I,n
(z) (n 1) of the buffer content at
the beginning of a random slot during the nth sleep period of an idle period can
then be found as follows. Considering such a slot, we know that it can have any
position j (0 j t
n
1) within the sleep period with equal probability 1/t
n
. Also,
assuming the position of this slot is j, we have that the matrix generating function
of the number of arrivals since the start of the sleep period is A(z)
j
. Moreover, the
buffer is empty at the beginning of the nth sleep period. Therefore, we nd
U
I,n
(z) = v

n
1
t
n
t
n
1

j=0
A(z)
j
= v

n
1
t
n
(I A(z)
t
n
)(I A(z))
1
. (6.17)
6-10 SLEEP MODE FOR CORRELATED DOWNLINK TRAFFIC
Hence, the vector generating function U
I
(z) of the buffer content at the beginning
of a random slot in an idle period is obtained as
U
I
(z) =

n=1
U
I,n
(z)
=
v
1
E[length idle]
(I V(z))(I A(z))
1
, (6.18)
where we have subsequently used the formulas (6.17), (6.16), (6.14) and (6.7), and
the fact that
0
= 0 and hence A(0)

0
= I.
Now, we have all the necessary parts to construct the vector generating function
U(z) of the buffer content at the beginning of a random slot. The fraction of busy
slots is equal to the load of the queueing system, = E[a] E[S]. Hence, the fraction
of idle slots is equal to 1. Therefore,
U(z) = U
B
(z) +(1)U
I
(z). (6.19)
Let E[length busy] denote the mean length of a busy period. As there is per
busy period exactly one service after which the buffer is empty (namely the last),
we can relate u
0
, E[S] and E[length busy] as follows:
u
0
=
E[S]
E[length busy]
. (6.20)
Also, can be written as
=
E[length busy]
E[length busy] +E[length idle]
. (6.21)
Using these two relations and expressions (6.12)(6.14) and (6.18) in equation
(6.19), we can nally write U(z) as
U(z) = v
1
(I V(z))
_
I z(zI S(A(z)))
1
(I S(A(z)))
_
(I A(z))
1
= v
1
(I V(z))(zI S(A(z)))
1
(z 1)S(A(z))(I A(z))
1
, (6.22)
with
v
1
=
v
1
E[length busy] +E[length idle]
. (6.23)
Note however, that we do not have to compute E[length busy] nor E[length idle]
separately, as normalisation of U(z) yields the right factor. Also note that v
1
has
an important probabilistic interpretation: it is the vector of probabilities that an
arbitrary slot is the rst slot of an idle period. From equation (6.22), the probability
mass function of the buffer content U at the beginning of an arbitrary slot can be
computed by an inverse z-transformation. Moments can also be calculated, in
function of the derivatives of U(z) with respect to z, for z = 1. For example, the
mean and the variance of the buffer content are obtained as follows:
E[U] = U

(1)1 and Var[U] = U

(1)1+E[U] E[U]
2
. (6.24)
CHAPTER 6 6-11
6.4 Analysis of the Delay
Our delay analysis follows a classical pattern: we rst derive the distribution of the
unnished work at the beginning of a randomslot, and then we use this preliminary
result to obtain the distribution of the delay.
6.4.1 Unnished Work
Usually, the unnished work of a queueing system at the beginning of a slot is
dened as the number of slots it takes to empty the system if no new arrivals would
occur from that slot onwards. In most systems, this concept of unnished work
measures the time until the system is ready to serve new packets. In the presence of
sleep periods, it obviously does not: the system can be empty and still not ready to
serve new packets. The most elegant way to adapt the denition of the unnished
work so that it does represent the time until readiness to serve, is to introduce so-
called sleep work: at the beginning of each sleep period, say of length t
n
, we put
t
n
slots of sleep work in the system, which makes that the server is at work during
every slot. Let w
N
be the unnished work at the beginning of an arbitrary slot N .
Note that due to the introduction of sleep work, w
N
> 0. We dene the auxiliary
random variable w

N +1
.
= w
N
+
N
1, where
N
is the amount of work arriving
during slot N , i.e. the sum of the service times of the arriving packets. Note that

N
has matrix generating function A(S(z)). The unnished work at the beginning
of the next slot, w
N +1
, depends on whether w

N +1
is zero or not. If w

N +1
>0 then
w
N +1
= w

N +1
, else w
N +1
=t
n
for a certain n.
The vector generating functions W(z) and W

(z) capture the distributions of


w
N
and w

N
respectively:
[W(z)]
i
=

k=1
P[w
N
= k, background = i]z
k
(6.25)
and
[W

(z)]
i
=

k=0
P[w

N
= k, background = i]z
k
. (6.26)
By denition of w

N
, we have that
W

(z) = W(z)
A(S(z))
z
. (6.27)
Now, considering slot N to be an arbitrary slot, we dene the probability vec-
tors w
n
(n 1) as follows. Let [w
n
]
i
be the probability that N is the rst slot of
the nth sleep period in an idle period and that the background state is i during slot
N . In other words,
[w
n
]
i
.
=P[N is the rst slot of sleep period n, background = i]. (6.28)
6-12 SLEEP MODE FOR CORRELATED DOWNLINK TRAFFIC
As the system only enters the nth sleep period if there have been no arrivals during
the previous n1 sleep periods, we have
w
n
= v
1
A(0)

n1
. (6.29)
Recall from equation (6.23) that v
1
is the vector of the probabilities that an arbi-
trary slot is the rst of an idle period. As w

N
is zero if and only if N is the starting
slot of a sleep period, we also have
W

(0) =

n=1
w
n
= v
1

n=1
A(0)

n1
. (6.30)
On the other hand, we express the relationship between w

N
and w
N
, as fol-
lows:
P[w
N +1
=k, background =i] =P[w

N +1
=k, background =i] +

n=1
[w
n
]
i

k
t
n
, k >0,
(6.31)
where
m
n
denotes Kroneckers delta. The second term on the right-hand side rep-
resents the additional sleep work. By multiplying both sides by z
k
and summing,
we get a second relation between the vector generating functions W(z) and W

(z):
W(z) = W

(z) W

(0) +

n=1
w
n
z
t
n
. (6.32)
This last equation, combined with (6.27), (6.29) and (6.30), leads to
W(z) = v
1

n=1
A(0)

n1
(z
t
n
1)z(zI A(S(z)))
1
. (6.33)
6.4.2 Packet Delay
In this section, we derive the distribution of the packet delay. This is the number
of slots that a random (but tagged) packet spends in the system, not including
its arrival slot, but including its departure slot. We can decompose the delay of a
random packet into two parts. In view of the considered FCFS discipline, the delay
of a packet consists of (1) the unnished work w at the beginning of the arrival slot
minus 1, and (2) the service times of all g packets that arrive during the arrival slot
of the tagged packet, but not later than the packet (thus including the tagged packet
itself). Hence, conditioning on the background state i during the arrival slot, we
have that the probability generating function D(z) of the delay must satisfy
D(z) =
N

i=0
E[z
w1
S(z)
g
1
i
], (6.34)
CHAPTER 6 6-13
where 1
i
is an indicator function, indicating the event that the arrival slot has back-
ground state i. The D-BMAP arrival process has the Markov property, i.e. the
arrival process from a certain slot onwards depends only on the state in that slot.
Thus, we get
D(z) =
N

i=0
1
z
E[z
w
1
i
] E[S(z)
g
[i]. (6.35)
The probability that an arrival occurs during a slot with a certain background state
is proportional to the mean number of arrivals during such a state. Hence, we can
relate the distribution of the random variable w to the distribution of the unnished
work w at the beginning of an arbitrary slot, as follows:
D(z) =
N

i=0
1
z
E[z
w
1
i
]
E[a[i]
E[a]
E[S(z)
g
[i], (6.36)
where the factor
1
E[a]
is a normalisation factor. The generating function E[z
g
[i] can
be shown to be equal to (see e.g. [18])
E[z
g
[i] =
z(A
i
(z) 1)
E[a[i](z 1)
. (6.37)
Combining (6.36), (6.37) and the denition of W(z), we get
D(z) =
N

i=1
z
1
[W(z)]
i
S(z)(A
i
(S(z)) 1)
E[a](S(z) 1)
, (6.38)
which combined with (10.1) leads to
D(z) = v
1

n=1
A(0)

n1
(z
t
n
1)(zI A(S(z)))
1
(A(S(z)) I)1
S(z)
E[a](S(z) 1)
.
(6.39)
The mean and the variance of the delay are computed as follows:
E[D] = D

(1) and Var[D] = D

(1) +D

(1) D

(1)
2
. (6.40)
It has been veried that the obtained results are in accordance with Littles famous
law [19], [20]:
E[U] = E[a] E[D]. (6.41)
6.5 Energy Consumption
As stated earlier, the goal of the sleep mode operation is to reduce the energy
consumption of a mobile device, by switching off its radio interface or antenna as
much as possible. For this analysis, we assume that the main energy cost of the
6-14 SLEEP MODE FOR CORRELATED DOWNLINK TRAFFIC
mobile device is to keep the radio interface activated, and that a listening slot (at
the end of each sleep period) costs just as much energy as a random busy slot. The
literature seems to suggest that these assumptions are fair (see e.g. [1]). Under
these assumptions, the antenna activity rate, P[active], i.e. the fraction of slots
that the antenna is active, is a good measure for the energy consumption. In the
reference case without sleep periods, this probability is equal to 1. In the general
case, we can derive the following expression for P[active]:
P[active] = +(1)
E[# sleep periods per idle period]
E[length idle]
. (6.42)
The rst term in (6.42) denotes the fraction of busy slots, while the second
denotes the fraction of time that the antenna is activated during idle periods, in
view of the fact that the number of slots with an active antenna during an idle
period equals the number of sleep periods in this idle period, as the antenna is
activated once at the end of each sleep period. We have that
E[# sleep periods per idle period] =

n=1
v
n
1, (6.43)
due to the denition of v
n
, and the property that for any discrete nonnegative ran-
dom variable Y, E[Y] =

n=1
P[Y n]. The mean E[length idle] is given by ex-
pression (6.15).
Note that the vector U
c
(0) is needed to calculate P[active], hence we cannot
determine the activity rate without a buffer analysis, at least not when the trafc is
correlated.
6.6 Numerical Examples
In this section, we will show some practical results of our model. Note that we will
use fairly simple assumptions, which makes it easier to interpret the results. Our
model is however not restricted to these assumptions, and can also handle more
powerful trafc models, service time distributions, etc.
We assume a so-called ON-OFF trafc model, i.e. the arrival process is mod-
ulated by a two-state Markov chain. When the Markov chain is in state 1 (the ON
state) during a slot, arrivals occur according to a Bernoulli process with parameter
p. When in state 2 (the OFF state), no arrivals occur. The Markov chain spends a
fraction of the time in the ON state. The coefcient of correlation between the
states in two subsequent slots is 1
1
K
, where K is a measure for the mean lengths
of both ON and OFF periods, given by
K
1
and
K

respectively. It can easily be


veried that the matrix generating function A(z) for this trafc model equals
A(z) =
_
1p+ pz 0
0 1
__
1
1
K
1
K

K
1

K
_
. (6.44)
CHAPTER 6 6-15
The case of uncorrelated states from slot to slot clearly corresponds to K = 1.
Larger values of K imply that ON slots and OFF slots occur more clustered to-
gether in time and the arrival process is more bursty. The correlation parameter K
is therefore also referred to as the burstiness factor or the burst-length factor of the
arrival process [12], [21].
Further, unless otherwise stated, we assume that the service times have a so-
called shifted geometric distribution, i.e. their probability generating function is
equal to
S(z) =
(1s)z
1sz
, (6.45)
where 0 s < 1 and E[S] =
1
1s
. For convenience, we introduce notations like
(2, 4, 8, 8, ) to denote a strategy in which the rst sleep period equals 2 slots, the
second 4 slots, and every sleep period thereafter equals 8 slots.
In Figs. 6.4 and 6.5, we have plotted respectively the mean delay E[D] and the
antenna activity rate P[active] versus the load for different values of the corre-
lation parameter K. We have chosen =
2
3
, E[S] =
5
3
and strategy (2, 4, 8, 16, 32,
32, ). The load = pE[S] is hence varied by changing the value of p. We ob-
serve that for high values of , the mean delay grows faster as the trafc correlation
increases. This is in line with the common wisdom that trafc correlation tends
to deteriorate the system performance. An observation which is more specic for
this model is that for low loads, the mean delay is fairly high, and then decreases
to reach a certain minimum, before increasing again when 1. This is due to
the fact that for very small loads, the system will almost always be in the longest
sleep period of length T
max
(in this example, T
max
= 32 slots) when a packet ar-
rives. The expected number of slots until the service of such a packet will start, is
equal to
T
max
1
2
, such that the expected delay for loads approaching zero is equal
to
T
max
1
2
+E[S]. Numerical examples conrm this observation. When we look at
the activity rate P[active], we see that regardless of the correlation, P[active] 1
for 1, and P[active]
1
T
max
for 0, which is what we expect. Perhaps more
remarkable is that trafc correlation lowers the activity rate, and hence is good for
the battery life.
Fig. 6.6 shows the inuence of the correlation parameter K on the mean delay
E[D] more directly. We have chosen =
2
3
, E[S] =
5
3
and strategy (2, 4, 8, 16, 32,
32, ), for different loads. We see that E[D] either converges to a certain value,
or keeps increasing rapidly. The determining factor is the load during ON periods,

1
= pE[S]. If
1
> 1, the buffer will build up dramatically during ON periods,
and, as long as itself is smaller than 1, build down again during OFF periods, but
these large uctuations in buffer occupancy result in a large average delay, which
becomes ever larger as the mean lengths of the ON and OFF periods increase (i.e.
as K increases).
In Fig. 6.7, we show the ratio between the mean delay for strategy (32, 64, 128,
6-16 SLEEP MODE FOR CORRELATED DOWNLINK TRAFFIC
K = 40
K = 10
K = 1
0 0.2 0.4 0.6 0.8 1
0
10
20
30
40

E
[
D
]
Figure 6.4: Mean delay versus the load for a (2, 4, 8, 16, 32, 32, ) strategy for different
Ks, where =
2
3
and E[S] =
5
3
.
K = 40
K = 10
K = 1
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1

P
r
[
a
c
t
i
v
e
]
Figure 6.5: Antenna activity rate versus the load for a (2, 4, 8, 16, 32, 32, ) strategy for
different Ks, where =
2
3
and E[S] =
5
3
.
CHAPTER 6 6-17
=
1
9
;
1
=
1
6
=
4
9
;
1
=
2
3
=
2
3
;
1
= 1
=
7
9
;
1
=
7
6
0 50 100 150 200
0
10
20
30
K
E
[
D
]
Figure 6.6: Mean delay versus the correlation parameter for a (2, 4, 8, 16, 32, 32, . . .)
strategy, for different loads, where =
2
3
and E[S] =
5
3
.
= 0.2
= 0.4
= 0.6
0 10 20 30 40 50
0
2
4
6
8
K
E
[
D
]
/
E
[
D

]
Figure 6.7: Ratio between the mean delay E[D] for strategy (32, 64, 128, 256, 256, ) and
the mean delay E[D

] for strategy (1, 1, ) (i.e. no sleep mode) versus the correlation


parameter K, for different loads, where p = 1, E[S] =
1
2
.
6-18 SLEEP MODE FOR CORRELATED DOWNLINK TRAFFIC
A
B
C
0 0.2 0.4 0.6 0.8 1
0
10
20
30

E
[
D
]
Figure 6.8: Mean delay versus the load for strategies A:(16, 16, ), B:(2, 4, 8, 16, 16, )
and C:(1, 2, 3, 4, , 15, 16, 16, ), where K = 30, =
2
3
and E[S] =
5
3
.
256, 256, ) and the mean delay for the reference case where no sleep mode op-
eration is used. We have chosen p = 1, E[S] =
1
2
and the load is varied by
changing the value of . The extra delay caused by the sleep mode mechanism
diminishes when the trafc correlation increases, which shows that the sleep mode
is especially effective for bursty trafc.
In Figs. 6.8 and 6.9, we show how well different sleep mode strategies fare,
in function of the load. We already know what the inuence of T
max
is when
the load approaches zero. We now investigate the inuence on the delay and the
activity rate if this maximum is reached directly i.e. strategy (T
max
, T
max
, ), or
exponentially as in the IEEE standard i.e. (2, 4, 8, ,
T
max
2
, T
max
, T
max
, ), or even
more slowly for example linearly (1, 2, 3, , T
max
1, T
max
, T
max
, ).
In Figs. 8 and 9, we have chosen K = 30, =
2
3
and E[S] =
5
3
. Note that
the activity rate versus the load shows a straight line between (0,
1
T
max
) and (1, 1)
when T
max
is reached directly. When T
max
is reached more slowly, the activity
rate increases, especially so for reasonable (i.e. 0.2 < < 0.6), hence such
slow strategies are less energy efcient. On the other hand, the mean delay in
the same load interval seriously benets from slowly starting strategies. The best
strategy depends of course on the trade-off one is willing to make between energy
consumption and quality of service, but it appears to us that sleep mode operation
provides powerful instruments to reach an acceptable compromise.
Finally, in Figs. 6.10 and 6.11 we have plotted respectively the mean delay
E[D] and the standard deviation of the delay StdDev[D], versus the load , for
three different sleep mode strategies, K = 20 and =
1
2
. Here we have assumed a
CHAPTER 6 6-19
A
B
C
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1

P
r
[
a
c
t
i
v
e
]
Figure 6.9: Antenna activity rate versus the load for strategies A:(16, 16, ),
B:(2, 4, 8, 16, 16, ) and C:(1, 2, 3, 4, , 15, 16, 16, ), where K = 30, =
2
3
and
E[S] =
5
3
.
constant-length service time of 2 slots. We observe that not only the mean delay,
but also its standard deviation is lower for more slowly starting strategies. Also
note that in both plots, the curves of the three strategies coincide, for small values
of . Fig. 11 also illustrates that we are not conned to rst-order moments and
that we can just as easily compute the higher order moments of the buffer content
and the delay.
6.7 Conclusion
In this paper, we have analysed a particular type of discrete-time queueing models
with vacations, which is well-suited to model the sleep mode operation of IEEE
802.16e. Unlike previous work, we have taken the trafc correlation into account
by using a D-BMAP. Our analysis technique provides efcient formulas for such
important characteristics as the packet delay and the antenna activity rate. Numer-
ical examples show that sleep mode operation is indeed a viable technique to get
considerable energy savings, at the cost of a moderate increase in delay. The pres-
ence of correlation in the carried trafc stream has a major impact on the behaviour
of the BS buffer and the efciency of the sleep mode mechanism.
6-20 SLEEP MODE FOR CORRELATED DOWNLINK TRAFFIC
A
B
C
0 0.2 0.4 0.6 0.8 1
0
10
20
30
40

E
[
D
]
Figure 6.10: Mean delay versus the load for strategies A: (64, 64, ), B:
(2, 4, 8, 16, 32, 64, 64, ) and C: (2, 4, 6, 8, , 62, 64, 64, ), for K = 20, =
1
2
and
constant service times of 2 slots.
CHAPTER 6 6-21
A
B
C
0 0.2 0.4 0.6 0.8 1
0
10
20
30
40

