GuangjunXie BaiduIncBeijingChina NankaiBaiduJointLabNankaiUniversityTianjinChina Emailxieguangjun1980163com GuangzhiXuGangWangXiaoguangLiuRuiCaoYanGao NankaiBaiduJointLabNankaiUniversi ID: 843998
Download Pdf The PPT/PDF document "hUBIAnOptimizedHybridMappingSchemeforNAN..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1 hUBI:AnOptimizedHybridMappingSchemeforNA
hUBI:AnOptimizedHybridMappingSchemeforNANDFlash-BasedSSDs GuangjunXie BaiduInc.,Beijing,China Nankai-BaiduJointLab,NankaiUniversity,Tianjin,China Email:xieguangjun1980@163.com GuangzhiXu,GangWang,XiaoguangLiu,RuiCao,YanGao Nankai-BaiduJointLab,NankaiUniversity,Tianjin,China Email:xgz.nku@mail.nankai.edu.cn,wgzwp@163.com,liuxg74@yahoo.com.cn Abstract NANDash-basedSSDshavebecomeattractive alternativestoharddiskdriversduetotheirhighrandomread performancesandlowpowerconsumptions.However,thepoor incommercialapplications. Inthispaper,weproposeanovelmappingschemecalled hybridmappingunsortedblockimages(hUBI).hUBIaimsfor (1)optimizedrandomwriteperformances,(2)lowwritelaten- cies,and(3)lowspaceconsumptions.Itoptimizestraditional hybridmappingschemestoachieve(1).InhUBI,themerge operationinvolvesonlyasingleblock.So,(2)isguaranteed.To avoidintroducingahighspacecostmaintainingthemetadata, hUBIputsitsmetadataintoout-of-band(OOB)areasofSSD pages,whichobtains(3).Ourexperimentalresultsshowthat holdslowwritelatenciesatthesametime. I.I NTRODUCTION OverthelastdecadetheCPUspeedhasincreaseddra- maticallywhiletheaccesstimeoftraditionaldiskshasonly improvedslowly[1].Thebottleneckseemsmoreapparent whenparallelworkloadsintroducedbyrapiddevelopingIT applicationsincrease.Asthetraditionalharddisksusing spinningdisksandmovableread/writeheadscannotmeet applicationsrequirements,peopleseekfornewstorage mediumthatcansolvethisproblem.Attheendoflastcen- tury,solid-statedisks(SSD)wereintroducedintothestorage traditionalharddisks[2]. SSDshavethepropertiesoflowpowerconsumptionsand highaccessspeeds.Now,mostSSDsarebasedonNAND ashchips.Alongwiththeattractivepropertiesoflowcosts andhighdensities,theNANDashchiphasaproblem thatblocksshouldbeerasedbeforebeingwrittenanderase operationsarequiteslow.Inconsiderationofwearleveling andread/writelatencies,dynamicmappingsareusedinmost SSDdrivers. Mappingschemescanbedividedintothreedimensions: Mappingschemes[3]maintainmappingsbetweenlogical blocksandphysicalblocks.Eachpageupdatewillcause ablockeraseandseveralpagecopyoperations.Block- Mappingschemesimplysimpleimplementationsandlow spacecostsfortheirmetadata.However,duetotheirhigh writelatencies,block-mappingschemeshavepoorrandom writeperformances.Toimprovethewriteperformance, page-mappingschemes[4]maintainne-grainedmappings betweenlogicalpagesandphysicalpages.Inpage-mapping schemes,pagewritesareperformedonfreepages.Com- paredtoblock-mappingschemes,page-mappingschemes writeperformances.However,page-mappingschemescon- sumealotofspacestostoretheirmetadata.Hybridmapping schemes[5][6][7]combinetheadvantagesofblock- mappingschemesandpage-mappingschemes.Theygroup blocksintodataareaandlogarea.Pagesinthelogarea areaddressedlikepage-mappingschemes.Pagewritesare performedonfreepagesinthelogarea.Whenthelog areahasnofreepages,amergeoperation[10]istriggered toeraselogareablocksandstoresvalidpagesindata areablocks.Bythismeans,hybridmappingschemesget