HAL Id: inria-00129782
https://hal.inria.fr/inria-00129782v3
Submitted on 13 Feb 2007
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
The Performance of Broadcasting with Network Coding in Dense Wireless Networks
Cédric Adjih, Song Yean Cho, Philippe Jacquet
To cite this version:
Cédric Adjih, Song Yean Cho, Philippe Jacquet. The Performance of Broadcasting with Network Coding in Dense Wireless Networks. [Research Report] RR-6120, INRIA. 2007. �inria-00129782v3�
inria-00129782, version 3 - 13 Feb 2007ISSN 0249-6399
a p p o r t
d e r e c h e r c h e
THÈME 1
INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE
The Performance of Broadcasting with Network Coding in Dense Wireless Networks
Cédric Adjih, Song Yean Cho, Philippe Jacquet
N° 6120
Février 2007
Unité de recherche INRIA Rocquencourt
Domaine de Voluceau, Rocquencourt, BP 105, 78153 Le Chesnay Cedex (France)
Téléphone : +33 1 39 63 55 11 — Télécopie : +33 1 39 63 53 30
in Dense Wireless Networks
Cédri Adjih, Song Yean Cho, Philippe Jaquet
Thème1Réseauxetsystèmes
ProjetHIPERCOM
Rapportdereherhe n° 6120 Février200715pages
Abstrat: We present aprotool for wireless broadasttransmission basedon network
oding. Theprotooldoesnotneedneighborsensingandnetworktopologymonitoring. We
giveananalysisoftheperformaneoftheprotoolinanunit graphwirelessnetworkmodel
withuniformdensity,andintheaseofasinglesoure. Weshowthatevenwithonesoure,
duetodensity,network odingoerssomegains.
Inpartiularweshowthatin1dimensional(1D)networktheperformaneofthesimple
protoolbasedonnetworkodingislosetotheoptimalooding(withoutnetworkoding)
byafatorarbitraryloseto 1whenthenetworkdensityinreases. In2dimensional (2D)
networkssimulationsshowthattheratiotooptimalissimilartooutperformingMultiPoint
Relay(MPR) ooding.
Key-words: wirelessnetworks,networkoding, broadasting,multi-hop
les réseaux sans l denses
Résumé: Nousprésentonsunprotoolepourladiusiondanslesréseauxsanslbasésur
leodagederéseau. Leprotoolen'apasbesoindedéouvertedevoisinsetni dedétetion
delatopologieduréseau. Nousdonnonsune analysedelaperformaneduprotooledans
le as d'une soure simple et dans un modèle de réseausans l qui est un graphe disque
unité ave unedensité uniforme. Nousmontrons quemême ave uneseule soure,grâe à
ladensitédesnoeuds, leodagederéseaupermet desgains. Enpartiulier,nousprouvons
quedans un réseauunidimensionnel (1D), laperformane duprotoole simplebasé sur le
odage de réseau est prohe de la performane de l'inondation optimale (sans odage de
réseau)parunfateurarbitrairementdeprohede1quandladensitéduréseauaugmente.
Dans réseaux en 2 dimensions, des simulations montrent que le rapport à l'optimal est
similairevoiremeilleur queelle del'inondationparrelaismultipoints(MPR).
Mots-lés : réseauxsansl,odagederéseau,diusion,multi-sauts
1 Introdution
Intraditionalommuniationsystems,nodesexhangedatathroughrelayingbyintermedi-
atenodes: thepakets areforwardedby theintermediateroutersunmodiedforbothtra-
ditionaluniastand multiast ommuniation. Reently, thepioneering work of Ahlswede
et al. [1℄ hasshown that network oding, where intermediatenodesmix information from
dierent ows, an ahieve higherthroughput for multiasting than lassial routingover
networks represented as direted graphs. The works of [1,2,3,4,5,6℄ set theoretial and
pratialfoundationsfor theframework ofmanyreentworksonerning wirelessnetwork
oding (seesetion2). Thewirelessmediaisinherentlyabroadastmediawhihhasruial
advantagesineitallowsoverhearing. Theinherentadvantageouldbeexploitedby wire-
lessnetwork oding, permittinggainsbyreduingthenumberoftransmissionsrequiredto
transmitthesameamountinformation(energy-eieny). Dueto goodmathingbetween
theinheritadvantageandgaining,wirelessnetworkodinghasbeensuessfullyappliedto
uniasttraof802.11networksforinstanein[8℄,andithasalsobeenappliedtomultiast
tra,suhasin [7℄,[5℄and[10℄.