S
t
d
D
e
v
[
D
]
Figure 6.11: Standard deviation of the delay versus the load for strategies A: (64, 64, ),
B: (2, 4, 8, 16, 32, 64, 64, ) and C: (2, 4, 6, 8, , 62, 64, 64, ), for K = 20, =
1
2
and
constant service times of 2 slots.
6-22 SLEEP MODE FOR CORRELATED DOWNLINK TRAFFIC
Appendix
In this appendix, we provide a short proof for the property that if two matrices
can be written as a power series of the same matrix, then the two matrices are
commutative. Let the matrices F and G be dened as
F
.
=

i
f
i
M
i
and G
.
=

i
g
i
M
i
, (6.46)
for a square matrix M and scalar sequences f
i
and g
i
. If we assume that the above
sums converge absolutely (which is the case if f
i
is a probability matrix and M
pgm evaluated inside the unit circle), then the product FG can then be written as
FG =

j
f
i
g
j
M
i+j
. (6.47)
The expression for GF can be written in exactly the same form, hence matrices F
and G are commutative.
References
[1] IEEE 802.16e-2005, Part 16: Air interface for xed and mobile broadband
wireless access systems Amendment 2: physical and medium access con-
trol layers for combined xed and mobile operation in licensed bands
Corrigendum 1, February 2006.
[2] M.S. Kuran, T. Tugcu, Asurvey on emerging broadband wireless access tech-
nologies, Computer Networks, Vol. 51, No. 11, 2007, pp. 30133046.
[3] Y. Xiao, Energy saving mechanismin the IEEE 802.16e wireless MAN, IEEE
Communications Letters, Vol. 9, No. 7, 2005, pp. 595597.
[4] Y. Zhang, M. Fujise, Energy management in the IEEE 802.16e MAC, IEEE
Communications Letters, Vol. 10, No. 4, 2006, pp. 311313.
[5] K. Han, S. Choi, Performance analysis of sleep mode operation in IEEE
802.16e mobile broadband wireless access systems, Proceedings of the IEEE
63rd Vehicular Technology Conference, VTC2006-Spring (Melbourne, 7-10
May 2006), Vol. 3, pp. 11411145.
[6] Y. Park, G.U. Hwang, Performance modelling and analysis of the sleep-mode
in IEEE 802.16e WMAN, Proceedings of the IEEE 65th Vehicular Technol-
ogy Conference, VTC2007-Spring (Dublin, 22-25 April 2007), pp. 2801
2806.
CHAPTER 6 6-23
[7] J.-B. Seo, S.-Q. Lee, N.-H. Park, H.-W. Lee, C.-H. Cho, Performance analy-
sis of sleep mode operation in IEEE 802.16e, Proceedings of the 60th Vehic-
ular Technology Conference, VTC2004-Fall (Los Angeles, 26-29 September
2004), Vol. 2, pp. 11691173.
[8] S.-J. Kwon, Y.W. Chung, D.K. Sung, Queueing model of sleep-mode opera-
tion in cellular digital packet data, IEEE Transactions on Vehicular Technol-
ogy, Vol. 52, No. 4, 2003, pp. 11581162.
[9] Y.-B. Lin, Y.-M. Chuang, Modeling the sleep mode for cellular digital packet
data, IEEE Communications Letters, Vol. 3, No. 3, 1999, pp. 6365.
[10] N.-H. Lee, S. Bahk, MAC sleep mode control considering downlink trafc
pattern and mobility, Proceedings of the IEEE 61st Vehicular Technology
Conference, VTC2005-Spring (Stockholm, 30 May1 June 2005), Vol. 3,
pp. 20762080.
[11] H. Bruneel, Queueing behavior of statistical multiplexers with correlated in-
puts, IEEE Transactions on Communications, Vol. 36, No. 12, 1988, pp.
13391341.
[12] S. Wittevrongel, H. Bruneel, Discrete-time queues with correlated arrivals
and constant service times, Computers & Operations Research, Vol. 26, No.
2, 1999, pp. 93108.
[13] C. Blondia, O. Casals, Statistical multiplexing of VBR sources: A matrix-
analytic approach, Performance Evaluation, Vol. 16, No. 13, 1992, pp. 5-20.
[14] A.T. Andersen, B.F. Nielsen, A Markovian approach for modeling packet
trafc with long-range dependence, IEEE Journal on Selected Areas in Com-
munications, Vol. 16, No. 5, 1998, pp. 719732.
[15] H.R. Gail, S.L. Hantler, B.A.Taylor, Spectral analysis of M/G/1 and G/M/1
type Markov chains, Advances in Applied Probability, Vol. 28, No. 1, 1996,
pp. 114165.
[16] H. Takagi, Queueing analysis, a foundation of performance evaluation, Vol.
3: Discrete-time systems, North-Holland, Amsterdam, 1993.
[17] I. Mitrani, Modelling of computer and communication systems, Cambridge
University Press, Cambridge, 1987.
[18] H. Bruneel, Buffers with stochastic output interruptions, Electronics Letters,
Vol. 19, No. 18, 1983, pp. 735737.
[19] J.D.C. Little, A proof for the queuing formula: L = W, Operations Re-
search, Vol. 9, No. 3, 1961, pp. 383387.
6-24 SLEEP MODE FOR CORRELATED DOWNLINK TRAFFIC
[20] D. Fiems, H. Bruneel, A note on the discretization of Littles result, Opera-
tions Research Letters, Vol. 30, No. 1, 2002, pp. 1718.
[21] S. De Vuyst, S. Wittevrongel, H. Bruneel, Mean value and tail distribution
of the message delay in statistical multiplexers with correlated train arrivals,
Performance Evaluation, Vol. 48, No. 14, 2002, pp. 103129.
7
Delay versus Energy Consumption of
the IEEE 802.16e Sleep-mode
Mechanism
Stijn De Vuyst, Koen De Turck, Dieter Fiems, Sabine Wittevron-
gel and Herwig Bruneel
published in IEEE Transactions on Wireless communications
Abstract. We propose a discrete-time queueing model for the evaluation of the
IEEE 802.16e sleep-mode mechanism of Power Saving Class (PSC) I in wireless
access networks. Contrary to previous studies, we model the downlink trafc by
means of a Discrete Batch Markov Arrival Process (D-BMAP) with N phases,
which allows to take trafc correlation into account. The tradeoff between energy
saving and increased packet delay is discussed. In many situations, the sleep-mode
performance improves for heavily correlated trafc. Also, when compared to other
strategies, the exponential sleep-period update strategy of PSC I may not always
be the best.
7.1 Introduction
We focus on the performance of the sleep-mode mechanism for Power Saving
Class (PSC) I, dened in the IEEE 802.16e standard for Broadband Wireless Ac-
7-2 DELAY VERSUS ENERGY CONSUMPTION FOR IEEE 802.16E
L L
L L L
T
0
T
0 0
2T 0
2T
0
4T
MS
BS
buffer
arriving
downlink
traffic
d
o
w
n
l
i
n
k
sleep
period
1
sleep
period
1
sleep
period2
sleep
period2
sleep
period3
SLEEP SLEEP AWAKE
busy busy idle
AWAKE
Figure 7.1: Operation of the IEEE 802.16e PSC I sleep-mode mechanism in case of
downlink trafc. Without loss of generality, we assume a sleep periodto include the
ensuing listening interval L, contrary to the denition in the standard. This assumption
facilitates the analysis, without affecting the validity of the results.
cess networks [1]. This standard aims to support mobile users in WiMAX (IEEE
802.16) wireless networks by addressing typical mobility-related issues. On the
MAC layer, these include enhanced security, faster handover protocols and energy
saving mechanisms, the latter of which are the topic of this letter.
Consider the downstream communication between a single Mobile Station
(MS) and its serving Base Station (BS) over a WiMAX wireless link. To extend its
battery life, the IEEE 802.16e sleep-mode mechanism enables the MS to turn off
its radio interface for a certain time whenever the BS has no packets in its queue
destined for this MS, see Fig. 7.1. Specically, if the BSs queue is empty, the
MS starts a sleep period during which it remains powered down and thus cannot
be reached by the BS. At the end of this period, the MS is briey reactivated (lis-
tening interval, indicated on the gure by L) to check whether there are packets
waiting for it at the BS. If not, the MS initiates a second sleep period, a third, and
so on. However, if any packets have arrived at the BS during the last sleep pe-
riod, the MS remains powered and enters awake mode. The BS then transmits all
the packets in its buffer exhaustively, after which the whole procedure is repeated.
Hence, the BS transmitter alternates between idle periods, consisting of a number
of contiguous sleep periods, and busy periods. In case of IEEE 802.16e PSC I,
an exponential update strategy is used. The rst sleep period of an idle period
has length T
0
while every subsequent sleep period doubles in length until a maxi-
mum T
max
is reached. After that, all additional sleep periods have length T
max
. We
denote this update strategy by (T
0
, 2T
0
, 4T
0
, . . . , T
max
).
The drawback of this mechanism is an inherent degradation of the QoS. As the
transmission of packets arriving at the BS must await the end of the current sleep
period, the overall delay of the packets increases. It is therefore important to care-
CHAPTER 7 7-3
fully tune the parameters of the sleep-mode mechanism in order to obtain a good
tradeoff between energy reduction and additional delay. To this end, the MS en-
ergy consumption has been studied in e.g. [3, 4] for sleep mode PSC I and in [5, 6]
for PSC II. Queueing models that also analyse the BS packet delay have been
presented in [79]. All of these models however, assume an uncorrelated mem-
oryless arrival process (i.e. a Poisson process) for the trafc arriving at the BS,
which is very restrictive in many cases [10]. It is known that Markov-modulated
models [11, 12] can capture the memory of trafc traces quite well. Hence, we
consider here the sleep-mode BS queueing behaviour assuming a Discrete Batch
Markov Arrival Process (D-BMAP) [13]. This is a far more general model that
can capture complex correlation structures. The model proposed is also generic
with respect to the update strategy of the sleep periods and not restricted to the
exponential strategy of IEEE 802.16e PSC I alone. This allows us to explore some
alternatives: in section 7.5, we compare the performance of the exponential strat-
egy with both constant and linear strategies.
The main purpose of this letter is to demonstrate that trafc correlation heavily
affects the performance of the sleep-mode mechanism, both in terms of achieved
energy efciency and packet delay. Neglecting this correlation in the analysis,
as in [39], may therefore lead to a very poor estimation of the true performance
characteristics. Additionally, we propose a convenient efciency metric that can be
used to optimise the design parameters. In section 7.2 we present our model of the
BS queue. The analysis is outlined in Sections 7.3 and 7.4. For an arrival process
with two phases, we compare the performance of different update strategies in
section 7.5. Section 7.6 concludes.
7.2 Model
We model the BS as a discrete-time single-server queue with innite capacity and
a rst-come-rst-served (FIFO) discipline [14]. Time is assumed to be divided
into xed-length intervals called slots. The service time of a packet is the number
of slots required to transmit it successfully from the BS to the MS. We assume
the service times of the packets to be independent random variables with com-
mon probability generating function (pgf) S(z) and mean E[s] =S

(1). This is a
reasonable assumption in case the wireless medium is fast fading [15].
Packets arrive in the queue according to a D-BMAP with N phases (or states)
[13]. In each slot, the arrival process resides in one of the N phases and is charac-
terised by the probabilities a(k, j[i) (k0, i, j 1, , N) that, given phase i in a
slot, there are k packet arrivals and the phase switches to j in the next slot. These
probabilities can be arranged in the NN probability generating matrix (pgm)
7-4 DELAY VERSUS ENERGY CONSUMPTION FOR IEEE 802.16E
A(z) with entries
[A(z)]
i j
=

k=0
a(k, j[i)z
k
.
Many methods exist to t this model to measured trafc data, see e.g. [16]. The
equilibrium distribution of the phases is found as the 1N probability vector
satisfying A(1) = and 1=1, where 1 is a N1 column vector with all entries
equal to 1. The mean number of arrivals per slot is then E[a] =A

(1)1 and the


load of the BS queue is =E[a]E[s].
The sleep-period update strategy is (t
1
, t
2
, . . .) where the positive integers t
n
(n 1) are the lengths of the subsequent sleep periods of the same idle period.
Further dene
n
=
n
i=1
t
i
(n1) and
0
=0 as the starting slots of the sleep periods
relative to the start of the idle period. We will assume that the sleep-period lengths
remain xed after the Jth sleep period, i.e. if t
n
=t
J
for n J. For the PSC I
mechanism, this clearly applies with t
J
=T
max
.
7.3 Queueing Analysis
In [17], we obtained the distribution of the BS queue content via some intermediate
steps. First the queue is considered at service completion instants only, i.e. at the
beginning of a slot following the transmission of a packet. If the ith entry of the
probability generating vector (pgv) U
c
(z) is the partial pgf of the queue content at
a service completion with phase i, then
U
c
(z) = U
c
(0)(V(z)I)
_
zIS(A(z))
_
1
S(A(z)), (7.1)
where I denotes the NN identity matrix and where S(A(z)) is the matrix power
series in A(z) with the same coefcients as the power series S(z) in z. The pgm
V(z) is calculated as
V(z) =
J1

n=1
A(0)

n1
_
A(z)
t
n
A(0)
t
n

+A(0)

J1
_
IA(0)
t
J
_
1
_
A(z)
t
J
A(0)
t
J

. (7.2)
In (7.1), the unknown vector U
c
(0) can be determined by invoking the bound-
edness of U
c
(z) for all [z[ 1 using the following procedure. First, note that
the queue content at subsequent service completions forms a stochastic process
of the M/G/1-type, see e.g. [18]. The transform method for solving such sys-
tems has been treated extensively in [19]. Specically, it is shown that the matrix
zIS(A(z)) on the right-hand side of (7.1) becomes singular in exactly N points
z =z
i
(i =1, . . . , N) in the unit disc, counting multiplicities. These points are the
zeroes of the function det(zIS(A(z))) and can be calculated numerically, along
CHAPTER 7 7-5
with their multiplicities, using a root-nding algorithm. In case the zeroes z
i
are
distinct, i.e. all with multiplicity one, it is possible to construct column vectors
i
satisfying
(z
i
IS(A(z
i
)))
i
= 0, (7.3)
for each i =1, . . . , N. Clearly, as the matrix z
i
IS(A(z
i
)) is singular, a non-trivial
solution
i
always exists. Now, note that in (7.1) the last two factors commute.
Substitution of z=z
i
in (7.1) and post-multiplying both sides with (7.3) yields
U
c
(0)(V(z
i
)I)
i
= 0, i = 1, . . . , N.
This is a set of N equations that are known to be independent [20] and thus allow
to solve for the entries of U
c
(0). The procedure is somewhat more involved in the
peculiar case that zeroes z
i
occur with multiplicity larger than one, see [19].
Once the distribution of the queue content at service completions (7.1) is known,
it is possible to obtain the distribution for arbitrary slot as well. Let the ith entry of
the pgv U(z) be the partial pgf of the queue content in an arbitrary slot with phase
i. In [17], we nd
U(z) = (z1)E[a]U
c
(z)
_
A(z)I
_
1
. (7.4)
Additionally, if u
c,0
=U
c
(0)1 is the probability that the queue is empty at a service
completion, the mean length of a busy period is E[B]=E[s]/u
c,0
, whereas the mean
length of an idle period is found as
E[I] =
U
c
(0)
u
c,0
_J1

n=1
t
n
A(0)

n1
+t
J
A(0)

J1
_
IA(0)
t
J
_
1
_
1, (7.5)
7.4 Mean delay and energy consumption
Many interesting characteristics of the distribution of the BS queue content can be
derived from (9.6), such as the moments up to any order or its tail asymptotics.
For instance, the mean queue content is E[u] =U

(1)1 and, by Littles law [21], the


mean delay experienced by an arbitrary packet is E[d] =E[u]/E[a]. With regard to
the energy consumption of the MS, we assume that the MS is active during every
busy slot, as well as during the last slot of each sleep period (i.e. the listening
interval). When active, it consumes one unit of energy, while no energy is used
during the other (idle) slots. Hence, the mean packet delay D and the energy E
consumed by the MS per slot are
D =
U

(1)1
E[a]
and E = +(1)

E[I]
, (7.6)
7-6 DELAY VERSUS ENERGY CONSUMPTION FOR IEEE 802.16E
where denotes the mean number of subsequent sleep periods in a single idle
period,
=
U
c
(0)
u
c,0
_J1

n=1
A(0)

n1
+A(0)

J1
_
IA(0)
t
J
_
1
_
1.
7.5 Examples and discussion
These results can be used to assess the performance of the sleep-mode mechanism
under a wide variety of trafc conditions, service-time distributions and sleep-
period update strategies. Some important phenomena can be demonstrated with as
little as two phases for the D-BMAP arrival process. We here assume an ON/OFF
trafc source:
A(z) =
_
1q+qz 0
0 1
__
p
1
1p
1
1p
2
p
2
_
. (7.7)
The second matrix contains the phase transition probabilities. The rst matrix is
chosen such that when a slot is in phase 1 (ON), an arrival occurs with probability
q, whereas there are no arrivals in a slot with phase 2 (OFF). The equilibrium dis-
tribution of the phases directly follows as =[ 1] where =
1p
2
2p
1
p
2
, which
means that the long-run fraction of slots in the ON- and OFF-phase is and 1
respectively. Also, the mean numbers of contiguous slots in the same phase are
K
1
and
K

respectively, with K=
1
2p
1
p
2
. Hence, higher values of K mean that
the arrival process remains in the same phase for longer periods of time and there-
fore exhibits more correlation. If K=1 however, the arrivals are uncorrelated. The
parameter K is referred to as the correlation factor or the burstiness factor of the
arrival process [22]. Let us assume ARQ retransmissions are used to make sure
a packet is received correctly by the MS. If each transmission requires one slot
and succeeds with probability 0.5, then the service times of the BS queue have a
shifted geometric distribution with mean E[s] =2. The load of the system is then
=qE[s]. Three types of strategies for updating the consecutive sleep-period
lengths are considered: constant, denoted as T
0
C, where every sleep period has
the same length T
0
; linear, denoted as T
0
L = (T
0
, 2T
0
, 3T
0
, 4T
0
, . . . , T
max
), where
each sleep period increases with the same amount T
0
up to a maximum T
max
; ex-
ponential, where the subsequent sleep periods double in length up to maximum
T
max
and which we denote as T
0
E = (T
0
, 2T
0
, 4T
0
, 8T
0
, . . . , T
max
). For the linear
and exponential strategies, we assume T
max
=256 slots.
Use of the sleep-mode mechanism increases both the packet delay (which is
bad) and the battery life of the MS (which is good). To make a fair tradeoff, one
needs to compare both delay and energy consumption to the reference case where
the MS is always activated. Let D

and E

=1 be the mean packet delay and MS


energy consumption in this reference case (they can be obtained from (7.6) using
CHAPTER 7 7-7

8C

32C

64C



8L


32L

64L

8E


32E

64E

D
constant

linear

exponential

K=50
40 30 20 10
K=1
2 2.2 2.4 2.6 2.8 3
0
2
4
6
8
10
12
14
Figure 7.2: Shift of (
E
,
D
) vs. K in case =0.25, =0.25 and different strategies.
strategy 1C). We dene

E
=
E

E
1, and
D
=
D
D

1, (7.8)
which respectively give the battery-life extension and the delay increase result-
ing from sleep-mode usage. The minimal value of both
E
and
D
is equal to
0, corresponding to the reference case of communication without the sleep-mode
mechanism. However, in case any of the above sleep-mode strategies is used, both
values will increase. For example,
E
=0.7 would indicate that the MSs battery
life is increased by 70% while the corresponding
D
>0 measures the inherent
delay cost.
In Fig. 7.2 we illustrate how the trafc correlation K typically affects the per-
formance of various sleep-period update strategies. Here, and are both xed at
0.25 and we show the shift of the point (
E
,
D
) as K increases from 1 (no trafc
correlation) to 50 (high correlation). Clearly, points in this graph represent a bet-
ter performance as they are situated lower (small
D
) and further to the right (high

E
). First observe that for the constant strategies, the energy consumption is insen-
sitive to K. In fact, for T
0
C we know that E =+(1)
1
T
0
because exactly one
out of every T
0
idle slots is a listening slot. However, the delay cost
D
is reduced
signicantly if K increases, especially so when the sleep period T
0
is larger. For
the linear and exponential strategies, the situation is less unambiguous, although
7-8 DELAY VERSUS ENERGY CONSUMPTION FOR IEEE 802.16E

K=1

K=10

K=50

K=1

K=10

K=50

D
constant 64C

exponential 64E

=0.03
0.1 0.2 0.3 0.4
=0.47
0 5 10 15 20 25
0
10
20
30
Figure 7.3: Shift of (
E
,
D
) vs. in case =0.25, K=1, 10, 50 and for strategies 64E
and 64C.
increased trafc correlation generally allows more energy to be saved (especially
for low initial sleep period T
0
). Sometimes, but not necessarily, this comes at the
price of a higher delay. The question which strategy performs optimally for a
given trafc pattern, is obviously a matter of how much energy saving is valued
over delay cost. For example, let us compare the shift of 8E and 64C in Fig.
7.2 in case of a delay-sensitive application. For K =1, 8E obviously performs
best, as its delay cost is about six times lower. For K=50 however, the delay of
64C is only half that of 8E while both strategies have almost the same energy
consumption.
Whether or not correlation is benecial for a certain strategy also depends on
the load , as we demonstrate in Fig. 7.3. The shift of (
E
,
D
) for strategies 64C
and 64E is plotted as the load changes from 0.03 to 0.47 for three values of the
trafc correlation factor K. This conrms our previous observation that correlation
is always benecial for a constant strategy. For the exponential strategy however,
correlation improves the performance only if the load is high, while the opposite
is true in case of low load.
In order to compare different sleep-mode strategies quantitatively, we need a
scalar metric for their performance. We propose to dene the efciency
c
of a
CHAPTER 7 7-9
8C
32C
8L
32L
8E
32E

0.5
0 0.1 0.2 0.3 0.4 0.5
5
0
5
10
Figure 7.4: Efciency
0.5
for different strategies in case K=50, =0.25.
strategy as

c
=
E
c
D
, (7.9)
where c is a delay penalty which reects to what extent low packet delay is
preferred over energy saving. For delay-sensitive applications, this penalty must
be taken relatively high. If the primary focus is to extend battery life however,
then c should be small. An example is given in Fig. 7.4 where the efciency

0.5
is shown as a function of the load in case K =50 and =0.25. According
to cost function (7.9), it is efcient to use a sleep-mode strategy as long as its

c
is positive. If
c
<0 however, the best solution is not to use the sleep-mode
mechanism at all (or equivalently, to use 1C). Particularly in Fig. 7.4, observe
that the constant and linear strategies outperform the exponential (IEEE 802.16e)
strategies for all but the smallest values of the load.
Clearly, more complex trafc models than (7.7) can be considered as well. For
example, the 4IPP model in [23], often used in IEEE 802.16 performance studies,
consists of the superposition of 4 independent ON/OFF-sources. Each of them can
be described as a 2-state D-BMAP of the form (7.7): source i (i = 1, 2, 3, 4) has
pgm A
i
(z) as in (7.7) with parameters
i
, q
i
and correlation factor (or timescale)
K
i
. The superposition of these four sources is again a D-BMAP, with pgm A(z) =

