《信息科学类专业英语》课件第2章

Vocabulary(词汇)ImportantSentences(重点句)QuestionsandAnswers(问答)Problems(问题)Lesson2ExtremeScaleComputing

(第二课超大规模计算)

Supercomputinghasbeenamajorpartofmyeducationandcareer,fromthelate1960swhenIwasdoingatomicandmolecularcalculationsasaphysicsdoctoratestudentattheUniversityofChicago,totheearly1990swhenIwasgeneralmanagerofIBM’sSP

familyofparallelsupercomputers.

Theperformanceadvancesofsupercomputersinthesepastdecadeshavebeenremarkable.ThemachinesIusedasastudentinthe1960sprobablyhadapeakperformanceofafewmillioncalculationspersecondormegaflops.Gigaflops(billions)peakspeedswereachievedin1985,teraflops(trillions)in1997,andpetaflops(a1followedbyfifteenzeros)in2008.

Thesupercomputingcommunityisnowaimingforexascalecomputing,1,000,000,000,000,000,000calculationspersecond.Thepursuitofexascale-classsystemswasahottopicattherecentSC09

supercomputingconference.

Inthequestforthefastestmachines,supercomputershavealwaysbeenattheleadingedgeofadvancesinIT,identifyingthekeybarrierstoovercomeandexperimentingwithtechnologiesandarchitecturesthatgenerallythenappearinmorecommercialproductsafewyearslater.

Throughthe1970sand1980s,thefastestsupercomputerswerebasedonvectorarchitecturesandusedhighlysophisticatedtechnologiesandliquidcoolingmethodstoremovethelargeamountsofheattheygenerated.

Bythelate1980s,thesecomplexandexpensivetechnologiesranoutofgas.Asthemicroprocessorsusedinpersonalcomputersandtechnicalworkstationswerebecomingincreasinglypowerful,youcouldnowbuildsupercomputersusingtheseCMOSmicrosandparallelarchitecturesatamuchlowerpricethanthepreviousgenerationsofvectormachines.Asimilartransitiontomicroprocessorcomponentsandparallelarchitecturestookplaceinthemainframesusedincommercialapplications.

Massivelyparallelarchitectures,usingtenstohundredsofthousandsofprocessorsfromthePCandUnixmarketshavedominatedsupercomputingoverthepasttwentyyears.Theygotusintotheterascaleandpetascaleranges.But,theywillnotgetustoexascale.AnothermassivetechnologyandarchitecturaltransitionnowloomsforsupercomputingandtheITindustryingeneral.[1]

Anticipatingthemajorchallengesinvolvedinthetransitiontoexascale,theDepartmentOfEnergy(DOE)andDARPAlaunchedaseriesofactivitiesaroundthreeyearsagotostartplanningforsuchsystems.

ThisDARPAExaScaleComputingStudyprovidesaverygoodoverviewofthekeytechnologychallenges.Thestudyidentifiedfourmajorchallengeswherecurrenttrendsareinsufficient,anddisruptivetechnologybreakthroughsareneededtomakeexascalecomputingareality.

TheEnergyandPowerChallengeispervasive,affectingeverypartofthesystem.Today’sleadingedgepetascalesystemsconsumebetween2-3Megawatts(MW)perpetaflop.Itisgenerallyagreedthatanexaflopsystemmustconsumearound20MWs,otherwisetheiroperatingcostswouldbeprohibitivelyexpensive.The1000-foldincreaseinperformancefrompetascaletoexascalemustthusbeachievedatnomorethana10-foldincreaseinpowerconsumption.

Suchstretchtargetswereactuallyachievedinthetransitionfromterascaleinthelate1990stopetascalenow.Butnoonebelievesitcanbedoneagainwithtoday’stechnologies,hencetheassumptionthatatechnologyandarchitecturaltransitionasprofoundastheonetwodecadesagoisnowrequired.

TheMemoryandStorageChallengeisamajorconsequenceofthepowerchallenge.Thecurrentlyavailablemainmemories(DRAM)anddiskdrives(HDD)thathavedominatedcomputinginthelastdecadeconsumewaytoomuchpower.Newtechnologiesareneeded.

TheConcurrencyandLocalityChallenge

isanotherconsequenceofthepowerchallenge.Overthepasttwentyyearswehavebeenabletoachieveperformanceincreasesthroughacombinationoffasterprocessesandhigherlevelsofparallelism.But,wearenolongerabletoincreasetheperformanceofasingleprocessingelementbyturninguptheclockrateduetopowerandcoolingissues.

Wenowhavetorelysolelyonincreasedconcurrency.

