38-An Introduction to Disk Drive Modeling.pdf

AnintroductiontodiskdrivemodelingChrisRuemmlerandJohnWilkesHewlett-PackardLaboratories,PaloAlto,CAMuchresearchinI/Osystemsisbasedondiskdrivesimulationmodels,buthowgoodarethey?Anaccuratesimulationmodelshouldemphasizetheperformance-criticalareas.ThispaperhasbeenpublishedinIEEEComputer27(3):1729,March1994.ItsupersedesHPLabstechnicalreportsHPL9368rev1andHPLOSR9329.Copyright1994IEEE.Internalorpersonaluseofthismaterialispermitted.However,permissiontoreprint/republishthismaterialforadvertisingorpromotionalpurposesorforcreatingnewcollectiveworksforresaleorredistributionmustbeobtainedfromtheIEEE.Toreceivemoreinformationonobtainingpermission,.Note:thisfilewasobtainedbyscanningandperformingOCRontheIEEEpublishedcopy.Asaresult,itmaycontaintypographicorothererrorsthatarenotinthepublishedversion.Minorclarificationsandupdateshavebeenmadetothebibliography.1Modernmicroprocessortechnologyisadvancingatanincrediblerate,andspeedupsof40to60percentcompoundedannuallyhavebecomethenorm.Althoughdiskstoragedensitiesarealsoimprovingimpressively(60to80percentcompoundedannually),performanceimprovementshavebeenoccurringatonlyabout7to10percentcompoundedannuallyoverthelastdecade.Asaresult,disksystemperformanceisfastbecomingadominantfactorinoverallsystembehavior.Naturally,researcherswanttoimproveoverallI/Operformance,ofwhichalargecomponentistheperformanceofthediskdriveitself.Thisresearchofteninvolvesusinganalyticalorsimulationmodelstocomparealternativeapproaches,andthequalityofthesemodelsdeterminesthequalityoftheconclusions；indeed,thewrongmodelingassumptionscanleadtoerroneousconclusions.Nevertheless,littleworkhasbeendonetodevelopordescribeaccuratediskdrivemodels.Thismayexplainthecommonplaceuseofsimple,relativelyinaccuratemodels.Webelievethereismuchroomforimprovement.Thisarticledemonstratesanddescribesacalibrated,high-qualitydiskdrivemodelinwhichtheoverallerrorfactoris14timessmallerthanthatofasimplefirst-ordermodel.Wedescribethevariousdiskdriveperformancecomponentsseparately,thenshowhowtheirinclusionimprovesthesimulationmodel.Thisenablesaninformedtrade-offbetweeneffortandaccuracy.Inaddition,weprovidedetailedcharacteristicsfortwodiskdrives,aswellasabriefdescriptionofasimulationenvironmentthatusesthediskdrivemodel.CharacteristicsofmoderndiskdrivesTomodeldiskdrives,wemustunderstandhowtheybehave.Thus,webeginwithanoverviewofthecurrentstateoftheartinnonremovablemagneticdiskdriveswithembeddedSCSI(SmallComputerSystemsInterconnect)controllers,sincethesearewidelyavailable.Diskdrivescontainamechanismandacontroller.Themechanismismadeupoftherecordingcomponents(therotatingdisksandtheheadsthataccessthem)andthepositioningcomponents(anarmassemblythatmovestheheadsintothecorrectpositiontogetherwithatrack-followingsystemthatkeepsitinplace).Thediskcontrollercontainsamicroprocessor,somebuffermemory,andaninterfacetotheSCSIbus.Thecontrollermanagesthestorageandretrievalofdatatoandfromthemechanismandperformsmappingsbetweenincominglogicaladdressesandthephysicaldisksectorsthatstoretheinformation.Below,welookmorecloselyateachoftheseelements,emphasizingfeaturesthatneedtobeconsideredwhencreatingadiskdrivemodel.Itwillbecomeclearthatnotallthesefeaturesareequallyimportanttoamodelsaccuracy.Therecordingcomponents.Moderndisksrangeinsizefrom1.3to8inchesindiameter；2.5,3.5,and5.25inchesarethemostcommonsizestoday.Smallerdiskshavelesssurfaceareaandthusstorelessdatathantheirlargercounterparts；however,theyconsumelesspower,canspinfaster,andhavesmallerseekdistances.Historically,asstoragedensitieshaveincreasedtowhere23gigabytescanfitonasingledisk,thenext-smallerdiameterintheserieshasbecomethemostcost-effectiveandhencethepreferredstoragedevice.2Increasedstoragedensityresultsfromtwoimprovements.Thefirstisbetterlinearrecordingdensity,whichisdeterminedbythemaximumrateoffluxchangesthatcanberecordedandreadback；currentvaluesarearound50,000bitsperinchandwillapproximatelydoublebytheendofthedecade.Thesecondcomesfrompackingtheseparatetracksofdatamorecloselytogether,whichishowmostoftheimprovementsareoccurring.Currentvaluesareabout2,500tracksperinch,risingtoperhaps20,000TPIbytheendofthedecade.Theproductofthesetwofactorswillprobablysustainagrowthrateabove60percentperyeartotheendofthedecade.Asinglediskcontainsone,two,orasmanyasadozenplatters,asshowninFigure1.Thestackofplattersrotatesinlocksteponacentralspindle.Although3,600rpmwasadefactostandardformanyyears,spindlerotationspeedhasincreasedrecentlytoasmuchas7,200rpm.Themedianrotationspeedisincreasingatacompoundrateofabout12percentperyear.Ahigherspinspeedincreasestransferratesandshortensrotationlatencies(thetimefordatatorotateunderthehead),butpowerconsumptionincreasesandbetterbearingsarerequiredforthespindle.Thespinspeedistypicallyquotedasaccuratewithin0.5to1percent；inpractice,thediskspeedsvaryslowlyaroundthenominalrate.Althoughthisisperfectlyreasonableforthedisksoperation,itmakesitnearlyimpossibletomodelthedisksrotationalpositionsome100-200revolutionsafterthelastknownoperation.Fortunately,manyI/Ooperationsoccurinbursts,sotheuncertaintyappliesonlytothefirstrequestintheburst.Eachplattersurfacehasanassociateddiskheadresponsibleforrecording(writing)andlatersensing(reading)themagneticfluxvariationsontheplatterssurface.Thediskdrivehasasingleread-writedatachannelthatcanbeswitchedbetweentheheads.Thischannelisresponsibleforencodinganddecodingthedatastreamintoorfromaseriesofmagneticphasechangesstoredonthedisk.Significantfractionsoftheencodeddatastreamarededicatedtoerrorcorrection.Theapplicationofdigitalsignalprocessingmaysoonincreasechannelspeedsabovetheircurrent100megabitspersecond.(Multichanneldiskscansupportmorethanoneread/writeoperationatatime,makinghigherdatatransferratespossible.However,thesedisksarerelativelycostlybecauseoftechnicaldifficultiessuchascontrollingthecrosstalkbetweentheconcurrentlyactivechannelsandkeepingmultipleheadsalignedontheirplatterssimultaneously.Thelatterisbecomingmoredifficultastrackdensitiesincrease.)Figure1:themechanicalcomponentsofadiskdrive.b.topview.a.sideview.armassemblyarmheadspindlesectortrackarmheadarmpivotplattercylinder3Thepositioningcomponents.Eachdatasurfaceissetuptostoredatainaseriesofconcentriccircles,ortracks.Asinglestackoftracksatacommondistancefromthespindleiscalledacylinder.Todaystypical3.5-inchdiskhasabout2,000cylinders.Astrackdensitiesincrease,thenotionofverticalalignmentthatisassociatedwithcylindersbecomeslessandlessrelevantbecausetrackalignmenttolerancesaresimplytoofine.Essentially,then,wemustconsiderthetracksoneachplatterindependently.Toaccessthedatastoredinatrack,thediskheadmustbemovedoverit.Thisisdonebyattachingeachheadtoadiskarmaleverthatispivotednearoneendonarotationbearing.Allthediskarmsareattachedtothesamerotationpivot,sothatmovingoneheadcausestheotherstomoveaswell.Therotationpivotismoreimmunetolinearshocksthantheolderschemeofmountingtheheadonalinearslider.Thepositioningsystemstaskistoensurethattheappropriateheadgetstothedesiredtrackasquicklyaspossibleandremainsthereeveninthefaceofexternalvibration,shocks,anddiskflaws(forexample,nonconcentricandnoncirculartracks).Seeking.Thespeedofheadmovement,orseeking,islimitedbythepoweravailableforthepivotmotor(halvingtheseektimerequiresquadruplingthepower)andbythearmsstiffness.Accelerationsof30-40garerequiredtoachievegoodseektimes,andtooflexibleanarmcantwistandbringtheheadintocontactwiththeplattersurface.Smallerdiameterdiskshavecorrespondinglyreduceddistancesfortheheadtomove.Thesediskshavesmaller,lighterarmsthatareeasiertostiffenagainstflexingallcontributingtoshorterseektimes.Aseekiscomposedofaspeedup,wherethearmisaccelerateduntilitreacheshalfoftheseekdistanceorafixedmaximumvelocity,acoastforlongseeks,wherethearmmovesatitsmaximumvelocity,aslowdown,wherethearmisbroughttorestclosetothedesiredtrack,andasettle,wherethediskcontrolleradjuststheheadtoaccessthedesiredlocation.Veryshortseeks(lessthan,say,twotofourcylinders)aredominatedbythesettletime(13milliseconds).Infact,aseekmaynotevenoccur；theheadmayjustresettleintopositiononanewtrack.Shortseeks(lessthan200400cylinders)spendalmostalloftheirtimeintheconstant-accelerationphase,andtheirtimeisproportionaltothesquarerootoftheseekdistanceplusthesettletime.Longseeksspendmostoftheirtimemovingataconstantspeed,takingtimethatisproportionaltodistanceplusaconstantoverhead.Asdisksbecomesmallerandtrackdensitiesincrease,thefractionofthetotalseektimeattributedtothesettlephaseincreases.“Average”seektimesarecommonlyusedasafigureofmeritfordiskdrives,buttheycanbemisleading.Suchaveragesarecalculatedinvariousways,asituationfurthercomplicatedbythefactthatindependentseeksarerareinpractice.Shorterseeksaremuchmorecommon,l,2althoughtheiroverallfrequencyisverymuchafunctionoftheworkloadandtheoperatingsystemdrivingthedisk.Ifdiskrequestsarecompletelyindependentofoneanother,theaverageseekdistancewillbeonethirdofthefullstroke.Thus,somesourcesquotetheone-third-strokeseektimeasthe“average”.Otherssimplyquotethefull-stroketimedividedbythree.Anotherwayistosumthetimesneededtoperformoneseekof4eachsizeanddividethissumbythenumberofdifferentseeksizes.Perhapsthebestofthecommonlyusedtechniquesistoweighttheseektimebythenumberofpossibleseeksofeachsize:Thus,thereareN1differentsingle-trackseeksthatcanbedoneonadiskwithNcylinders,butonlyonefull-strokeseek.Thisemphasizestheshorterseeks,providingasomewhatbetterapproximationtomeasuredseek-distanceprofiles.Whatmatterstopeoplebuildingmodels,however,istheseek-time-versus-distanceprofile.Weencouragemanufacturerstoincludetheseintheirdiskspecifications,sincetheonlyalternativeistodeterminethemexperimentally.Theinformationrequiredtodeterminehowmuchpowertoapplytothepivotmotorandforhowlongonaparticularseekisencodedintabularforminthediskcontroller.Ratherthaneverypossiblevalue,asubsetofthetotalisstored,andinterpolationisusedforintermediateseekdistances.Theresultingfine-grainedseek-timeprofilecanlookratherlikeasawtoothThermalexpansion,armpivot-bearingstickiness,andotherfactorsoccasionallymakeitnecessarytorecalibratethesetables.Thiscantake500-800milliseconds.Recalibrationsaretriggeredbytemperaturechangesandbytimers,sotheyoccurmostfrequentlyjustafterthediskdriveispoweredup.Insteady-stateconditions,recalibrationoccursonlyonceevery1530minutes.Obviously,thiscancausedifficultieswithreal-timeorguaranteed-bandwidthsystems(suchasmultimediafileservers),sodiskdrivesarenowappearingwithmodifiedcontrollerfirmwarethateitheravoidsthesevisiblerecalibrationscompletelyorallowsthehosttoscheduletheirexecution.Trackfollowing.Fine-tuningtheheadpositionattheendofaseekandkeepingtheheadonthedesiredtrackisthefunctionofthetrack-followingsystem.Thissystemusespositioninginformationrecordedonthediskatmanufacturingtimetodeterminewhetherthediskheadiscorrectlyaligned.Thisinformationcanbeembeddedinthetargetsurfaceorrecordedonaseparatededicatedsurface.Theformermaximizescapacity,soitismostfrequentlyusedindiskswithasmallnumberofplatters.Astrackdensityincreases,someformofembeddedpositioningdatabecomesessentialforfine-grainedcontrolperhapscombinedwithadedicatedsurfaceforcoarsepositioningdata.However,theembedded-datamethodaloneisnotgoodatcopingwithshockandvibrationbecausefeedbackinformationisonlyavailableintermittentlybetweendatasectors.Thetrack-followingsystemisalsousedtoperformaheadswitch.Whenthecontrollerswitchesitsdatachannelfromonesurfacetothenextinthesamecylinder,thenewheadmayneedrepositioningtoaccommodatesmalldifferencesinthealignmentofthetracksonthedifferentsurfaces.Thetimetakenforsuchaswitch(0.5-1.5ms)istypicallyonethirdtoonehalfofthetimetakentodoasettleattheendofaseek.Similarly,atrackswitch(orcylinderswitch)occurswhenthearmhastobemovedfromthelasttrackofacylindertothefirsttrackofthenext.Thistakesaboutthesametimeastheend-of-seeksettlingprocess.Sincesettlingtimeincreasesastrackdensityincreases,andthetracksondifferentplattersarebecominglesswellaligned,head-switchingtimesareapproachingthosefortrackswitching.Nowadays,manydiskdrivesuseanaggressive,optimisticapproachtoheadsettlingbeforeareadoperation.Thismeanstheywillattemptareadassoonastheheadisneartherighttrack；afterall,ifthedataareunreadablebecausethesettlehasnotquitecompleted,nothinghasbeenlost.(Thereisenougherrorcorrectionandidentificationdatainamisreadsectortoensurethatthedataarenotwronglyinterpreted.)Ontheotherhand,ifthedataareavailable,itmightjustsaveanentirerevolutionsdelay.Forobviousreasons,5thisapproachisnottakenforasettlethatimmediatelyprecedesawrite.Thedifferenceinthesettletimesforreadsandwritescanbeasmuchas0.75ms.Datalayout.ASCSIdiskappearstoitsclientcomputerasalinearvectorofaddressableblocks,eachtypically256-1,024bytesinsize.Theseblocksmustbemappedtophysicalsectorsonthedisk,whicharethefixed-sizedata-layoutunitsontheplatters.Separatingthelogicalandphysicalviewsofthediskinthiswaymeansthatthediskcanhidebadsectorsanddosomelow-levelperformanceoptimizations,butitcomplicatesthetaskofhigherlevelsoftwarethatistryingtosecond-guessthecontroller(forexample,the4.2BSDUnixfastfilesystem).Zoning.Tracksarelongerattheoutsideofaplatterthanattheinside.Tomaximizestoragecapacity,lineardensityshouldremainnearthemaximumthatthedrivecansupport；thus,theamountofdatastoredoneachtrackshouldscalewithitslength.Thisisaccomplishedonmanydisksbyatechniquecalledzoning,whereadjacentdiskcylindersaregroupedintozones.Zonesneartheouteredgehavemoresectorspertrackthanzonesontheinside.Therearetypically3to20zones,andthenumberislikelytodoublebytheendofthedecade.Sincethedatatransferrateisproportionaltotherateatwhichthemediapassesunderthehead,theouterzoneshavehigherdatatransferrates.Forexample,onaHewlett-PackardC22403.5-inchdiskdrive,thebursttransferrate(withnointertrackheadswitches)variesfrom3.1megabytespersecondattheinnerzoneto5.3MBpsattheoutermostzone.3Trackskewing.Fastersequentialaccessacrosstrackandcylinderboundariesisobtainedbyskewinglogicalsectorzerooneachtrackbyjusttheamountoftimerequiredtocopewiththemostlikelyworst-casehead-ortrack-switchtimes.Thismeansthatdatacanbereadorwrittenatnearlyfullmediaspeed.Eachzonehasitsowntrackandcylinderskewfactors.Sparing.Itisprohibitivelyexpensivetomanufactureperfectsurfaces,sodisksinvariablyhavesomeflawedsectorsthatcannotbeused.Flawsarefoundthroughextensivetestingduringmanufacturing,andalistisbuiltandrecordedonthediskforthecontrollersuse.Sothatflawedsectorsarenotused,referencestothemareremappedtootherportionsofthedisk.Thisprocess,knownassparing,isdoneatthegranularityofsinglesectorsorwholetracks.Thesimplesttechniqueistoremapabadsectorortracktoanalternatelocation.Alternatively,slipsparingcanbeused,inwhichthelogicalblockthatwouldmaptothebadsectorandtheonesafteritare“slipped”byonesectororbyawholetrack.Manycombinationsoftechniquesarepossible,sodiskdrivedesignersmustmakeacomplextrade-offinvolvingperformance,expectedbad-sectorrate,andspaceutilization.AconcreteexampleistheHPC2240diskdrive,whichusesbothformsoftrack-levelsparing:slip-tracksparingatdiskformattimeandsingle-trackremappingfordefectsdiscoveredduringoperation.Thediskcontroller.Thediskcontrollermediatesaccesstothemechanism,runsthetrack-followingsystem,transfersdatabetweenthediskdriveanditsclient,and,inmanycases,managesanembeddedcache.Controllersarebuiltaroundspeciallydesignedmicroprocessors,whichoftenhavedigitalsignalprocessingcapabilityandspecialinterfacesthatletthemcontrolhardwaredirectly.Thetrendistowardmorepowerfulcontrollersforhandlingincreasinglysophisticatedinterfacesandforreducingcostsbyreplacingpreviouslydedicatedelectroniccomponentswithfirmware.InterpretingtheSCSIrequestsandperformingtheappropriatecomputationstakestime.Controllermicroprocessorspeedisincreasingjustaboutfastenoughtostayaheadoftheadditionalfunctionsthe6controllerisbeingaskedtoperform,socontrolleroverheadisslowlydeclining.Itistypicallyintherange0.3-1.0ms.Businterface.Themostimportantaspectsofadiskdriveshostchannelareitstopology,itstransferrate,anditsoverhead.SCSIiscurrentlydefinedasabus,althoughalternativeversionsarebeingdiscussed,asareencapsulationsofthehigherlevelsoftheSCSIprotocolacrossothertransmissionmedia,suchasFibreChannel.MostdiskdrivesusetheSCSIbusoperationssynchronousmode,whichcanrunatthemaximumbusspeed.Thiswas5MBpswithearlySCSIbuses；differentialdriversandthe“fastSCSI”specificationincreasedthisto10MBpsacoupleofyearsago.Disksarenowappearingthatcandrivethebusat20MBps(“fast,wide”),andthestandardisdefinedupto40MBps.ThemaximumbustransferrateisnegotiatedbetweenthehostcomputerSCSIinterfaceandthediskdrive.ItappearslikelythatsomeserialchannelsuchasFibreChannelwillbecomeamorepopulartransmissionmediumatthehigherspeeds,partlybecauseitwouldhavefewerwiresandrequireasmallerconnector.BecauseSCSIisabus,morethanonedevicecanbeattachedtoit.SCSIinitiallysupporteduptoeightaddresses,afigurerecentlydoubledwiththeuseofwideSCSI.Asthenumberofdevicesonthebusincreases,contentionforthebuscanoccur,leadingtodelaysinexecutingdatatransfers.Thismattersmoreifthediskdrivesaredoinglargetransfersoriftheircontrolleroverheadsarehigh.Inadditiontothetimeattributedtothetransferrate,theSCSIbusinterfacesatthehostanddiskalsorequiretimetoestablishconnectionsanddeciphercommands.OnSCSI,thecostofthelow-levelprotocolforacquiringcontrolofthebusisontheorderofafewmicrosecondsifthebusisidle.TheSCSIprotocolalsoallowsadiskdrivetodisconnectfromthebusandreconnectlateronceithasdatatotransfer.Thiscyclemaytake200sbutallowsotherdevicestoaccessthebuswhilethedisconnecteddeviceprocessesdata,resultinginahigheroverallthroughput.Inolderchannelarchitectures,therewasnobufferinginthediskdriveitself.Asaresult,ifthediskwasreadytotransferdatatoahostwhoseinterfacewasnotready,thenthediskhadtowaitanentirerevolutionforthesamedatatocomeundertheheadagainbeforeitcouldretrythetransfer.InSCSI,thediskdriveisexpectedtohaveaspeed-matchingbuffertoavoidthisdelay,maskingtheasynchronybetweenthebusandthemechanism.SincemostSCSIdrivestakedataoffthemediamoreslowlythantheycansenditoverthebus,thedrivepartiallyfillsitsbufferbeforeattemptingtocommencethebusdatatransfer.Theamountofdatareadintothebufferbeforethetransferisinitiatediscalledthefence；itssizeisapropertyofthediskcontroller,althoughitcanbespecifiedonmodernSCSIdiskdrivesbyacontrolcommand.Writerequestscancausethedatatransfertothedisksbuffertooverlaptheheadrepositioning,uptothelimitpermittedbythebufferssize.TheseinteractionsareillustratedinFigure2.Cachingofrequests.Thefunctionsofthespeed-matchingbufferinthediskdrivecanbereadilyextendedtoincludesomeformofcachingforbothreadsandwrites.Cachesindiskdrivestendtoberelativelysmall(currently64kilobytesto1megabyte)becauseofspacelimitationsandtherelativelyhighcostofthedual-portedstaticRAMneededtokeepupwithboththediskmechanismandthebusinterface.Read-ahead.Areadthathitsinthecachecanbesatisfied“immediately,”thatis,injustthetimeneededforthecontrollertodetectthehitandsendthedatabackacrossthebus.Thisisusuallymuchquickerthanseekingtothedataandreadingitoffthedisk,somostmodernSCSIdisksprovidesome7formofreadcaching.Themostcommonformisread-aheadactivelyretrievingandcachingdatathatthediskexpectsthehosttorequestmomentarily.Aswewillshow,readcachingturnsouttobeveryimportantwhenitcomestomodelingadiskdrive,butitisoneoftheleastwellspecifiedareasofdisksystembehavior.Forexample,areadthatpartiallyhitsinthecachemaybepartially