4
i=1
A
i
(z), and the analysis of sections 7.3 and 7.4 can be applied.
7-10 DELAY VERSUS ENERGY CONSUMPTION FOR IEEE 802.16E
7.6 Conclusions
We provided an analytic framework to estimate the performance of the (IEEE
802.16e) sleep-mode mechanism under a wide range of conditions for the traf-
c model (D-BMAP), the service-time distribution and the sleep-period update
strategy. We show that in many situations the sleep-mode mechanism performs
signicantly better in case of highly correlated trafc. Another important obser-
vation is the fact that the exponential sleep-period update strategy used in IEEE
802.16e PSC I is not always the best choice for certain trafc patterns between BS
and MS.
References
[1] IEEE 802.16e-2005, Part 16: Air interface for xed and mobile broadband
wireless access systems Amendment 2, Corrigendum 1, February 2006.
[2] B. Kim, J. Park, Y.-H. Choi, Power saving mechanisms of IEEE 802.16e:
sleep mode vs. idle mode, Lecture Notes in Computer Science, vol. 4331
(ISPA 2006 Ws), pp. 332340, 2006.
[3] Y. Xiao, Energy saving mechanism in the IEEE 802.16e wireless MAN,
IEEE Commun. Lett., vol. 9, no. 7, pp. 595597, 2005.
[4] Y., Zhang, M. Fujise, Energy management in the IEEE 802.16e MAC,
IEEE Commun. Lett., vol. 10, no. 4, pp. 311313, 2006.
[5] Y.-H. Han, S.-G. Min, D. Jeong, Performance comparison of sleep mode
operations in IEEE 802.16e terminals, Lecture Notes in Computer Science,
vol. 4490 (ICCS 2007, Part IV), pp. 441448, 2007.
[6] J.-R. Lee, Energy efcient sleep interval decision method considering delay
constraint for VoIP trafc, Lecture Notes in Computer Science, vol. 4003
(NEW2AN 2006), pp. 142151, 2007.
[7] K. Han, S. Choi, Performance analysis of sleep mode operation in IEEE
802.16e mobile broadband wireless access systems, in Proc. IEEE 63rd Ve-
hicular Techn. Conf. VTC 2006-Spring, vol. 3, pp. 11411145.
[8] Y. Park, G.U. Hwang, Performance modelling and analysis of the sleep
mode in IEEE 802.16e WMAN, in Proc. IEEE 65th Vehicular Techn. Conf.
VTC 2007-Spring, pp. 28012806.
[9] J.-B. Seo, S.-Q. Lee, N.-H. Park, H.-W. Lee, C.-H. Cho, Performance anal-
ysis of sleep mode operation in IEEE 802.16e, in Proc. IEEE 60th Vehicular
Techn. Conf. VTC 2004-Fall, vol. 2, pp. 11691173.
CHAPTER 7 7-11
[10] V. Paxson, S. Floyd, Wide-area trafc: the failure of Poisson modeling,
IEEE/ACM Trans. Netw., vol. 3, no. 3, pp. 226244, 1995.
[11] A.T. Andersen, B.F. Nielsen, A Markovian Approach for Modeling Packet
Trafc with Long-Range Dependence, IEEE J. Sel. Areas Commun., vol.
16, no. 5, 1998.
[12] H. Michiel, K. Laevens, Teletrafc engineering in a broad-band era, Proc.
IEEE, vol. 85, no. 12, pp. 20072033, 1997.
[13] C. Blondia, O. Casals, Statistical multiplexing of VBR sources: A matrix-
analytic approach, Perform. Evaluation, vol. 16, no. 13, pp. 5-20, 1992.
[14] H. Bruneel, Performance of discrete-time queueing systems, Comp. &
Oper. Res., vol. 20, no. 3, pp. 303320, 1993.
[15] C.C. Tan, N.C. Beaulieu, On rst-order Markov modeling for the Rayleigh
fading channel, IEEE T. Commun., vol. 48, no. 12, pp. 20322040, 2000.
[16] D. Fiems, V. Inghelbrecht, B. Steyaert, H. Bruneel, Markovian character-
isation of H.264/SVC scalable video, in Proc. 15th Int. Conf. on Analyti-
cal and Stochastic Modelling Techniques and Applications, ASMTA 08 (46
June 2008, Nicosia, Cyprus).
[17] K. De Turck, S. De Vuyst, D. Fiems, S. Wittevrongel, Performance analysis
of the IEEE 802.16e sleep mode for correlated downlink trafc, Telecom-
mun. Syst., DOI 10.1007/s11235-008-9119-8.
[18] M.F. Neuts, Structured stochastic matrices of M/G/1 type and their applica-
tions, Marcel Dekker, New York, 1989.
[19] H.R. Gail, S.L. Hantler, B.A. Taylor, Spectral analysis of M/G/1 and G/M/1
type Markov chains, Adv. Appl. Probab., vol. 28, no. 1, pp. 114165, 1996.
[20] H.R. Gail, S.L. Hantler, Linear independence of root equations for M/G/1
type Markov chains, Queueing Syst., vol. 20, no. 34, pp. 321339, 1995.
[21] D. Fiems, H. Bruneel, A note on the discretization of Littles result, Oper.
Res. Lett., vol. 30, no. 1, pp. 1718, 2002.
[22] S. Wittevrongel, H. Bruneel, Discrete-time queues with correlated arrivals
and constant service times, Comp. & Oper. Res., vol. 26, no. 2, pp. 93108,
1999.
[23] C.R. Baugh, J. Huang, R. Schwartz, D. Trinkwon, Trafc model for 802.16
TG3 MAC/PHY simulations, Technical report, IEEE 802.16 Broadband
Wireless Access Working Group, 2001.
8
Performance Analysis of Sleep Mode
Mechanisms in the Presence of
Bidirectional Trafc
Koen De Turck, Stijn De Vuyst, Dieter Fiems, Herwig Bruneel,
Sabine Wittevrongel
submitted to IEEE Transactions on Wireless Communications.
Abstract. Short battery life being a major impediment to a more widespread
use of wireless devices, industry standards like WiMAX and 3GPP Long Term
Evolution (LTE) embed power saving mechanisms in their protocol stack. Trad-
ing in network delay performance for extended battery life, these mechanisms re-
quire a careful balance to optimize overall network performance. Therefore, this
chapter assesses the impact of these mechanisms on network delay performance
and battery life, thereby rening existing performance studies of the power sav-
ing mechanisms. In particular, this study explicitly takes into account presence of
both downlink and uplink trafc and their inuence on the power saving mecha-
nism. This contrasts with previous studies, where the analyses only accounted for
downlink trafc. We obtain numerically efcient procedures to compute packet
delay as well as energy efciency characteristics under Markovian trafc. Then,
performance of power saving is investigated by some numerical examples. Our
results clearly indicate that uplink trafc heavily impacts the energy efciency of
8-2 SLEEP MODE WITH BIDIRECTIONAL TRAFFIC
the power saving mechanism.
8.1 Introduction
The IEEE 802.16e standard [1] backed by the WiMAX Forum [2] and the Long
Term Evolution (LTE) [3] of the Universal Mobile Telecommunications System
(UMTS) are emerging standards that have the potential to become the major stan-
dard for broadband wireless communication in the near future. These standards
regulate communication between mobile subscriber stations (MSS) and base sta-
tions (BS) user equipment (UE) and evolved Node-Bs (eNB) in LTE terminol-
ogy in a metropolitan area wireless network.
Short battery life is one of the main impediments to a more widespread use of
wireless devices. Hence, understandably, a lot of research is directed at solving or
at least mitigating this problem. Foremost this can be done by improving on the
efciency of batteries, but lately there is also a lot of interest in including energy-
saving measures in the communication protocols themselves. On that account, it is
no wonder that various energy-saving elements are incorporated in both WiMAX
and LTE. Energy saving mechanisms are referred to as sleep mode and idle mode
in WiMAX [4] and as discontinuous reception (DRX) in LTE [5].
In WiMAX, a MSS in sleep mode is still registered to a BS and still performs
hand-off procedures. In contrast, idle mode operation (which is optional in current
WiMAX standards) goes further and allows the MSS to be completely switched off
and unregistered with any BS, while still receiving broadcast trafc. In LTE, DRX
can be initiated when the UE is registered with an eNB (RRC CONNECTED: ra-
dio resource control connected state) or when it is not (RRC IDLE: radio resource
control idle state). In the former case, the UE scans the neighboring eNBs when
it detects signal quality degradation with respect to the serving eNB and performs
handover when required by the signal quality. Here, we mainly consider the sleep
mode mechanism and DRX in RRC CONNECTED. In either case, the MSS (or
UE) turns itself off for predetermined periods of time which are negotiated with
the BS (eNB). In accordance with WiMAX terminology, these periods are referred
to as sleep windows in the remainder of this chapter.
Three sleep mode classes are dened by the WiMAX standard. When in Sleep
Mode Class I, the sleep window is progressively doubled in size from a prenegoti-
ated minimum to a prenegotiated maximum value. Having reached the maximum
sleep window length, this sleep window is repeated uninterruptedly, until trafc
arrives. This class is considered to be suitable for best-effort and non-real-time
trafc. Class II features xed-length sleep windows, that is the same sleep win-
dow size is repeated continuously until there is uplink or downlink trafc to be
transmitted. This class is typically employed for UGS (unsolicited grant services)
trafc. Finally, class III negotiates only a one-time sleep period. This is typically
CHAPTER 8 8-3
used for management trafc, when the MSS knows when upcoming trafc is to be
expected.
Although class I is most interesting from a modeling point of view, the present
analysis is general enough to encompass all three classes. In particular, our model
assumes that a general sequence of sleep windowsizes t
i
, i 1, has been negotiated
between MSS and BS. In addition, this sequence is kept xed during the entire
operation of the system, instead of being negotiated at every start of a sleep period.
For LTE, the DRXcycle is initiated if there is neither download nor upload traf-
c for a predened amount of time or a predened number of frames. The choice
of the lengths of the sleep windows differs considerably from WiMAX. First, LTE
powers down the UE for short periods, a small sleep window is chosen. If there
is still no trafc after a predened number of these short sleep windows, DRX
switches to longer sleep windows. Choosing a short window size rst, allows the
UE to wake up fast in case of unexpected data arrival immediately after the DRX
cycle is enabled. As shown further, the general sequence of sleep window sizes
introduced above, is sufciently general to study DRX for any xed parameter set-
ting. As such, this chapter presents a unied analysis of power saving operations
in WiMAX and LTE.
Literature review
Power saving mechanisms have received a lot of interest lately from the perfor-
mance modeling community, most authors focusing on the sleep mode mecha-
nism of WiMAX. In [4], the average energy consumption of the MSS is obtained
in case of Poisson downlink trafc, as well as an approximate expression for the
mean packet delay. Retaining the Poisson assumption, Zhang and Fujise [6] in-
vestigate the energy consumption of the MSS accounting for both downlink and
uplink trafc. For DRX, Bontu and Illidge [5] calculate the energy consumption
of the UE for Poisson trafc. These authors also calculate the average delay that is
experienced by a packet that arrives during a sleep window.
An accurate assessment of the delay experienced at the BS (or eNB) buffer
however, requires a queueing model. For IEEE 802.16e, Han and Choi [7] model
the BS buffer as a continuous-time nite-capacity queue with a Poisson arrival
process and deterministic service times. A semi-Markov chain analysis leads to
expressions for the mean packet delay and the mean energy consumption by the
MSS. The analysis in [8] is based on an M/G/1/K queueing model with multiple
vacations and exhaustive service. For queueing systems with vacations, the server
leaves for a vacation whenever there are no packets in the queue. If there are no
packets in the queue upon returning from the vacation, the server immediately
leaves for another vacation, and so on [9]. Clearly, the vacations can be used to
model the sleep windows of the MSS. As the buffer space is nite, [8] also studies
8-4 SLEEP MODE WITH BIDIRECTIONAL TRAFFIC
packet loss at the BS. Similar work can also be found in [10], where the length of
a vacation is assumed to depend on the previous vacation length.
Predating UMTS, Cellular Digital Packet Data (CDPD) also provides a power
saving mechanism. As in LTE and contrasting with IEEE 802.16e, the subsequent
sleep windows do not increase in length, apart from the rst sleep window which
is shorter. The system can therefore be modeled as a queueing system with mul-
tiple vacations and exceptional rst vacation length. See [11] where CDPD sleep
mode is investigated for nite buffers and Poisson trafc. Delay, loss and energy
efciency are investigated. In addition to queueing models, Lin and Chuang [12]
assess the performance of CDPD sleep mode by simulation.
Investigation of optimal sleep window lengths have not been limited to the
restrictions imposed by the various standards. E.g., an alternative to the exponen-
tial increase of the sleep windows is evaluated by simulation in [13]. Finally, in
our previous work on the sleep mode mechanism of WiMAX [15], we considered
a general D-BMAP arrival process, and we found that trafc correlation, which
was hitherto neglected in almost every study, has an important impact on the sleep
mode performance.
The present chapter investigates the effects of the presence of uplink trafc
on the performance of the sleep mode operation, which is another aspect that has
received little attention in existing studies. In fact, only in [6], uplink trafc has
thus far been given attention, although in that paper it was restricted to Poissonian
trafc only and performance assessment was limited to energy efciency. Nev-
ertheless, the inuence of uplink trafc is considerable, as the MSS immediately
ceases the sleep mode operation whenever there is uplink trafc to be transmitted.
Thus, during uplink transmission, there is neither extra sleep mode delay experi-
enced by incoming downlink trafc nor any power saving, as the communication
circuitry is already switched on. We model the uplink activity by a nite but oth-
erwise general Markovian background process and derive buffer content statistics
as well as packet delay statistics for the downlink activity. The background pro-
cess can be tailored to model Markovian uplink trafc, which may queue up in
a nite capacity buffer. Note that the power saving does not affect uplink delay
and loss performance. Standard queueing models can be used to investigate uplink
performance which is therefore not considered here. We keep the derivation of the
performance measures concise by considering various types of embedded points.
Organization
The structure of the remainder of this chapter is the following. In Section 8.2, we
expound on the details of the performance model. In this section we also introduce
the notational conventions used throughout. Here and in the remainder, we avail
ourselves of WiMAX terminology in the clear understanding that the methodology
CHAPTER 8 8-5
at hand also applies to LTE. The performance analysis is spelled out in Section
8.3. We then apply our queueing theoretical results to a realistic WiMAX scenario
in Section 8.4 and illustrate our approach by some numerical examples. Finally,
conclusions are drawn in Section 8.5.
8.2 Model Description
For the purpose of our performance analysis, we consider a discrete-time stochas-
tic model, that is, time is divided into xed-length intervals called slots. The BS
is equipped with a buffer for the temporary storage of trafc in the downlink di-
rection, i.e. from BS to MSS. This BS buffer (from now on denoted as simply
the buffer) is modeled as a discrete-time queue with innite capacity and a rst-
come-rst-served (FCFS) server discipline.
To illustrate the dynamics of this buffer, a sample evolution of the system is
shown in Figure 8.1. The downward arrows indicate indicate the arrival and de-
parture of downlink packets at the BS buffer. The evolution of the buffer content
is shown in light gray. The upward arrows indicate the presence of uplink trafc.
During the dark gray intervals the MSS is in sleep mode.
( ( ( (
o
1
J
(
o
2
B ( (
/
J
(
/ / o
1
BS
MS
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
o
2
/ /
B
o
1
J
(
o
2
o
3
B ( (
/
J
(
/ / /
BS
MS
31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
B
/
J
(
/ / o
1
o
2
o
3
B ( (
o
1
J
(
o
2
BS
MS
61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90
Figure 8.1: The bidirectional sleep mode model at work over 90 slots. The sleep period
lengths are t
1
= 2, t
2
= 4, t
3
= 6, t
4
=t
max
= 8. The special sets of instants A, B, C, I , and
S
n
are indicated.
During the consecutive slots, downlink packets (from now on simply denoted
as packets) arrive in the buffer in batches according to an arrival process that
is detailed later on. The service times of these packets (i.e. the times needed to
8-6 SLEEP MODE WITH BIDIRECTIONAL TRAFFIC
transmit a packet to the MSS) are considered to be independent and identically
distributed (iid) random variables. Let the random variable s denote a service
time (expressed as a number of slots); its probability generating function (pgf) is
denoted by S(z)
.
=

i=1
P[s = i]z
i
.
During the course of our analysis we refer to slots during which there is uplink
trafc as active slots; the rest of the slots are called passive. The beginning
of an active or passive slot is an active or passive time instant, respectively. Both
the uplink trafc and the downlink trafc are modulated by the same (background)
Markov chain with M phases, characterized by the transition matrix P, with entries
as follows:
[P]
i j
.
=P[
k
= j[
k1
= i], (8.1)
where
k
is the random variable that denotes the phase (or background state) of the
Markov chain during slot k. The probability that a slot is active is phase dependent;
during phase i, this probability is q
i
. The distribution of the number of (downlink
packet) arrivals is also modulated by the background Markov chain. The pgf of
the batch size during phase i is given by A
i
(z). For convenience, we dene three
diagonal matrices with entries as follows:
[Q]
i j
.
=
i
j
q
i
, (8.2)
[Q]
i j
.
=
i
j
(1q
i
), (8.3)
[A(z)]
i j
.
=
i
j
A
i
(z), (8.4)
where the notation
i
j
denotes Kroneckers delta (i.e.
i
j
= 1 if i = j, and
i
j
= 0
otherwise. Note that this trafc scheme is very versatile and permits the uplink
trafc to be correlated to the downlink trafc. The load is dened as the average
number E[a] of packet arrivals per slot times the average service duration E[s]. The
system admits a stationary distribution if the condition < 1 is fullled.
The sleep-mode mechanism is incorporated in the queueing model as follows.
When the BS buffer is empty and the system is active (in the presence of uplink
trafc), the operation is not different from usual queueing models: the server re-
sumes work in the slot immediately after the rst packet arrival. However, when
the buffer becomes empty and the system is passive, it enters sleep mode: an in-
ternal timer is then started and the server awakes to check the buffer content after
a sleep period of t
1
time slots. When upon awaking the server nds that there are
still no packets, it goes to sleep again, this time for a sleep period of t
2
slots. The
next sleep period would be of length t
3
, and so on. In general, t
n
, n 1, denotes
the length (expressed in slots) of the nth sleep period after the buffer has become
empty. If after a sleep period, the server nds a non-empty buffer, it serves all
packets present at that point and also all new packets that arrive while the server
is working, until the buffer becomes empty again and the whole procedure is re-
peated. We call the sequence t = [t
i
]

i=0
of sleep lengths a sleep schedule. The set
CHAPTER 8 8-7
of all sleep schedules is denoted by S, and the set of sleep schedules that have a
repetitive tail is denoted by R
.
=t S : j, k, i > j : t
i
=t
i+k
. This subset plays
an important r ole in the following.
When the system becomes active during a sleep period, sleep mode is immedi-
ately abandoned, and the system is ready to serve incoming packets immediately.
Note that, when the system becomes passive again while the buffer is still empty,
sleep mode is resumed, but starting again with a sleep period of length t
1
, and so
on.
Let us agree upon some terminology: the system alternates busy periods (dur-
ing which the downlink trafc is being processed) with idle periods. Idle peri-
ods consist of a number of sleep periods and/or idle but active periods. Note
that some of our terminology is standard in queueing theory, but may have an-
other meaning in the IEEE community. For example, the term idle period is
fairly standard queueing theory terminology but should not be confounded with
the idle-mode mechanism of IEEE WiMAX. Secondly, we have opted for the neu-
tral term slots but will explain in Section 8.4 that the most natural translation
into a WiMAX context is a WiMAX frame. Lastly, the sleep periods we con-
sider include the listening intervals at the end. Hence, sleep periods correspond to
WiMAX sleep intervals plus the corresponding listening interval.
8.3 Analysis of Buffer content
In this section, we aim to determine the stationary distribution of the buffer content
at slot boundaries. More specically, we will derive the 1 M vector generating
function U(z) whose jth entry is dened as follows:
[U(z)]
i
=

k=0
z
k
P[u = k, = j]z
k
, (8.5)
Note that the Markov chain associated with the system has a rather large and
complex structure. Indeed, we must keep track of: (1) the number of packets in the
system; (2) the state of the background chain; (3) the elapsed service time of the
current packet (if applicable); (4) the number of packets in the system; (5) whether
or not the uplink channel is active; (6) how much time is elapsed if the system
resides in sleep mode. Although it is possible to write down the one-step transition
probabilities between states of this enormous state space and subsequently solve
the balance equations, we take a different route and use the following approach.
We rst dene various sets of embedded points. For notational convenience, we
tag the embedded points by special marks, such as I , S
n
and so on (see below).
Next, we establish relationships between the stationary distributions of the buffer
content at different sets of embedded points. Combination of these relationships
8-8 SLEEP MODE WITH BIDIRECTIONAL TRAFFIC
then nally leads to U(z). We are essentially solving the same Markov chain as
before, but we look at more convenient larger steps between embedded points.
For our rst group of embedded points we look inside idle periods. We mark
the beginning of an idle period by tag I , the beginning of the nth sleep period
by S
n
, n > 0, and the beginning of an active but idle slot by tag A. Note that,
necessarily, the buffer at such time instants is empty, and hence, the equilibrium
state during such time instants consists only of the phase variable. The probability
vector
I
has entries dened as follows:
[
I
]
j
.
=P[ = j, I ], (8.6)
that is, the jth entry of
I
is the probability of being in phase j during a slot whose
start instant is marked with tag I . Similarly, we also dene the probability vectors

A
and
S
n
, n 1. We now state three relationships involving these probability
vectors.
A. A time instant is marked by tag S
1
(i.e. it is the start of the rst slot of the
rst sleep period) if one of two (mutually exclusive) conditions is fullled:
(1) the time instant is marked with tag I and it is a passive time instant as
well; (2) the previous slot is active but idle, no arrivals occur in the previous
slot, and the current time instant is passive. This leads to the following
relationship:
[
S
1
]
j
= [
I
]
j
(1q
j
) +
M

i=1
[
A
]
i
A
i
(0)[P]
i j
(1q
j
).
or in vector notation,

S
1
=
I
Q+
A
A(0)PQ. (8.7)
B. A time instant is marked by tag S
n+1
if and only if the time instant t
n
slots
earlier is marked by tag S
n
, there are no arrivals during those t
n
slots and the
system remains passive up until the current time instant. Hence, we have