However,thefoushasoftenbeenonmultiplesoureswherethegainoriginatespreisely
from theexisteneof dierentdiretional ows, forinstane in [8℄foruniast, in [9℄(with
pakets andanti-pakets pairs),andin [5℄.
In ontrast, in this artile, we fous on broadastingwith a single soure. Our goalis
notsomuhtoahievemaximumperformanewithwirelessnetworkoding,but ratherto
illustratethatwirelessnetworkodinganahievegainsevenwithasinglesoure(asinthe
general,non-wireless,aseof[1℄),thankto density,andtogiveaninsightonthedynamis
ofnetworkodingin suhanenvironment. Thereforeourapproahfollowsthespiritofthe
algebrai gossip [11℄,but fousingontheidiosynrasiesofthewirelessmedium.
Todoso,wefousonasimpleprotoolfornetworkoding,whihdoesnotneedneighbor
sensing or network topology monitoring. We initially onsider its behavior in the one-
dimensional(1D)ase,i.ethe1Dunitdiskgraphmodel. Wedesribehowthepaketows
fromasoureevolveinwaves,andbystudyingthepropagationofthewaves,weareableto
quantifythe performaneof the simpleprotool in an unit graphwireless network model,
with an asymptoti model. Our analysis shows that, in 1D, as the density inrease the
performaneoftheprotoolisloseto theoptimal broadastbyafatorarbitraryloseto
1. Simulationsonrmtheresultsoftheasymptotimodel.
Thenweonjeturethebehaviorofthenetworkwithonesourein1Dtothatin2Dunit
diskgraphmodel. Thebehaviorin2Dshouldbeessentiallythesameasin1D(withthesame
oneptofwaves,onlywithmoreomplexwaves),withsimilarperformaneaspreditedby
themodelin1D.Simulationsresultsonrmthisinsight. Theartileisorganizedasfollows:
insetion2,wedesribethesimplenetworkodingalgorithmforwirelessbroadasting. In
setion3wedesribethewavepropagationanalysisthatwillbethebasisofourperformane
analysis. Insetion3wedeveloptheanalytialmodelsandourresults. Insetion5weshow
thesimulationresultswhih onrmandextendouranalytialresults.
RR n°6120
2 Related Work, Framework and Algorithms
2.1 General Framework of Wireless Network Coding
Thestartingpointofnetwork oding is theelebratedwork from [1℄,showingthatoding
innetworksouldahievemaximumbroadastapaity(givenbythemin-ut),whileinthe
general ase, it is out of reah of traditional transmission methods (i.e. without network
oding).
Subsequently, ithas been shownthat a simpleformof oding, linear oding [2℄,(using
linearombinationsofdatasymbolsbelongingtoGaloiseldsFp),issuienttoahievethe
boundsof[1℄. Furthermore,[3℄presentedonemethod whih doesnotrequireoordination
of (the oding at) the nodes, by introduing random linear oding and by showing that
suient eld size results in high probability of suess. With randomlinear oding, the
oding inside thenetwork isno longerpredetermined, sineit usesrandomoeientsfor
thelinearombinations. Thisapproahwasrenedby[4℄,intheontextofpaketnetworks:
theinformationisontainedinindividualpaketsandtherandomizedoeientsanthen
beinluded in a speial additionalpaket header. Relying on thefundamental broadast
nature of the wireless media, [6℄ introdued the rst pratialimplementation of network
odingin awirelessenvironment,anddemonstratedgainsforrossinguniastdatatras.
Theseworkssetthepathtopratialfoundations,whiharedesribedforinstanein[6℄,
andthatweareusingin thisartile.