efciencyatthesametime.However,intraditionalhybrid mappingschemes,complexmergeoperationsmaycause highwritelatencies,whichcouldbecatastrophicintime- sensitiveapplications. Inthispaper,wepresentanovelhybridmappingscheme calledhybridmappingunsortedblockimages(hUBI).hUBI contributesinseveralaspects: hUBIusesaoptimizedhybridmappingalgorithmgiv- inghighrandomwriteperformances. whichleadstoaguaranteedlowwritelatency. hUBImaintainsitsmetadataintheout-of-band(OOB) areas,whichreducesthespaceconsumptionandthe numberofI/Ooperations. ExperimentalresultsshowthathUBIprovidesanconsid- erablerandomwriteperformanceandalowtimeoutrate undermixedreadandw
2 riteworkload. 2011 International Joint C
riteworkload. 2011 International Joint Conference of IEEE TrustCom-11/IEEE ICESS-11/FCST-11978-0-7695-4600-1/11 $26.00 © 2011 IEEEDOI 10.1109/TrustCom.2011.1391015 Thefollowingpartsofthispaperareorganizedasfollow: Section2introducesotherrelatedworks.Section3de- scribesthedesignofhUBIindetails.Section4showsour analysisandexperimentsofhUBI.AndnallySection5 givestheconclusion. II.R ELATED W ORK UnsortedBlockImages(UBI)[11]isamappingscheme usedintheLinuxkernel.Ithasalayeredarchitectureand mostotherschemesadoptitsdesign.UBIisablock-mapping scheme.Asdiscussedabove,eachpageupdateinUBIbrings ablockeraseandseveralpagecopyoperations.Blockerase andpagecopyarebothtimeconsumingoperations.Sopage updatesinUBIhaveahightimecost. Therehavebeenresearchesonhybridmappingschemes. Kim[5]presentedalogblockschemeinwhicheachlog block(i.e.,ablockinlogarea)isdedicatedtoonedata block(i.e.,ablockindataarea).Thismaycauselowspace utilizationoflogblocks.e.g.,thereisanupdatetoonedata block.Althoughthelogareastillhaslotsoffreerooms,if thelogblockitownsisfullyoccupied,amergeoperation willstillbenecessary.Tofullyutilizethelogareaspace, Lee[7]proposedamappingschemecalledfullyassociative sectortranslation(FAST).InFAST,apageinlogblockcan bemappedtoanydatablocks.FASTmergesfewertimes than[5],butitbringscomplexmergelogic.Amergein[5] onlyinvolvestwoblocks.Whileattheworstcase,amergein FASTcaninvolveasmanyas n +1 blocks( n isthenumber ofpagesinalogblock).Park[8]introducedacompromising solutionthatmaps K logblocksto N datablocks.Heran alotofworkloadstryingtondpropervaluesfor K and N .Liu[9]triedtoapplyhybridassociationsbetweenlog blocksanddatablocks.Foronehotdatablock(i.e.,theblock thataccessedfrequently),aexclusivelogblockisassigned, whileothercolddatablocksusefullyassociationsdescribed inFAST.However,themergeproblemremains. PartitionbasedUBI(pUBI)isanimprovedhybridmap- pingscheme.InpUBI,blocksaregroupedintopartitions. Eachpartitionisacontiguousregionoftheaddressspace. Blocksinapartitionaredividedintodatablocks,logblocks andm-blocks.Updatestoapartitionwillbeperformedon itscorrespondinglogblocks.Inonepartition,thereisonly oneactivem-blockstoringitsmetadata.Amergeoperation iscalledwhentherearenofreepagesinthelogarea. ComparedwithUBI,pUBIprovidesabetterrandomwrite performance.However,themergeoperationinpUBImay involveseveraldataandlogblocks,thereforeitcanbringa numberofblockeraseandpagecopyoperations.Insome circumstance,suchatimeconsumingmergeoperationmay causeadataupdatetimeout. hUBIadoptsthelayeredarchitectureofUBIbutdesigns anewhybridmappingmechanism.Tosolvethetimeout probleminpUBI,hUBIfocusesonoptimizingmergeop- erations.ExperimentalresultsshowthathUBIsolvesthis !"! # $%# &%# Figure1.OverviewArchitecture ' ( )) ) ) *+ ,,*- Figure2.HuaweiSSDPageLayout problemeffectivelyandprovidesacomparablerandomwrite performancewithpUBI. III.D ESIGNOFH UBI ThemajorobjectiveofhUBIisloweringthewritelaten- ciesandspaceconsumptionsofhybridmappingschemes. OurworkisbasedontheNA
