mobicom11-peng-greenbs-final

Traffic-DrivenPowerSavinginOperational3GCellularNetworksACMMobicom2011LasVegas,Nevada,USAChunyiPeng1,Suk-BokLee1,SongwuLu1,HaiyunLuo∗,HewuLi21UniversityofCalifornia,LosAngeles2TsinghuaUniversityUCLAWiNGSurgingEnergyConsumptionin2G/3G0.5%ofworld-wideelectricitybycellularnetworksin2008[Fettweis]~124Twhin2011(expected)[ABI]CO2emission,comparableto¼bycarsOperationcost,e.g.,$1.5BbyChinaMobilein2009Risingenergyconsumptionat16-20%/yearMoore’slaw:2xpowerevery4~5yearsby2030Mobicom20112CPeng(UCLA)[Fettweis]:G.FettweisandE.Zimmermann,ICTenergyconsumption-trendsandchallenges,WPMC’08.[ABI]:ABIResearch.Mobilenetworksgogreen–minimizingpowerconsumptionandleveragingrenewableenergy,2008.UCLAWiNGEnergyConsumptioninCellularNetworks0.1wX5B=0.5GW1~3kwX4M=8GW10kwX10K=0.1GW90%(~99%)CellularInfrastructure10%(~1%)MobileTerminals~80%byBSesThekeytogreen3GisonBSnetworkMobicom20113CPeng(UCLA)Source:NokiaSiemensNetworks(NSN)UCLAWiNGOutlineOverviewProblemandrootcauseExistingsolutionsOursolutionCharacterizing3GdynamicsExploitingdynamicsindesignWorkingwith3GstandardsEvaluationSummaryandInsightsMobicom20114CPeng(UCLA)UCLAWiNGCaseStudyinaRegional3GNetworkNon-energy-proportionality(Non-EP)totrafficloadMobicom20115CPeng(UCLA)IdealCurrentLoad:(#linkin15min)Power(Kw)Power-loadcurveinabigcitywith177BSes(3GUMTS)UCLAWiNGRootCauseforEnergyInefficiencyMobicom2011CPeng(UCLA)6EachBSisnon-EPPBSPtxPmiscPtxPmiscPmiscLargeportionofconsumedenergyeven@zerotrafficloadaslongastheBSison.PtxPower(w)loadl500l0005002000UCLAWiNGRootCauseforEnergyInefficiencyTrafficishighlydynamicFluctuateovertimeBeunevenatBSesMobicom2011CPeng(UCLA)7Largeenergyoverheadatlighttraffic=non-EP.TurnoffBScompletelytosavemoreenergy!LowusageatnightUCLAWiNGGoalsandChallenges1.System-wideenergyproportionality(EP)HowtodesignEPnetworkwithnon-EPBScomponents?1.NegligibleperformancedegradationHowtomeetlocation-dep.coverage&capacityrequirements?2.3GstandardcomplianceHowtosupportenergyefficiencyw/ochanging3Gstandard?Mobicom20118CPeng(UCLA)UCLAWiNGExistingSolutionsOptimization-basedapproachPracticalissuesunaddressedTheoreticalanalysisonlyComponent-basedapproache.g.,oncooling,poweramplifierNosystem-widesolutionComplementourapproachCleanslatedesigne.g.,C-RANRe-architectthe3GinfrastructureCommunicationandcomputationintensiveminE(x)subjecttoC1,C2…constraintsMobicom20119CPeng(UCLA)UCLAWiNGOurSolutionRoadmapCharacterizingmulti-dimensionaldynamicsExploitingdynamicsindesignWorkingwith3GstandardsEvaluationMobicom201110CPeng(UCLA)UCLAWiNGTemporalDynamicsisPervasiveLowaverageutilizationunderdynamicloadPeak-to-idletrafficis5at40~80%BSesLargesavingpotentialforquiethoursMobicom201111CPeng(UCLA)UCLAWiNGTemporalDynamicsisStableTemporalpatternisnear-termstableTrafficateachBSisquitestableonadailybasisAutocorrelationwith24-hourlagis0.92at70%BSesDay-to-dayvariation(|Curr–Prev|/Prev)is0.2at70%BSesMobicom201112CPeng(UCLA)Region1Region2Region3Region470%BS0.920.930.940.9490%BS0.830.830.900.90Autocorrelationwith24-hour-lagTrafficispredictable.CasefortrafficprofilingUCLAWiNGSpatialDynamicsDeploymentvariesatlocationsDenseinbigcities20+neighbor(1KM)Mobicom2011CPeng(UCLA)13RichBSredundancyensurescoverage.UCLAWiNGSpatial-temporalDynamicsTrafficisalsodiverseatvariouslocationsPeakhoursaredifferentMultiplexinggain~2atpeakhoursLowerboundfortheratioofcapacitytotrafficMobicom2011CPeng(UCLA)14Multiplexinggain:sum(maxTraffic)/sum(traffic)LargesavingpotentialevenatpeakhoursUCLAWiNGRoadmapCharacterizingmulti-dimensionaldynamicsExploitingdynamicsindesignWorkingwith3GstandardsEvaluationMobicom201115CPeng(UCLA)UCLAWiNGIssueI:HowtoSatisfyLocation-dependentCoverage&CapacityConstraints?OnceaBSturnsoff,clientsinitsoriginalcoverageshouldstillbecoveredMobicom2011CPeng(UCLA)16✗✗✗✗✗✗✔✔✔✗✗Evenifthetotalcapacityisenough,itmayfailtoservemobileclientsduetocoverageissue.providelocation-dependentcapacityUCLAWiNGSolutionI:BuildingVirtualGridsDivideintoBSvirtualgridsBSeswithinagridcovereachotherDecouplecoverageconstraintLocation-dependentcapacitymeetslocation-dep.trafficVirtualBSGridsMobicom201117CPeng(UCLA)turnon/offBSess.t.cap=loadjiri+d(i,j)Rirj+d(i,j)Rj✔✗✗✗✗✗✗✗✗✗✗✔✔✔UCLAWiNGIssueII:HowtoEstimateTrafficLoad?Atwhattimescaleistrafficloadpredictable?Exploitnearperiodicityoverconsecutivetime-of-the-dayWhattoestimate?Instantaneoustrafficloadvs.trafficupper-envelopeChoicesbetweenaccuracyandover-estimateTradeoffbetweenenergyefficiencyandmiss-rateMobicom2011CPeng(UCLA)18UCLAWiNGSolutionII:ProfilingEstimatetrafficenvelopeviaprofilingLeveragenear-termstabilityReduceruntimecomputation&communicationReducemissrateviatrafficenvelopeestimationMobicom201119CPeng(UCLA)Sum24intervalsStatEstimateS,D,EVOutput?S(i,k)(1)?S(i,k1)S(i,k)?D(i,k)(1)?D(i,k1)|S(i,k)?S(i,k)|EV(i,k)?S(i,k)?D(i,k)UCLAWiNGIssueIII:HowtoMinimizeOn/OffSwitches?Frequenton/offswitchingisundesirableLargeramp-uptimewhenonReducedlifetimeforcoolingandothersubsystemsHowoftentoswitchon/off?Over24-hourp