S
n+1
=
S
n
(PQA(0))
t
n
. (8.8)
C. The third relation claries the occurrence of active but idle time instants. A
time instant is marked by A if one of three mutually exclusive conditions is
fullled: (1) the time instant is marked with tag I and it is an active time
instant; (2) the beginning of the previous slot is marked by A, no arrivals
occur during the previous slot and the current time instant is active; (3) at
the current time instant, the system becomes active in a sleep period, that is,
there exists a time instant marked with S
1
, no arrivals have occurred in the
CHAPTER 8 8-9
interval up until the current time instant, and all but the current time instant
in the enclosed interval are passive. Based on these observations, we obtain

A
=
I
Q+
A
A(0)PQ
+
S
1

i=0
(A(0)PQ)
i
A(0)PQ
=
I
Q+
A
A(0)PQ
+
S
1
(I A(0)PQ)
1
A(0)PQ, (8.9)
where I denotes the identity matrix.
Next we focus our attention on sets of special time instants inside busy periods.
We discern three types: on the beginning of a busy period (marked by tag B), on
the time instants at which a new service is started (marked by tag b); and on the
time instants at which a service is completed (marked by tag C). We denote by
U
B
(z) the vector generating function at an instant with tag B, with the following
entries:
[U
B
(z)]
j
=

k=1
P[u = k, = j, B]z
k
. (8.10)
The vector generating functions U
b
(z) and U
C
(z) of the buffer content at b or C
instants respectively, are dened analogously. Below we establish three relations
for these vector generating functions.
D. The system has three distinct possibilities of entering a busy period at a
time instant: (1) the previous slot is active but idle, and there is at least
one arrival during the previous slot; (2) the system becomes active while
being in or at the end of a sleep period during which at least one arrival
has occurred; (3) the current time instant is the end of a sleep period during
which arrivals have occurred and the current time instant is passive. Each of
the possibilities corresponds to a term in the following equation:
U
B
(z) =
A
(A(z) A(0)) P
+

n=1

S
n
t
n
1

i=0
_
(A(z)PQ)
i
PQA(z)
(A(0)PQ)
i
A(0)PQ
_
+

n=1

S
n
(A(z)PQ)
t
n
A(0)PQ)
t
n
). (8.11)
E. The evolution of the buffer content from the beginning to the completion of
a service is as follows. In a single-server FCFS system, from the start to the
8-10 SLEEP MODE WITH BIDIRECTIONAL TRAFFIC
completion of a service, only one packet leaves the buffer. The number of
arriving packets during a service time depends on the length of a service.
Note that when the service length is k, (A(z)P)
k
is a matrix whose entry
[(A(z)P)
k
]
i j
is the partial pgf of the number of arrivals during a service time
followed by a slot with phase j, given a phase i in the rst slot of the service
time. Therefore, we have
U
C
(z) = U
b
(z)
1
z

k=1
P[S = k](A(z)P)
k
=
1
z
U
b
(z)S(A(z)P), (8.12)
where S(A(z)P) is a matrix-generating function that represents the number
of packet arrivals during a service time.
F. The set of time instants that is marked as C but not as I is equal to the set of
time instants that is marked as b but not as B. Hence, the distribution of the
buffer content at such a subset of time instants must be equal as well:
U
C
(z)
I
= U
b
(z) U
B
(z). (8.13)
Finally, there are two relations that give the distributions of the buffer content
at the beginning of a random slot and a random busy slot.
G. It is a consequence of the censored Markov chain property (1.7) that:
U(z)(I A(z)P) = (1z)U
C
(z). (8.14)
H. The vector generating function of the buffer content during a random busy
slot is given by
U

(z)(I A(z)P) = U
b
(z)(I S(A(z)P)). (8.15)
We can solve each of the occurring functions and variables in terms of the
constant vector
I
, which can subsequently be found by standard techniques to
compute the boundary vector of an M/G/1-type queueing model [17]. We can
reduce the innite sums to a closed-form expression whenever the sleep schedule
has a repetitive tail, that is, t R.
We can then get such performance measures as the moments of the buffer
content at various time instants, as well as probabilities that slots of a certain type
occur. As we will see, this is crucial for the computation of power consumption
measures. By Littles law, we can get from the mean buffer content also the mean
packet delay.
CHAPTER 8 8-11
8.3.1 Packet delay
The stationary distribution of the packet delay is determined via the stationary dis-
tribution of the unnished work during busy time instants. This is a modication
of the stochastic process of the buffer content, in which we do not record the num-
ber of packets present in the buffer, but rather the total work (in number of slots)
that the present population of packets represents. Hence, when k packets arrive
in the buffer, the amount of work is the sum of k iid service times, and hence has
probability generating function S(z)
k
. Denote by W

(z) the vgf of the unnished


work during busy slots. Notice that we know the stationary distributions of the
unnished work at the beginning and at the end of idle periods from the analysis
of the buffer content: they are respectively given by the probability vector
I
and
the vgf U
B
(S(z)). During a random busy slot, transitions occur according to an
additive process with associated matrix-generating function
1
z
PA(S(z)).
W

(z) = U
B
(S(z))(I
1
z
PA(S(z))
1
. (8.16)
Now, we derive the distribution of the delay. Note that the delay of a tagged
packet arriving during slot t consists of three parts: (1) the unnished work present
in the system at time t, or in case the system is idle, the time until the sleep period
is ended; (2) service times of the packets that arrive during slot t but before the
tagged packet; (3) the service time of the tagged packet itself.
The pgf of the packet delay is equal to:
D(z) =
1

_
W

(z) +
A
+

n=1

S
n
t
n
1

j=0
[PQA(S(z))]
j
t
n
j1

k=0
(PQz)
k
PQ+(PQz)
t
n
j
_

S(z)
S(z) 1
(PA(S(z)) I)1.
8.3.2 Power consumption
Having determined the probabilities that slots of all different types occur, it is
straightforward to calculate the average power consumption during stationarity.
E[P] =

n=1
u
S
n
1((t
n
1)E
s
+1E

)
+E
rt
U

(1)Q1+E
r
U

(1)Q1+E
t
u
A
1, (8.17)
where 1 is a column vector of length M, with all entries equal to 1, and E
s
, E

, E
r
,
E
t
, E
rt
denote the energy consumption during respectively sleep slots, listening
8-12 SLEEP MODE WITH BIDIRECTIONAL TRAFFIC
slots, receive slots (when downlink trafc is received), transmit slots (when uplink
trafc is being transmitted), and slots with both uplink and downlink trafc.
8.4 Numerical examples
We consider a queueing system with two buffers: an innite buffer in the downlink
direction, and a nite uplink buffer. This makes sense from a practical point of
view as memory is scarce in mobile stations but less so in the base stations.
As we focus on the Best Effort (BE) and Non Real-Time Polling Service
(nrtPS) trafc classes only, we select the packet arrival model for the downlink
trafc according to IEEE recommendations [19]. In the Individual Subscriber
Internet model from [19] the packet arrivals are assumed to conform to the In-
terrupted Poisson Process (IPP). This is a two-state Markov chain with a Poisson
arrival process in the ON state, and no arrivals in the OFF state. The model is char-
acterized by three parameters: an arrival rate , parameter denoting the fraction
of the time that the process resides in the ON-state and K measures the average
time between two ON periods. Unless indicated otherwise, we take = 0.5. We
assume that a packet transmission has a duration of 4 slots. As for the values for
the energy consumption, we assume that listening, receiving and sending slots con-
sume 1 unit of energy, while sleep slots. It is hard to nd data on energy usage that
hold true for many different protocols and appliances, but [20] indicates that the
these assumptions are not too far off. Hence we take E
s
= 1; E

= 1; E
r
= 1; E
t
= 1
and E
rt
= 2.
We look at the exponential sleep mode updating strategy of IEEE 802.16e char-
acterized by a t
min
and t
max
, as well as at the LTE sleep mode updating strategy,
which consists of a closedown period (i.e. n
c
sleep periods of length 1), followed
by n

short periods of length slots, followed in turn by as many longer periods of


length L as needed.
(t
i
)
i=0, ,
= (1, , 1
. .
n times
, , ,
. .
n

times
, L, L, ). (8.18)
The uplink buffer has n = M1 places and has a tridiagonal transition matrix
P:
[P]
0,0
= 1
u
;
[P]
i,i+1
=
u
;
[P]
i1,i
=
u
;
[P]
i,i
= 1
u

u
;
[P]
n,n
= 1
u
; (8.19)
CHAPTER 8 8-13
The load of this buffer is
u
=
u
/
u
. The parameter c
u
=
u
+
u
indicates how
fast the buffer level uctuates. Note that 0 < c
u
1 is a necessary and sufcient
condition. A downlink sleep mode is interrupted whenever the uplink buffer is
non-empty. The probability of an empty uplink buffer is independent of c
u
and
is equal to (1
u
)/(1
n+1
u
). We will see that in spite of this, c
u
has also a
signicant inuence on the performance of the downlink buffer.
The rst two plots, Figs. 8.2 and 8.3 show the inuence of the amount of
uplink trafc. When
u
= 0, then the system behaves as in the unidirectional case;
whereas for
u
= 1, the effects of sleep mode are almost completely gone away.
Notice that even a tiny amount of uplink trafc mitigates the sleep-mode effects to
a high degree.
u = 0
u = 0.05
u = 1.0
0 0.1 0.2 0.3 0.4 0.5
10
20
30
40

d
M
e
a
n
d
e
l
a
y
Figure 8.2: Mean delay versus the load
d
on the downlink channel for different values of
the load
u
on the uplink channel. Other parameters include
n = 40; c
u
= 0.7; K = 100; t = (1, 1, 1, 8, 8, 8, 64).
In the next two plots, Figs. 8.2 and 8.3 , we look at the effects of downlink
arrival correlation and of LTE-style versus 802.16e-style sleep strategies. When
looking at the packet delay, we see that for low downlink loads, the effects of the
strategy are dominant, whereas for high loads, the trafc correlation is dominant.
It is also interesting to note that the effects of arrival correlation on the energy
consumption is negligible.
We look more directly at the inuence of the uplink load
u
in Figs. 8.6 and
8.7. Here we see how the timescale parameter c
u
of the uplink buffer matters: if
the uplink buffer level uctuates slowly (c
u
<< 1), then the effects of sleep mode
are seen even for higher values of
u
, whereas in the opposite case, only for small
values of
u
a sleep mode penalty in the mean packet delay is seen.
8-14 SLEEP MODE WITH BIDIRECTIONAL TRAFFIC
u = 0
u = 0.05
u = 1.0
0 0.1 0.2 0.3 0.4 0.5
0
0.5
1
1.5
2

d
E
n
e
r
g
y
c
o
n
s
u
m
p
t
i
o
n
Figure 8.3: Energy consumption versus the load
d
on the downlink channel for different
values of the load
u
on the uplink channel. Other parameters include
n = 40; c
u
= 0.7; K = 100; t = (1, 1, 1, 8, 8, 8, 64).
A generally valid observation is that the optimal trade-off between increase in
average packet delay and decrease in energy consumption shifts towards longer
sleep lengths.
8.5 Conclusion
We analyzed the sleep mode mechanism of IEEE 802.16e WiMAX protocol, by
means of a queueing model that takes into account both downlink and uplink traf-
c. We are thus able to compute efciently such crucial performance measures
as the mean packet delay and the expected power consumption. We applied our
theoretical model to a realistic scenario and found that even a modest amount of
uplink trafc has a tremendous inuence on the system performance. The versatil-
ity of the chosen downlink trafc model allows us to model a multitude of trafc
situations, and to assess the inuence on the delay and the power consumption.
Acknowledgements
The third author is a Postdoctoral Fellow with the Research FoundationFlanders
(FWOVlaanderen), Belgium.
CHAPTER 8 8-15
S1; K = 10
S1; K = 100
S2; K = 10
S2; K = 100
0 0.2 0.4 0.6
0
10
20
30
40

d
M
e
a
n
d
e
l
a
y
Figure 8.4: Mean delay versus the load
d
on the downlink channel for different values of
K and for 802.16-style sleep schedule S
1
: (1, 2, 48, 16, 32, 64) and LTE-style schedule
S
2
: (1, 1, 1, 4, 4, 4, 64). Other parameters include n = 40; c
u
= 0.7;
u
= 0.02.
References
[1] IEEE 802.16e-2005, Part 16: Air interface for xed and mobile broadband
wireless access systems Amendment 2: physical and medium access con-
trol layers for combined xed and mobile operation in licensed bands
Corrigendum 1, February 2006.
[2] B. Li, Y. Qin, C.P. Low, C.L. Gwee, A survey on Mobile WiMAX, IEEE
Communications Magazine, Vol. 45, No. 12, pp. 7075, 2007.
[3] 3GPP TS 36.300 E-UTRAN Overall description Stage 2, v. 10.1.0,
September, 2010.
[4] Y. Xiao, Energy saving mechanism in the IEEE 802.16e wireless MAN,
IEEE Communications Letters, Vol. 9, No. 7, 2005, pp. 595597.
[5] C. Bontu and E. Illidge, DRX mechanism for power saving in LTE, IEEE
Communications Magazine, Vol. 47, No. 6, pp. 4855, 2009.
[6] Y. Zhang and M. Fujise, Energy management in the IEEE 802.16e MAC,
IEEE Communications Letters, Vol. 10, No. 4, 2006, pp. 311313.
[7] K. Han and S. Choi, Performance analysis of sleep mode operation in
IEEE 802.16e mobile broadband wireless access systems, Proceedings of
8-16 SLEEP MODE WITH BIDIRECTIONAL TRAFFIC
S1
S2
0 0.2 0.4 0.6
0.1
0.2
0.3
0.4

d
E
n
e
r
g
y
c
o
n
s
u
m
p
t
i
o
n
Figure 8.5: Energy consumption versus the load
d
on the downlink channel for different
values of K and for 802.16-style sleep schedule S
1
: (1, 2, 48, 16, 32, 64) and LTE-style
schedule S
2
: (1, 1, 1, 4, 4, 4, 64). Other parameters include n = 40; c
u
= 0.7;
u
= 0.02.
the IEEE 63rd Vehicular Technology Conference, VTC2006-Spring (Mel-
bourne, 7-10 May 2006), Vol. 3, pp. 11411145.
[8] Y. Park and G.U. Hwang, Performance modelling and analysis of the sleep-
mode in IEEE 802.16e WMAN, Proceedings of the IEEE 65th Vehicular
Technology Conference, VTC2007-Spring (Dublin, 22-25 April 2007), pp.
28012806.
[9] H. Takagi, Queueing Analysis: A Foundation of Performance Evaluation,
Volume 1: Vacation and priority Systems. Amsterdam: North Holland, 1991.
[10] J.-B. Seo, S.-Q. Lee, N.-H. Park, H.-W. Lee, and C.-H. Cho, Performance
analysis of sleep mode operation in IEEE 802.16e, Proceedings of the
60th Vehicular Technology Conference, VTC2004-Fall (Los Angeles, 26-29
September 2004), Vol. 2, pp. 11691173.
[11] S.-J. Kwon, Y.W. Chung, and D.K. Sung, Queueing model of sleep-mode
operation in cellular digital packet data, IEEE Transactions on Vehicular
Technology, Vol. 52, No. 4, 2003, pp. 11581162.
[12] Y.-B. Lin and Y.-M. Chuang, Modeling the sleep mode for cellular digital
packet data, IEEE Communications Letters, Vol. 3, No. 3, 1999, pp. 6365.
[13] N.-H. Lee and S. Bahk, MAC sleep mode control considering downlink traf-
c pattern and mobility, Proceedings of the IEEE 61st Vehicular Technology
CHAPTER 8 8-17
cu = 0.01
cu = 0.3
cu = 1.0
0 0.2 0.4 0.6 0.8 1 1.2
5
10
15
20

u
M
e
a
n
d
e
l
a
y
Figure 8.6: Mean delay versus
u
for different c
u
and
n = 20, = 0.5, K = 50, t = (1, 2, 4, 8, 16, 32, 64)
Conference, VTC2005-Spring (Stockholm, 30 May1 June 2005), Vol. 3, pp.
20762080.
[14] K. De Turck, S. De Vuyst, D. Fiems, and S. Wittevrongel, An analytic model
of IEEE 802.16e sleep mode operation with correlated trafc, Proceedings
of NEW2AN 2007, pp. 109120.
[15] K. De Turck, S. De Vuyst, D. Fiems, and S. Wittevrongel, Performance
analysis of the IEEE 802.16e sleep mode for correlated downlink trafc,
Telecommunication Systems, Vol. 39, No 2, pp. 145156.
[16] S. Asmussen, Applied Probability and Queues, second edition Springer,
Berlin, 2003.
[17] H.R. Gail, S.L. Hantler, and B.A.Taylor, Spectral analysis of M/G/1 and
G/M/1 type Markov chains, Advances in Applied Probability, Vol. 28, No.
1, 1996, pp. 114165.
[18] L. Berlemann, C. Hoymann, G.R. Hiertz, and S. Mangold, Coexistence and
interworking of IEEE 802.16 and IEEE 802.11(e), IEEE 63rd Vehicular
Technology Conference (VTC), 1:2731, 2006.
[19] C.R. Baugh, J. Huang, R. Schwartz, and D. Trinkwon, Trafc model for
802.16 tg3 mac/phy simulations, Technical report, IEEE 802.16 Broadband
Wireless Access Working Group, 2001.
8-18 SLEEP MODE WITH BIDIRECTIONAL TRAFFIC
cu = 0.01
cu = 0.3
cu = 1.0
0 0.2 0.4 0.6 0.8 1 1.2
0
0.2
0.4
0.6
0.8
1

