multimedia

multimediastreaming

JulienBourgeois,AdrienHenriet

Laboratoired’Informatiquedel’Universit´edeFranche-Comt´e(LIFC)

CentredeD´eveloppementMultim´edia25201Montb´eliardCedex,France

{Julien.Bourgeois,Adrien.Henriet}@univ-fcomte.fr

JeromeLacan,TanguyP´erennou

ENSICA-LAAS,

1,placeEmileBlouin,31056Toulousecedex5{jerome.lacan,tanguy.perennou}@ensica.fr

Abstract

Thisarticlepresentsanoriginalmethodtovalidatetheef-ficiencyoftheshortlengthmaximumseparabledistanceforwarderrorcodeswhenusingbroadcastingmodein802.11b/gmultimediastreaming.Ourmethodisbasedonrealtracesonwhichtheforwarderrorcodewillbemapped.Then,thequalityofthestreamingisevaluatedbycalculatingthePeakSignal-to-NoiseRatio(PSNR).

KEYWORDS

Multimedia,FEC,errorcodes,simulation,wireless,stream-ing

Introduction

Thedevelopmentofwirelessnetworksbandwidthischang-ingthewaytoaccesstomultimediacontents.WiFi,alsoknownas802.11b/g,isthemostpopularlocalwirelessnet-work.WiFicancommunicateaccordingtotwomodes:pointtopointandbroadcast.Pointtopointcommunicationsofferreliabilitybyensuringretransmissionoflostpackets.ThelostpacketsretransmissionismanagedbytheMAClayerof802.11b/g.Whenusingbroadcastingmode,theMAClayerdoesnotmanageretransmissionoflostpackets.Therefore,inavideostreamingsession,errorswilloccurandvideoqual-itywilldecrease.Inthisarticle,weproposetoevaluatethebenefitofusingadaptiveForwardErrorCode(FEC)bycal-culatingthePeakSignal-to-NoiseRatio(PSNR)ofthetrans-mittedvideos.Thisexperimentbasedonrealtracesensuresaresultveryclosetothereality.

Inthefollowingsection,wepresentconventionalerrorcon-trolmechanisms.Section,discussestheefficiencyofshort-lengthMDSFEC.Thisefficiencyiscalculatedusingrealex-perimentsthatgeneratetraces.Simulationsusethesetracestoobtainresults.Sectionconcludesthearticleanddescribesfutureworktobedone.

ConventionalErrorControlMechanisms

Inordertoensurepacketstransmissionreliability,twoessen-tialtypesofmechanismscanbedistinguished[3]:•reactivemechanisms(ARQ:AutomaticRepeatre-quest):whenapacketislost,thetransmitterretransmits

it.Thepacketlossindicationcanbecarriedoutbytheemissionofpositiveornegativeacknowledgments.•proactivemechanisms(FEC:ForwardErrorCorrec-tion):thetransmittersendsredundancypacketsapriori,makingitpossibleforthereceivertorecovercertainlostpacketsthankstotheredundancypackets.Thesepack-etsareusuallycalculatedbyusingblockerrorcorrectingcodes.Thesecodesgeneraten−kredundancypacketsstartingfromasetofkinformationpackets(tobepro-tected)sothatthereceivercanrecoverthekinformationpacketsassoonasithasreceivedasufficientnumberofpacketsamongallthesentpackets.

ItshouldbepointedoutthatthesetwomechanismsareoftencombinedbyusingacknowledgmentstoadjustthequantityofredundancyofFECcodes.SuchacombinationisoftencalledHybridARQoftypeI(packetsalreadyreceivedbythereceiverarenotcombinedwithredundancypacketsreceivedaftertheretransmissionrequest)oroftypeII(allpacketsthatrelatetothesameblockofkpacketsareusedbythedecoder).Beingrelativelyflexibleandeasytoimplement,reactivemechanismsarethemostusedinpackettransmissions(inTCPforexample).Proactivemechanismsareparticularlyusedinmulticasttransmissions(wherearedundancypacketallowstocompensateforthelossofadifferentpacketforeachofthereceivers).Inaddition,real-timetransmissionswhichdonotallowdelaysduetoretransmissions,oftenuseproactivemechanisms.

