1Copyright2015EMCCorporation.
Allrightsreserved.
RAIDShield:Characterizing,Monitoring,andProactivelyProtectingAgainstDiskFailuresPresenter:AoMaAjointworkwithFredDouglis,GuanlinLu,DarrenSawyerSurendarChandra,WindsorHsu2Copyright2015EMCCorporation.
Allrightsreserved.
DiskfailuresarecommonplaceWhole-diskfailurePartialfailureRAIDiswidelydeployedProtectdataagainstfailureswithredundancyPervasiveRAIDProtection3Copyright2015EMCCorporation.
Allrightsreserved.
StoragesystemisevolvingEscalateduseoflessreliabledrivescausesmorewhole-diskfailuresIncreasingdiskcapacityresultsinmoresectorerrorsSolutionAddextraredundancy(RAID5,RAID6,…)–EnsuredatareliabilityatthecostofstorageefficiencyRAIDOverviewIsaddingextraredundancyanefficientsolution4Copyright2015EMCCorporation.
Allrightsreserved.
Analyzed1millionSATAdisksandrevealedFailuremodesdegradingRAIDreliabilityReallocatedsectorsreflectdiskreliabilitydeteriorationDiskfailureispredictableBuiltRAIDSHIELD,anactivedefensemechanismReconstructfailingdiskbeforeit'stoolate!
PLATE:single-diskproactiveprotection–Deploymenteliminates70%ofRAIDfailuresARMOR:diskgroupproactiveprotection–RecognizevulnerableRAIDgroupsWhatWeDid5Copyright2015EMCCorporation.
Allrightsreserved.
BackgroundDiskfailureanalysisRAIDSHIELD:IdentifyfailureindicatorReallocatedSector(RS)characterizationSinglediskproactiveprotectionDiskgroupproactiveprotection5Outline6Copyright2015EMCCorporation.
Allrightsreserved.
Diskfailuredoesnotfollowafail-stopmodelTheproductionsystemsstudieddefinefailureasConnectionislostAnoperationexceedsthetimeoutthresholdWritefailsWhole-diskFailureDefinition7Copyright2015EMCCorporation.
Allrightsreserved.
EachdiskdrivemodelisdenotedasRelativesizeswithinafamilyareorderedbythecapacitynumber–E.
g.
A-2islargerthanA-1DiskModelPopulation(Thousands)FirstDeploymentLogLength(Months)A-13406/200860A-216511/200860B-110006/200848C-19310/201036C-225312/201036D-138409/201121DiskDataCollection8Copyright2015EMCCorporation.
Allrightsreserved.
WhatDoRealDiskFailuresLookLike9Copyright2015EMCCorporation.
Allrightsreserved.
0001424462031010203040500-612-1824-3036-4248-54A-2MonthDistributionofLifetimeofFailedDrives001241534291130102030400-612-1824-3036-4248-54A-1MonthAlargefractionoffaileddrivesarefoundatasimilaragePercentage(%)Percentage(%)10Copyright2015EMCCorporation.
Allrightsreserved.
ThenumberofaffecteddiskskeepgrowingAbout10%ofdisksgetsectorerrorsatthe3rdyearSectorerrornumbersincreasescontinuouslyAverageerrorcountincreases25%to300%yearoveryearIncreasingFrequencyofSectorErrors11Copyright2015EMCCorporation.
Allrightsreserved.
DrivefailingatasimilarageFailurerateisnotconstantAhighriskofmultiplesimultaneousfailuresIncreasingfrequencyofsectorerrorsExacerbateriskofreconstructionfailuresPassiveRedundancyisInefficientEnsuringreliabilityintheworstcaserequiresaddingconsiderableextraredundancy,makingitunattractivefromacostperspective12Copyright2015EMCCorporation.
Allrightsreserved.
MotivationEnsuredatasafetywithminimalredundancyProactivelyrecognizeimpendingfailuresandmigratevulnerabledatainadvanceMethodologyIdentifyindicatorofimpendingfailureIndicatorcharacterizationProactiveprotectionRAIDSHIELD,TheProactiveProtection13Copyright2015EMCCorporation.
Allrightsreserved.
PotentialindicatorsVariousdiskerrorsCriteriaofagoodindicatorIthappensmuchmorefrequentlyonfaileddisksratherthanworkingdisksApproachQuantifythediscriminationbetweenerrorvalueonfaileddisksandworkingones–DecilescomparisonisusedIdentifyFailureIndicator14Copyright2015EMCCorporation.