Generation: rst,thepaketsaresupposedtobedividedintogenerations. Ageneration
isasetofpakets(ingeneral,fromoneorseveralsoures),thatouldbemixedtogetherby
linearombination. Intherestof thepaper,wewill assumethatallpaketsbelongto one
generationfromasoureand denoteagenerationsize thedimension, D.
Vetors: seond, the pakets are equally sized and are divided into bloks of sym-
bols over a eld Fp : ontent = (s1, s2, ..., sh). As in [4℄, the pakets inlude a header
whih isthelistofoeients. Hene thepaket formatis atuallyavetoroftheformat:
(g1, g2, . . . , gD;s1, s2, . . . , sh).
Transmission: atanypointoftime,anodeofthenetworkhasalistof vetors,linear
ombinations of initial soure pakets. When the node transmits, it generates a random
linearombinationofthevetorsv0, v1, ..., vk iturrentlyhas: P
iαivi (wherethe(αi)are
randomoeientsofFp),andtransmititbywirelessbroadasting.
Deoding: oneanodehasreeivedDlinearlyindependentvetors,itisabletodeode theDpaketsofthegeneration.
Withinthislassialframework,thetwonalquestionsare: whendoesthesouresenda
paket? and: whendotheothernodes(non-soure)maketransmissions? Thistransmission
sheduling istheheartofthealgorithm.
INRIA
Algorithm1: NCAL1
Soure sheduling: poisson soure;thesouretransmitssequentiallytheD vetors
1.1
(pakets)ofagenerationwithanexponentialinterarrival.
Nodes'start and stoponditions: thenodesstarttransmittingwhentheyreeive 1.2
therstvetor. Theyontinuetransmittinguntiltheyhaveenoughvetorstoreover
theD sourepakets. Thentheystilltransmitoneadditionalpaket(after usual
exponentialdelay).
Nodes'sheduling: poissonretransmission;thenodesretransmitlinear 1.3
ombinationsofthevetorsthattheyhave,withanexponentialinterarrival
2.2 Algorithms for Transmission Sheduling
2.2.1 Single soure
Thetransmission sheduling used in ourartile is the algorithm NCAL1. Essentially, the
nodestransmitrandomombinationsofthevetors,withexponentialinterarrival,untilthey
havesuientvetors. This algorithmanbeomparedto thealgorithmsused in [5℄,[10℄
and[11℄forinstane. Thealgorithm1isaversionofRandomLinearCoding(RLC)withpush
forthewireless ontext(hene, withfousonspatial gossip). It is alsoavariantof Blind-
Forwarding with NetworkCoding (BF-NC)from[10℄. Finally,thealgorithm1omparesto
the algorithm NC3 of [5℄, exept that we use asimple stop ondition. By hoie, in this
artile,wedidnotattempttouseopportunisti([6℄)mehanismsalsopresentedintheited
works.
Inalgorithm1,theintentoftheadditionalpaket transmissionat1.2istogiveneigh-
bors onemorehanetogetthelasttransmissionompletingtheD dimensionspae.
However, thestopondition in NCAL1 still doesnotguarantee that allinformation is
disseminated to all nodes, hene for referene we used another algorithm, NCAL2 whih
assumesreeption reports andstoptransmittingonlywhenallneighborhaveallpaketsto
restorealloriginalpakets.
Algorithm2: NCAL2
Soure sheduling: asin 1.1.
2.1
Nodes'start and stoponditions: asin 1.2,thenodesstarttransmittingwhen 2.2
theyreeivetherstvetorbut theyontinuetransmittinguntilthemselvesand
their neighborshaveenoughvetorstoreovertheD sourepakets.
Nodes'sheduling: asin 1.3 2.3
RR n°6120
2.2.2 Extension to multiple soures
Thetwo algorithm(NCAL1 and NCAL2) anbeextended formultiple soures. Considering
paketsfromdierentsouresarein thesamegeneration,synhronizationisneessaryand
isnotaddressedhere.
