搜索
1System-LevelPowerManagement:AnOverviewAliIranliandMassoudPedramUniversityofSouthernCaliforniaDeptofElectricalEngineeringLosAngelesCAAbstractOneofthekeychallengesofcomputersystemdesignisthemanagementandconservationofenergy.Thischallengeisevidentinanumberofways.Thegoalmaybetoextendthebatterylifetimeofaportable,battery-powereddevice.Theprocessingpower,memory,andnetworkbandwidthofsuchdevicesareincreasingquickly,resultinginanincreaseindemandforhigherpowerdissipation,whilethebatterycapacityisimprovingatamuchslowerpace.Othergoalsmaybetolimitthecoolingrequirementsofacomputersystemortoreducethefinancialcostofoperatingalargecomputingfacilitywithahighenergybill.Thischapterfocusesontechniqueswhichdynamicallymanageelectronicsystemsinordertominimizeitsenergyconsumption.Ideally,theproblemofmanagingtheenergyconsumedbyelectronicsystemsshouldbeaddressedatalllevelsofdesign,rangingfromlow-powercircuitsandarchitecturestoapplicationandsystemsoftwarecapableofadaptingtotheavailableenergysource.Manyresearchandindustrialeffortsarecurrentlyunderwaytodeveloplow-powerhardwareaswellasenergy-awareapplicationsoftwareinthedesignofenergy-efficientcomputingsystems.Ourobjectiveinthischapteristoexplorewhatthesystemsoftware,vis-à-vistheoperatingsystem(OS),candowithinitsownresourcemanagementfunctionstoimprovetheenergyefficiencyofthecomputingsystemwithoutrequiringanyspecialized,low-powerhardwareoranyexplicitassistancefromapplicationsoftwareandcompilers.TherearetwoapproachestoconsiderattheOS-levelforattackingmostofthespecificenergy-relatedgoalsdescribedabove.Thefirstistodevelopresourcemanagementpoliciesthateliminatewasteoroverheadandallowenergy-efficientuseofthedevices.Thesecondistochangethesystemworkloadsoastoreducetheamountofworktobedone,oftenbychangingthefidelityofobjectsaccessed,inamannerwhichwillbeacceptabletotheuseroftheapplication.Thischapterprovidesafirstintroductiontothesetwoapproacheswithappropriatereviewofrelatedworks.BackgroundAsystemisacollectionofcomponentswhosecombinedoperationprovidesausefulservice.Typicalsystemsconsistofhardwarecomponentsintegratedonsingleormultiplechipsandvarioussoftwarelayers.Hardwarecomponentsaremacro-cellsthatprovideinformationprocessing,storage,andinterfacing.Softwarecomponentsareprogramsthatrealizesystemandapplicationfunctions.Sometimes,systemspecificationsarerequiredtofitintospecificinterconnectionsofselectedhardwarecomponents(e.g.,Pentiumprocessor)withspecificsystemsoftware(e.g.,WindowsorLinux)calledcomputationalplatforms.Systemdesignconsistsofrealizingadesiredfunctionalitywhilesatisfyingsomedesignconstraints.Broadlyspeaking,constraintslimitthedesignspaceandrelatethemajordesigntrade-offbetweenqualityofservice(QoS)versuscost.QoSiscloselyrelatedtoperformance,i.e.,systemthroughputand/ortasklatency.QoSrelatesalsotothesystemdependability,i.e.,toaclassofsystemmetricssuchasreliability,availability,andsafetythatmeasuretheabilityofthesystemtodeliveraservicecorrectly,withinagiventimewindowandatanytime.Designcost2relatestodesignandmanufacturingcosts(e.g.,siliconarea,testability)aswellastooperationcosts(e.g.,powerconsumption,energyconsumptionpertask.)Inrecentyears,thedesigntrade-offofperformanceversuspowerconsumptionhasreceivedlargeattentionbecauseof:(i)thelargenumberofsystemsthatneedtoprovideserviceswiththeenergyprovidedbyabatteryoflimitedweightandsize,(ii)thelimitationonhigh-performancecomputationbecauseofheatdissipationissues,and(iii)concernsaboutdependabilityofsystemsoperatingathightemperaturesbecauseofpowerdissipation.Herewefocusonenergy-managedcomputer(EMC)Systems.Thesesystemsarecharacterizedbyoneormorehigh-performanceprocessingcores,largeon-chipmemorycores,variousI/Ocontrollercores.Theuseofthesecoreswillforcesystemdesignerstotreatthemasblackboxesandabandonthedetailedtuningoftheirperformance/energyparameters.Ontheotherhand,variousI/Odevicesareprovisionedinthesystemleveldesigntomaximizetheinteractionbetweentheuserandthesystemand/oramongdifferentusersofthesamesystem.Dynamicpowermanagement(DPM)isafeatureoftherun-timeenvironmentofanEMCsystemthatdynamicallyreconfiguresitselftoprovidetherequestedservicesandperformancelevelswithaminimumnumberofactivecomponentsoraminimumactivitylevelonsuchcomponents.DPMencompassesasetoftechniquesthatachieveenergy-efficientcomputationbyselectivelyturningoff(orreducingtheperformanceof)systemcomponentswhentheyareidle(orpartiallyunexploited.)ThefundamentalpremisefortheapplicabilityofDPMisthatsystems(andtheircomponents)experiencenon-uniformworkloadsduringoperationtime.Suchanassumptionisvalidformostsystems,bothwhenconsideredinisolationandwheninter-networked.AsecondassumptionofDPMisthatitispossibletopredict,withacertaindegreeofconfidence,thefluctuationsofworkload.Inthischapterwepresentandclassifydifferentmodelingframeworksandapproachestodynamicpowermanagement.ModelingEnergyManagedComputers(EMC)AnEMCmodelstheelectronicsystemasasetofinteractingpowermanageablecomponents(PMC’s)controlledbyoneormorepowermanagers(PM’s.)WemodelPMC’sasblackboxes.WearenotconcernedonhowPMC’saredesigned;insteadwefocuson