Allrightsreserved.
FaileddiskshavemoremediaerrorsthanworkingonesThediscriminationisnotsignificantenoughMediaErrorComparison2359152232478611123471330020406080100123456789faileddiskworkingdiskDecilesA-2MediaErrorCount15Copyright2015EMCCorporation.
Allrightsreserved.
2238718732752281212422025000001262905001000150020002500123456789faileddiskworkingdiskA-2DecilesReallocatedSectorCountA-2DecilesRSisstronglycorrelatedwithdiskfailuresReallocatedSector(RS)Comparison16Copyright2015EMCCorporation.
Allrightsreserved.
MostfaileddrivestendtohavealargernumberofRSthanworkingonesRSisstronglycorrelatedwithwhole-diskfailures,followedbymediaerrors,pendingsectorerrorsanduncorrectablesectorerrorsCorrelationBetweenSectorErrorsAndWhole-diskFailureRSisastrongindicatorofimpendingdiskfailure17Copyright2015EMCCorporation.
Allrightsreserved.
LargerRScountimplieshigherfailurerateintwo-monthwindowDiskFailureRateGivenDifferentRSCount1.
767758083858689909091929393949502040608010004080120160200240280320360400440480520560600DiskFailureRate(%)RScountA-2RSCharacterization(1)18Copyright2015EMCCorporation.
Allrightsreserved.
LargerRScount,fastertofail10%25%median75%90%RSCountTimeMargin(Days)DiskFailureTimeGivenDifferentRSCountRSCharacterization(2)19Copyright2015EMCCorporation.
Allrightsreserved.
RScountindicatesthedegreeofdiskreliabilitydeteriorationUsetheRScounttopredictimpendingdiskfailureinadvancePLATE:SingleDiskProactiveProtection20Copyright2015EMCCorporation.
Allrightsreserved.
70.
166.
66461.
859.
952.
14742.
63936.
94.
52.
82.
11.
71.
40.
80.
70.
40.
30.
27010203040506070809010020406080100200300400500600failurespredictedfalsepositivePercentage(%)BoththepredictedfailureandfalsepositiveratesdecreaseasthethresholdincreasesSimulationResult:FailuresCapturedRateGivenDifferentRSThresholdRSthreshold21Copyright2015EMCCorporation.
Allrightsreserved.
551515801070020406080100WithoutProactiveProtectionWithProactiveProtectionHardwareFailuresOthersTripleFailuresEliminatedTripleFailuresSingleproactiveprotectionreducesabout70%ofRAIDfailures,equivalentto88%ofthetriple-diskfailuresPLATEDeploymentResult:CausesofRecoveryIncidentsPercentage(%)22Copyright2015EMCCorporation.
Allrightsreserved.
10%remainingtriplefailuresPLATEmissesRAIDfailurescausedbymultiplelessreliabledrives,whoseRScountshaven'texceedthethresholdTriagePrioritizediskgroupswithhighestriskMotivationofARMOR:TheRAIDGroupProactiveProtection23Copyright2015EMCCorporation.
Allrightsreserved.
1-11-21-31-42-12-22-32-43-13-23-33-44-14-24-34-4XXXXThreatofFailureImminentFailureGoodDiskHealthyDG1ImminentFailureofDG2ProtectedDG3PossibleFailureofDG4Singlediskprotection:Replace2-3,2-4,3-4(PLATE)Can'tidentifyDG4northedifferencebetweenDG2andDG3Groupprotection:ReplaceDG4orincreaseredundancy(ARMOR)ProtectDG4andrecognizethedifferencebetweenDG2andDG3DiskGroupProtectionExample24Copyright2015EMCCorporation.
Allrightsreserved.
CalculatethesinglediskfailureprobabilityConditionalprobabilitythroughBayesTheoremCalculatetheprobabilityofavulnerableRAIDCombinationofthosesinglediskprobabilitiesthroughjointprobabilityARMORMethodology25Copyright2015EMCCorporation.
Allrightsreserved.
ThediscriminationshowsARMORiseffectivetorecognizeendangeredDGsInpractice,itidentifiesmostDGfailuresthatarenotpredictedbyPLATEProbabilityDecilesdistributionEvaluation0.
250.
330.
390.