Inthecontextconsideredhere,i.e.real-timemulticastvideotransmissions,theuseofFECisclearlythebestsolution.TwomainfamiliesofFECcanbedistinguished:theMaxi-mumSeparableDistance(MDS)codesandtheLow-DensityParity-Check(LDPC)ones.TheMDScodes,oftenbasedonReed-Solomoncodes,havethepropertythattheinformationpacketscanberecoveredassoonasanykpacketsamongthensentonesarereceived.However,forlonglengths(n>500),thesecodes,whichhavequadraticencodinganddecodingcomplexities,becometoocostlyinCPU.Ontheotherhand,LDPCcodeshavelinearcomplexityandthencanbeusedwithlonglengths[4,7].However,theyneedmorereceivedpackets(between5and20%[6])thanMDSonestorecovertheinformationpackets,especiallyforshortcodes.AstrongconstraintonFECusedforreal-timetransmissionsisthattheymusthaveashortlength.Indeed,thedecoder

mustnecessarilywaitthereceptionofthelastpacketstobe-ginthedecoding.ThenitappearsthattheFECcodeswhichmustbeconsideredhereareshort-lengthMDScodes.NextSectionspresentanevaluationofthesecodes.Evaluatingtheshort-lengthMDSFECefficiencyToevaluatetheefficiencyofFECcodes,andextremelyim-portantpointistoconsiderarealisticrepresentationofthechannelintermsofpacketlossrateandlossesdistribution.Evenifsomemodelingofthe802.11channellosseswerealreadyproposed(seee.g.[8]),wepreferredconsiderrealtracestoevaluatethecodes.Experiments

Aplatformofwirelesstestswasinstalledinordertocol-lectIPbroadcastpacketslossmeasurements.Itreliesonan802.11binfrastructuremodenetwork,whichcomprisesoneaccesspoint,10nodeswithan802.11binterface,andonenodewithanEthernetinterface,directlyconnectedtotheaccesspoint.Varioustypesof802.11binterfaces(PCI,USBandPCMCIA)andmaterialarchitectures(PC,laptopPC,PocketPC)wereused.

Methodology

First,aserieof5experimentswascarriedoutduringwhichallthenodesremainedmotionless.Thenodeswereplacedatvariousplaces,withadistanceof0to4wallswiththeaccesspoint.Inasecondseriesof6experiments,4ofthenodesweremobile:eachonewascarriedbyanoperatorwalkingwithoutaspecifiedtrajectorywhilealwaysremainingintherangeoftheaccesspoint.

Ineachexperiment,thefixednodepacmantransmitsamulticastMPEG1/RTPvideo-onlystream,whichconsistedof300to1400bytespacketswithameanbandwidthof500kbps.Eachexperimentisapproximately90secondslong.ThevideostreamisproducedbytheJavaMediaStudioapplicationandisbroadcastedinmulticastonthewirelessnetwork.Thetransmitterisfixedinalloftheexperimentsandremainsclosetotheaccesspoint,withaminimalatten-uationofthesignalonthecorrespondinglink.Theresult-ingMACpacketsareautomaticallyre-emittedbytheaccesspointtowards9nodesoverthewirelessnetwork,and1nodeovertheEthernet.

Thus,eachexperimentmadeitpossibletorecordthese-quencenumbersoftheRTPpacketsactuallyreceivedbyeachreceiver.In802.11b,thereisnoacknowledgementnorretransmissionintheMAClayerforbroadcastedpack-ets[1].Hence,areceivedUDPpacketwastransmittedonlyoncefromthetransmittertowardstheaccesspointandthentowardsthereceivers.Bymakingtheassump-tionthatnopacketislostontheEthernet,theEthernetre-ceiverserveur-mitvallowsthemeasurementofMAClossesovertheuplinkbetweenthetransmitterandtheaccesspoint,thewirelessreceiversallowingtomeasuretheaddi-tionalMAClossesoverthedownlink.