u
E
n
e
r
g
y
c
o
n
s
u
m
p
t
i
o
n
Figure 8.7: Energy consumption versus
u
for different c
u
and
n = 20, = 0.5, K = 50, t = (1, 2, 4, 8, 16, 32, 64)
[20] E.-S. Jung and N.H. Vaidya, Improving IEEE 802.11 power saving mecha-
nism, Wireless Networks, Vol. 14 No. 3 pp. 375391, 2004.
9
Efcient performance analysis of
newly proposed sleep-mode
mechanisms for IEEE 802.16m in case
of correlated downlink trafc
Koen De Turck, Stijn De Vuyst, Dieter Fiems, Herwig Bruneel,
Sabine Wittevrongel
submitted to Performance Evaluation.
Abstract. There is a considerable interest nowadays in improving energy ef-
ciency of wireless telecommunications. The sleep-mode mechanism in WiMAX
(IEEE 802.16) and the discontinuous reception (DRX) mechanism of LTE are
prime examples of energy saving measures. Recently, Samsung proposed some
modications on the sleep-mode mechanism, scheduled to appear in the forth-
coming IEEE 802.16m standard, aimed at minimizing the signaling overhead. In
this work, we present a performance analysis of this proposal and clarify the dif-
ferences with the standard mechanism included in IEEE 802.16e. We also propose
some special algorithms aimed at reducing the computational complexity of the
analysis.
9-2 SLEEP MODE ANALYSIS FOR 802.16M
9.1 Introduction
The IEEE 802.16e standard (WiMAX) [1] has the potential to be the next big
standard in wireless communications. It regulates the communication between
mobile stations (MS) and base stations (BS) in a metropolitan area wireless net-
work. Energy-saving mechanisms in wireless communications are currently a hot
topic. Short battery life is one of the main impediments to a more widespread
use of wireless devices. Hence, understandably, a lot of research is directed at
solving or at least mitigating this problem. In the rst place, this can be done by
making more efcient batteries, but lately there is also a lot of interest in including
energy-saving measures in the communication protocols themselves. It is therefore
no wonder that the WiMAX committee has opted to incorporate various energy-
saving elements which are commonly referred to as sleep mode and idle mode.
Power saving is generally achieved by turning off parts of the MS in a controlled
manner when there is neither trafc from the MS (uplink trafc) nor to the MS
(downlink trafc). Whereas a MS in sleep mode is still registered to a BS and still
performs hand-off procedures, idle mode operation (which is optional in current
WiMAX standards) goes further and allows the MS to be completely switched off
and unregistered with any BS, while still receiving broadcast trafc.
The focus on energy efciency present in the WiMAX standard is shared by the
other major standardization effort LTE (Long Term Evolution). This is a successor
of UMTS standard (Universal Mobile Telecommunications System), adapted for
the heterogeneous demands
In this paper, we mainly consider the sleep-mode mechanism, in which the MS
turns itself off for predetermined periods of time which are negotiated with the BS.
For the upcoming IEEE 802.16m standard [2], two companies have proposed a
new sleep-mode mechanism, Samsung [3] and Nokia Siemens Networks [4]. The
difference between the customary WiMAX sleep-mode mechanism and the new
proposals is that the former must be renegotiated every time the trafc volume is
low, while in case of the latter mechanisms, trafc is organized in a structure of
cycles. When the system goes into sleep mode, it does so for the rest of the cycle.
This requires very few signaling messages. As the Nokia Siemens proposal is a
special case of the one championed by Samsung, we will mainly consider the latter.
To the best of our knowledge, no equivalent mechanism has yet been considered
for the LTE framework, although it seems as reasonable for this standard as it is in
the WiMAX case.
Sleep-mode operation has received quite some attention lately from the perfor-
mance modeling community. In [5], the average energy consumption of the MS
is obtained in case of downlink trafc only, as well as an approximate expression
for the mean packet delay. The energy consumption of the MS in case of both
downlink and uplink trafc is considered in [6]. Both [5] and [6] model the in-
CHAPTER 9 9-3
coming (and outgoing) trafc as a Poisson process. An accurate assessment of the
delay experienced at the BS buffer however, requires a queueing model. For IEEE
802.16e, in [7] the BS buffer is modeled as a continuous-time nite-capacity queue
with a Poisson arrival process and deterministic service times. A semi-Markov
chain analysis leads to expressions for the mean packet delay and the mean energy
consumption by the MS. The analysis in [8] is based on an M/G/1/K queueing
model with multiple vacations and exhaustive service, where the vacations repre-
sent the sleep intervals. Similar work can also be found in [9], where the length of
a vacation is assumed to depend on the previous vacation length. In [10], the sleep-
mode operation in Cellular Digital Packet Data (CDPD) services is evaluated. The
difference with IEEE 802.16e is that the subsequent sleep intervals do not increase
in length. The system can thus be modeled as a queueing system with multiple va-
cations and exceptional rst vacation. The loss probability in both [8] and [10] is
obtained as well. A simulation study of CDPD sleep-mode performance is found
in [11]. An alternative to the exponential increase of the sleep interval lengths is
evaluated by simulation in [12]. Finally, in our previous work on the sleep-mode
mechanism [13, 14], we considered a general D-BMAP arrival process, and we
found that trafc correlation, which was hitherto neglected in almost every study,
has an important inuence on the sleep-mode performance.
A preliminary version of this paper can be found in [15]
The structure of this paper is as follows. In Section 9.2, we describe Samsungs
proposal and design a mathematical model for it. Next, the analysis of the buffer
content and power consumption is spelled out in Section 9.3. We develop a com-
putationally efcient algorithm in Section 9.4, specically tailored to this model.
Next, we provide an analysis of the packet delay in Section 9.5. We employ our
analysis for a WiMAX scenario in Section 9.6 together with some numerical re-
sults. Finally, we draw some conclusions in Section 11.6.
9.2 Model
Let us rst sketch the sleep-mode mechanism that is included in the current IEEE
802.16e protocol. Whenever there is no data to be transmitted between the mobile
station (MS) and the base station (BS), the MS can negotiate to go into sleep mode,
that is, it will switch off its radio interface for some time, so that the BS cannot
transmit packets during this time. There are three types of sleep mode foreseen in
the protocol. The simplest is the so-called class III sleep mode, in which a single
sleep interval is negotiated. For class II sleep mode, the MS checks at the end of
each sleep interval if the BS has already packets to be transmitted to the said MS.
If not, then a new sleep interval of the same length is started automatically. The
last, and perhaps the most well-known is the class I sleep-mode mechanism, which
is very similar to class II, except that the subsequent sleep intervals do not stay the
9-4 SLEEP MODE ANALYSIS FOR 802.16M
same but double each time in length until a certain maximum is reached.
In principle, sleep-mode parameters must be renegotiated every time when
there is no data trafc between BS and MS (i.e. at the end of each busy period, in
queueing parlance). This results in a certain signaling overhead, and it is this short-
coming that the new 802.16m proposal addresses. To this end, data transmission
is organized in cycles, consisting of a number of frames. When at a certain point
during the cycle, the trafc between BS and MS is interrupted, then a so-called
closedown interval is started. If during this closedown interval no new trafc ap-
pears, the mobile station goes into sleep mode until the end of the current cycle.
If packets arrive during the closedown interval, then they are transmitted without
delay. The advantage of this scheme is that once this sleep mode is in order, it does
not have to be renegotiated.
We model the buffer in the base station (from now on denoted as simply the
buffer) as a discrete-time queue with innite capacity and a rst-come-rst-served
(FCFS) server discipline. This buffer is fed by the trafc in the downlink direc-
tion (i.e. from the base station to the mobile station). In this work, as in most
other papers on sleep-mode operation, we do not take into account the uplink traf-
c. We adhere to queueing terminology and denote by a slot, what is in the IEEE
community usually referred to as a frame. The durations of a cycle, of a listening
interval at the beginning of a cycle, and of a closedown interval are denoted by N,
L, and C respectively. As with all time-related quantities in our analysis, these pa-
rameters are expressed as an integer number of slots. We will further assume that
0 L C <N, which is conform to most practical situations. Packets arrive at the
buffer according to a discrete-time batch Markovian arrival process (D-BMAP).
That is, the arrival process is determined by a nite-state Markov chain (with M
states), which is usually referred to as the background Markov chain. The arrival
process is dened by the values a(k, j[i); k 0; i, j 1, , M, denoting the
probability that if the background state is i during a slot, there are k arrivals during
this slot and the background state during the next slot is j. We put these probabili-
ties in a matrix generating function A(z) with dimension MM, which is dened
as follows:
A(z)
.
=

k=0
A
k
z
k
, where [A
k
]
i j
.
= a(k, j[i). (9.1)
Delayed access is assumed, i.e. arriving packets do not enter the buffer until the
end of their arrival slot. Packets are transmitted one by one according to a FCFS
discipline, and every packet requires exactly one slot of transmission time. The
model at hand differs from a normal single-server queueing model when the buffer
gets empty, as in that case a closedown interval of duration C is started. When dur-
ing the closedown interval new packets arrive at the BS side, then the closedown
interval is abandoned and the packets are served during the next slot. Otherwise,
the system goes into sleep mode for the rest of the cycle, and newly arriving pack-
CHAPTER 9 9-5
ets accumulate in the buffer, waiting to be transmitted during the next cycle. If
the closedown interval gets interrupted by the beginning of the next cycle, it gets
abandoned as well. At the beginning of each cycle there is a listening interval of
L slots, which behaves identical to a closedown interval. An example of how the
system might evolve under the model as described above is depicted in Fig. 9.1.
In this gure closedown, listening and sleep slots are indicated by C, L and S re-
spectively. The light grey area shows the evolution of the queue content in the BS
buffer. Arrows indicate arrivals and departures of packets.
C C C C C
L L
C C C C
S S S S
BS
MS
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
queue
content
end of
cycle
L L L
C C C C
S S S S S
BS
MS
31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
C C C
L L L S S S S S S S S S S S S
BS
MS
61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90
C C C C C
L L
C C C C C C C
S
BS
MS
91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120
Figure 9.1: Evolution of the system according to our model for N = 15,C = 4 and L = 3.
Closedown, listening and sleep slots are indicated by C, L and S respectively.
The difference between the proposal of Samsung [3] and the one of Nokia
Siemens Networks [4] is that the latter does not have a closedown mechanism, and
9-6 SLEEP MODE ANALYSIS FOR 802.16M
hence corresponds to the special case C = 0.
9.3 Analysis of Buffer Content and Power Consump-
tion
In this section, we determine the probability generating function U(z) of the buffer
content during an arbitrary slot. The rst step in the analysis is the choice of a con-
venient set of embedded points and the corresponding Markovian state description.
After some trial and error, we settled for the following embedded points: either the
beginning of the rst slot of a cycle, or the end of a departure slot (excluding the
last slot of a cycle), that is, a slot in which a packet is transmitted. Further in our
analysis, we will tag these embedded points by the special mark A. As there is
a fairly large number of possible transitions, we will walk through them, one by
one. The Markovian state description consists of the buffer content u during the
slot, the background state s during the slot and the position n within the cycle at
the beginning of a slot, where we denote the beginning of the cycle with 0, and
the end of the different slots with 1, 2, , N1. In the next paragraphs, we detail
all possible transitions from system state (u, n, s) to (u

, n

, s

), along with the cor-


responding transition probabilities. These steps take a certain amount of care, as
off-by-one errors are easily committed.
The rst case concerns the functioning of the system in active mode : when
there are packets available to be sent (u > 0), then one of them is sent and the
embedded Markov chain jumps to the next cycle position (that is, we add one
modulo N)
From To Probability
(u, n, s) (u+k 1, n+1 mod N, s

) [A
k
]
ss
if u > 0
For the second case, we look at the closedown intervals, more specically at
closedown intervals that are interrupted because of a packet arrival. For this to
happen, we must start from an empty buffer, then there must be a number of slots
i in which no arrivals occur, followed by a slot during which there is at least one
arrival. The next embedded instant is at the end of the slot thereafter. The notation
[z
k
]F(z) denotes the coefcient of power series F(z) in z
k
.
From To Probability
(0, n, s) (k, n+i +2, s

) [z
k+1
][A(0)
i
(A(z) A(0))A(z)]
ss

if 0 < n < Ni 2, 0 i <C


Now we turn our attention to situations where the system enters sleep mode.
In that case, we again start from an empty buffer (u = 0), followed by a period of
C slots in which no arrivals occur. Then we have a sleep interval of NnC slots
in which arrivals accumulate. The next embedded instant is at the cycle boundary.
CHAPTER 9 9-7
From To Probability
(0, n, s) (k, 0, s

) [z
k
][A(0)
C
A(z)
NnC
]
ss
if 0 < n NC
If the cycle boundary is reached during a closedown interval, then it is not
continued, but rather a listening interval is started. Hence the embedded process
just jumps to the cycle boundary.
From To Probability
(0, n, s) (0, 0, s

) [A(0)
Nn
]
ss
if NC < n < N
The last two cases concern the listening interval at the beginning of a cycle.
As already noted in the previous section, a listening interval functions in exactly
the same way as a closedown interval, except that it occurs only at the beginning
of a cycle. Due to that reason, we need not consider the interruption of a listening
interval by a cycle boundary. This leads to the last two cases:
From To Probability
(0, 0, s) (k, 0, s

) [z
k
][A(0)
L
A(z)
NL
]
ss

(0, 0, s) (k, i +2, s

) [z
k+1
][A(0)
i
(A(z) A(0))A(z)]
ss
if 0 i < L
Note that the buffer content u decreases at most by one, i.e. the Markov chain
is skip-free to the left in u, which means that the Markov chain can be written as
an M/G/1 process with level set u : u 0 and phase set (n, s) : 0 n < N, 1
s M. We give a short account of the solution method in the next section, along
with some improvements that are tailored to this specic model.
We now derive relations between the buffer content distribution at embedded
points and the distributions during closedown, sleep and listening intervals, in or-
der to eventually establish the distribution at an arbitrary time instant. Essentially,
we must ll in the gaps that can occur between two embedded time instants. Let
us rst give some denitions. The (row) vector generating function U
A
n
(z) records
the buffer content distribution at an embedded time instant with position n for each
background state. Specically, the ith entry of this vector is the partial pgf
[U
A
n
(z)]
i
=

k=0
z
k
P[u = k, s = i, n, A]. (9.2)
where P[u = k, s = i, n, A] denotes the probability of having a buffer content k and
a state i during a slot and a position n within the cycle, as well as an embedded
instant (marked with tag A) at the beginning of the slot. As embedded time instants
are just one type of slot boundaries, this means that
n
U
A
n
(z) is a partial vector
generating function, i.e.
n
U
A
n
(1)1 < 1 where 1 denotes a column vector with all
entries equal to 1.
Likewise, let U
S( j)
n
(z), U
C( j)
n
(z) and U
L( j)
n
(z) denote the vector generating func-
tions corresponding to the end of the jth slot of a sleep, closedown and listening
interval respectively. We nd the following equations:
U
L( j)
j
(z) = U
A
0
(0)A(0)
j1
A(z) for 1 j L, (9.3)
9-8 SLEEP MODE ANALYSIS FOR 802.16M
U
C( j)
n
(z) = U
A
nj
(0)A(0)
j1
A(z) for 1 j C, j < n < N, (9.4)
U
S( j)
n
(z) =
_
U
A
0
(0)A(0)
L
A(z)
j
, for n = j +L, 0 < j < NL,
U
A
njC
(0)A(0)
C
A(z)
j
for j +C < n < N, 0 < j < NC1.
(9.5)
Note that equation (9.3) is based on the fact that the system will reach the
end of the jth slot of a listening interval if the BS buffer is empty at the start of
the cycle and no packet arrivals occur during the rst j 1 slots of the cycle; the
buffer content at the end of the jth slot of the listening interval then equals the
number of arrivals during this slot. Similarly, equation (9.4) expresses that starting
from an empty buffer at an embedded time instant with position n j, the initiated
closedown interval will consist of at least j slots if no packets arrive during the rst
j 1 closedown slots and the buffer content at the end of the jth closedown slot
corresponds to the packet arrivals in this jth slot. Finally, equation (9.5) follows
from the observation that the system enters sleep mode if the buffer is empty at the
end of a listening or closedown interval; the buffer content at the end of the jth
sleep slot equals the total number of arrivals during the j sleep slots (with matrix
generating function A(z)
j
).
By taking the summation of the generating functions over all types of slots,
we nd the following expression for the generating function U(z) of the buffer
content:
U(z) =
N1

n=0
U
A
n
(z)1+U
A
0
(0)
_
X
(L)
(0) +A(0)
L
X
(NL1)
(z)
_
1
+
NC1

j=1
U
A
j
(0)
_
X
(C)
(0) +A(0)
C
X
(NCj)
(z)
_
1
+
N1

j=NC
U
A
j
(0)X
(Nj)
(0)1, (9.6)
where the matrix generating function X
(k)
(z) is dened as
X
(k)
(z) =
k

i=1
A(z)
i
= A(z)(I A(z)
k
)(I A(z))
1
. (9.7)
Apart from the buffer behavior, the other quantity that is important for a proper
evaluation of the sleep-mode mechanism is the expected power consumption E
over a slot. Let external parameters E
A
, E
L
, E
C
and E
S
denote the power consump-
tion during an active, listening, closedown and sleeping slot respectively, then we
CHAPTER 9 9-9
nd that
E =E
A
n
_
U
A
n
(1)1U
A
n
(0)1
_
+U
A
0
(0)
_
E
L
X
(L)
(0) +E
S
A
(L)
(0)X
(NL1)
(1)
_
1
+
NC1

j=1
U
A
j
(0)
_
E
C
X
(C)
(0) +E
S
A
(C)
(0)X
(NCj)
(1)
_
1
+E
C
N1

j=NC
U
A
j
(0)X
(Nj)
(0)1. (9.8)
9.4 Computational method
In the previous section, we showed that, by choosing the set of embedded points
carefully, we can reduce the problemto the study of an M/G/1-type Markov chain,
which has a transition matrix with the following block structure:
P =
_
_
_
_
_
_
_
B
0
B
1
B
2

A
0
A
1
A
2

A
0
A
1
A
2

A
0
A
1
A
2

.
.
.
.
.
.
.
.
.
_
_
_
_
_
_
_
. (9.9)
The block matrices A
k
and B
k
of order MN can be found using the tables in the
previous section. Numerical solutions of this type of Markov chain revolve since
the pioneering work of Neuts [16] around the computation of the fundamental
matrix G, which is the smallest nonnegative solution of the matrix equation
G =

k=0
A
k
G
k
(9.10)
The stationary probability vector = [
0
,
1
, ] that satises = P, can
subsequently be found by the equations

0
=
0
B

0
and
n
=
_

0
B

n
+
n1

i=1

i
A

ni+1
_
(I A

1
)
1
, for n > 0, (9.11)
where
A

n
=

i=n
A
i
G
in
, B

n
=

i=n
B
i
G
in
. (9.12)
Expression (9.11) is known as Ramaswamis formula, after another pioneer of
the matrix-analytic method. There exist a lot of algorithms that aid in computing
the desired quantities of interest. Even a straightforward translation of formula
(9.10) constitutes a solid and numerically stable computational method, especially
9-10 SLEEP MODE ANALYSIS FOR 802.16M
when the Markov chain possesses a strong negative drift. There exist more sophis-
ticated methods as well, for an overview see e.g. [17]. For all these methods, the
computational complexity increases very fast with respect to the size of the phase
set (here equal to MN). Indeed, as we discern a lot of matrix-matrix multiplica-
tions of size MN in the above formulas, the complexity will most often be cubic
in MN, that is O(M
3
N
3
). As the cycle length N can be substantial, this would
seem bad news for the scalability of our approach. However, the structure that is
present in the block matrices A
k
allows us to formulate variants of the customary
algorithms that are signicantly more efcient. Indeed, block matrices A
k
have
the following form:
A
k
=
_
_
_
_
_
_
_
_
_
0 A
k
0 0
0 0 A
k
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0
0
.
.
.
.
.
.
0 A
k
A
k
0 0 0
_
_
_
_
_
_
_
_
_
= C
N
A
k
, (9.13)
where denotes the Kronecker product and C
N
is the cyclic permutation matrix
that maps the vector (v
1
, v
2
, , v
N1
, v
N
) to (v
N
, v
1
, v
2
, , v
N1
).
Hence the matrix A
k
is block circulant, which means that there exists a block
row vector a such that [A
k
]
i j
= [a]
ij mod N
. In this case, the block vector a is
equal to [0, A
k
, 0, , 0]. Apart from being efciently storable (only the rst block
row or block column must be stored), block circulant matrices are closely related
to the discrete Fourier transform, and that fact speeds up the basic operations con-
siderably.
Consider two block circulant matrices M
1
and M
2
with rst block rows equal
to resp. m
1
and m
2
. The product M
1
M
2
is again block circulant [20] with rst
block row equal to IDFT(DFT(m
1
) DFT(m
2
)), where DFT(.) and IDFT(.) de-
note the discrete Fourier transform and the inverse discrete Fourier transform re-
spectively, and the operation denotes the component-wise product of the blocks:
let c = [C
1
, , C
N
] and d = [D
1
, , D
N
] denote two block vectors of length N,
with MM block sizes, then c d is equal to
c d = [C
1
D
1
, , C
N
D
N
]. (9.14)
In order to derive an efcient procedure to compute the fundamental matrix G,
we rst note that because the matrices A
k
are block circulant, G is block circulant
as well. This can be intuited from the stochastic interpretation of the matrix G.
Entry [G]
i j
denotes the probability that, starting from level n in phase i, the process
returns to level n1 in nite time, and that j is the rst phase visited in level n1.
This probability does not depend on the cycle positions at the beginning and end
CHAPTER 9 9-11
of such excursions, but merely on the difference (modulo N) between the cycle
positions.
All this means we can perform the iterative computation of G entirely in the
DFT domain. Indeed, let g
.
= [

G
i
]
i=1, ,N
denote the DFT of the rst block row of
G, and let a
k
denote the DFT of [0, A
k
, 0, , 0], then we nd from relation (9.10):
g =

k=0
a
k

k times
..
g g. (9.15)
Note that we can perform the iteration on each element of the block vector sepa-
rately. Thus we must solve N nonlinear matrix equations of the form:

G
i
=

k=0

i
A
k

G
k
i
, 0 i < N. (9.16)
where = exp(2i/N) denotes the Nth root of unity. Note that we can solve these
equations with a variety of techniques ranging from the traditional functional iter-
ation over subspace methods to cyclic reduction. The former are simpler to imple-
ment, but cyclic reduction provides better convergence when the negative drift of
the queueing system is close to zero. Although the sequence

A
k
.
=
i
A
k

k=0, ,
does not consist of non-negative matrices (unless i = 0), many algorithms do not
require this fact [17]. Rather, the matrix-valued power series
k
z
k

A
k
must belong
to a Wiener algebra, by which is meant that
k
[

A
k
[ < . As
k
[

A
k
[ =
k
[A
k
[,
this condition merely asserts that the DBMAP arrival process characterized by A
k
is well-behaved.
When we take the functional-iteration algorithm as an example, then the above
transformation of the system leads to a reduction from KN
3
M
3
to KNM
3
oating
point operations per iteration, where K denotes the index at which the matrix power
series is cut off.
We have found no earlier references to this variant of the standard M/G/1 tech-
nique. It may have applications beyond the ones in this paper, more precisely
whenever the operation of the system is organized in both small time-units called
slots and larger deterministic time-units, called cycles in IEEE, but frequently
called frames in other literature.
We tie the loose ends together and show the consecutive steps one has to take
to determine the unknown vector U
A
(0) that appears in Eqs. (9.6) and (9.8).
1. Compute