440.
460.
50.
630.
730.
930.
150.
20.
230.
250.
270.
280.
30.
310.
3200.
20.
40.
60.
81123456789GroupswithmorethanonefailureGroupswithoutfailure26Copyright2015EMCCorporation.
Allrightsreserved.
GooglereportsSMARTmetricssuchasreallocatedsectorstronglysuggestanimpendingfailure,buttheyalsodeterminethathalfofthefaileddisksshownosucherrors[Pinheiro'07]DifferentworkloadandRAIDrewriteDiskfailurepredictionAveragemaximumlatency[Goldszmidt'12]SMARTfailureprediction[Murray'05,Hughes'02]RelatedWork27Copyright2015EMCCorporation.
Allrightsreserved.
Weanalyzed1millionSATAdrivesObservefailuremodesdegradingRAIDreliabilityRevealRScountreflectsthediskreliabilitydeteriorationDiskfailureispredictableWebuiltRAIDSHIELD,anactivedefensemechanismPLATE:singlediskproactiveprotection–Deploymenteliminates70%ofRAIDfailuresARMOR:diskgroupproactiveprotection–RecognizevulnerableRAIDgroups–HopetodeployinfutureIsaddingextraredundancyanefficientsolutionUseasmuchredundancyasneededtoensureavailabilityProactivereplacementshoulddecreasethelevelneededSummary28Copyright2015EMCCorporation.
Allrightsreserved.
RAIDShield:Characterizing,Monitoring,andProactivelyProtectingAgainstDiskFailuresQuestionsAcknowledgementAndreaArpaci-DusseauandRemziArpaci-DusseauDataDomainengineerteam,membersofADandCTOoffice,StephenManley29Copyright2015EMCCorporation.
Allrightsreserved.
CalculatethesinglediskfailureprobabilityCalculatetheprobabilityofavulnerableRAID
说明一下:gcorelabs的俄罗斯远东机房“伯力”既有“Virtual servers”也有“CLOUD SERVICES”,前者是VPS,后者是云服务器,不是一回事;由于平日大家习惯把VPS和云服务器当做一回事儿,所以这里要特别说明一下。本次测评的是gcorelabs的cloud,也就是云服务器。 官方网站:https://gcorelabs.com 支持:数字加密货币、信用卡、PayPal...
7月4日是美国独立日,大致就是国庆节的意思吧。hostodo今年提前搞了个VPS大促销活动,4款便宜VPS,相当于7折,续费不涨价,本次促销不定时,不知道有多少货,卖完为止。VPS基于KVM虚拟,NVMe阵列,1Gbps带宽,自带一个IPv4+/64 IPv6,solusvm管理,送收费版DirectAdmin授权,VPS在用就有效! 官方网站:https://www.hostodo.com ...
搬瓦工怎么样?这几天收到搬瓦工发来的邮件,告知香港pccw机房(HKHK_1)即将关闭,这也不算是什么出乎意料的事情,反而他不关闭我倒觉得奇怪。因为目前搬瓦工香港cn2 GIA 机房和香港pccw机房价格、配置都一样,可以互相迁移,但是不管是速度还是延迟还是丢包率,搬瓦工香港PCCW机房都比不上香港cn2 gia 机房,所以不知道香港 PCCW 机房存在还有什么意义?关闭也是理所当然的事情。点击进...
阵列卡为你推荐
brandoff香港购物在哪里硬盘工作原理简述硬盘的工作原理。百度关键词价格查询在百度设置关键字是怎么收费的冯媛甑夏如芝是康熙来了的第几期?百度关键词分析怎样对关键词进行分析和选择5xoy.comhttp://www.5yau.com (舞与伦比),以前是这个地址,后来更新了,很长时间没玩了,谁知道现在的地址? 谢谢,sodu.tw台湾人看小说的网站是www.cn12365.orgwww.12365china.net是不是真的防伪网站300373一搓黑是真的吗www.diediao.com跪求鸭王2鹤城勿扰黑龙江省的那个 城市是被叫做鹤城?
域名查询工具 租服务器价格 河南vps burstnet 主机 香港机房 韩国电信 typecho 长沙服务器 cdn联盟 vip购优惠 台湾谷歌 支付宝扫码领红包 能外链的相册 爱奇艺会员免费试用 中国电信网络测速 永久免费空间 ssl加速 umax 密钥索引 更多