Results

Figure1.Tracesofpacketslosswithfixedsta-tions

Figure2.Tracesofpacketslosswithmobilesstations

The5experiments,whereallthenodeswerefixed,producedsimilartraces,justasthe6experimentswithmobilenodes.Wecanseeherethatthemobilityisn’tasignificantfactorwhoinfluateinpacketlost.Figures1and2givesthetracesofoneoftheexperimentswithfixedstationsandonewithmobilestations(pda,miura,montardoandharfang).Thex-axiscorrespondstothesequenceoftransmittedpack-etsandthey-axisillustratestheburstinessofpacketlossesforeachstation.Thereceivingstationsarelistedontheleft.Thepercentagegivenafterthestationnameisthepacketlossrateobservedforthisstationinthisexperiment.Eachlineindicatesifpacket#xwasreceived(lowestposition)orlost(higherpositions)bythecorrespondingstation.Loss

burstsarerepresentedbytheirlengthalongthey-axis.Togetanideaofthelengthofthebursts,considerthatstationlosslinesare10lossesapart.Inexperiment#1forexample,serveur-mitvdoesnotloseanypacket,whichresultsinastraightline.Incontrastnetlookexperiencesabigburstof32packetsstartingatpacket#15andafewotherlessimportantbursts:thisresultsinanalmoststraightlinewithahighpeakforthe32-packetlongburstandafewsmallerpeakscorrespondingtoburstsof1,2,3or5packets.Finally,pdaexperiencesalotmorelossbursts(255burstshavealengthgreaterthan1,generally2,3,4,or5packetslong),whichresultsinaveryirregularline.Thetenreceivingsta-tionsareclassifiedonthey-axisfromthenearesttothemostdistantwithrespecttotheaccesspoint,exceptformeta-bwhichisclosetotheaccesspointbutreceivesfewpacketsduetohardwarelimitations.

Theexperimentsshowthatpracticallynodataislostontheuplink,andthatmanylossesareobservedonthedownlink.Theyarerelatedtotheusedmaterial,thepositionofthefixedreceiverortomobility.Somesimultaneousburstsattesttheoccurrenceofimportantdisturbancesonthedownlink,whichcouldnotbediagnosed.Thepeaksofveryimportantampli-tuderecordedonmeta-brevealadisconnectionphaseduetoahardwareconfigurationproblem.

Moregenerally,aconsiderablequantityofpacketsarelostinbursts,whichcanbetemporallycorrelatedforagivenre-ceiverand/orspatiallycorrelatedforseveralreceivers.ThepeaksobservedforeachindividualreceiverofFigure2de-notetemporalcorrelation.Spatialcorrelationcanbeob-servedtoowhenpeaksoccurforthesamepackets,asine.g.experiment#1forpackets#2168to#2088,whicharesimul-taneouslylostbyallstationsexceptserveur-mitvandnetlook.Ananalysisofthesecorrelationscanbefoundin[5].

Simulationdrivenbytraces

ToevaluatetheefficiencyoftheFECcodesappliedtoavideosequence,onesolutionistocounttheratiobetweenwell-receivedandlostspackets.Ifthissolutioncanshowthethe-oriticalefficiencyoftheFECcodes,itdoesnotshowtherealefficiencyoftheFECcodesappliedtoavideosequence.TheideaistocalculatethePSNRofthereceivedvideo.Indeedofthemethodology,andthetypeoftrace,wecan’tknowthesizeofeachpacket,andso,can’thaveamorerealisticsim-ulationbasednotonlyonpacket,butondata.Thisisn’talimitation,becauseofthestructureisthesame,andberela-tivelyneararealsituation.