G
i
by formula (9.16).
2. Compute the IDFT of

G
i
and construct the matrix G.
3. Compute
0
by means of formula (9.11). This vector is equal to U
A
(0) up
to a normalizing constant.
9-12 SLEEP MODE ANALYSIS FOR 802.16M
4. We nd this normalizing constant by substituting z =1 in Eq. (9.6) and using
U(1) = 1.
Note. An alternative derivation of the fast recursive scheme (9.16) can be made
by observing that what we really want to know is the joint distribution of the length
of a sub-busy period and the background state at the ned of the sub-busy period
given the initial state at the sub-busy period start. Remember that the concept of
a sub-busy period is used to denote the rst passage time to level k 1, given that
the process starts at level k.
This joint distribution can be encoded in the matrix generating function G(z)
with dimension MM, dened as follows:
[G(z)]
i j
=

k=1
z
k
P[S
k
= j, = k[S
0
= i]. (9.17)
where denotes the rst return time to a lower level and where S
0
and S
k
denote the
D-BMAP background state at the start and end of this sub-busy period respectively.
An argument by induction shows that G(z) satises the following expression:
G(z) = z

k=0
A
k
G(z)
k
. (9.18)
We do not in fact need all information contained in G(z) but only the distribu-
tion of the sub-busy period lengths modulo N. These can be found by evaluating
G(z) at the Nth roots of unity and by taking the inverse Fourier transform.
9.5 Analysis of Packet Delay
We can also derive the distribution of the packet delay, i.e. the time that an arbi-
trary packet resides in the system. The rst step in our analysis is to derive the
distribution of the unnished work at the beginning of an arbitrary slot. In the
considered model, service times equal one slot, and thus during busy periods, the
unnished work is exactly equal to the buffer content. During sleep periods, we
take the convention as in [13] that the unnished work is the number of slots that
the system still must reside in sleep mode. The idea is that at the beginning of a
sleep period of length l, the system receives l slots worth of sleep work. Like-
wise, one unit of work is added at the beginning of a slot in closedown mode. In
this way, one unit of work is performed regardless of the state the system is in,
thus making the analysis more uniform. With these conventions the unnished
work represents the time until the system is ready to serve new packets.
After some manipulations, we nd that the vector generating function W(z) of
CHAPTER 9 9-13
the unnished work at the beginning of an arbitrary slot is equal to:
W(z) =
N1

n=0
U
A
n
(z) +
N1

n=0
N1

j=0
z
Nn
U
S( j)
n
(z)
+
L

j=1
zU
L( j)
j
(z) +
N1

n=1
C

j=1
zU
C( j)
n
(z). (9.19)
The delay of an arbitrary packet is equal to the sum of the unnished work at
the beginning of an arbitrary arrival slot minus 1, plus the service times of the
packets that arrive in the arrival slot of the tagged packet but no later than the
packet (that is, including the tagged packet itself). Note that the distribution of the
number of packets during the arrival slot of an arbitrary packet is different from
the distribution for a random slot, for example some background states may have
higher arrival rates, thus skewing the original distribution. We nd in fact the same
relation between the distributions of the unnished work and the packet delay as
in [13]:
D(z) =
1
(1z)
1
W(z)(I A(z))1. (9.20)
where is the average number of arrivals during a random slot.
9.6 Practical Example
In this section, we show some examples of what is possible with the model de-
veloped in the present papers. In all examples, we will use an interrupted Poisson
arrival process, conform to guidelines of the IEEE community [22] for the 802.11
standard. This is a two-state (on/off) Markov model with Poissonian arrivals with
rate
ON
in the on-state and no arrivals in the off-state. The Markov chain spends
a fraction of the time in the on-state. The coefcient of correlation between
the states in two subsequent slots is 1
1
K
, where K is a measure for the mean
lengths of both on and off periods, given by
K
1
and
K

respectively. It can easily


be veried that the matrix generating function A(z) for this trafc model equals
A(z) =
_
e

ON
(z1)
0
0 1
__
1
1
K
1
K

K
1

K
_
. (9.21)
The case of uncorrelated states from slot to slot clearly corresponds to K = 1.
Larger values of K imply that on-slots and off-slots occur more clustered together
in time and the arrival process is more bursty. The correlation parameter K is
therefore also referred to as the burstiness factor or the burst-length factor of the
arrival process [25], [26]. As transmission times have a xed length of 1 slot, the
load is equal to =
ON
.
For the power consumption during the different types of slots we use the values
summarized in Table 9.1. As there are as yet no consumer-end WiMAX devices,
9-14 SLEEP MODE ANALYSIS FOR 802.16M
we have used the corresponding WiFi data [24] as a guideline. The performance
measures we will look at are the mean packet delay and the mean power consump-
tion.
Parameter Value Description
E
S
0.045 W Sleep mode power
E
L
1.15 W Listening mode power
E
C
1.15 W Power during closedown
E
A
1.65 W Transmit mode power
Table 9.1: Power consumption parameters
In the rst pair of plots, Fig. 9.2, we look at the inuence of the trafc cor-
relation on the mean packet delay and mean power consumption. We see a con-
rmation of the well-known fact that trafc correlation deteriorates the queueing
performance tremendously. Correlation has a small but marked improvement on
the average power consumption. Also note that the delay gets larger for very light
loads, which is intuitively clear from the fact that in that case the system goes into
sleep mode for longer time periods, and hence incoming packets must wait for a
longer time.
The next pair of plots, Fig. 9.3 shows the inuence of the length C of the
closedown interval on the performance measures, for loads =0.2 and =0.4. As
can be expected, larger values of C signify an increase of the power consumption
and a decrease of the mean delay.
We see the same story, but to a larger degree, in the last two plots, Fig. 9.4,
where we show the inuence of the cycle length N on the performance.
We conclude with an illustration of the variance of the packet delay contrasted
with the mean packet delay, for different values of K.
9.7 Conclusion
We analyzed Samsungs proposal of a new sleep-mode mechanism to be included
in the next version of the IEEE 802.16m WiMAX protocol, by means of a queue-
ing model. From the numerical results, we see that the proposal shows the same
qualitative behavior as the customary (802.16e) sleep-mode mechanism. That is,
we observe that the performance of the sleep-mode mechanism, both in terms of
packet delay and energy consumption, is severely inuenced by correlation present
in the arrival process. Properly taking into account such correlation structures is
therefore crucial in order to achieve accurate performance estimates.
Also, calculations are expedited signicantly by exploiting the circulant matrix
of the blocks in the transition probability matrix explicitly.
CHAPTER 9 9-15
K = 40
K = 10
K = 1
0 0.2 0.4 0.6 0.8 1
0
10
20
30
40

E
[
D
]
(a) Mean packet delay
K = 1
K = 10
K = 40
0 0.2 0.4 0.6 0.8 1
0
500
1000
1500

E
[
E
]
i
n
m
W
(b) Energy consumption
Figure 9.2: Mean packet delay versus the load for different levels of correlation, where
N = 32, L = 1,C = 8, and = 0.2. The load is varied from 0 to 1 by increasing
ON
from 0 to its maximal value
1
= 5.
Acknowledgements
The third author is a Postdoctoral Fellow with the Research FoundationFlanders
(FWOVlaanderen), Belgium.
References
[1] IEEE 802.16e-2005, Part 16: Air interface for xed and mobile broadband
wireless access systems Amendment 2: physical and medium access con-
trol layers for combined xed and mobile operation in licensed bands
Corrigendum 1, February 2006.
[2] The Draft IEEE 802.16m System Description Document, IEEE 802.16m-
08/003r4, July 2008.
9-16 SLEEP MODE ANALYSIS FOR 802.16M
= 0.2
= 0.4
2 4 6 8 10 12 14 16 18
15
20
25
30
35
C
E
[
D
]
(a) Mean packet delay
= 0.2
= 0.4
2 4 6 8 10 12 14 16 18
300
400
500
600
700
800
900
C
E
[
E
]
i
n
m
W
(b) Energy consumption
Figure 9.3: Mean packet delay and the energy consumption versus the closedown time for
different values of the load, where N = 64, L = 1, and = 0.2.
K = 1
K = 50
20 40 60 80 100 120
10
20
30
40
N
E
[
D
]
(a) Mean packet delay
K = 1
K = 50
20 40 60 80 100 120
750
800
850
900
N
E
[
E
]
i
n
m
W
(b) Energy consumption
Figure 9.4: Mean packet delay and the energy consumption versus the cycle length for
different levels of correlation, where L = 1,C = 8, = 0.2, = 0.2.
[3] Sleep mode operation for IEEE802.16m, C802.16m-08/721r1, July 2008.
[4] Keep-awake mechanism for 802.16m sleep mode, C802.16m-08/718, July
2008.
[5] Y. Xiao, Energy saving mechanism in the IEEE 802.16e wireless MAN,
IEEE Communications Letters, Vol. 9, No. 7, 2005, pp. 595597.
[6] Y. Zhang and M. Fujise, Energy management in the IEEE 802.16e MAC,
IEEE Communications Letters, Vol. 10, No. 4, 2006, pp. 311313.
[7] K. Han and S. Choi, Performance analysis of sleep mode operation in
IEEE 802.16e mobile broadband wireless access systems, Proceedings of
CHAPTER 9 9-17
0 0.2 0.4 0.6 0.8 1
0
10
20
30
40

E
[
D
]
(a) Mean packet delay
0 0.2 0.4 0.6 0.8 1
0
100
200
300
400

V
a
r
[
D
]
(b) Variance of packet delay
Figure 9.5: Mean packet delay and variance of packet delay versus the load for different
levels of correlation, where L = 1,C = 8, = 0.2, = 0.2.
the IEEE 63rd Vehicular Technology Conference, VTC2006-Spring (Mel-
bourne, 710 May 2006), Vol. 3, pp. 11411145.
[8] Y. Park and G.U. Hwang, Performance modelling and analysis of the sleep-
mode in IEEE 802.16e WMAN, Proceedings of the IEEE 65th Vehicular
Technology Conference, VTC2007-Spring (Dublin, 2225 April 2007), pp.
28012806.
[9] J.-B. Seo, S.-Q. Lee, N.-H. Park, H.-W. Lee, and C.-H. Cho, Performance
analysis of sleep mode operation in IEEE 802.16e, Proceedings of the
60th Vehicular Technology Conference, VTC2004-Fall (Los Angeles, 2629
September 2004), Vol. 2, pp. 11691173.
[10] S.-J. Kwon, Y.W. Chung, and D.K. Sung, Queueing model of sleep-mode
operation in cellular digital packet data, IEEE Transactions on Vehicular
Technology, Vol. 52, No. 4, 2003, pp. 11581162.
9-18 SLEEP MODE ANALYSIS FOR 802.16M
[11] Y.-B. Lin and Y.-M. Chuang, Modeling the sleep mode for cellular digital
packet data, IEEE Communications Letters, Vol. 3, No. 3, 1999, pp. 6365.
[12] N.-H. Lee and S. Bahk, MAC sleep mode control considering downlink traf-
c pattern and mobility, Proceedings of the IEEE 61st Vehicular Technology
Conference, VTC2005-Spring (Stockholm, 30 May1 June 2005), Vol. 3, pp.
20762080.
[13] K. De Turck, S. De Vuyst, D. Fiems, and S. Wittevrongel, Performance
analysis of the IEEE 802.16e sleep mode for correlated downlink trafc,
Telecommunication Systems, Vol. 39, No 2, pp. 145156.
[14] S. De Vuyst, K. De Turck, D. Fiems, S. Wittevrongel, H. Bruneel, Delay
versus energy consumption of the IEEE 802.16e sleep-mode mechanism,
IEEE Transactions on Wireless Communications, Vol. 8, No. 11, pp. 5383
5387 (2009).
[15] K. De Turck, S. De Vuyst, D. Fiems, S. Wittevrongel, H. Bruneel, Perfor-
mance of the sleep-mode mechanism of the new IEEE 802.16m proposal for
correlated downlink trafc. Proceedings of Netcoop 2009, Lecture Notes in
Computer Science, pages 152-165, Eindhoven, The Netherlands, 2009.
[16] M.F. Neuts, Structured stochastic matrices of M/G/1 type and their appli-
cations, volume 5 of Probability: Pure and Applied. Marcel Dekker Inc.,
New York, 1989.
[17] D.A. Bini, G. Latouche, B. Meini, Numerical methods for structured
Markov chains, Oxford University Press, 2005.
[18] H. R. Gail, S. L. Hantler, B. A. Taylor. Non-skip-free M/G/1 and G/M/1
type Markov chains Advances in Applied Probability, Vol. 29, No. 3 (Sep.
1997), pp. 733758.
[19] B. Meini. An improved FFT-based version of Ramaswamis formula.
Stochastic Models, Vol. 13, No. 2 (1997) pp. 223238.
[20] C. Van Loan. Computational frameworks for the fast Fourier transform,
volume 10 of Frontiers in Applied Mathematics. Society for Industrial and
Applied Mathematics (SIAM), Philadelphia, PA, 1992.
[21] L. Berlemann, C. Hoymann, G.R. Hiertz, and S. Mangold, Coexistence and
interworking of IEEE 802.16 and IEEE 802.11(e), IEEE 63rd Vehicular
Technology Conference (VTC), Vol. 1, pp. 2731, 2006.
[22] C.R. Baugh, J. Huang, R. Schwartz, and D. Trinkwon, Trafc model for
802.16 tg3 mac/phy simulations, Technical report, IEEE 802.16 Broadband
Wireless Access Working Group, 2001.
CHAPTER 9 9-19
[23] D. Sivchenko, B. Xu, V. Rakocevic and J. Habermann, Internet trafc per-
formance in IEEE 802.16 networks, European Wireless, European Wireless
2006, Athens, Greece, 2-5 April 2006.
[24] E.-S. Jung and N.H. Vaidya, Improving IEEE 802.11 power saving mecha-
nism, Wireless Networks, Vol. 14 No. 3 pp. 375391, 2004.
[25] S. Wittevrongel, H. Bruneel, Discrete-time queues with correlated arrivals
and constant service times, Computers & Operations Research, Vol. 26, No.
2, 1999, pp. 93108.
[26] S. De Vuyst, S. Wittevrongel, H. Bruneel, Mean value and tail distribution
of the message delay in statistical multiplexers with correlated train arrivals,
Performance Evaluation, Vol. 48, No. 14, 2002, pp. 103129.
10
Performance of sleep-mode
mechanisms under light-trafc
conditions
Koen De Turck, Dieter Fiems, Sabine Wittevrongel, Herwig Bruneel
presented at BWWQT 2009
Modern standards for wireless telecommunication, such as IEEE 802.16e (WiMAX),
foresee in several power-saving mechanisms. One of the main mechanisms is
sleep-mode operation, which allows the mobile device to switch off the antenna
during a negotiated time (it goes to sleep), thus reducing energy consumption.
However, trafc that arrives at the base station (BS) incurs an extra delay, because
its delivery can only start after the sleep period has ended. The sleep-mode mech-
anism in WiMAX has provisions to vary the lengths of subsequent sleep periods,
which allows to exploit correlation in the arriving trafc. We depart from our ear-
lier analyses and study the model of the sleep-mode mechanism specically under
light-trafc conditions. This results in relatively short and simple formulas that
give a lot of insight.
10-2 SLEEP MODE UNDER LIGHT TRAFFIC
10.1 Introduction
We study the performance of the energy saving mechanism dened in the IEEE
802.16e [1] standard for Broadband Wireless Access networks. Consider the
downstream communication between a Mobile Station (MS) and its serving Base
Station (BS) over a wireless link. To allow the MS to extend its battery life, the
standard provides a sleep-mode mechanism that allows the MS to turn off its radio
interface for a certain time whenever the BS has no packets in its queue. Specif-
ically, if the BSs queue is empty, the MS starts a sleep period during which it
remains powered down and thus cannot be reached by the BS. After this period,
the MS is briey reactivated to check whether there are packets waiting for it at
the BS. If not, the MS initiates a second sleep period, a third, and so on. However,
if any packets arrived at the BS during the last sleep period, the MS remains pow-
ered and enters awake mode. This allows the BS to transmit all the packets in its
buffer exhaustively, after which the whole procedure is repeated. In case of IEEE
802.16e, an exponential update strategy is used, where the sleep periods double in
length every time until a certain maximum is reached. However, in this paper, we
allow more general sequences of sleep period lengths.
Quite a few authors have in recent times investigated the performance of the
IEEE sleep-mode mechanism. We refer to Chapter 6 for an overview.
The structure of the paper is as follows. We elaborate on the model in sec-
tion 10.2. The light-trafc technique is explained in section 10.3. We give the
main results in section 10.4. Finally, we numerically verify the exact results we
obtained in a previous chapter 6 with the light-trafc results we derived here in
section 10.5.
10.2 Model and Problem Statement
We model the BS as a discrete-time single-server queue with innite capacity and a
rst-come-rst-served discipline. Time is divided into xed-length intervals called
slots such that changes in the system can occur at slot boundaries only. The service
(i.e. transmission) times of the packets, expressed as a number of slots, are inde-
pendent and have common probability generating function (pgf) S(z) and mean
E[S]. Packets arrive in the queue according to a D-BMAP with M phases. This
arrival process is characterised by the values a(k, j[i), (k 0, i, j 1, , M),
denoting the probability that, given phase i in a slot, there are k packet arrivals
and the phase switches to j in the next slot. These probabilities can conveniently
be arranged in the MM probability generating matrix (pgm) A(z) with entries
[A(z)]
i j
=

k=0
a(k, j[i)z
k
. The stationary probabilities of the phases are given by
the entries of a row vector a that satises a = aA(1) and a1 = 1, where 1 is a
column vector of ones of appropriate length. The mean number of arrivals per slot
CHAPTER 10 10-3
is given by = aA

(1)1.
The sleep-period update strategy is denoted by a sequence (t
1
, t
2
, . . .) where the
integers t
n
(n 1) are the lengths of the subsequent sleep periods of the same idle
period. Let us also dene
n
=
n
i=1
t
i
(n 1) and
0
= 0 indicating the starting
slots of the sleep periods relative to the start of the idle period. We assume that the
lengths of the sleep periods become equal above a certain value J, i.e. t
n
= t
J
, for
n J.
We cite here the results we obtained in [2] regarding the virtual waiting time
and the energy consumption. We also derived the distributions of the queue con-
tent and the packet delay, but in this paper we focus on the two aforementioned
quantities. The vector generating function W(z) is a row vector of M generating
functions with entries dened as follows:
[W(z)]
i
=

k=0
P[w = k, b = i]z
k
, (10.1)
where b and w are random variables denoting the phase and the virtual waiting
time at the beginning of a random slot in stationary regime. The virtual waiting
time at a time instant is dened as the time until the server is ready to serve a
virtual packet that arrives exactly at that time instant. We found in [2] that
W(z) = w

n=1
A(0)

n1
(z
t
n
1)(zI A(S(z)))
1
, (10.2)
where w is the 1M boundary vector that must in general be determined analyti-
cally.
We dene the energy consumption E of the base station as the probability that
during a random slot the antenna is switched on. Note that the antenna is switched
on during every busy slot and also during every last slot of a sleep period. We
found that
E = +w

n=0
A(0)

n
1. (10.3)
Although the above sketched solution technique is very powerful, there are a
couple of drawbacks to it. Firstly, computing the vector w is a numerically com-
plex process, especially if the number of background states M is large. Secondly,
the exact inuence of arrival correlation is not evident from the equations. We try
to circumvent these problems by considering a Taylor expansion in the region that
for this application is most interesting from a practical point of view, namely
where the trafc is low.
10-4 SLEEP MODE UNDER LIGHT TRAFFIC
10.3 Light-trafc technique
Although perhaps less well-known than their heavy-trafc counterparts, light-trafc
approximations have a long history, and very different approaches have been de-
veloped to tackle a wide variety of systems. For an excellent overview, see e.g. [4].
It is very important how light-trafc conditions are reached: different ways
of reaching low trafc can give different results. Here we take the so-called -
thinning approach: we start from a normal arrival process, and then consider
the family of related processes where each arrival is retained with a probability ,
independent from anything else. Light-trafc conditions are then reached in the
limit 0.
We want to obtain the mean virtual waiting time and the mean power con-
sumption under light trafc. Generally, these quantities are studied by relating the
distributions at different time instants and then, by invoking stationarity, we get an
equation for the desired quantity. Under light trafc, we can take the more direct
approach of averaging over all sample paths.
For this technique to work, we assume that the operation of the system has
started at time t =, and hence, assuming that the system is stable, has reached
stationarity in slot 0. Let S be the set of sample paths. A sample path s for this
model records the arrival times and the service durations of each packet, resulting
in two sequences i
1
, i
2
, and j
1
, j
2
, , as well as a variable g signifying the
number of slots until the service of the very rst packet starts. Note that g is
uniformly distributed between 0 and t
J
1. Let be the quantity of interest,
for example the virtual waiting time W at the beginning of slot 0 or the energy
consumption E during slot 0. Then we can compute the expectation of this quantity
under -thinning as follows:
E

[] =

(ds) f (s), (10.4)


where

(.) is the probability measure for sample paths under -thinning, and
function f (.) computes the quantity for given sample path s. As the variable
g and the service times are independent of the arrival process, and moreover not
subject to -thinning, average over them separately. From henceforth we only
include the arrival times in the sample path.
Such probabilities are functions of , of which we can make a Taylor expan-
sion. Under light trafc, we consider only coefcients in, say, the n lowest powers
of . As we will see, this leads to an enormous reduction in the amount of sample
paths we need to consider. Note that the probability matrix generating function of
the arrival process under -thinning is given by:
A

(z) = A(1+ z), (10.5)


CHAPTER 10 10-5
which gives rise to the following Taylor-expansion in = 0:
A

(z) =

n=0
1
n!
A
(n)
(1)
n
(z 1)
n
=

k=0
z
k

n=k
(1)
nk
_
n
k
_
1
n!
A
(n)
(1)
n
=

k=0
z
k
C
k
(). (10.6)
Note that the probability of a sample path can written as an innite product of
transition matrices C
k
(), premultiplied by vector a and postmultiplied by vector
1, where each k signies the number of arrivals during a certain slot. From this we
observe that the probability of a sample path with n arrivals is a power series in
where the lowest non-zero coefcient is in
n
.
The following theorem shows the formula for a second order light-trafc ap-
proximation of a quantity . The proof follows the lines of the expositions in [3]
and [5], applied
Theorem 10.3.1. The expectation of a quantity with associated function f (.)
under -thinning has the following second-order approximation:
E

[] = f () +
+

i=
( f (i) f ()) (10.7)
+
2
+

i
1
,i
2
=
R([i
1
i
2
[)( f (i
1
, i
2
) f (i
1
) f (i
2
) + f () +O(
3
),
(10.8)
where i
1
, i
2
, denotes a sample path with arrivals in slots i
1
, i
2
and so on. The
function R(.) is dened as
R(0) =
1
2
aA

(1)1,
R(n) =
1
2
aA

(1)A(1)
n1
A

(1)1, where n 1. (10.9)


We can view R(.) as a sort of autocorrelation function.
10.4 Main results
The functions f
E
(s) and f
W
(s) that determine respectively the energy consump-
tion and the unnished work during slot 0 for a given sample path s follow quite
straightforwardly from the system equations. For two arrivals, the complete def-
initions are already quite lengthy, and hence we omit them in favor of the results
themselves.
10-6 SLEEP MODE UNDER LIGHT TRAFFIC
After some tedious manipulations, we nd from formula (10.8) that the energy
consumption is approximated by
E

[E] =
1
t
J
+
_
E[S] +J 1
E[S] +
J1
t
J
_
+
+2
2
J1

j=0
X
j
_

J1

j
t
J
J +1+ j
_
R(0)
_

J1

j
t
J
J +1
_
+O(
3
),
(10.10)
where the X
j
s are dened as
X
0
=
1
t
J
t
J
1

i=0

k=1
P[S = k]
k+i

n=0
R(n) (10.11)
X
j
=
1
t
J
t
J
1

i=0

k=1
P[S = k]
k1

m=0
R(k +m+i +
j1
), if 0 < j < J. (10.12)
This expression for P[E] has an intuitive basis. The constant term tells what hap-
pens when there are no arrivals at all: the system will forever be in the sleep
period with length t
J
, and hence have an energy consumption
1
t
J
. In the rst order
derivative we see the amount of extra energy consumption induced by one arrival.
The second order derivative shows the inuence of the overlap of two packet life
times (service time plus subsequent sleep periods). The terms X
j
are such that

2
X
j
gives the probability of having a sample path with two arrivals that have
exactly j sleep periods in between. For the mean virtual waiting time we nd:
E

[W] =
t
J
1
2
+
_
1
2
E[S(S1)] +T
J1

J1
t
J
1
2
_
+
+2
2
_
J1

j=0
X
j
_
t
J
1
2

J1
T
J1
+E[S]
_
+Y E[S] X
0
E[S]
2
+
+
J1

j=1
X
j
_
t
J
1
2
(E[S]
j
) T
j
E[S]t
j
_
_
+O(
3
), (10.13)
where
Y =
1
t
J
t
J
1

i=0

k=1
P[S = k]

J1
+k1

n=0
nR(n) (10.14)
T
j
=
j

i=1
t
i
(t
i
1)
2
(10.15)
CHAPTER 10 10-7
10.5 Numerical example
0 0.005 0.01 0.015 0.02
20
22
24
26
28
30
32

[
W
]
0 0.005 0.01 0.015 0.02
0.02
0.04
0.06
0.08

[
E
]
Figure 10.1: Plots showing the exact curve of E

[E] and E

[W] vs.

