Notes on FEC supported congestion control for one to many reliable multicast. Working Draft
- 格式:pdf
- 大小:1.77 MB
- 文档页数:32
ANGEWANDTEMATHEMATIKUNDINFORMATIK
UNIVERSIT¨ATZUK¨OLN
ReportNo.98.334
FECSupportedCongestionControlinOnetoManyReliableMulticast
by
FrankBrockners
August,1998
CenterforParallelComputing(ZPR)
UniversityofCologne
Weyertal80,D–50923K¨olnKeywords:CongestionControl,ForwardErrorCorrection(FEC),ReliableMulticastFECSupportedCongestionControl
inOnetoManyReliableMulticast
FrankBrockners
CenterforParallelComputing(ZPR)
UniversityofCologne
brockners@zpr.uni-koeln.de
TechnicalReportZPR-TR98-334
August,1998
Abstract
ThispaperdescribesthedesignofanewFECfueledrateand
congestioncontroller(RCR)whichistargetedprimarilyatone
tomanyreliablebulkmulticastdatatransfer.Theconceptof
RCRisvalidatedbyanalyticalandsimulationresults.The
heuristicanalysisisbasedonanewextendedmodelforflows,
whichimplementacongestioncontrolalgorithmsimilarto
TCP.ThemaingoalwastodevelopanalgorithmwhichisTCP-
friendlyandcompeteswiththecontrolloopofTCP.Largede-
laysinthecontrolcircuit,scalingissuesandhighpacketloss
probabilitiesareamongthemajorchallengesofareliablemul-
ticastratecontrollerimplementingTCP-faircongestioncon-
trol.Thecontrollerpresentedinthispapershowsthatlayered
forwarderrorcorrection(FEC)redundancycodingnotonlyre-
ducesthecontrol-andretransmissionsinreliablemulticasten-
vironmentsusingautomaticrepeatrequests(ARQ)totrigger
retransmissionsbutalsohelpsacongestioncontrollertocom-
petewithTCP.FECpermitsthereceiverstotoleratelossesina
waythattheyonlyneedtosignalizelosslevelswhicharesig-
nificantforthem.Thishigh-passfilteringofthelosssignalre-
vertsthedisadvantageofreactingslowlytolossintoanadvan-
tageindirectcomparisontoTCP.Thegreaterresponsiveness
withrespecttorateincrementensuresthatTCPalwaysreceives
itsshareofthetotalavailablebandwidth.Simulationsandana-
lyticalanalysisformulticastaswellasunicastsetupsshowthat
alreadyaverymoderatelevel(some%)ofredundancysuffices
tostrengthenaconnectionsufferingfromlongdelaysandhigh
lossprobabilities.
Keywords:CongestionControl,ForwardErrorCorrection,
ReliableMulticast.1Introduction
Congestioncontrolisvitalfortheinternetandisevenmore
importantforreliablemulticastsetups.Besidesfairnessand
throughputconsiderations,flowandcongestioncontrolhas
beenrecognizedasamajorproblemforreliablebulkmulticast
datatransfer.
Existingapproachesforcongestioncontrolschemesinreliable
multicastcanbedevidedintoclosed-loopandopen-loopcon-
trolschemes.Allclosedloopcontrolschemeshaveincommon
thatreceiverssendsomekindoffeedbackinformation(either
positiveornegativeacknowledgements)backtothesource.
Thesourceevaluatesthefeedbackinformationandaccordingly
adjustsitsemissionrate.Agreatvarienceofthepresentre-
liablemulticastframeworksuseclosed-loopcongestioncon-
trol.Inopen-loopscenariosacontrolpathfromthereceiversto
thesourcedoesnotexist.Open-loopsolutionsemployapre-
configuredrate,highlevelsofredundancycodingorrepeated
transmissions.Repeatedtransmissionsdenotesthesendingof
thesameinformationelementsmultipletimes.Thismayei-
therbeatdifferenttimesortodifferentmulticastgroups.The
controllerpresentedinthispaperprimarilyreliesonclose-loop
congestioncontrolbutalsoemploysaspectsofopen-loopcon-
trol,asitutilizesredundancycodingtoincreaseprotocoleffi-
ciency.
Wetakeashortglanceatthedifferentapproachesforflow
andcongestioncontrolforreliablemulticastbeforestudy-
ingourapproachwhichintegratesTCP-fairnessandTCP-
competitiveness.
Sender-initiatedapproachesforflowcontrolthatuseposi-
tiveacknowledgments(ACKs)alwaysfacetheACKimplo-
sionproblemwhilereceiverorientedschemesdosotoointhe
absenceofanNAKaggregationorsuppressionscheme.In
[1]Towsleyetal.demonstratedthatreceiver-initiated(NAK
1based)errorrecoveryusuallyissuperiortosender-initiateder-
rorrecovery.Dependingonthewayhowreceiverinitiatedpro-
tocolssolvetheNAKimplosionproblem,theycanbedivided
into(1)timerbased,(2)deterministicand(3)structure(tree-
)basedapproaches.ExamplesincludeSRM[2]foratimer
based,DTRM[3]fordeterministicandRMTP[4]aswellas
TMTP[5]fortreebasedapproaches.