3 NDash-basedSSDsproduced byHuaweiInc.and
NDash-basedSSDsproduced byHuaweiInc.andcanbeappliedtootherSSDs.Overview architectureofhUBIisshowninFigure1. Therawmemorytechnologydevice(MTD)driverpro- videsaphysicalerasableblock(PEB)viewwithoutwear leveling.ThehUBIdriverimplementsahybridmapping algorithmandprovidesalogicalerasableblock(LEB)view. Finally,touseSSDsasnormalblockdevices,theMTD blockdriveremulatesasectorviewforupperapplications. A.MetadataLayout 1)DiskLayout: hUBIdividespagesineachblockinto logareaanddataarea.Thepage-mappinginfoforlogarea pagesisstoredondisktorecordtherelationshipsbetween logicalpagesandphysicalpages.Sincethepage-mapping infochangesalongwithpageupdates,bystoringthepage- mappingalongwiththepagedata,onlyoneI/Ooperation isneededwhenwritingboththepagedataandthepage- mappinginfo.Figure2showsthepagelayoutofHuawei MLCSSDs.Theout-of-band(OOB)areaisusedforerror controlandthersttwobytesofOOBareusedasabadblock ag.Only2bitsoftheagareactuallyused,thereforewe canusethesecondbytetoholdthepage-mappinginfo,as showninFigure3. Twomainadvantagesaregainedbytheabovedesign:(1) ThenumberofI/Ooperationsisreduced,asthepagedata andthepage-mappingcanbewrittentodisksinoneI/O operation.(2)ByusingtheOOBareastostorethemetadata, hUBIdoesnotconsumeextrastoragespaces. 1016 ' ( )) ) ) *+ ,,*- ( . $&!. Figure3.hUBIPageLayout 2)MemoryLayout: Therearetwoalternativestomain- tainthein-memorypage-mappingtable(PMT).Oneofthem storesthephysicalpagenumber(PPN)foreachpageinthe logarea.Foraread/writeaccess,iftherequiredpageis inthelogarea,thePPNcanbegottenbysearchingthe PMT.Otherwise,iftherequiredpageisinthedataarea,as dataareapagesarestoredintheorderoflogicaladdresses, thePPNisequaltothelogicalpagenumber(LPN).By usingthissolution,fewermemoryspacesareconsumedto maintainthemetadata.However,areadoperationmaycause asmanyas n comparisonstogetthecorrespondingPPN ( n isthetotalnumberoflogpagesinoneblock).The othermethodmaintainsthepage-mappinginfoforbothlog anddataareapages,whichissimilartothepage-mapping scheme.Bythisway,readoperationsavoidPMTsearching. But,morespaceswillbeoccupiedandthingsgetevenworse whenthesizeofstoragegrowslarger. Tobalancebetweenmemoryconsumptionsandperfor- mances,wecomeupwithacompromisingmethod.hUBI optimizestherstdesignbyaddingabitmap.Onebitinthe bitmapcorrespondstoauniquelogicalpageandindicates whetheritisinthelogareaornot.Whenaccessingapage, thecorrespondingbitinthebitmapisrsttested.Sincebit testisafastoperation,wepaylittleperformancepenaltyto savealotofmemoryspace. AnoverviewofmetadatalayoutforhUBIisshownin Figure4.hUBIextendstheeraseblockmappingtable(EMT) usedbyUBItosupporthybridmapping.EachentryofEMT storesthemetadataforalogicalblock.Itcontainsfourparts: theblock-mappinginfo,apointer,anindexandabitmap. ThepointerpointstothePMT.Theindexrecordsthenext freelogareapageinPMT.Thebitmapindicateswhethera pageisinthelogarea.e.g.,inFigure4,LEB1ismappedto PEB10.InPEB10,anupdateoperationisperformedafter amergeoperation.TheLPNofthewrittenpageis2. B.ReadandWriteAlgorithms Inthissection,wedescribethereadandwritealgorithms ofhUBIindetails. Thekeylogicofthereadalgorithmistheaddresstrans- lation.InhUBI,alogicaladdresscanbedividedintothree parts:aLEB(logicalerasableblock)number,aLPN(logical pagenumber)intheLEBandanoffsetinthepage.Usually, readrequestsarealignedbypages,thuswecanas
