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

Table5:Dynamicandleakagepowerestimatesofanon-chiprouteranditslinks.

Figure2:Leakagepowerdistributionacrossthemajor

functionalunitsofanon-chiprouter:bu ers,arbiters,

crossbarandlinks.Arbiterleakagepowerisnegligibleandnotvisibleinthe gure.

Toexplorethepotentialofpower-awarebu ers,we rstcharacterizenetworkbu erutilizationwiththetra cmodelproposedin[8],withPoissontaskinter-arrivalrate,andself-similarpacketinter-arrivalrateswithineachtaskses-sion.Thisworkloadexhibitsthehightemporalandspatialvariancepresentinmanyreal-lifenetworks.Wesimulatethechip-to-chipnetworkdescribedinSec.3(2virtualchan-nelsperport).Fixed-lengthpacketsof20 itsareassumed.Fig.3graphstheaverageandminimumnumberofidlebu ersastra cincreases.Asexpected,alargenumberofbu ersisleftidleatlowinjectionrates.Interestingly,whiletherearerouterswhosebu ersarefully-occupied(minimumnumberofidlebu ers=0)athighnetworkload,averagebu erutilizationremainsratherlow,withabout85%idlebu ers.Thisisre ectiveofthehighvarianceinthework-loadthatresultsinalargegapbetweenaverageandmax-imumnetworkutilizationthatisinherentinmanyactualworkloads.Clearly,placingtheseidlebu ersinaninactivemodethatuseslessleakagepowerwillresultinsigni cantleakagepowersavings.

4.1Power-awarebufferpolicydesign

Arouterbu erisutilizedinastream-likefashion.Whena itentersarouter,itgetswrittenintoanunoccupiedbu er,andsitstherewhileaseriesofrouteroperationsistriggered:routing,virtual-channelallocationandswitchal-location.Whenitisscheduledtoleavetherouter,the itisreadfromthebu erpool,andthebu eristhenmarkedasunoccupiedandreleasedbacktothefreelist,readytobereusedwhenanew itenterstherouter.

Wetermapolicythatturnsabu ertoinactivemodeonlywhenit’sunoccupiedsingleandonethatswitchesabu ertoinactivemodeanytimeit’snotbeingaccessed,i.e.whenit’sbothunoccupiedandoccupied,double.Toevalu-atethee ectivenessofanypolicy,weneedayardstick–wede netwotheoreticallyideal,thoughunachievable,policies:

Figure3:Averageandminimumnumberofidlebu ers

outof128 its/bu er.

Ideal-Single,thatreducesleakagepowertozeroinstantlyforbu ersthatareunoccupiedwithnoadditionalpowerover-head,andIdeal-Double,thatdoessosimilarlyforbu erswhentheyarebothunoccupiedandnotbeingaccessed.Apower-awarebu erpolicycanbeoblivious,i.e.itdoesnottakecurrentbu erutilizationorworkloadintoaccount;oradaptive,tuningthepolicyaccordingtocurrentutiliza-tion.Itcanalsobeconservative,makingsurenetworkper-formanceisnotimpacted,vs.aggressive,targetingasmuchleakagepowersavingsaspossible,evenifthiscomesattheexpenseofnetworkperformance.

Fig.4showsthedesignspaceofpower-awarebu erpoli-ciesthatweenvision,andseveralsimplepoliciesthatweproposeateachdesignpoint.Eachpolicycantargeteithersingleordoubleleakagepowersavings.First,weproposeaconservativepolicy,Lookahead,thatobliviouslyplacesbu ersinlow-leakagemodeandwakesthemupNcyclesbeforetheyareaccessed.Whena itisreadfromthebu erqueue,thatbu erwillbeswitchedtothelow-leakagein-activemode(iftherearemorethanNfreebu ers),andwhena itarrivesandiswrittenintoarouterbu eratthetailofthequeue,thebu erthatisNcellsaheadwillbeswitchedtonormaloperatingmode.Thepolicyisconser-vativeasitsetsthelookaheadwindowofNtothenumberofcyclesneededtoswitchabu erfrominactivetoactivemode(transitiondelay),soafreebu erwillalwaysbeavail-ablewhen itsarrive,andnetworkperformancewillneverbea ected.Clearly,ifthebu ersizeBislessthanN,ourpolicywillresultinnoleakagepowersavings.Anaggres-sivevariantofthispolicy,Lookahead-AggsimplyshortensNtolessthanthetransitiondelay,tradingo performanceforhigherleakagepowersavings.OurimplementationofLookaheadinsertsanewlyfreedbu erbackattheheadofthefreelist,soanactivebu erhasthehighestchanceofreuse,minimizingtheimpactonnetworkperformancesigni cantly.Asimpleadaptivepolicy,wecallPredictive,usespriorbu erutilizationhistorytopredictfutureusage,adjustingthelookaheadwindowNaccordingly.Weusea