Thetopterascalesystemsoftenyearsagohadroughly10,000processingelements.Today’spetscalesystemisupinthelow100,000s.But,because,theonlywaytonowincreaseperformancetowardanexascalesystemismassiveparallelism,anexaflopsupercomputermighthave100sofmillionsofprocessingelementsorcores.Suchmassiveparallelismwillrequiremajorinnovationsinthearchitecture,softwareandapplicationsforexascalesystems.ThisDARPAExascaleSoftwareStudyprovidesagoodoverviewofthesoftwarebreakthroughsrequired.

Finally,wehavetheResiliencyChallenge,thatis,theabilityofasystemwithsuchhugenumberofcomponentstocontinueoperationsinthepresenceoffaults.Anexascalesystemmustbetrulyautonomicinnature,constantlyawareofitsstatus,andoptimizingandadaptingitselftorapidlychangingconditions,includingfailuresofitsindividualcomponents.TheexascaleresiliencychallengesarediscussedinthisDARPAreportonSystemResiliencyatExtremeScales.

Therearevastbusinessimplicationstosuchamassivetechnologyandarchitecturaltransition.Forone,theecosystemofthepasttwentyyears,wherePCshaveprovidedthecomponentsforparallelsupercomputers,isnowgivingwaytoanewbusinessecosystem.Consumerelectronics,mobiledevicesandembeddedsensorsarenowthenewpartnersoftheextremescalesupercomputingcommunity,becausetheysharethesamerequirementsforplentiful,powerfulandinexpensivecomponentsthatconsumelittlepower.

Thistransitiontoanewecosystemalreadystartedaboutfiveyearsago.IBM’sBlueGenefamilyusesrelativelylowpower,embeddedcoresasitsprocessingelements,andRoadrunner’shybriddesignincludestheCellprocessorsoriginallydevelopedforSony’sPlayStation3.

ThemostpowerfulsupercomputersintheUShavegenerallybeendevelopedfor,andfirstinstalledatDOE’snationallabs,eitheraspartofitsAdvancedScientificComputingResearch(ASCR)programinsupportofenergyandenvironmentalresearch,ortheAdvancedSimulationandComputingprograminsupportofnuclearweaponsresearch.TheseDOElabstypicallyworkcloselywiththevendorsintherequirementsanddesignofsuchleadingedgesupercomputers.

Tobegintounderstandtherequirementsforexascalemachines,ASCRsponsoredaseriesoftownhallmeetings,whichheldopendiscussionsonthemostcriticalandchallengingproblemsinenergy,theenvironmentandbasicscience.Thesemeetingswherethenfollowedbyaseriesoftechnicalworkshops,eachfocusingonaspecificscientificdomain.

TheDOEtownhallsandworkshopshaveidentifiedtheopportunityforexascalecomputingtorevolutionizethewayweapproachchallengesinenergyresearch,environmentalsustainabilityandnationalsecurity.Theyalsoidentifiedtheimpactofexascalecomputingonkeyscienceareaslikebiology,astrophysics,climatescienceandnuclearphysics.

Oneoftheirmostcompellingconclusionsisthatwithexascalecomputing,wearereachingatippingpointinpredictivescience,anareawithpotentiallymajorimplicationsacrossawholerangeofnew,massivelycomplexproblems.Letmeexplain.

HighendsupercomputersaregenerallydesignedforeithercapabilityorcapacitycomputingCapabilitysupercomputersdedicatethewhole(ormostofthe)machinetosolveaverylargeproblemintheshortestamountoftime.Capacitysupercomputers,ontheotherhand,supportlargenumbersofuserssolvingdifferentkindsofproblemssimultaneously.

Whilebothkindsofsupercomputingareveryimportant,initiativesdesignedtopushtheenvelope,likeDOE’sexascaleproject,tendtofocusonthedevelopmentofcapabilitymachinestoaddressGrandChallenge

problemslikethosementionedabove,thatcouldnotbesolvedinanyotherway.

Capabilitycomputinghasbeenprimarilyappliedtowhatissometimesreferredtoasheroiccomputations,wherejustaboutthewholemachineisappliedtoasingletask.And,withoutadoubt,therearequiteanumberofproblemsthatwewillbeabletoaddresswithmachines1000timesmorepowerful.

But,atleastasexciting,isthepotentialforexascalecomputingtoaddressaclassofhighlycomplexproblemsthathavebeenbeyondourreach,notjustduetotheirsheersize,butbecauseoftheirinherentuncertaintiesandunpredictability.Thewaytodealwithsuchuncertaintyistosimultaneouslyrunmultipleensemblesorcopiesofthesameapplications,usingmanydifferentcombinationsofparameters,andthusbeabletoexplorethesolutionspaceoftheseotherwiseunpredictableproblems.Thiswillletussearchforoptimalsolutionstomanyproblemsinscienceandengineering,aswellasenableustocalculatetheprobabilitiesofextremeevents.

Thisnewstyleofpredictivemodelingwillhelpusapplymorescientificmethodologiestomanykindsofproblems,fromclimatestudiestothedesignofsafenuclearreactors.Beyondscienceandengineering,therearemanydisciplinesthatwillbenefitfromsuchpredictivecapabilities,fromeconomicsandmedicinetobusinessandgovernment.