Methodology

Asdescribedinfigure3,oursimulationtooltakestwoinputs:amovieandtheexperimenttracefiledescribedinsection.ThemovieistheForeman,areferencesequencefortestingmultimediaplatform.TheaimofourtoolistocomputethePSNRofthemoviewhenpacketsarelostaccordingtothetracefile.ThisPSNRrepresentstheavailablequalityfortheclientforeachdecodedframeofthestream.Indeed,during

multimediastreamingthelostpacketswillnothavethesameeffectonthevideoquality.Somepacketswillseriouslyaffectthevisualqualityofthevideowhereasotherswillnot.Thisdifferencedependsonvariousparameterslikethekindoflostpacket,themomentinthegroupofpictures(GOP)thepacketislost,etc.Infact,ifthelastP-frameofaGOPislost,thequalitywillnotbeaffectedbecause,justafterthisimage,anintraimagewillbedecoded.AsthelastP-frameofaGOPandthefirstI-frameofthefollowingGOPdonothaveanytemporaldependances,onlyoneframewillbedamaged.AnanotherexampleiswhenapacketislostonanP-framejustbeforeacameramovement.Theresultingimagewillbedamagedbutthecameramovementwilldeletethiserror.ThatiswhyitisnecessarytocalculatethePSNRandnotjusttocountthelostpacketstoevaluatetheresultingvideoquality.

OursimulationtoolcancomputethePSNRwithorwithouttheshort-lengthMDSFEC.Infact,itcanapplyanykindofprocessingonthepackets.Oursimulationtooliscomposedoftwomainboxes:theSimulationBoxandthePSNRBox.TheSimulationBoxiscomposedofmplayeranopen-sourcemovieplayerandtheDamagerwhichisabletoalterthepack-etstransmittedbymplayer.TheDamagerdoesnotrandomlyalterthepackets,ittakesasinputthetracefilegeneratedinsection.Whenapacketisgeneratedinmplayer,theDam-ageraltersitornotaccordingtothetracefile.Itcanalsotakeanerrorcodeinformationintoaccount.Here,weuseourshort-lengthMDSFECasapluginfortheDamager.Thispluginmodifiesthetracefilebycorrectingthelostpacketswhenitispossible.

ThePSNRBoxreceivesonlyimages,notpackets.MPlayerreceivespacketsandgeneratesPortableNetworkGraphic(PNG)images.TheseimagesarethengatheredbythePSNRBoxwhichconvertsthemintoarawformatandthencom-putesthePSNR.TheoutputofoursimulationtoolisatextfilewhichcontainsthePSNRforeachgeneratedimage.Results

Thethreefigures4,5and6showthePSNRofthemovieForemanvisualizedrespectivelybytheclient“coucouyaya”,“meta-b”and“harfang”inthemobileenvironment.StrongorlightFECsrepresenttheminimumsizeoftheredundancyinformationinthecode.ThesolidcurverepresentsthePSNRofvideosequencewithoutanyerrors.ThedottedcurveshowsthePSNRcalculatedwhenstrongFECsareappliedtothepacketswhereasthedashedcurvepresentsthePSNRcalculatedwhenlightFECsareappliedtothepackets.Thedotted-dashedcurveshowsthePSNRofthetransmissionwithouterrorcodes.

“Coucouyaya”receivesthepacketswithfewerrors,thatiswhythethreecurves,PSNRwithstrongFECs,lightFECsandwithoutFECsareveryclose.Theinterestingthingtoseeisthattheuseofstrongcodesorlightcodeswhentherearenolongburstsoferrorsisnearlyequivalent.

“Meta-b”and“harfang”havenearlythesameerrorprofile:lotsoferrors,andlongburstsoferrors.Forthesestation,itisclearthattheFECsenhancethevisualqualityofthevideo.

Figure3.Thesimulationtool

But,whentheburstsaretoolong,eventhestrongFECscannotcorrecttheerrorsandretreivingtheoriginalcontentofthepackets.This,isthesecondinterestingfacttaughtbytheseexperiments.