3 Casading Wave Model
Figure1: Wavemodel
The entral fous of this artile is the study the behavior of the above algorithms in
wireless networks with are modeled with disk-unit graphs. The dynami behavior of the
networkisrepresentedingure1: thenodes(whitedisks)arerandomlydistributedinaline
topology,thesoureisoriginatingnewvetors,(additional dimensionsof thevetorspae)
and these new dimensionspropagate in the network as waves. We represented the eet
ofsome suessivetransmissions t0, t1, t2, t3 andt4. The eet ofonetransmission anbe summarized by howit inreases (or does not inrease)the dimension spae of the nodes.
Here,initialtransmissiont0andthesuessivetransmissiont1fromthesoure,inreasethe
dimensionoftheneighborsofthesoure(whihbeame1). Transmissiont2fromaneighbor
ofthesoure,propagatesfurther alinearombinationoft0 andt1. Thisproessontinues,
and d transmission from the soure in generalpropagates as seen in gure 1. Hene we
have a asade of waves. Note that this ours in our ase, beause we haveone soure
andaline topology, whereanyvetorthat anodereeivemustbealinearombinationof
thevetorsalreadyreeivedatanyleftneighbor. Preiseanalysisofthepropagationofthe
INRIA
waveisparamounttotheharaterizationoftheperformaneofalgorithm1. Suhanalysis
ispresentedinsetion4.
4 Analysis of performane
4.1 Wave propagation model
In this setion, we provide an analysis of the propagation waves in the network oding
shemedesribedin setion2.2.1. Eahwaveisdenedbytheborderbetweenareaswhere
nodes have reeived samevetor spae ombination of pakets, asshown on gure 1. In
partiular thekth waveindiates the limitbetweenthe areawhere nodes havereeived a
vetorspaeof kdimensionwith theareaofthenodes whih havereeivedavetorspae
ofk−1 dimension.
In1D, thewavepositionsare determinedby theposition of theborderon thepositive
realaxis: xk(t)is theabsissaofthekthwaveat timet, withoriginx= 0at theloation
ofthesoure. Wealsoassumethattheorderingxk(t)> xk+1(t)atalltimes. Themodelis
thefollowing:
Soure transmission: thesouretransmitspakets atrateλν where ν is thenode
densityandλ <1 (averageneighborsizeisM = 2ν).
Neighborsofthesoure: notethatthenewreatedwavesatuallydepartfrompo-
sitionx= 1,sinetheallthediretneighborsofthesourereeiveeverytransmission ofthesoure.
Node retransmission: eah node retransmitsat rate1. Hene note that in aarea
oflengthunitandtimeunit,therelaynodestransmitat aPoissonrateofdensityν.
Wave propagation: whenarelay nodeat position y transmits avetor, twoases
an our. If there atually exists a losest wave border xi suh that xi > y and xi−y <1,thenthetransmissionmovesthewaveborderatpositiony+ 1. Otherwise
transmissionhasnoeet onexisting waves.
Notie that wavepropagation may break the order of the xk, in whih ase the indexing
wouldneedto beadjusted.
4.2 Atual performane
From now we restrit our analysis to 1D unit disk graph model. Let onsider anode at
distane x from the soure node. Let∆(x, Q) bethe average time between the rst wave
and thelast wave. Forexampleifx≤1 we have∆(x) = 2Qλ , sine whenneighboringthe soure,thenodeexperieneanaveragetimeof
2
λM betweentwowavegenerations(i.e. two datapakettransmissionsbythesoure)wherethegenerationsize DisMQ.
Theorem1 The average number of paket transmission by a node at distane x to the
soureis∆(x, Q) + 1.
RR n°6120
Proof: The node starts to shedule enoded paket when it reeives the rst wave and
ontinueaslongasit hasnotreeivedthelastwavewhih arrivesin average∆(x, Q)time
unit later. Sine the enoded paket generation is Poisson of rate 1, the average number
ofenoded paket transmittedduring∆(x, Q)isexatly∆(x, Q). Thereforethenumberof
paketsheduledis∆(x, Q) + 1. The+1denotesthelastpaketwhihissheduledduring
the∆(x, Q)periodbut isatuallytransmitted aftertheperiodaordingtoalgorithm.