Ensemblecomputinghasattributesofbothcapabilityandcapacitycomputing.Itdevotesthewholemachinetooneproblem,butitdoessobyrunningmanycopiesoftheprobleminparallelwithdifferentinitialconditions.Innovativetechniquesarealreadyemergingtohelpdevelopersbetterprogramandmanagesuchensemble-orientedapplications.

Finally,itisimportantnottounderestimatetheimpactofexascalebreakthroughstomorecapacityorientedmachines,aswellastosmallermachinesthatsharethesametechnologies,architecture,softwareandapplications.Manyoftheinnovationthatwillenableustodevelopexascaleclasssupercomputerswillyieldrelativelyinexpensivepetascaleclasssystemsaswellassmallerones.Thewidertheaccesstosuchfamiliesofsystems,therichertheoverallecosystemincludingapplications,usersandtechnologies.

Inaddition,thesesameexascaleinnovationswillfindwideusageinthemorecommerciallyorientedcloudcomputingsystems.Thetechnologyrequirementsarequitesimilar,especiallytheneedforlowpower,lowcostcomponents.Theyalsosharesimilarrequirementsforhighlyefficient,autonomicsystemmanagement.Onecanactuallyviewcloud-basedsystemsasakindofexascaleclasssupercomputersdesignedtosupportembarrassinglyparallelworkloads,suchasmassiveinformationanalysisorhugenumbersofsensorsandmobiledevices.

InitsStrategyforAmericanInnovation,theObamaadministrationlistedexascalecomputingamongtheGrandChallenges

ofthe21stcenturyinscience,technologyandinnovation,that“willallowthenationtosetandmeetambitiousgoalsthatwillimproveourqualityoflifeandestablishthefoundationfortheindustriesandjobsofthefuture.”Itexplicitlycalledfor.[2]

“Anexascalesupercomputercapableofamilliontrillioncalculationspersecond–dramaticallyincreasingourabilitytounderstandtheworldaroundusthroughsimulationandslashingthetimeneededtodesigncomplexproductssuchastherapeutics,advancedmaterials,andhighly-efficientautosandaircraft.”[3]

Thisisatrulyexcitingandimportantchallenge.1．molecularadj.adv.n.分子状态的;分子式;分子量。

2．remarkableadj.值得注意的；不寻常的。

3．loomn.织布机vi.隐约出现;阴森地逼近。

4．stretchvt.&vi.n.adj.伸展;拉紧;延伸;尽力,全力以赴。

5．resiliencyn.跳回,弹性。

6．ecosystemn. <生>生态系统。

7．sustainabilityn.持续性;能维持性;永续性。Vocabulary

8．compellingadj.引人入胜的;扣人心弦的;非常强烈的;不可抗拒的;令人信服的。

9．simultaneouslyadv.同时地。

10．ensemblen.系集，整体,总效果;合奏,合唱;合奏组。

11．callfor去接(某人),去取(某物);需要;要求。

12．therapeuticsn.[pl.]治疗学,疗法。

13．FLOPS(FLoatingpointOperationsPerSecond)每秒浮点操作数，用于度量浮点计算机速度的单位，见下表。

本文中的exascale是计算机科学家又定义的一个新词，exa来源于上面的表格中，暂且翻译成超大规模计算。

[1]Massivelyparallelarchitectures,usingtenstohundredsofthousandsofprocessorsfromthePCandUNIXmarketshavedominatedsupercomputingoverthepasttwentyyears.Theygotusintotheterascaleandpetascaleranges.But,theywillnotgetustoexascale.AnothermassivetechnologyandarchitecturaltransitionnowloomsforsupercomputingandtheITindustryingeneral.

ImportantSentences近20年来，在PC和UNIX市场上使用几十万个处理器的大规模并行体系机构占主导地位。这使得我们进入了万亿和千万亿计算的范围。但是还不能使我们进入极大规模计算。对于超级计算和信息产业来说，其他的大规模技术和体系结构的转变目前还不明朗。

[2]InitsStrategyforAmericanInnovation,theObamaadministrationlistedexascalecomputingamongtheGrandChallengesofthe21stcenturyinscience,technologyandinnovation,that“willallowthenationtosetandmeetambitiousgoalsthatwillimproveourqualityoflifeandestablishthefoundationfortheindustriesandjobsofthefuture.”Itexplicitlycalledfor.

在“美国创新战略”中，奥巴马政府把超大规模计算作为21世纪科学、技术和创新的重大挑战之一。他们明确号召：这将使得我们建立并达到改善生活质量、建立未来产业和工作岗位的基础的宏伟目标。

[3]Anexascalesupercomputercapableofamilliontrillioncalculationspersecond–dramatically