Ineachoftheseapproachestheemissionrateiseitherre-
strictedtoasmalltolerableleveloraTCP-likecongestion
controlisimplementedinordertopreventorresolveconges-
tion.RMP[6]andRMTPserveasexamplesforTCP-like
windowcontrolledcongestioncontrolschemes.Bothrelyon
positiveACKs,eithertothetokesite(RMP)ortothedesig-
natedreceivers(RMTP).NAKbasedschemessignificantlyre-
duceprotocolcontroltrafficbutlackacontrolsignaltoclock
themselves(likeACKsinTCPdo).Insteadtheyusuallyem-
ploysomekindofofratecontrol.Explicitrateadjustmentis
commonformostoftheapproaches.Beitclose-loopcon-
gestioncontrol(RMTP[4],TMTP[5])oropen-loopcon-
trolwithafixedemissionratelikeSRM.Layeredmulticast
schemes,includingtheonebasedonFECwhichtransfersthe
TCP-principlesofcongestionwindowincreaseanddecrease
togroupjoinsandleavesofthereceivers[7]areexamplesfor
open-loopcongestioncontrol.Aproblemcommontoallcon-
trollersinterpretinglossasacongestionsignalis,thatpacket
lossisverylikelytooccur:Foraheuristicargumentonemay
thinkofuncorrelatedflowsondisjointpathsinthenetwork,
eachwithapacketlossprobabilityof1%.Theoverallpacket
lossprobabilityamountstoasmuchas40%.Lossratesofthis
magnitudedestroyanykindofTCP-likecongestioncontrol[8]
–Aratecontrollerwouldendupwiththeminimumemission
ratepermitted.[9]and[10]giveanindepthanalysisoftheloss
patternoftheMBone.Apossiblesolutionofthisproblemisto
tolerateacertainamountofloss.[11]discussesamonitoring
basedapproachwhereasLTRC[12]directlymeasurestheloss
rateandadjuststherateincrease-decreasefunctionofthecon-
troller.
Theflow-andcongestioncontrollerRCRpresentedinthispa-
peralsousesaTCP-likeemissionratecontrolandredundancy
(FECcoding)toimplementlosstolerance.FECenableseach
receivertofilterthelosssignallocallyandtoreacttotheoc-
curanceoflossindividually.Theactiveuseofredundancyin
thecontrolcircuitcompromisesbetweenTCP-friendlinessand
TCP-competitiveness.
Thispaperisstructuredasfollows:Afterashortoverview
aboutthedistributionmodel,section2focusesonsomeaspects
ofreliablecongestioncontrolanddiscussesthefairnessaspect.
Thediscussionofthedesignofthecontrollerinsection3is
followedbyananalyticalmodelofthecontrollerforthelater
analysisinsection5.Section5comparestheresultsobtainedfromthemodelwiththoseofaprototypicalimplementation
withinthens-2[13]networksimulationtool.
1.1Overviewofthedistributionmodel
Wefollowourprotocoldesigngivenin[14]ofaonetomany
distributionmodelwhichrunsontopoftheIP-multicastser-
viceandistreebased.Itisreliableinawaythatareceiver
eitherreceivesalldatacorrectlyorisabletodetectanunre-
coverablesituation.Theprotocolachievesreliabilitybyusing
both-redundancycoding(FEC)andautomatedrepeatrequests
(ARQ).Itfollowsacompletereceiverinitiatedretransmission
schemeandusesthettlhopcounttoformlocalgroups.The
burdenforcorrectreceptionofallpacketsiscarriedbythere-
ceivers.Bymeansofanexpandingringsearcheachreceiver
findsaso-calledleader,whoisclosertothesourceandhasthe
capabilitytolocallyretransmitpacketswithalimitedhopcount
value.ThisimpliesthattheIPmulticasttreehastobesymmet-
ric.Receiverfeedback(e.g.ARQ)issenttotheleaderonly,
whodecideswhethertoforward,takeactions,orsimplydis-
cardit.Thesenderdoesnotkeepanystateinformationabout
thesetofreceivers.
Thecontrollerthispaperisaboutonlyuses[14]asframework.
Thetinyandstraightforwardratecontrolalgorithmdescribed
in[14]wascompletelyreplaced.Therefore,theresultsand
conclusionsofthispapercaneasilybetransferredtoothermul-
ticastandevenunicastscenarios.
2Somedesignissues
2.1Fairnessconsiderations
Stabilityandrobustnessoftheinternetheavilydependoncon-
gestioncontrolledflows.Acontrollerhastointerpretpacket
lossasacongestionsignalandbackoffitsemissionrateincon-
ditionsofloss.Withaconstantlygrowinginternet,end-to-end
congestioncontrolbecomesavitalfactor.In[15]Floydand
Fallshowthenegativeeffectsofflowswhichareunrespon-
sivetolossandemphasizetheneedforcontinuingtheuseof
end-to-endcongestioncontrol.Therefore,theyproposerouter
mechanismstorestrictthebandwidthofbesteffortflowstoa
disproportionateshareoftheoverallavailablebandwidth.Up
to90%oftoday’sinternetWANtrafficusesTCP[16]astrans-
portprotocol.ThismotivatesnewprotocolstotakeTCPasref-
erenceandadjusttheircongestioncontroltobeTCP-friendly.
[15]definesaflowtobeTCP-friendly,ifthearrivalrateof
theflowdoesnotexceedthebandwidthofacorresponding
TCPconnectionunderthesamecircumstances.Basicallythis
maybeachievedbyacontrolalgorithmsimilartothemethod
employedbyTCP,i.e.byreducingtheemissionratetohalf
2