:= as well as
their rst and second order approximations (in straight lines, dashes and long dashes
respectively).
We veried the obtained light-trafc limits with our earlier results for an on/off ar-
rival source and for sleep strategy (4, 8, 16, 32, 64, ) and found indeed an agree-
ment.
10.6 Conclusion
Light-trafc approximations are a very interesting tool for the analysis of power-
saving protocols. We can not only reconstruct our previous work under light trafc,
but all sorts of generalizations are possible as well. It is noteworthy that the rst
order approximation is insensitive to trafc correlation. It would be interesting to
nd out whether this is due to either the applied thinning scheme or the Markovian
trafc correlation, or whether this can be established more generally.
Postscriptum
In this chapter, we sought to nd light-trafc characteristics by adapting the sample-
path method of [3]. After having gained more experience in this domain (eg.
in [6]), the author feels that the approach that makes use of the deviation ma-
trix is at least as promising and uses less ad-hoc reasoning. As we mentioned in
the introduction, this approach makes use of the expansion formula:

=
0

k=0
[(P

P
0
)Z]
k
,
10-8 SLEEP MODE UNDER LIGHT TRAFFIC
where Z is the deviation matrix associated with the no-arrival queueing system.
It is possible to extend this light-trafc analysis to many other systems such as
the sleep-mode mechanism considered in chapter 9, systems with several mobile
stations etcetera.
The plots indicate that the light trafc limit does not converge very quickly, in
that respect we had more promising results with light-correlation expansions, than
with light-trafc expansions [6].
References
[1] IEEE 802.16e-2005 Part 16: Air interface for xed and mobile broadband
wireless access systems // 2006.
[2] De Turck K., De Vuyst S., Fiems D., Wittevrongel S. Performance analysis of
the IEEE 802.16e sleep mode for correlated downlink trafc // Telecommu-
nication systems 2008. V. 39. No 2. P. 145156.
[3] Reiman M., Simon B. Open queueing systems in light trafc // Mathematics
of operations research, 1989. V. 14. No. 1. P. 2659.
[4] Kovalenko I. Rare events in queueing theory. A survey // Queueing systems
1994. V. 16. No 1. P. 149.
[5] Tsoukatos. K. Heavy and light trafc regimes for M/G/ trafc models //
PhD thesis. 1999.
[6] D. Fiems, B. Prabhu, K. De Turck, Analytic approximations of queues with
lightly- and heavily-correlated autoregressive service times, Annals of Oper-
ations research, to appear.
11
Large-Deviations Analysis for
Energy-Saving Mechanisms in
Wireless Networks
Koen De Turck, Dieter Fiems, Stijn De Vuyst, Herwig Bruneel,
Sabine Wittevrongel
Accepted for ValueTools 2011
Abstract. In this chapter, we make use of large-deviations techniques to study
the performance of sleep mode mechanisms. These allow us to look at a general
multi-user model, for which we obtain a sample path large-deviations principle. As
the system equations exhibit discontinuous boundaries, there are some mathemat-
ical obstacles to overcome. The results shed light on the trade-off between buffer
overow and energy consumption. We illustrate the results in scenarios with one
and multiple mobile stations.
11.1 Introduction
Energy-saving mechanisms in wireless communications are currently a hot topic.
Short battery life is one of the main impediments to a more widespread use of wire-
less devices. Hence, understandably, a lot of research is directed at solving or at
least mitigating this problem. Evidently, this can be done by improving on battery
11-2 LARGE DEVIATIONS FOR SLEEP MODE
efciency, but lately there is also a lot of interest in reducing the energy consump-
tion of the wireless devices by including energy-saving measures in the commu-
nication protocols themselves. On that account, it is no wonder that two major
standardization efforts for next-generation wireless communications that is, the
IEEE-backed 802.16 committee (also known as WiMAX), and the so-called long-
term evolution (LTE) of UMTS have opted to incorporate various energy-saving
elements which are referred to as sleep mode and idle mode in the WiMAX
context and as discontinuous reception (DRX) in LTE terminology. Power saving
in WiMAX is achieved by turning off parts of the MS (mobile station) in a con-
trolled manner when there is neither trafc from the MS (uplink trafc) nor to the
MS (downlink trafc). Whereas a MS in sleep mode is still registered to a BS
(base station) and still performs hand-off procedures, idle mode operation (which
is optional in current WiMAX standards) goes further and allows the MS to be
completely switched off and unregistered with any BS, while still receiving broad-
cast trafc. In LTE, similar functionality is present, with user equipment (UE)
fullling the role of MSs and evolved Node-Bs (eNB) as the BSs.
A considerable number of scientic papers have been devoted to the perfor-
mance evaluation of sleep-mode mechanisms in the context of wireless networks.
A wide variety of tools have been employed: simulations [7] or Markov-chain
based (with or without transform-domain methods) [5, 6]. A feature which sets
our paper apart is that we look at the impact of sleep mode in a multiple mobile
stations scenario. The scenario with multiple mobile stations is too complex to
be solved with transform-based or matrix-analytic techniques, therefore we rely
on large-deviations tools [9, 10], which allow the asymptotic computation of small
probabilities on an exponential scale. The rather specic nature of the results (only
rare events are considered) is compensated by the broad generality in which the re-
sults can be developed. Furthermore, in telecommunication networks, rare events
may have a larger impact on performance than average behavior, as it is rare
events which cause huge delays and subsequently user discomfort and loss of per-
ceived performance.
The model in this paper is not continuous at the boundaries, a condition which
complicates large-deviations analyses by a fair amount. Various techniques to
work around these discontinuities have been applied, such as the idempotent prob-
ability method of Puhalski [8], or the contraction-mapping framework of [10]. In
this paper, we opt for an adaptation of the latter (see also [13, 15, 16] for other
papers that t within the same paradigm). Its main idea is the transformation of a
large deviation principle of the arrival process into a large-deviations principle of
the process of interest by means of the powerful contraction mapping theorem.
Finally, we would like to note that although (the rst part of) this article is
fairly theoretical in nature, its aim is a practical one: we investigate whether large-
deviations techniques can offer quantitative recommendations to network engi-
CHAPTER 11 11-3
neers as to how to organize sleep mode in wireless devices. For example, we would
like to determine which sleep mechanism is superior from a rare-event overow
probability point of view, the one as dened in the 802.16e protocol [1] or the
newly proposed [3]. The examples in Section 11.5 indeed offer a starting point for
such suggestions to the protocol designer.
The outline of the remainder of this paper is as follows. In Section 11.2, we
expound on the modelling assumptions used in this paper. Section 11.3 contains
the main result and its proof. We detail on a numerical computation scheme in
Section 11.4 and illustrate important special cases in Section 11.5. Finally, we
draw conclusions in Section 11.6.
11.2 Setting
Consider a number M of mobile stations (MS) operating under the same base
station. As is common in wireless protocols, all operations occur in a slotted
(discrete-time) manner. Packets arrive at the base station destined for a specic
MS m according to a stationary arrival process X
m
(t), t N, 1 mM, and are
temporarily buffered at the base station while awaiting transmission, in a dedicated
buffer for each MS. In the following, we will assume that the arrival streams X
m
to the different mobile stations are mutually independent but possibly distributed
according to a different law. We denote the mean rate of X
m
by
m
.
For energy-saving purposes, mobile stations may reside in so-called sleep mode.
The MS then consumes less energy, but is unable to receive packets from the BS.
When not in sleep mode, the MS is said to be active. We assume that the BS can
send packets at a constant maximum rate C which is shared over all active MS. The
transmission rate as seen by a single MS is thus greater if fewer stations are active.
A number of possible mechanisms of how to organize sleep and active periods in
the different MSs have been proposed. We will take a closer look at two of these
mechanisms.
S
1
: When the buffer dedicated to an MS, say MS m, gets empty, then MS m
goes into sleep mode, and a timer is started. The content of the buffer dedicated to
trafc for MS m is checked at the end of a series of instants t
i
, with service being
resumed at the rst such instant that the buffer content is non-empty. This is a gen-
eralization of the sleep-mode mechanism described in WiMAX. In the WiMAX
standard [1], there are three possible sleep-mode classes, of which the most com-
monly researched is class III, for which the t
i
satisfy the following relation:
t
i
= max(2
i1
t
min
, t
max
),
where t
min
and t
max
are parameters negotiated between the mobile station and the
base station. In LTE, the DRX mechanism is slightly different, resulting in a dif-
ferent sequence of t
i
parameters: at rst, there is a so-called close-down period, in
11-4 LARGE DEVIATIONS FOR SLEEP MODE
which the antenna and decoding device is not yet turned off, then a series of short
sleep periods of identical length take place, followed by a series of long sleep
periods of identical length.
S
2
: Sleep and active periods are organized in cycles of a xed length L. When
the buffer of a certain MS m gets empty, it goes into sleep mode for the rest of the
ongoing cycle. The technological advantage of this scheme is that it decreases sig-
naling overhead. It has been proposed for the forthcoming IEEE802.16m wireless
standard [2, 3]. To the best of our knowledge, no equivalent mechanism has yet
been considered for the LTE framework, although it seems as reasonable for this
standard as it is in the WiMAX case.
Note that for both mechanisms the system is stable under the natural condition

m
< C. Before we develop a mathematical formulation of this model, we
introduce some notation.
Let X denote the space of discrete-time R
M
-valued processes. Let X denote a
process in X and its truncation to [0, t) is denoted by X[0, t). Also, X(t) denotes
the value of the process at time t, and X[0, t)
.
=
t1
i=0
X
i
, with X[0, 0) = 0. We are
mainly interested in two things: (1) the distribution of the buffer content (and its
close relative the packet delay) and (2) the expected energy consumption of an MS
station. We develop a set of recursive equations which relate the arrival process to
the buffer content process and to the status (active or in sleep mode).
The two mechanisms S
1
and S
2
satisfy the following system equations:
Q
m
(t +1) = (Q
m
(t) +X
m
(t) r
m
(S
1
(t), , S
M
(t)))
+
. (11.1)
where Q
m
and X
m
denote the buffer content process and the input process of MS m
respectively. Also, the auxiliary processes S
m
, 1 m M indicate when MS m is
sleeping. The functions r
m
determine the rate at which the BS sends to the MS m
depending on the state of the other MS. This is a versatile description that allows
to model a variety of situations: (1) schemes in which the MS always send at the
same rate, regardless of the status of the other MS (no bandwidth redistribution);
(2) schemes with bandwidth redistribution. For example, the available bandwidth
is divided in equal parts among the active MSs; (3) priority mechanisms where
some MS only get bandwidth when other MSs are not working. This setup is of
course more general than what we can reasonably explore in Section 11.5, hence
we will only highlight a few scenarios. We just aim to develop a large-deviation
result with a natural generality. Note that a setup with no bandwidth redistribu-
tion effectively uncouples the queues, so that the single MS case is sufcient for
assessing performance.
For S
1
we dene another auxiliary process

S
m
which counts how long the sys-
tem has the system been residing consecutively in sleep mode. Its evolution is as
follows (we denote by and denote logical or and and respectively):

S
m
(t +1) = (

S
m
(t) +1)1

S
m
(t)/ T Q
m
(t)>0
(11.2)
CHAPTER 11 11-5
where T denotes the set of time instants t
i
(i.e. on which the dedicated buffer is
checked). The sleep process S
m
for S
1
is then dened as:
S
m
(t) = 1

S
m
(t)>0
(11.3)
For sleep mechanism S
2
, new cycles start for mobile station m at time instants
belonging to the set LN+
m
.
=Ln+
m
: n N (hence
m
1, , M denotes
the offset of MS m). The sleep process S
m
for S
2
satises the following recursion:
S
m
(t +1) = 1
Q
m
(t)=0(t / LN+
m
S
m
(t))
(11.4)
Finally we make some assumptions on the energy consumption. Each MS
consumes a xed amount E
a
during a slot in which it is active, and an amount E
i
during an idle slot. At the end of each sleep interval, an additional amount E

is
consumed ( stands for listening).
11.3 Large Deviations Analysis
11.3.1 Motivation
The scenario with multiple mobile stations is too complex to be solved with trans-
form-based or matrix-analytic techniques. Therefore, we opt for a large-deviations
analysis. The so-called many-sources or many-ows scaling [10] yields the most
interesting results in this context, and hence we adopt this scaling throughout this
paper. We consider a sequence of processes X
L
, which are interpreted as the aver-
age of L independent processes each distributed like X and speed up the transmis-
sion rate by a factor L as well. The many-ows large-deviations limit describes
what happens when the number of ows is very large. In particular, it provides the
exponential decay of rare events associated with this limit.
The choice of the many-ows scaling requires some justication. Indeed, in
contrast to for example trafc in the backbone of a network, the number of different
ows that an individual user sees may not be very large. However, we believe that
this modelling choice is justied for the following reasons: (1) Next-generation
platforms such as WiMAX aim to operate at a larger, metropolitan area scale, and
hence aggregate more trafc; (2) it is one of the few modeling techniques that
can still provide answers in the rather general framework of this paper. Fast-time
scaling, the other frequently used large-deviations scaling is too crude, as it lters
away the effects of sleep-mode. Indeed, if the operation of the system is sped up,
then sleep periods are also shortened, and vanish in the limit. It is worth noting
that if we try to x this, by scaling the sleep periods with a factor L as we speed
up time L times, then we essentially get a simplied many-ows scaling (with no
time correlation in the trafc ows).
11-6 LARGE DEVIATIONS FOR SLEEP MODE
Let us briey recall the notion of a large-deviations principle (LDP). We refer
the reader to a.o. [9] and [10] for references on large deviations. A sequence of
random variables X
L
in a Hausdorff space X with Borel sigma algebra B is said to
satisfy a large deviations principle with good rate function I if for any B B,
inf
xB
o
I(x) liminf
L
1
L
logP(X
L
B)
limsup
L
1
L
logP(X
L
B) inf
x

B
I(x)
where I : X R has compact level sets, and B
o
and

B denote the interior
and closure of the set B respectively. This is called a sample path LDP when X is
a space of processes.
In the following, we will be concerned with the buffer content of one mobile
station (say, without loss of generality, MS 1). We denote by q
S
i
, i = 1, 2 two
functions which map the arrival processes onto the buffer content of MS 1 at time
instant 0 under sleep-mode mechanism S
i
. Analogously, mappings w
S
i
: X R,
i = 1, 2 denote the energy consumption of MS 1 during time instant 0.
11.3.2 Limiting process
It is useful at this point to investigate the behavior of the sleep-mode mecha-
nisms in the many-ows uid limit. The difference with the large-deviations
limit should be clear: whereas the limiting process concerns itself with the al-
most sure behavior when the number of ows approaches innity, large deviations
aims to quantify the (small) probabilities that the system deviates from this almost
sure path. Consider the sequence of processes Q
L
m
(t) which result from plugging
X
L
m
(t) =
1
L

L
=1
X
()
m
(t) into (11.1), where X
()
m
, 1 L are independent and
identically distributed versions of the arrival process.
For many queueing systems, this limiting process (for L ) converges to
a deterministic process which moreover converges almost surely to zero in nite
time [14]. The sleep-mode models considered in this paper are peculiar in that
they reduce to periodic deterministic processes. We will show this without being
entirely rigorous. Indeed, as the number of ows approaches innity, under certain
regularity conditions we can apply a law of large numbers argument, which says
that
X
L
m
(t) , a.s., for L .
Thus, the amount of trafc arriving at the systemduring a slot approaches , almost
surely. For mechanism S
1
, we thus nd a periodic behavior in which each mobile
station goes to sleep for the duration of t
1
slots, collect an amount of t
1
of trafc,
which subsequently decreases again; see Figure 11.1. The sleep periods of the
different MSs do not necessarily coincide, and depend on the initial conditions of
CHAPTER 11 11-7
the system. This already highlights a problem we will meet later on, namely that
the system does not forget completely its past state. We omit the proof that the
limiting process is necessarily periodic. Note that for this mechanism the sleep
periods have in the limit a xed length while the lengths of the working periods
depend on the initial condition. We therefore explicitly keep track of the offsets
with which the MSs go into sleep mode in the limiting process in the random
variables D
(1)
m
0, , t
1
1. More precisely, D
(1)
m
denotes the last time that MS
m goes from active mode to sleep mode, relative to the last time MS 1 went into
sleep mode. Obviously, D
(1)
1
=0. Note that this choice keeps the D
m
time-invariant
for the limiting process Q

m
().
Figure 11.1: The limiting process Q

m
for the mechanism S
1
. The parameter t
1
and the
random variable D
2
are illustrated. Remark that as Q

m
is a discrete-time process, the
straight lines should in fact be replaced by discrete dots.
For the mechanism S
2
, the situation is slightly different in that the sleep periods
in the limit do not have a xed length but the combined sleep and working periods
are xed to the cycle length or a multiple thereof (see Figure 11.2). Also for this
mechanism we need to keep track of the periodic cycle starts, which we summarize
into the variables D
(2)
m
0, , L1.
11.3.3 Outline
In this section, we formulate assumptions under which a large-deviations principle
for the arrival processes can be transformed by means of the contraction principle
11-8 LARGE DEVIATIONS FOR SLEEP MODE
Figure 11.2: The limiting process Q

m
for the mechanism S
2
. The parameters L and are
illustrated. Also here, as Q

m
is a discrete-time process, the straight lines should in fact be
replaced by discrete dots.
[10] into a large-deviations principle of the queue content process.
The contraction principle pushes through LDPs from one topological space to
another, under the condition that the mapping between the spaces is continuous.
The usual programme (as elucidated for example in [10]) for deriving an LDP for
a new situation is thus as follows:
1. Take the LDP for the input processes (arrival processes).
2. Dene a mapping between the inputs and the output of interest. Examples
of the latter are buffer content at time 0, the buffer content sample path is a
possible output as well, but a bit harder to obtain.
3. Tweak the topology and/or the mapping such that the mapping is indeed
continuous. Recall that a mapping f : X Y is continuous if the inverse
image of every open set of Y is an open set in X. A possible choice is the
so-called initial topology [9].
4. Strengthen the LDP such that it holds for the topology found in the previous
point. We often observe a trade-off: the stronger the topology for X, the
more mappings that are continuous, but the harder it is to strengthen the
LDP. In essence, we must establish exponential tightness of the sequences
of measures under the topology.
CHAPTER 11 11-9
5. Application of the contraction principle gives us the desired LDP in the topo-
logical space Y .
For the LDP results presented here, we need the following assumption on the
scaled arrival processes X
L
m
.
Condition 1. (Finite-time characteristics) For R
t
, dene

