Guaranteeing long processor lifetimes by limiting the on-set of wear-out or hard failures is a critical requirement for all processor manufacturers [3, 4]. In addition, lifetime predictability is also crucial from the point of view of service replacements,
AReliabilityOdometer-LemonCheckYourProcessor!JayanthSrinivasan,SaritaV.Adve,PradipBose,JudeA.RiversDepartmentofComputerScience,UniversityofIllinois,Urbana-ChampaignIBMT.J.WatsonResearchCenter,YorktownHeights,NYsrinivsn,sadve@cs.uiuc.edu,pbose,jarivers@us.ibm.com
TheProblem
Guaranteeinglongprocessorlifetimesbylimitingtheon-setofwear-outorhardfailuresisacriticalrequirementforallprocessormanufacturers[3,4].Inaddition,lifetimepredictabilityisalsocrucialfromthepointofviewofser-
vicereplacements,avoidingdown-time,etc.Thisisparticu-larlytrueforvendorsofhigh-endserverswherepredictablereliabilityisanimplicitrequirement.However,processorsrunningdifferentapplicationsexhibitalargerangeoflife-timesbasedonapplicationcharacteristicslikepowercon-sumptionandtemperature.Moreimportantly,thisrangeisrapidlyincreasingaswescaletodeepsub-microntechnolo-gies[4].Hence,duetothevarietyofapplicationsrunonatypicalprocessor,andtheinherentlyunpredictablenatureoftheseapplications,guaranteeingprocessorlifetimesisbe-comingincreasinglydif cult,ifnotimpossible.Onecurrentapproachtohandlingthisunpredictabilityistoover-designsystemsfromareliabilitystandpoint.However,thisisex-pensive,andcurtailsperformanceandyield,makingdesignmoreexpensiveandeatingintopro ts.Thereisaclearneedtodynamicallytrackprocessorwear-outandpredictharderroroutages.Inhigh-endserverclassprocessors,thiscanbeusedtopreemptservicefail-uresminimizingexpensiveperiodsofdown-time.Eveninlower-endsystems,trackingreliabilitystatisticsdynami-callycanbeausefulfeature.Forexample,inlargeproces-sorfarms,bymonitoringwear-out,wecanbetterbalancelifetimesacrossmanyprocessorsbyintelligentlydistribut-ingapplications.Additionally,hardwareadaptationtech-niqueslikeDynamicReliabilityManagement(DRM)[3]havebeenproposedtoallowthemicroarchitecturetotrade-offperformanceandreliabilityonthe y.Suchtechniquesalsorequiredynamicreliabilitymonitoring.Finally,thereisalargemarketforusedhardware,particularlyamongnetworkprocessorsandhigh-endserverprocessors[1,2].Knowledgeofpre-existingwear-outandusageinformationforsuchhardwarewouldbeimmenselyusefulinpricingandwarrantycerti cation.thereliabilityodometerwoulddynamicallymonitorproces-sorusageandstorewear-outinformation.Everyeventintheprocessor’shistorywouldincrementtheodometerappropri-ately.TheodometercanbeimplementedeitherintermsofFailures-in-Time(FITS)orMTTF.In[3],wedescribeRAMP,amicroarchitecturalmodelforprocessorreliabilitymeasurement,thatmodelstheimpactofapplicationcharacteristicsondifferentproces-sorfailuremechanisms(electromigration,stressmigration,gate-oxidebreakdown,andnegative-biastemperatureinver-sion).RAMPishighlymodularindesign-asaresult,otherfailuremechanisms,ifimportanttothehardwarebe-ingmonitored,caneasilybeincluded.Additionally,RAMPalsomodelsapplication-independentwear-outeventslikethermalcycleswhicharisefromtheprocessorbeingturnedonandoff.RAMPcanbedirectlyappliedintheuseoftherelia-bilityodometer.MostofthehardwarerequiredbyRAMP(orthereliabilityodometer),liketemperaturesensorsandactivitycounters,alreadyexistincurrentprocessors.Theonlyadditionalhardwarerequirementswouldbesomesim-plecombinatoriallogicandalittlestorage.Evenifprecisefailuremechanismmodelsarenotavail-able,asimplerversionofthereliabilityodometercanbeimplementedusingthermalsensors.Mostwear-outfailuremechanismsfollowanexponentialdependenceonproces-sortemperaturebasedontheArrheniusrelationship[4].Asimplecounterwhichincorporatesthisexponentialeffectoftemperaturewouldgiveareasonableestimateofpriorpro-cessorusage.Suchareliabilityodometerwouldnotonlybeofusetolemoncheckusedhardware,butalsoinadaptivecontrolwhichcanbudgetreliabilitybasedonapplications,perfor-mancerequirements,andcostconsiderations.References
TheSolution
Toaddresstheseproblems,weproposeareliabilityodometer.Muchliketheodometerfoundinautomobiles,
ThisworkwasperformedwhileJayanthSrinivasanwasasummerin-ternatIBMT.J.WatsonResearchCenter.[1]DataleaseSystems.doc.guandang.net.[2]NetworkHardwareResale.doc.guandang.net.[3]J.Srinivasanetal.TheCaseforLifetimeReliability-AwareMicroprocessors.InProc.ofthe31stAnnualIntl.Symp.onComp.Architecture,June2004.[4]J.Srinivasanetal.TheImpactofTechnologyScalingonLifetimeReliability.InProceedingsofthe2004InternationalConferenceonDependableSystemsandNetworks,June2004.
1
Check with your car insurance company to make ... o o o o Consumer Reports for reliability ...odometer Closing the Deal with a Dealership ...
Check with your car insurance company to make ... o o o o Consumer Reports for reliability ...odometer Closing the Deal with a Dealership ...
热门文档