ABSTRACT Leakage Power Modeling and Optimization in Intercon(5)

Power will be the key limiter to system scalability as interconnection networks take up an increasingly significant portion of system power. In this paper, we propose an architectural leakage power modeling methodology that achieves 95-98 % accuracy agains

AdaptiveOblivious

Conservative

Aggressive

Figure4:Designspaceofpower-awarebu erpolicies.simplestatistic–whentherearemorewritesthanreadstoabu erinatimewindowW,Nisincrementedtillithitsanupper-boundNhigh.Otherwise,itisdecrementedtoalower-boundNlow.Theintuitionisthatwhenbu erwritesoutnumberreads,thebu erpoolisbuildingup,withfewerandfewerfreebu ers,soanadaptivepolicyshouldbelessaggressiveinswitchingbu erstoinactivemodeinordertoenhancenetworkperformance.Conversely,whenmore itsareleavingratherthanenteringtherouter,anadaptivepol-icycanmoreaggressivelyswitcho bu ers,guessingthatfewerwillbeneeded.

4.2Circuit-levelmechanisms

Power-awarebu ersrequirecircuit-levelmechanismsthatallowbu erstobeputintoinactivemodeforleakagepowersavings.Severalcircuit-levelmechanismshavebeenpro-posedforleakagepowersavingsinSRAMs[4,7],targetedformicroprocessorcaches.Sincerouterbu ersareusuallyconstructedwithSRAMs,thesecanbereadilyappliedtopower-awarebu ers.

Thecharacteristicsofcircuit-levelmechanismsthatarecriticaltopower-awarebu ersare:(1)transitiondelay-thetimeittakestoswitchabu erbetweenthenormalop-eratingmodeandtheinactivemode;(2)transitionenergy-thedynamicenergyincurredeachtimetoe ectatran-sition;(3)leakagepowersavings-thedi erencebetweentheleakagepowerincurredatnormaloperatingmodeandthatatinactivemode;and(4)datapreservation-whethertheinactivemodepreservesthecontentsoftheSRAMs,i.e.whetherthiscircuittechniquecanbeappliedtobothsingleanddoublepower-awarebu erpolicies.

Inthispaper,wechoosetwocircuit-levelmechanismswithfairlydi erentcharacteristics–Drowsy[4],andGatedVddSRAMs[7].DrowsySRAMshavefastertransitiondelaysthanGatedSRAMs,preservesdatacontent,butdeliverslessleakage2energysavingsintheinactivemodeasshowninTable6.Bothtechniqueshavenegligiblee ectontheaccesstime.

5.EXPERIMENTALRESULTS

WeextendaC++networksimulatortoinvestigatethepower-performanceofpower-awarebu ers[5].Thesamesetofrouterparametersasthatinsection3withan8-by-8meshin0.07µmtechnologyisused.Hereaveragelatency,tencyreferstothetimefromthecreationofthe rst itofthepackettilltheejectionofitslast itfromthenetworkatthedestination,throughputreferstotheinjectionrateatwhichaveragenetworklatencyexceedstwicethelatencyatzeronetworkload,andleakage2

asWhileweassumedthecharacteristicsofitpublished,thatitdoesnotpreservedataGatedininactiveVddSRAMSmode,poorercanbetransitionsizedtodelay.

ensuredatapreservationthoughwithapowersavingsisexpressedasapercentageofthetotalleak-agepowerconsumedbyrouterbu ers.Simulationsarerunfor1millioncycles.

LeakagepowersavingsofLookaheadpolicy.Fig.5comparesthee ectivenessoftheconservativeLookaheadpolicy(N=10forGatedVdd,and1forDrowsycells)againsttheidealpolicies.Ideal-Doublesavescloseto100%ofbu erleakagepower,sinceitonlykeepsabu eractiveduringaccesses.Ideal-SinglegetssavingsclosetothatofIdeal-Doubleatlowtra cworkloadsas itsdonotstayinbu ersforlong.Astra cincreases,however,notshut-tingbu erso whentheyareoccupiedin-betweenwritesandreadsresultinalmost10%lessleakagepowersavings.Asimilardi erenceisobservedbetweenLookahead-SingleandLookahead-DoublewithDrowsycells.

With256 it-bu ersateachrouterinputport,Lookahead-SinglesavesmoreleakagepowerwithGatedVddratherthanDrowsycells.WhilethelongtransitiondelayofGatedVddresultsinalargeN=10,potentiallyleadingtoupto9fewerbu ersturnedinactive,thisisoverwhelmedbytheremain-ingsubstantialnumberofbu ersthatcanstillbeleveraged.Thus,withlargebu ers,thehigherleakagepowersavingsperSRAMcellofGatedVddleadstohigheroverallnetworkpowersavingsascomparedtoDrowsySRAMs.

Theconverseishowevertruewithsmallerbu ers(Fig.6).Here,thelargeNofLookahead(GatedVdd)constrainsthenumberofbu ersthatcanbeturnedinactive,andthelowtransitiondelayofDrowsycellswinover.Notethatastraf- crateincreases,however, itsoccupybu ersforalongertime,soLookahead-Single(Drowsy)isunabletoexploititsfasttransitiondelay.Lookahead-Double(Drowsy)howeverleveragesthisforhigherleakagepowersavingsathightra cinjectionrates.

Leakagepowersavingsofaggressiveandpredic-tivepolicies.WesimulatedsingleLookahead-Aggpoli-cies,withalookaheadwindowNshortenedfrom10to4and2,forGatedVdd.ThePredictivepolicysimulatedhasW=10,Nlow=1,Nhigh=2.Fig.7showsthatasexpected,Lookahead-AggimprovestheleakagepowersavingsofLooka-head,pushingsavingsupto81%atlowtra c.Predictivepushesitevenfurther,upto88%savingsatlowtra c.Evenatveryhightra cloads,Predictivestillsaves71%leakagepower,asitbetteradaptstoactualutilization.Thisshowsthatevenasimpleadaptivepolicycanoutperformobliviouspolicies.

Performanceimpactofpower-awarebu erpoli-cies.Lookahead,beingaconservativepolicy,doesnothaveanimpactonperformanceasitalwaysensurestherewillatleastbeanactivebu eravailableawaitinganarriving it.However,theaggressiveLookahead-AggandPredictivepoli-ciescanpotentiallycauseperformancepenalties.Fig.8sim-ulatesthelatency-throughputperformanceofthesetwopoli-cies,showingnegligibleperformancedegradationforbothpoliciesascomparedtoanetworkwithnopower-awarebu ers.

6.CONCLUSIONS

Wehaveproposedamethodologyformodelingleakagepoweronthearchitecturelevel.Tofacilitatetheuseof

thismethodology,wewilldistributetheI

leaktablesonline.WeherealsoincorporatedournetworkarchitecturalleakagepowermodelsintoOrion[9]soarchitectscaneasilyfactorindynamicandleakagepowerestimateswhenevaluating