Corollary1 Thetotalaveragenumberofpakets(dataandenoded)transmittedbyallthe
nodesbetween distane0anddistaney isequal toQM+M2 Ry
0(T(x, Q) + 1)dx.
Proof: The term QM denotes the data pakets transmitted by the soure. The other paketsaretheenodedpaketgeneratedbytherelaynodeswherethedensityν= M2.
Whenthenodeisnotneighborofthesoure,i.e. x >1,wehavetheexpression∆(x, Q) =
2Q
λ + (x−1) (sl−sf). where sf istheaverageslownessofrstwavepropagationandsl is
theaverageslownessofthelast wave.
Theorem2 Wehavesf ≥ M2 andsl≤ (116λ
−λ2)M
Beauseλ <1 thetimegapbetweentherstwaveandthelast waveisproportionalto
β
M. WhenM → ∞thetimegapbeomeszero,thusthewavepropagationproessdoesnot bringmuheetontheperformane.
Corollary2 The total average number of paket transmittedby nodes between distane 0
anddistaney isequal toQM+ (2Qλ + 1)M y2 +O(y1−1
−λ)whenM → ∞.
Corollary3 Theaveragenumberofretransmissionsperdatapaketbetweendistane0and
distaney isequal to1 + (2λ+Q1)y2+O((1y−1
−λ)QM).
If thenetworkis limited todistane y to thesoure: N = yM2 assuming thesoure isthe
leftmost node on the network, we have a retransmission ratio equal to 1 + (λ2 + Q1)MN + O((1 N
−λ)QM2)whihisarbitrarilyloseto2MN whenN, M→ ∞andλ→1.
4.3 Wave propagation slowness
Inthis setion we useν = M2 theuniform densityof node perunit lengthin the 1Dunit
graphmodel.
4.3.1 Propagation ofthe rst wave
Theorem3 Therst wavemovesataverage speedvf = ν2 lengthunit pertimeunit.
INRIA
Proof: the relay nodes that move the rst wave are those whih are betweenx1(t)−1
and x1(t). This nodes reate a transmission rateof ν transmissions per time unit. Eah transmissionmovesthewavebyanaveragetranslationof
1
2 unitlength.
Corollary4 The average slowness of the rst wave satises sf ≥ 2ν timeunit per length
unit.
Proof: wehavetheinequalitysf ≥v1
f
sinetheinversefuntionisonvex.
4.3.2 Propagation ofthe last wave
Theorem4 Thelastwave propagatesatan average slownesssmaller than
8λν
1−λ2 timeunit
perlengthunit.
The proof of this theorem is based on lemmas that study an upper bound of the atual
system.
4.3.3 The upper-bound system
Byupperboundsystemweonsiderasystemwherewavespositionsyk(t)arealwayssmaller
thanatualwaveposition: yk(t)< xk(t)foralltandk. Wedesribesuhsystem.
Letℓbeanumbersuhthat 1ℓ <1−λ. Weuttherealaxisin equalintervaloflength
1
ℓ. Wewillforethewavepositionsyk oftheupper-bound systemto beintegermultipleof
1 ℓ.
(unhanged)Thesouretransmitspaketsatrateλνwhere M2 isthenodedensityand λ <1
Thenewreatedwavesareat absiss1−1ℓ.
(unhanged) Relay nodes transmit at a Poisson rate of ν per length unit and time
unit.
When a relay node transmits at absiss y ∈]k−ℓ1,kℓ] for k < ℓ, it moves a wave at
position
ℓ−1
ℓ toposition1 +k−ℓ1. Iftherewasnowaveatposition ℓ−ℓ1,thentherelay
nodetransmissionhasnoeet onwavepositions.
When a relay node transmits at absiss y ∈]k−ℓ1,kℓ] for k ≥ ℓ, it moves a wave at
position
k
ℓ toposition1 +k−ℓ1. Iftherewasnowaveatposition kℓ,thentherelaynode
transmissionhasnoeet onwavepositions.
Notie that thewavesin theupper bound systemmoveless frequently than in theatual
system,andwhentheymove,itisonashorter distane. Thereforethisisanupper-bound
system.
RR n°6120