4 sumethe offsetinthepagetobezeroanduseatu
sumethe offsetinthepagetobezeroanduseatupleLEB,;-369;LPN $& !/0 1(2! $ & 3 4 + .+ . . 555 555 555 555 &1 & 6 331 0 &% 4 + . + . $% 1( & $& !/0 $&! + . 7 4 47 4 " $ " - 8 * *7 9 Figure4.hUBIMetadataLayout torepresentalogicaladdress.Inthefollowingparts,we usethisformoflogicaladdresstodemonstratethereadand writealgorithms. Algorithm1showsthereadalgorithmofhUBI.Toread apage,thealgorithmrstlocatesthepage-mappingtable foragivenLEB.Thenthecorrespondingbitinthebitmap istestedtocheckwhethertherequiredpageisinthelog area.Forareadrequesttologareapages,thealgorithm willsearchthepage-mappingtablesequentiallytondthe correspondingPPN.Otherwise,therequiredpageisinthe dataarea,thePPNisequaltotheLPN. Algorithm1: hUBIReadAlgorithm Input : leb , lpn begin emt _ entry =get_entry_from_emt( leb ); peb = emt _ entry.peb ; pmt = emt _ entry.pmt ; bitmap = emt _ entry.bitmap ; if is_in_log_area( lpn , bitmap )== true then ppn =get_ppn_from_pmt( lpn , pmt ); else ppn = lpn ; read_page( peb , ppn ); end Algorithm2showsthewritealgorithmofhUBI.The algorithmrstwritesthedataandthepage-mappinginfoto afreelogareapage.ThenanewentryisaddedtothePMT andthecorrespondingbitinthebitmapisalsoset.When therearenofreepagesinthelogarea,amergeoperationis calledtocleanthelogarea. Algoritm3describesthemergelogic.Whenthereareno freelogpagesinthelogareaofcurrentPEB P (LEB L ),a newPEB P isallocatedfromthefreePEBlist.Thenvalid 1017 Algorithm2: hUBIWriteAlgorithm Input : leb , lpn , data begin emt _ entry =get_entry_from_emt( leb ); peb = emt _ entry.peb ; pmt = emt _ entry.pmt ; bitmap = emt _ entry.bitmap ; ppn =get_next_free_log_page( pmt ); set_oob_mapping( oob , lpn ); write_physical_page_with_oob( peb , ppn , oob , data ); set_bitmap( lpn , bitmap ); set_pmt( lpn , pmt ); if nofreelogpages then merge( leb ); end end pagesarecopiedfrom P to P .Afterthat, P willbeput ontheeraselistwaitingtobeerasedbyagarbagecollector. Finally,theEMTwillbeupdatedtomap L to P .e.g.,in Figure4,accordingtoalgorithm2,theseconddataupdate toPEB3willtriggeramergeoperation,thenanewPEB K willbeallocatedandvalidpageswillbecopiedtoPEB K .Thenentry0inEMTwillbeupdatedtomapLEB0to PEB K . Algorithm3: hUBIMergeAlgorithm Input : leb begin emt _ entry =get_emt_from_emt( leb ); peb = emt _ entry.peb ; peb =alloc_free_peb(); bitmap = emt _ entry.bitmap ; pmt = emt _ entry.pmt ; foreach ppninpmt do copy_page( peb , peb , ppn ); end foreach ppninDataarea do if is_in_log_area( ppn , bitmap )== false then copy_page( peb , peb , ppn ); end end emt _ entry.peb = peb ; add_to_erase_queue( peb ); end IV.A NALYSISAND E XPERIMENTS A.Analysis Byusingbitmap,readalgorithmsofhUBIbringtinyextra timecost.Sotheanalysisbelowisfocusedonthewrite algorithm. Twoperform
5 ancemetricsareconcernedinevaluatingthe w
ancemetricsareconcernedinevaluatingthe writeperformanceofSSDs. MeanTimeToPageWrite(MTTPW) showstheaverage timecosttowritebackapage.Fromthismetric,theaverage latencyofwriteoperationscanbederived. MeanTimeToMerge(MTTM) showsthetimecostof onemergeoperation.Sincewriterequestsareblockedwhen merging,theworstcaseofwritelatencycanbederivedfrom thismetric. Assumethetimecostofreadingapageas t r ,thetime costtowriteapageas t w ,thetimecosttoeraseablockas t e andthetimecosttodeliverapagefromupperapplication toSSDstorageas t x .Forasingleblock, d standsforthe numberofpagesinthedataareaand l standsforthenumber ofpagesinthelogarea.SinceinhUBI,every l continuous writerequestswilltriggeramergeoperation,andamerge operationneedsablockcopyandanblockeraseoperations, wehave: MTTM h = d ( t r + t w +2 t x )+ t e (1) MTTPW h = MTTM h l + t w + t x (2) WecomparehUBIwithtwoUBI-basedmappingschemes, UBIandpUBI. UBIisatypicalblock-mappingscheme.Apageupdate inUBIwillcauseamergeoperation,whichthenneeds ablockread,ablockwriteandablockeraseoperations. Assumethereare k pagesinablock,thenthe MTTPW and MTTM ofUBIare: MTTM u = k ( t r + t w +2 t x )+ t e (3) MTTPW u = MTTM u + t w + t x (4) pUBIisahybridmappingversionofUBI.InpUBI,blocks aregroupedintopartitions.Blocksinapartitionaredivided intodatablocks,logblocksandm-blocks.Writerequests ofapartitiongotoitslogblocksanddatablocksstorethe mergeddata.Eachpartitionstoresitsmappinginfoinam- block.ToanalyzetheperformanceofpUsBI,weintroduce twometrics. Localityratio :Supposethatthereare D datablocks, L logblocksinapartitionand k pagesinablock.AspUBI usesFAST[7]onitslogblocks,every L × k writerequests causeamergeoperation.Wedene astheaverageratioof differentLEBscoveredby L × k continuouswriterequests toalldatablocksinapartition. Sequentialratio :Whenmerging,ifalogblockcovers exactlyoneLEB,pUBIwillturnthislogblockintoadata block.Wedene astheaverageratiooflogblocksthatcan beturnedintodatablockstoalldatablocksinapartition. 1018 Weknowthatevery L × k writerequestscauseamerge operation.Fortheserequests,thereare D × differentLEBs covered,and D × LEBscanbeturnedintodatablocks directly.Therefore,inamergeoperation, D × k × ( ) pagesarecopiedand D × ( )+ L blocksareerased. Inaddition,aseverypartitioninpUBIstoresitsmapping infoinam-blockpageandmappingsarechangedduring merging,eachmergeoperationwillcauseapagewriteto am-block,whichthenimpliesa 1 /k eraseoperation.From theaboveanalysis,wecangetthe MTTW and MTTPM ofpUBI: MTTM p = D ( kt r + kt w +2 kt x + t e )( )+ Lt e + t w + t e k (5) MTTPW p = MTTM p kL + t w + t x (6) Tooptimizethesequentialwriteperformanceofmapping schemes,MTDblockdriverimplementsawrite-backcache. ThesizeofthecacheisequaltothesizeofaLEB.For awriterequest,iftherequiredLEBiscontainedinthe buffer,thewriteoperationisperformeddirectlyonthe pagecache.Otherwise,thecachedLEBisdiscarded,dirty pagesareushedtothediskandtherequiredLEBis fetched.Tofullyutilizethiscache,hUBIoptimizesitswrite algorithmintwoways:(1)Whenthenumberofushed pagesisbelowthesizeofalogarea,eachwriterequestis performedusingthewritelogicdescribedinAlgorithm2. (2)Whenthenumberofushedpagesisnotlessthan thesizeofalogarea,therequiredblockwillbemerged andvalidpageswillbedirectlycopiedtothedataareaof anewlyallocatedblock.Fromtheaboveanalysis,under sequentialwriteworkloads,the MTTPMs ofthethree mappingschemescanbechangedinto: MTTPW h = d ( t w +2 t x + t r )+ t e d (7) MTTPW p = Lt e + t w + t e k kL + t w + t x (8) MTTPW u = k ( t w +2 t x + t r )+ t e k (9) Theperformancemetricsfors
