文献翻译功耗优化从超低功耗到耗电系统.docx
《文献翻译功耗优化从超低功耗到耗电系统.docx》由会员分享,可在线阅读,更多相关《文献翻译功耗优化从超低功耗到耗电系统.docx(7页珍藏版)》请在冰豆网上搜索。
文献翻译功耗优化从超低功耗到耗电系统
本科毕业设计
外文文献及译文
文献、资料题目:
POWEROPTIMIZATIONS
文献、资料来源:
期刊
文献、资料发表(出版)日期:
2006.3.25
院(部):
专业:
班级:
姓名:
学号:
指导教师:
翻译日期:
外文文献:
POWEROPTIMIZATIONS:
FROMULTRA-LOWPOWERTOPOWER-HUNGRYSYSTEMS
Inthissection,weprovideanoverviewofpoweroptimizationstargetedatfourtypesofsystems:
Energy-scavengingultralowpowersystems,batteryoperatedembeddedandmobilesystems,plugged-indesktopmachinesandhigh-endservers.Whilesomeoftheoptimizationsareusefulacrossthedifferentcategoriesofsystems,domain-specificpoweroptimizationsareeffectiveinbalancingtheoftenconflictingdesignconstraints.Becauseprototypehardwarecanbeexpensivetodevelop,researchanddevelopmentinpower-awaredesigntypicallyrequiressimulation.Dependingonthetechnique,thismayrequireanythingfromdetailedcircuitsimulations,tocycle-by-cyclepipelinesimulations,tocoarse-grained,event-drivensystem-eventsimulations.
UltraLowPowerSystems
Theneedforultra-lowpowerdeviceshasemergedinvariousapplicationssuchasunmonitoredsensornetworksandmedicalprosthetics,wheretheamountofenergythatcanbetransferredtotheimplanteddevicefromanexternalpowersourceislimited.Insuchultra-lowpowerdevices,energymustoftenbescavengedfromtheenvironmentviathermalormechanicalmeanssimilartoold-fashionedself-windingwatches.Theamountofpoweravailableinthiswayissolowthat,inadditiontotheneedforlowdynamicenergyconsumptionduringoperationalphase,itisimperativethatthestaticpowerconsumptioninidlestateisalsonegligible.Ifthisconditionisnotmet,mostofenergyscavengedwillbewastedinstaticpowerdissipationandtherewillbeinsufficientenergytosustainthedeviceoperation.Consequently,ultra-lowpowerdevicesneedtodeployintelligentstrategiestoshutdowncomponentsorportionsofthesystemintolowerstaticpowermodestoconservepowerconsumption.Forexample,[F11]presentsamultiplierdesignthatcandynamicallyshutdownportionsofthecomponentbasedonthebitwidthoftheoperands.Itisalsopossibletoreducesamplingratesofanalogtodigitalconvertersorreducetheirprecisiontoreducepowerconsumptionbasedonqualityofservicerequirements.Significantpowerreductionisalsopossiblebycustomizingthecomponentsforthetargetedapplication.Forexample,anarchitectureforanevent-drivensensorapplicationwithhardwaresupportforinterruptprocessingcanbemorepowerefficientthanasimplemicrocontrollerthatreliesonoperatingsystemkernelsforhandlinginterrupts.Further,hardwareaccelerationblockscanbeusedforcommonsensoroperationssuchaspacketprocessingandforwardingtoimproveperformanceandreducepowerconsumption.Inadditiontocomputationalblocks,memoryorganizationcanalsobecustomized,asillustratedin[F11].
Inmanyoftheultra-lowpowersensingapplications,theperformancerequirementsarequitemodestincomparisontodesktopmachines.IncontrasttothefewGHzclockfrequenciesdeployedindesktopprocessors,thesesystemscanmeetfunctionalrequirementswhileoperatingatspeedsintherangeofafewKHz.Thislowspeedrequirementmakesitpossibletousearchitecturesbasedonsub-thresholdcircuits.Insuchsystems,thecircuitsuseleakagecurrentstochargeanddischargeloadcapacitancesanddramaticallyreduceoverallpowerconsumptionatexpenseofslowoperatingspeeds.Findinganoptimalvoltagetooperatethesystemshouldstrikeabalancebetweentheadditionalstaticpowerconsumedduetothesloweroperatingsystemwiththequadraticpowersavingsduetosupplyvoltagereduction.Inaddition,thedesignofsub-thresholdarchitecturesposevariouschallengesastheyaremoresusceptibletoexternalnoisesuchassofterrorsandaremoreinfluencedbyprocessvariation.