L
t
() =
1
L
logEexp(L X
L
(0, t]).
Assume that for each t and , the limiting cumulant generating function

t
() = lim
L

L
t
()
exists as an extended real number, and that the origin belongs to the interior of the
effective domain of
t
, and that
t
is an essentially smooth, lower-semicontinuous
function.
Due to the G artner-Ellis Theorem [9], the nite-time truncations over [0, t)
satisfy an LDP with good rate function
I(x) = sup

x
t
(). (11.5)
Some useful further assumptions include (which simplify matters for example
in the numerical computations of Section 11.4):
1. The arrival process has (asymptotically) independent increments, that is, the
rate function I(x) can be written as:
I(x) =

k
I(x
k
), (11.6)
for a certain good rate function I : R
M
R
+
. Examples include (but are
not limited to) the case that the original arrival process is indeed time-
independent and identically distributed.
2. Asymptotically Markov increments: there exists a space Z and good rate
functions K
0
(.), K(.[z), I(.[z) : Z R
+
, such that
I(x) = infK
0
(y
0
) +

j
I(x
j
[z
j
) +

j
K(z
j
[z
j1
)[z
j
Z. (11.7)
This kind of structure arises for example when taking the many-ows limit
for many Markovian arrival ows, see eg. [12].
11-10 LARGE DEVIATIONS FOR SLEEP MODE
11.3.4 Main Result
Let X be the set of arrival processes to the M buffers over a nite interval [0, T],
let Y
0
be the set representing the possible initial states (buffer content plus time
spent in sleep mode), and let Y be the set of the output process: the combined
buffer content and energy consumption evolutions in interval [0, T]. The function
q : Y
0
X Y maps the input process together with the initial condition to to the
output process. Assume that Y
0
and X satisfy LDPs over the Euclidean topology
with good rate functions J
0
(.) and I(.) respectively.
Theorem 11.3.1. An LDP holds for the queue content and energy consumption
processes over nite time interval [0, T] and nite buffer content b, with good rate
function J : Y R
+
,
J(y) = infJ
0
(y
0
) +I(x)[y = q(y
0
, x), x X, y
0
Y
0
(11.8)
Proof. Let
1
be the initial topology over Y
0
X with respect to the mapping q,
that is, the weakest topology that makes the function q(., .) continuous. We rst
establish an LDP over Y
0
X under this topology, and then the desired LDP over
Y follows by the contraction principle. In contrast to similar proofs, the initial
topology
1
is neither weaker nor stronger than the Euclidean topology. Hence we
must take a detour via their intersection topology
I
. This is the topology whose set
of open sets is the intersection of the open sets of the other two topologies. To push
the LDP from the Euclidean topology
e
down the coarser
I
, an application of the
contraction principle with the identity function sufces. To push it back up to the
stronger topology
1
with the help of the inverse contraction principle takes a little
more work. The identity function is again the mapping of choice, but now we must
also show that the sequence (Y
L
0
, X
L
) is exponentially tight. This is automatically
fullled if the space (Y
0
X,
1
) is compact. While this is not the case (consider
a sequence with the queue content approaching zero while the MS stays active),
we instead compactify the space and prove that the rate function is innite in the
added points. As (Y
0
X,
1
) is a semimetric topology, we may verify compactness
by showing that any sequence in Y
0
X has a convergent subsequence. We do so
inductively: prove that we can nd a subsequence that is convergent for the rst
time instant. If T = 1 then the proof is complete, otherwise we have reduced the
problem to a shorter interval.
11.4 Numerical method
Relatively few publications on large deviations (especially in telecommunications)
have a focus on the numerical solution of the LDP as a variational problem. In
the domain of statistical physics, such numerical methods have been investigated
CHAPTER 11 11-11
among others in [11]. This is partly understandable as it runs counter to the of-
ten cited promise of large-deviations to provide solutions where other methods fall
short, especially numerical methods suffering from the curse of dimensionality. In
this work however, we have implemented a few algorithms that solve a discretized
version of the variational problem (11.8), partly in order to check the optimal solu-
tions that were obtained by purely analytical means, and partly in order to extend
the range of scenarios that we can tackle. The results we nd are generally promis-
ing.
In order to stratify the discussion we focus on a specic large deviation event,
namely the case that one specic MS overows (i.e. reaches an overow level b).
Also, we restrict our attention to arrival processes satisfying either (11.6) or (11.7).
In the former case, we start out by discretizing the state space X into a discrete set

X with the help of a discretizing function : X



X. It is to be expected (but hard
to prove rigourously) that a sufciently ne grid leads to a solution that is close to
the solution of the original problem. The variational problem then reduces to the
problem of nding the shortest path in a graph with an edge from node v to node
v

having a cost infI(x) : v

= (q(v, x)).
In case of (11.7), we discretize the space Z of the arrival process as well leading
to a discretized space

X

Z.
This can be solved by means of various existing shortest-path algorithms, of
which Dijkstras algorithm is perhaps the most famous. The time complexity of
this class of algorithms is typically O(E +V logV), while space complexity is
O(V), where E is the number of edges and V is the number of vertices. Keeping
the number of edges reasonably small is key to a reasonably low computation time.
In the numerical examples, we restricted the rate function I of the arrival process
to a nite region D when the state space got too large (e.g. for scenarios with
multiple mobile stations):

I(x) =
_
I(x), if x D
, otherwise.
(11.9)
We can also replace the standard Dijkstra algorithm by the so-called A

algo-
rithm or a variant thereof. This algorithm can exploit heuristic bounds on the path
length. Especially in structured problems such as the one at hand, huge gains can
be made.
Note that even if we resort to numerical techniques to solve the variational
problem (11.8), its computational cost is cheaper than the more straightforward
method of extracting information from a Markov chain: direct computation of the
stationary vector has (in absence of further structural properties) a O(V
3
) time
complexity.
11-12 LARGE DEVIATIONS FOR SLEEP MODE
11.5 Practical Application
The LDP as described in Theorem 11.3.1 involves a variational problem that is in
general hard to solve. In this section, we look at a number of specic situations.
For reasons of analytical tractability, we limit ourselves to time-independent (un-
correlated) arrival processes. Of course, the variational formulation also holds for
a broader class of both short-range and long-range dependent arrival processes, but
in those cases analytical expressions are more scarce.
We attempt to nd the optimal (i.e. least unlikely) path for the buffer associated
with MS 1 to reach a certain level b. The state of the system at time instant 0 is
given by the buffer contents Q
m
(0), 1 m M.
There are a couple of heuristic rules to which an optimal path often adheres.
Of course, the buffer of MS 1 overows because there temporarily is a higher
amount of trafc than can be served. But by the presence of multiple MSs and by
sleep mode, this principal event can be strengthened or weakened. For example, in
case of mechanism S
1
, it may be cheaper to have an interval in which no arrivals
occur rst, such that a longer sleep interval will be initiated during which the buffer
content can increase faster. Secondly, there is the role of the other MSs. If there are
many MSs active, the service rate as seen by MS 1 is lower, so it may be benecial
to have more MSs active than average during the overowing path. A last feature
that will help us is that because of the time-independence of the arrival processes,
optimal paths are usually piecewise linear (linear geodesics). We will illustrate the
various scenarios by making use of the Brownian motion arrival process, as this
typically leads to nice closed form solutions. The reason for this is that the rate
function has a quadratic form:
I(x) =
(x )
2
2V
. (11.10)
11.5.1 Single Mobile Station under S
1
We take a look at the relatively simple system with one mobile station. First, we
look at mechanism S
1
. We are given the buffer content Q(0) at time instant 0 and
S(0), the amount of slots that MS 1 is in sleep mode at time instant 0, and T, the
interval in which the overow to level b must take place. Let us rst assume that
Q(0) = 0 and S(0) = 0. It is intuitively clear that an optimal path must have a
form as shown in Figure 11.3. It consists of three linear segments, a rst of length

i1
:=
i1
k=1
t
k
during which the arrival rate is zero. Secondly, a segment of length
t
i
during which the MS resides in the ith sleep interval and the arrival rate is given
by the unknown x, and lastly a segment where the MS is active and during which
the buffer content b is reached, and which has an unknown length t.
CHAPTER 11 11-13
The rate function associated with this path is equal to:
I
q
(b) = inf
i,x,y

i1
I(0) +t
i
I(x) +tI(y). (11.11)
From the gure we see the following relation between the different variables:
b = xt
i
+(y C)t,
so that we can eliminate t. It is noteworthy that the optimal slopes x

and y

generally do not depend on the optimal value i

. Indeed, as the rate function I is


differentiable, we can nd the optimum by differentiating the above expression to
x and y. We nd:
_
_
_
t
i
_
I

(x)
I(y)
yC
_
= 0
bxt
i
(yC)
2
((y c)I

(y) I(y)) = 0.
(11.12)
from each of which t
i
vanishes. For the special case of Brownian motion, with
mean and variance V, we nd explicitly the following simple expressions for x

and y

:
x

= y

= 2C.
It is interesting to note that the optimal arrival rate y

in the last segment is the


same as predicted in a normal single server queue. Our analysis also shows that
when overow level b is very large, the last segment will dominate and we get
the same tail behavior as in a system without sleep mode. This agrees with our
intuition about large buffer asymptotics. Of course, in wireless systems with lots
of delay-intolerant data, we expect buffers to be rather small. The optimization
problem is thus reduced to nding the minimum among the i, which is a nite
number for most practical situations.
Figure 11.3: The optimal path for a single MS situation under mechanism S
1
.
11-14 LARGE DEVIATIONS FOR SLEEP MODE
11.5.2 Single Mobile Station under S
2
Next, we investigate the situation in which a single mobile station operates under
mechanism S
2
. We assume that the buffer is empty at the beginning of the ex-
cursion and the current sleep cycle is projected to end during slot L
0
. There are
two candidate optimal paths, both shown in Figure 11.4. For path A, there are no
arrivals during the rst L
0
slots, so that the buffer can ll during an entire sleep
cycle. For path B, the buffer immediately starts to ll. It is intuitively clear that
when L
0
is small, path A will be better, and when L
0
is close to L, then path B will
be better. The rate functions of the paths are respectively
I
A
(b) = inf
x,y
L
0
I(0) +LI(x) +tI(y),
where b = xL+(y C)t, and
I
B
(b) = inf
x,y
L
0
I(x) +tI(y),
where b = xL
0
+(y C)t. The optimal arrival rates are again independent of b, L
and L
0
, and take the same values for Brownian motion as in the previous scenario.
Having found the optimal values x

and y

, we easily derive that path A is the


optimal path if
L
0
L

I(y

) (y

C)I(x

)
(y

C)(I(0) I(x

)) +x

I(y

)
,
and otherwise path B is better.
Figure 11.4: The optimal paths for a single MS situation under mechanism S
2
.
11.5.3 Scenarios with Two Mobile Stations (under mechanism
S
2
)
Now we move on the situation with 2 MS competing for the same bandwidth. As
we will see, the path to overow is considerably more complex than in the single
CHAPTER 11 11-15
MS case. We especially pay attention to the inuence of the offset between
the two sleep cycles. The excursion starts with two empty buffers. We discern
again a buildup during a sleep period and then a second phase in which the MS
actually reaches level b. The difference is that there is now a second MS that can
help make the overow more probable by being more active than average. The
unknowns x
1
and x
2
denote the arrival rates during the different phases of MS 1,
and y
1
, y
2
denote the arrival rates for MS 2. Unknown s denotes the fraction of the
time that the MS 2 is in sleep mode during the last phase (which has length t). As
MS 1 is continuously active in this last phase, MS 2 sees a service rate C/2, and
hence it will be active for a fraction 1s if the total trafc during this period is a
fraction (1 s) of the amount of trafc that the system can consume during this
period (namely +
1
2
Ct). Because of the linear geodesic property, this rate must
be constant. Hence we have
I
q
(b) = I(y
1
) +(L)I(y
2
)I((1s)(+
1
2
Ct)/t) +LI(x
1
) +tI(x
2
). (11.13)
We can derive two further relationships between the unknowns: b = x
1
L+t(x
2

C+
1
2
sC) and = y
1
+(y
2
C)(L).
It is possible to nd closed-form expressions for the case of Brownian motion
also in this case, but the expressions get dauntingly large. We can however see the
impact of the offset : if the two cycles would be synchronized, then the expen-
sive segment corresponding to rate y
2
(it is expensive because the BS is serving
at full rate C) vanishes, which results in more likely overows.
11.5.4 Energy consumption
What does the large-deviations analysis tell us about the other factor in the trade-
off, namely the energy consumption ? Firstly, there is the (quite crude) law-of-
large-numbers result, that says that the average energy consumption for a large
number of sources will be as observed in the limiting paths.
Secondly, information can also be drawn from the large deviations on the av-
erage energy consumption over a long period of time. This measures how likely it
is to deviate from the energy consumption as predicted by the limit results.
11.5.5 Numerical illustrations
In this section, we show some of the results we obtained by numerical means.
The rst example we look at considers a 1 MS scenario, with t
i
= i for i C and
t
i
= C +(i C)L for i > C. This boils down to having a closedown period of
length C followed by a possibly innite series sleep periods of the same length
L. We assume Brownian arrival processes with = 0.7 and V = 1.0. The plots
are in accordance with the analytically computed optimal paths in the last section.
11-16 LARGE DEVIATIONS FOR SLEEP MODE
Figure 11.5: The optimal path for a scenario with 2 MS under mechanism S
2
.
In case of a non-empty starting condition, we observe a behavior that is quite
different from systems without energy-saving mechanism: instead of approaching
the overow level directly, a cheaper path is chosen that visits the boundary and
thus prots from the sleep mode mechanism. Of course, if the initial level is very
close to the nal level, then the direct path may be optimal.
In Figure 11.7, we plot the decay rate I associated with optimal paths to level
b = 128 against the length L of the nal sleep periods, for different closedown
period lengths C. We observe three different regimes. Firstly, for small L, the
closedown period is too large for optimal overow paths to go through a sleep
mode phase and hence the direct path is better, resulting in a decay rate that is
independent of the C and L. Next, there is a region in which is it protable to
go into sleep mode, but overow is reached after the sleep period. For the nal
regime, the length L is sufciently large for the system to reach overow during
one sleep period of length L, which cause the curves to atten out for large L.
Next, in Fig. 11.8 we look at the most probable overow path for a 2 MS
scenario with sleep mechanism S
1
and bandwidth redistribution. We observe an
interesting synchronization phenomenon: before travelling to overow, the sleep
periods of the two MSs rst synchronize their work and sleep cycles, so that we
have a longer period of shared bandwidth, which makes overow of one buffer
more likely. This synchronization phase is quite messy to analyze in closed form.
This synchronization phenomenon is much less effective in the 2 MS system
CHAPTER 11 11-17
0 10 20 30 40 50
0
100
200
300
time k
Q(k)
Figure 11.6: The optimal path for a scenario with 1 MS under mechanism S
1
. S
1
.
10 20 30 40
0
20
40
60
80
100
120
140
sleep length S
d
e
c
a
y
r
a
t
e
I
Figure 11.7: The decay rate I of reaching level b = 128 against L for C = 2, 8, 16.
with mechanism S
2
(at least when the offset parameters
m
are chosen in a dis-
persed manner, for example 0 and L/2).
11.6 General observations and conclusions
The paths to overow get progressively more complex as the number of MS in-
creases. There are however a number of interesting observations which we can
extract from the variational formulation, even though we either cannot solve it
explicitly or the explicit solution is too complex to be useful.
In the previous subsection, we already touched upon the fact that the offset
11-18 LARGE DEVIATIONS FOR SLEEP MODE
0 10 20 30 40
0
20
40
60
80
100
120
time k
Q(k)
Figure 11.8: An optimal path for the two MS case with sleep mechanism S
2
.
with which the sleep modes of the different MS arise are a very important factor.
We see signicantly worse performance when the different MS have synchronized
sleep and active periods, both in terms of energy consumption and in terms of
overow probability. This is an area where mechanism S
2
has a marked advantage
compared with mechanism S
1
, as in the former the offsets are xed and can be
chosen by the base station so as to guarantee a good spread. For mechanism S
1
,
the sleep periods might drift and end up more or less synchronized (although this
drift is a rare event under the many-sources limit), thus making the overow more
likely. There is another feature that can be exploited under S
1
but not under S
2
,
namely the fact that we can have a long stretch without arrivals, so that the MS
enters a long sleep interval, during which the buffer content can rise to high levels.
This effect plays an even larger role for bursty trafc.
Although we cannot yet draw denitive conclusions, it appears that from a
many-sources point of view, mechanism S
2
seems preferable over the other.
References
[1] IEEE 802.16e-2005, Part 16: Air interface for xed and mobile broadband
wireless access systems Amendment 2: physical and medium access con-
trol layers for combined xed and mobile operation in licensed bands Cor-
rigendum 1, February 2006.
[2] The Draft IEEE 802.16m System Description Document, IEEE 802.16m-
08/003r4, July 2008.
[3] Sleep mode operation for IEEE802.16m, C802.16m-08/721r1, July 2008.
CHAPTER 11 11-19
[4] Keep-awake mechanism for 802.16m sleep mode, C802.16m-08/718, July
2008.
[5] K. De Turck, S. De Vuyst, D. Fiems, and S. Wittevrongel, Performance anal-
ysis of the IEEE 802.16e sleep mode for correlated downlink trafc, Telecom-
munication Systems, Vol. 39, No 2, pp. 145156.
[6] K. De Turck, S. De Vuyst, D. Fiems, S. Wittevrongel and H. Bruneel Perfor-
mance of the Sleep-Mode Mechanism of the New IEEE 802.16m Proposal for
Correlated Downlink Trafc Proceedings of NetCOOP 2009, Eindhoven, Oct.
2009
[7] N.-H. Lee, S. Bahk, MAC sleep mode control considering downlink trafc
pattern and mobility, Proceedings of the IEEE 61st Vehicular Technology
Conference, VTC2005-Spring (Stockholm, 30 May1 June 2005), Vol. 3, pp.
20762080.
[8] A. A. V. Anatolii A. Puhalskii. A large deviation principle for join the shortest
queue. Mathematics of Operations Research, 32(3):700710, august 2007.
[9] A. Dembo and O. Zeitouni. Large deviations techniques and applications,
volume 38 of Applications of Mathematics (New York). Springer-Verlag, New
York, second edition, 1998.
[10] A. Ganesh, N. OConnell, and D. Wischik. Big Queues. Springer, 2004.
[11] C. Giardin` a, J. Kurchan, and L. Peliti. Direct evaluation of large-deviation
functions. Phys. Rev. Lett., 96(12):120603, Mar 2006.
[12] M. Mandjes and A. Ridder. A large deviations analysis of the transient of
a queue with many Markov uid inputs: approximations and fast simulation.
ACM Trans. Model. Comput. Simul., 12:126, January 2002.
[13] N. OConnell. Large deviations for queue lengths at a multi-buffered re-
source. Journal of Applied Probability, 35(1):240245, 1998.
[14] P. Robert. Stochastic Networks and Queues. Stochastic modelling and Ap-
plied Probability. Springer verlag, Berlin, 2003.
[15] D. Wischik. The output of a switch, or, effective bandwidths for networks.
Queueing Systems, 32:383396, 1999.
[16] D. J. Wischik. Sample path large deviations for queues with many inputs.
The Annals of Applied Probability, 11(2):379404, 2001.
[17] P.E. Hart, N.J. Nilsson, B. Raphael. B. (1968) Aformal basis for the heuristic
determination of minimum cost paths IEEE Transactions on Systems Sciences
and Cybernetics, SSC-4(2): 100-107, 1968.
12
Conclusions
In this chapter, we end our journey and sum up what we have learned. We also
provide an outlook on possible future research.
12.1 Overview of the main results
In Chapter 2 and 3, we captured the behavior of a retransmission protocol
in a probability generating matrix, the service matrix. This allows for a
straightforward derivation of the buffer behavior of protocols that do not
need a resequencing buffer. In Chapter 2, we considered the well-known
go-back-N protocol, and with the help of some ideas on formal languages,
we expanded this to a slew of protocols that preserve the packet ordering
during transmission.
We identied, apart from the throughput, another fundamental performance
characteristic, i.e. the asymptotic variance of the service process, which is
the most important ingredient in the heavy-trafc limit of both the packet
delay and the buffer content.
We performed an asymptotic buffer analysis for a retransmission system
with a Gaussian fading (non-Markovian) error model. Our model also al-
lowed to consider (possibly heavily) correlated packet arrival processes.
This allows to identify when the correlation of the error process is a domi-
nant factor, and when the arrival correlation dominates.
12-2 CONCLUSIONS
We analysed the ARQ resequencing buffer with the help of large-deviations
techniques and showed a rather unexpected link with the famous coupon-
collector problem.
To the best of our knowledge, we were the rst to study the sleep mode
mechanism with a correlated arrival process. However, it is intuitively clear
that correlated arrivals have an inuence on the best way to organize sleep
mode. This is also evidenced by our numerical examples (Chapter 6).
We studied the optimization of the sleep mode parameters (and also gave
some thoughts on what a fair denition of optimality is in this context) in
Chapter 7.
The inuence of uplink trafc on the sleep mode mechanism was also inves-
tigated (Chapter 8).
During our investigations, we encountered a proposal for a new sleep mode
mechanism backed by some well-known companies and we performed an
analysis for this new proposal as well (Chapter 9).
We found a scaling that effectively captures the sleep mode mechanism and
performed a large-deviations analysis on it.
We investigated a numerical method for solving the optimization problem
that arises from the large-deviations analysis. Numerical methods are (a
bit surprisingly) rarely studied, especially in a telecommunications context.
The approach we chose was to discretize the problem space and perform
Dijkstras algorithm on the resulting grid.
As for the different methodologies we have considered, we notice the slightly
paradoxical circumstance that techniques which offer a greater deal of in-
tuition (large deviation, light trafc with sample paths,...) also demand a
greater level of mathematical care to make the arguments completely rigor-
ous. On the other hand, we have techniques that impose perhaps less rigour
but more calculating skills.
12.2 Future work
As the two big wireless broadband protocols are still evolving, an important
part of the future work will be to evaluate newly proposed energy measures
as they come along.
Control of telecommunication networks is likely to increase in importance.
One can argue that the successful scaling methods and analysis techniques
of the future are the ones that play nicely with control techniques.

You might also like