6 sdsusedinourexperiment areshowninTableI
sdsusedinourexperiment areshowninTableI TableI H UAWEI MLCSSDP ERFORMANCE M ETRICS Operation Time(us) 256KBBlockErase t e 1500 2KBPageRead t r 50 2KBPageWrite t w 800 2KBDataTransfer t x 50 FromTableIandeqs.(1)-(9).wecomputethetheoretical writeperformancesforUBI,pUBIandhUBI(TableII). Assumeeachblockcontains128pages( k =128).ForpUBI, eachpartitionhas16datablocks( D =16)and4logblocks ( L =4).ForhUBI,eachblockcontains100pagesfordata area( d =100)and26pagesforlogarea( l =26),twopages areremainedfortheblockmetadata. TableII T HEORETICAL W RITE P ERFORMANCE UBI pUBI hUBI SW(SequentialWrite)RW(RandomWrite) SWMTTPW 0.961ms 0.863ms( = ) 0.965ms RWMTTPW 123.1ms 4.71ms( =1 , =0 ) 4.561ms MTTM 123.1ms 6.65ms-1976.32ms 96.5ms AswecanseefromTableII,undersequentialworkloads, allthreeschemeshavesimilarperformances.AndashUBI mergesmoretimesthanpUBI,pUBIgetsahighersequen- tialwriteperformance.Whileunderrandomworkload,the complexmergeoperationofpUBIeffectsitsperformance andhUBIachievesalittlebetter.Moreover,intheworst case,pUBIwilltakenearly2stoperformamergeoperation, whichisunacceptableforsometimesensitiveapplications. Inaword,hUBIprovidesaconsiderablewriteperformance andholdsasteadywritelatency. B.Experimentalresults Aoverviewofourexperimentenvironmentisgivenin TableIII. TableIII E XPERIMENT E NVIRONMENT OS 64-bitRedhatLinuxAS5withkernel2.6.18 Host CPU IntelXenon5250 × 2 Memory DDR2FBD16GB HDD SAS146GB × 8RAID5 SSD Manufacturer HuaweiInc. ChipType 2GBNANDMLC BlockSize 256K pageSize 2K Tools Iometer Ver2006_07_27(x86_64) MySQL Ver14.7forRedhat-Linux-GNU(i686) First,weuseiometer[24]toevaluatethereadperfor- manceofhUBI(seeFigure5andFigure6).Wecanseethat hUBIperformsmuchbetterthantraditionalHDDs.Andas readoperationsinthethreeschemesdonottriggertime 1019 + 8+ .++ .8+ ++ 8+ 4++ 48+ -++ -8+ -:.*4*-.:8*8..+- ; 3 / 3/ =/ ' Figure5.SequentialReadPerformance + .++++ ++++ 4++++ -++++ 8++++ *++++ -:.*4*-.:8*8..+- ; 3 ,& =/ ' Figure6.RandomReadPerformance consumingwrite/eraseoperations.Theirreadperformances donotshowmuchdifference. Then,wesimulatedthreecommonscenariostoevaluate thewriteperformanceofhUBI:(1)asinglethreadedwget program[25]downloadslesfromalibraryofwebindexes, (2)amodiedmulti-threadedwgetprogramdownloadsles fromalibraryofwebindexesand(3)aprogramperforms randomupdatestoaMySQLdatabase.(1)isasequential writescenario,while(2)and(3)providerandomwrite workload.For(1)and(2),thetotalsizeofthelibraryis 80G,thereare76leswith1Gsizeandalotofsmallles. For(3),thereare5millionrecordsinthedatabase.The averagelengthofrecordsis497bytes.Resultsareshownin Figure7. FromFigure7,wecanseethatundersequentialworkload- s,thethreemappingschemeshavesimilarperformances. ThisisbecauseawritecacheisusedintheMTDblock driverandtheresultsarewithinourexpectation.While inrandomwritescenarios,hUBIandpUBIperformmuch better,becausehUBIandpUBImergemuchlesstimesthan UBI. Then,weuseamixedreadandwriteworkloadtoevaluate + + -+ *+ :+ .++ .+ .-+ 6. 6-/3( 3 ? Figure7.WriteBenchmark + 8++ .++
7 + .8++ +++ 8++ ;
