ocarina - ict-arcfire.eu
TRANSCRIPT
OCARINA("OptimizationstoCompelAdoptionofRINA")
MichaelWelzlRINAWorkshop– IndustryDay
i2CATBarcelona22.5.20181
Projectoverview• 5-yearprojectfundedbyNorwegianresearchcouncil,started1October2016;1postdoc+3Ph.D.students– Focusedonperformance.Assumptions:1. RINAneedsto showfantasticperformance,2. RINAcan showfantasticperformance!
• 3mainWPs:cong.control,routing,Internetdeployment– RINAforcesustothinkdifferentlyaboutnetworkalgorithmssuchasroutingandcongestioncontrol
– E.g.,Internet-like"end-to-end"congestioncontrolcouldbeimplementedinaRINAnetwork,butthatwouldbeaverystrangeconfiguration 2
IssuesofInternetCC.• Twomajormistakes:
1. Firstproducecongestion,thenreacttoit• Solution: createameaningful"load"signalthatdoesnotembedaveryspecificalgorithm(givesomefreedomtodesigners)
2.Cluelessaboutunderlyinginfrastructure,bydesign• Solution: useper-DIFloops,workwithback-pressure
• ECNisbroken– Costincurredinthenetworkisadditiveperhop(seeNUMtheory),butcan'tre-mark amarkedpacket
– Better"load"signalinDCTCP-styleusage:instantaneousqueuemarking,countmarks/RTT,orbeforeaqueueevengrows(virtualQ)
3
ProblemsofusingECNas“load”
• ModerncontrollerssuchasDCTCPconvergeathighmarkingprobabilities.
• Thetheory(e.g.NetworkUtilityMaximization(NUM))needsanadditivesignal;aproductvaluedeviatesmuchinhighmarkingprobabilities(>0.04)!
4
OurSolution• ExtendingtheTheory
– Withalotofmathandstabilityanalysisofcourse…• Results:(assumingalogarithmicutility)
Advantages:1. Newsignalisaprettygeneralsolution;justconveys"load",andcould
(relatively)easilybeextendedtomulti-bit2. Newsignalisprobablygoodinputtoload-basedroutingtoo3. REDasanalready-deployedsolutioncanbeused;onlysmallchangesat
sendersandreceivers5
x(1):avg.rateofafive-hopflowx(2):avg.rateofaone-hopflow
Simulationresults
deviation(previoustheory)
Numericalresults
ourmethod
Per-DIFloops:PRISTINEbackground
• AsequenceofDIFsdoingTCPCC.ismuchlikeasequenceofsplit-TCPPEPs→canbebeneficial[1]
• However,controlsusingrecursivequeuebasedfeedbackcanhavestabilityissues(+delayfrommultiplequeues)[2]
• Envisiontoaddressthiswithlogisticgrowthbasedcontrol[3]+new"fixed"ECN
6
1. PeymanTeymoori,MichaelWelzl,SteinGjessing,EduardGrasa,RobertoRiggio,KewinRausch,DomenicoSiracusa:"CongestionControlintheRecursiveInterNetworkingArchitecture(RINA)",IEEEICC2016,KualaLumpur,Malaysia,23-27May2016.
2. DavidHayes,PeymanTeymoori,MichaelWelzl:"FeedbackinRecursiveCongestionControl",13thEuropeanWorkshoponPerformanceEngineering(EPEW2016),Chios,Greece,5-7October2016.
3. PeymanTeymoori,DavidHayes,MichaelWelzl,SteinGjessing:"EvenLowerLatency,EvenBetterFairness:LogisticGrowthCongestionControlinDatacenters",IEEELCN2016,Dubai,UAE,7-10November2016.
WiFiuplink(e.g.videoconference)
• TCP's"sawtoothtest"doesnotmakemuchsensehere– 802.11MACis thishop'scongestioncontrol...butfocusedonsending1frame,nottellingusasendrate
• Couldusebufferdrainrate– perhapsBBRwouldworkwellhere?
• However:knowingbufferdrainraterequirestoalwayssend→workingonamodel-basedapproach 7
Deployment
• WecanconsiderRINA-under-IP,RINAoverlay,andRINA-IPgateways...
• Butwecanalsoconsider"switchingover"!– OnceahostdiscoversthatthewholepathtotheotherendisRINA-enabled,switch
– Today,often,pathsareshort(Google,FB,...arenotfarawayfromyou)
– TCP/IPareonlyrendez-vousprotocols– SomerecentIETFstandardscouldhelp
• AlittleironicJ 8
RelevantIETFworknotstrictlyOCARINA,butstill...
• TransportServices(TAPS)WG:makesappsprotocol-independent– Finished surveyingandcondensingservicesprovidedby:TCP,MPTCP,UDP,UDP-Lite,SCTP,LEDBAT
– NowworkingonAPI+implementationguidance,withAppleamongothers;implementations:Apple,NEAT(opensource)
• ProvisioningDomains(PvDs)(INTAREAWG):– RouterAdvertisement(RA)optionfromfirst-hoprouterconveysFQDNthathostcanusetoretrieveextrainfoaboutnetworkaccesscharacteristicsviaHTTPoverTLSquery
– Applicationsthenselect(vialocalIPaddress)whichPvDtouse,andcanlearnconfig.paramsfortransportlayerandabove 9
Conclusion
Thankyou!
13