EmbeddedandMobileSystems
Thenextclassofdevicesisthebattery-operatedlowpowerembeddedandmobiledevicesthatcanberechargedorreplacedperiodically.Thesedevicesneedtoprolongbatterylifewhilemeetingotherconstraintssuchassystemcostandqualityofservicerequirements.Someoftheembeddedsystemstradeoffqualityofservicewiththepowerconsumed.Forexample,portablevideoplayerscanreducethevideoqualityorsizeforreducingpowerconsumption.Sincemanyembeddedsystemsaredrivenbyaconstantthroughputrequirementratherthanoverallcomputationtime,itispossibletoexploitthevariabilityinworkloadstoreducepowerconsumption.Forexample,thecharacteristicsofavideostreamthatcanbedecodedbythevideoplayerataratefasterthan30fpscanbeusedtoreducesupplyvoltageorincreasethresholdvoltage.Whilepipeliningandparallelismhavebeenusedinhigh-performancedomainsforimprovingperformance,theycanbeusedforreducingpowerinconstantthroughputapplications,Theadditionalslackobtainedthroughreducedlogicdepthusingdeeperpipeliningcanbeexploitedbyreducingsupplyvoltageinsteadofreducingthefrequency.Theoptimaldepthofpipeliningforpowerconsumptiondependsonvariousfactorssuchasoverheadsofpipelineregister,activityfactorofthecircuit,theclockgatingmechanismusedtoreduceactivityinthepipelineregistersandtheratiobetweenleakageanddynamicenergy.Aparallelarchitecturecanbeusedsimilarlytoreducepowerbyoperatingmultipleinstancesatlowervoltagesinsteadofoperatingasingleunitfasterusingahighersupplyvoltage
Theutilizationofthesystemcomponentsvariesbasedontheapplicationworkload.Consequently,mostcomponentsofembeddedsystemssupportmultiplepowermodes.Typicallyamodewithlowerpowerconsumptionalsoincursmoreperformancepenalty.Forexample,aCPUoperatingatalowervoltagerequiresalongercycletime.Themodetransitionpoliciescanbeimplementedeitherdirectlyinhardware,orwhenmoreflexibilityisrequired,insoftware.Batteryoperatedsystemsalsorelyonimprovingthesynergybetweenthesoftwareandthehardwarecharacteristicsforprolongingbatterylife.Exampleoptimizationsincludeshapingthecurrentdrawnbythesystemtomaximizethebatteryefficiencyandchangingmemorydataaccesspatternsforplacingmorememorymodulesinlowerpowermodes.Similartoultra-lowpowersystems,embeddedsystemsalsorelyonhardwareaccelerationforcommonlyexecutedtaskstoreducepowerconsumption.Forexample,mostportableMPEGplayersincludeanASICfordedicated,efficientMPEGdecoding.Sincemanyembeddedapplicationsexecuteaknownsetofapplications,useofdeterministicsoftware-managedcaches(scratchpadmemories)ispreferredtotraditionalhardwaremanagedcachestoobtainbestperformanceandconsequent,energysavings.
DesktopSystems
Desktopsystemshavetraditionallyfocusedonperformanceasthemaindesignmetric.Consequently,numerousenhancementsforperformanceimprovementhavebeendeveloped,butsometimesatthecostofincreasedpower.Techniquessuchasbranchprediction,valueprediction,pre-fetchingandpredicationhavethepotentialtoimproveexecutiontime,reducingenergypertask,butontheotherhandmayincreasetheamountofwastedenergyduetoexecutionofinstructionsinthewrongpathduetomis-speculationorpre-fetchingblocksthatwillneverbeused.Theaggressivenessofthesetechniquesmustbebalancedagainsttheirpotentialwastefulness.Runtimefeedback-directedtechniquesarealsopossible,butrobust,stablecontrolschemesarenotyetavailable.
Superscalarprocessorsusedinworkstationsandpersonaldesktopsystems(includinghigh-performancelaptops)alsoprovidemanyopportunitiesforruntimeadaptationtoworkloadcharacteristics.[F10]providesasurveyofsomeofthesetechniques.Coarse-grainedadaptation,intheformofDVFSandsleepmodes,isalreadyavailableinmanypersonalcomputers,aswellasserversystems,andisgovernedbytheACPIstandard(AdvancedConfigurationandPowerInterface).Finer-grainadaptationispossibleaswell.Thesizeandcomplexityofmanymicroarchitecturalcomponentssuchastheissuequeue,registerfile,cachesandfunctionalunitsareusuallydesignedtoaccommodateburstsofactivity,butcanoftenbereducedbasedontheirpredictedusageoroccupancy.Asmallerstructurehelpsreducethedynamicenergyduetoasmallerloadcapacitanceandalsoreduceleakageenergyasfewertransistorsareleaking.Breakinglargeorheavilymulti-portedstructuresintodecoupledclusterscanalsohelpreducedynamicenergyaswellaspossiblyimprovingaccesslatency.Dynamicadaptationtechniquesneedtocarefullyevaluatetheoverheadforsupportingadaptationandtheresultinginfluenceonthepowerconsumptionoftherestofthecomponents.Furtherdynamicadaptationtechniquescanreducecostbypermittingthethermalandcoolingmechanismstobedesignedbasedontheaveragecasebehavior.Forexample,on-chipthermalsensorscantriggerpowerreductiontechniquessuchasinstructionthrottlingandvoltagescalingwhenthechiptemperatureincreasesbeyondaspecifiedcriticallevel.
High-EndServers
Incontrasttotheultra-low-power,battery-operateddevicesthatexhibitlongsleeptimesinterspersedwithsparseactivecomputationmodes,serversystemsarepluggedin,usuallyconsistofamultitudeofcomponentsystems,altogetherconsistingofamultitudeofCPUs,disks,memory,board-levelandsystem-levelinterconnectandrouters,etc.,andcansustainprolongedburstsofactivity.Thismeansthatserversystemsareoftenprovisionedforpeakload.Burstscanoccurunexpectedly(consideranewssiteaftersomemajorevent),butkeepingtheentireserverclusteratfullpowerisunacceptable.Electricutilitycostsinadatacenter(includingc