Thankstotheseresults,wecansaythatevenstrongFECscannotcopewithlongburstsasthestreamwillbeinterrupted.StrongFECsdelaytheinterruptionbutdonotavoidit.Thesecondfactisthatwhentherearenolongbursts,lightFECsaresufficient.Asthestrongcodesneedalotmorebandwidththanthelightones,theideaistopreferablyusethem.But,theproblemoflongburstsremains.Solutionswillbeinves-tigatedlater.

Conlusionandperpectives

Wehavepresentedourmethodtotesttheefficiencyoftheshortlengthmaximumseparabledistanceforwarderrorcodes.Oursimulationtoolisabletocalculatethequalityofanystreamedvideobyusingtracefilesofprevioustransmis-sion.TheresultsshowthattheFECscannotcopewithlongburstsoferror.Theycandelaytheerrorbutatlast,thevideoqualitywilldramaticallydecrease.

Futureworkswillhavetodealwiththisproblem.Theideawewillhavetoexperimentisthefollowing.LightFECsarekepttodealwithshortburstsoferrorandanRTPmixer[2]capableofregeneratinganI-frameisaddingbetweentheserverandtheclientstoovercomethelongburstsoferror.Whenainterruptionoccurs,theRTPmixerwilltransformaP-FrameintoaI-Frame,andthenthevideocancontinuethankstothisnewGOP.Withthismixedsolution(reactiveandproactive),thebiggerpartoftheerrorswillbetreatedwithoutaddingtoomuchtrafficonthewirelesschannel.REFERENCES

[1]ANSI/IEEEStd802.11.WirelessLANMACandPHYSpeci- 40 35 30PSNR (dB) 25 20 15 10 5 0

PSNR referencePSNR with high FECPSNR with low FECPSNR without FEC

400 600 800

1000Image

1200 1400

Figure4.PSNRcomputationforcoucouyaya

fications,1999.

J.Bourgeois,E.Mory,andF.Spies.Videotransmissionadapta-tiononmobiledevices.JournalofSystemsArchitecture,49(10-11):475–484,November2003.

S.LinandD.Costello,Jr.ErrorControlCoding:Fundamen-talsandApplications.PrenticeHall,EnglewoodCliffs,NJ.,1983.

M.Luby,M.Mitzenmacher,M.A.Shokrollahi,andD.A.Spielman.Efficienterasurecorrectingcodes.IEEETransac-tionsonInformationTheory,47(2):569–584,2001.

´valuationdeT.Perennou,J.Lacan,andH.Elnatour.e

m´ecanismesdecontrˆoled’erreurpourdestransmissionsmulti-pointssurdesr´eseauxdemobiles.TSI,Toappearin2005.J.S.PlankandM.G.Thomason.Apracticalanalysisoflow-densityparity-checkerasurecodesforwide-areastorageappli-cations.InDSN-2004:TheInternationalConferenceonDe-pendableSystemsandNetworks.IEEE,June2004.

[2]

[3]

[4]

[5]

[6]

40 35 30PSNR (dB) 25 20 15 10 5 0

PSNR referencePSNR with high FECPSNR with low FECPSNR with low FEC

400 600 800

1000Image

1200 1400

Figure5.PSNRcomputationformetab

40 35 30PSNR (dB) 25 20 15 10 5 0

PSNR referencePSNR with high FECPSNR with low FECPSNR without FEC

400 600 800

1000Image

1200 1400

Figure6.PSNRcomputationforharfang

[7]V.Roca,Z.Khallouf,andJ.Laboure.Designandevaluation

ofalowdensitygeneratormatrix(ldgm)largeblockfeccodec.InFifthInternationalWorkshoponNetworkedGroupCommu-nication(NGC’03),Munich,Germany,September2003.

[8]C.TangandP.McKinley.ModelingMulticastPacketLosses

inWirelessLANs.InProceedingsofMSWiM’03,2003.[9]Mplayer:opensourcemovieplayerforLinux

http://www.mplayerhq.hu.