+ .8++ +++ 8++ ; 3 @/3( " 31 / Figure8.MixedReadandWriteBenchmark theaveragewritelatencyofhUBI.Weperformconcurrent selectandupdatequeriestotheMYSQLdatabaseusedin theabovebenchmark.Figure8showstheaverageresponse timeofselectqueries.Intheexperiment,thenumberof selectqueriestentimesthenumberofupdatequires.Aswe cansee,theperformanceofUBIfallssharplywhenupdate quiresinvolvewhileUBIandpUBIremainasteadyresponse time. ToevaluatetheworstcaseofwritelatenciesforhUBI,we observethedistributionsoftheresponsetimeintheabove experiment.Figure9andTableIVshowthedetailedstatistics fortimeoutrequests.Aswecansee,foronehundredthou- sandselectqueries,pUBIhasaboutfourhundredqueries thattakeover20ms.So,thetimeoutrateofpUBIisabout 0.4%.However,inhUBI,thetimeoutrateisonly0.001%. ThisisbecausethemergeoperationsinhUBIinvolvesonly oneblock.WhileinpUBI,amergeoperationcaninvolve severalblockeraseandanumberofpagecopyoperations. V.C ONCLUSION Nowadays,NAND-basedSSDsarewidelyusedbothin thecommercialareaandthepersonalcomputing.However, thepoorrandomwriteperformanceofSSDshasprovento beaseriousproblem.Onewaytosolvesuchaproblemis 1020 + .++++ ++++ 4++++ -++++ 8++++ *++++ 9++++ :++++ A++++ .+++++ +B. C.B8 C8B.+ C.+B+ C+B8+ C8+B.++ D.++ 6 6 @/3( 36 36 @/3( 6 6 @/3( =/ 31 1 Figure9.ResponseTimeDistribution TableIV S TATISTICSOF T IMEOUT R EQUESTS [20,50) [50,100) 100 UBI-select 3 0 0 UBI-select+update 1089 921 389 pUBI-select 0 0 0 pUBI-select+update 123 15 250 hUBI-select 0 0 0 hUbI-select+update 1 0 0 optimizingthemappingalgorithmsofSSDs.Manyworks haveexploredthisarea.Inthispaper,weintroducea novelhybrid-mappingschemeforNAND-basedSSDscalled hUBI.hUBIcontributesinthreeaspects: hUBIusesaoptimizedhybridmappingalgorithm achievinghighrandomwriteperformance. ThemergeoperationinhUBIisperformedonasingle block,whichprovidesaguaranteedlowwritelatency. hUBIstoresthemappinginfoinOOBareasofSSD pageswhichholdsspace-efciency. Inaword,forNAND-basedSSDs,hUBIimprovestheir randomwriteperformancesandprovidesguaranteedlow writelatencies.However,asahybridmappingscheme,the logareaofhUBIcausesawasteofstoragespaces.Our futureworkwillfocusonreducingthespacecostofmapping schemesandprovidingabetterwriteperformanceatthe sametime. A CKNOWLEDGMENTS ThisworkwassupportedinpartbytheNationalHigh TechnologyResearchandDevelopmentProgramofChina (2008AA01Z401),NSFCofChina(60903028,61070014, 61170301),ScienceandTechnologyDevelopmentPlanof Tianjin(08JCYBJC13000),andKeyProjectsintheTianjin ScienceandTechnologyPillarProgram. R EFERENCES [1]MendelRosenblum,JohnK.Ousterhout. TheDesignandIm- plementationofaLog-StructuredFileSystem InProceedings ofthe13thACMSymposiumonOperatingSystemsPrinciples, 1991. [2]F.Douglis,R.Caceres,F.Kaashoek,K.Li,B.Marsh,JA. A. Tauber:StorageAlternativesforMobileComputers In Proceedingsofthe1stSymposiumonOperationSystems DesignandImplementation,1994. [3]A.Ban. FlashFileSystem UnitedStatesPatent,No. 5,404,485,April1995. [4]A.Gupta,Y.Kim,andB.Urgaonkar. DFTL:AFlashTrans- lationLayerEmployingDema
8 nd-basedSelectiveCachingof Page-levelAdd
nd-basedSelectiveCachingof Page-levelAddressMappings InProceedingofthe14th internationalconferenceonArchitecturalsupportforprogram- minglanguagesandoperatingsystems(ASPLOS09),March 2009. [5]JesungKim,JongMinKim,SamH.Noh,SangLyulMin, YookunCho. ASpace-EfcientFlashTranslationLayerfor CompactFlashSystems IEEETransactionsonConsumer Electronics,Vol.48,No.2,2002. [6]Tae-SunChung,Dong-JooPark,Sang-WonPark,Dong-Ho Lee,Sang-WonLee,Ha-JooSong. SystemSoftwareforFlash Memory:ASurvey InProceedingsofthe2006IFIPInterna- tionalConferenceonEmbeddedAndUbiquitousComputing, August2006. [7]S.Lee,D.Park,T.Chung,D.Lee,S.Park,andH.Song A LogBufferbasedFlashTranslationLayerUsingFullyAsso- ciativeSectorTranslation IEEETransactionsonEmbedded ComputingSystems,6(3):18,2007.ISSN1539-9087. [8]C.Park,W.Cheon,J.Kang,K.Roh,W.Cho,andJ.-S.Kim. ARe-congurableFTL(FlashTranslationLayer)Architecture forNANDFlash-basedApplications ACMTransactionson EmbeddedComputingSystems,7(4),2008. [9]Z.Liu,L.Yue,P.Wei,P.Jin,andX.Xiang. AnAdaptive Block-SetbasedManagementforLarge-ScaleFlashMemory InProceedingsofthe2009ACMSymposiumonApplied Computing,pages1621-1625,2009. [10]L.P.ChangandT.W.Kuo. AReal-timeGarbageCollection MechanismforFlashMemoryStorageSysteminEmbedded Systems The8thInternationalConferenceonReal-Time ComputingSystemsandApplications,2002. [11]ThomasGleixner,FrankHaverkamp,andArtemBityutskiy. UBI-UnsortedBlockImages IBMTechReport,2006-06-09. [12]Y.H.Bae. Designofahighperformanceashmemory-based solidstatedisk JournalofKoreanInstituteofInformation ScientistsandEngineers,25(6),2007. [13]Yuan-HaoChang,Jen-WeiHsieh,Tei-WeiKuo. Endurance enhancementofash-memorystoragesystems:anefcient staticwearlevelingdesign Proceedingsofthe44thannual conferenceonDesignautomation,June04-08,2007,San Diego,California. [14]Hyun-jinCho,DongkunShin,YoungIkEom. KAST:K- associativesectortranslationforNANDashmemoryinreal- timesystems DATE2009:507-512. 1021 [15]HyunJinChoi,Seung-HoLim,KyuHoPark. JFTL:A ashtranslationlayerbasedonajournalremappingforash memory ACMTransactionsonStorage(TOS),v.4n.4,p.1-22, January2009. [16]Jen-WeiHsieh,Li-PinChang,Tei-WeiKuo. Efcienton-line identicationofhotdataforash-memorymanagement Pro- ceedingsofthe2005ACMsymposiumonAppliedcomputing, March13-17,2005,SantaFe,NewMexico. [17]Li-PinChang,Tei-WeiKuo. Efcientmanagementforlarge- scaleash-memorystoragesystemswithresourceconservation ACMTransactionsonStorage(TOS),v.1n.4,p.381-418, November2005. [18]IntelCorporation. UnderstandingtheFlashTranslationLay- er(FTL)Specication TechReportDecember1998. [19]J.Kang,H.Jo,J.Kim,andJ.Lee. ASuperblock-basedFlash TranslationLayerforNANDFlashMemory InProceedings oftheInternationalConferenceonEmbeddedSoftware(EM- SOFT),pages161 ´ lC170,October2006.ISBN1-59593-542-8. [20]J.U.Kang,J.S.Kim,C.Park,H.Park,andJ.Lee. A multi-channelarchitectureforhigh-performancenandash- basedstoragesystem JournalofSystemsArchitecture, 53(9):644 ´ lC658,2007. [21]A.Kawaguchi,S.Nishioka,andH.Motoda. Aash-memory basedFileSystem ProceedingsoftheUSENIXTechnical Conference,1995. [22]Tei-WeiKuo,Jen-WeiHsieh,Li-PinChang,Yuan-Hao Chang. Congurabilityofperformanceandoverheadsinash management Proceedingsofthe2006conferenceonAsia SouthPacicdesignautomation,January24-27,2006,Yoko- hama,Japan. [23]Gye-JeongKim,Seung-CheonBaek,Hyun-SookLee,Han- DeokLee,MoonJeungJoe. LGeDBMS:asmallDBMSfor embeddedsystemwithashmemory Proceedingsofthe32nd internationalconferenceonVerylargedatabases,September 12-15,2006,Seoul,Korea. [24] IometerProject http://www.iometer.org/. [25] GNUWgetProject http://www.gnu.org/s/wget