You are on page 1of 21

1/6/13 Supercomputers

www.thocp.net/hardware/supercomputers.htm

Guidetoonlineschools

IntroductiontoSupercomputers
EditorCornelisRobat

Makeyoupdffilesmore manageableonyourcomputer.

Buysoftwarethatwillmake convertingpdftodocfastand
simple.Withapdfconverteryou canconvertpdftowordinno time! Pricerunner.com:

Lookingtopurchaseanew computer?Checkoutthischeap

laptopofferfromthisgreatprice comparisonsitetoday!Lookingfor
informationonotherelectronics? Checkoutaffordablemp3players andCanonDigitalCamerashere today!Togetgreatcomputertips, checkoutthisinformationright away!

Introduction|Chronology|Construction|Operatingsystems|Programming|Pioneers|Companies Introduction We'veallheardofsupercomputers.Butwhenyouthinkofasupercomputer,youusuallythinkofHAL2000andother sciencefictionstuff.Right? Unlikemainframesandminicomputers,supercomputersareusedfortheheavystufflikeweathermaps,construction ofatombombs,findingoil,earthquakeprediction,andscienceswherealotofcalculationsmustbedone.Theyare alsousedtohelpgovernmentseavesdroponanythingpassingthroughtelephone,datalines,email,orradiowaves andanythingthatiswritten,etc. Asyoucanseeinthetablebelow,therankingofasupercomputerisnearthetopofthecomputerspectrum. Quantumcomputers* Gridcomputers Supercomputers Mainframes Minicomputers Microcomputers Terminals

www.thocp.net/hardware/supercomputers.htm

1/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

Embeddedcomputers *Note:Quantumcomputersonlyexistonpaper,apartfromsomeexperimentalsetups.Gridcomputersare comparablewithsupercomputersbutarealotcheaper. Historically,asupercomputerisassociatedwiththefastestcomputeravailableorthelargestinsize.Supercomputing means"masscomputingatultrahighspeed."

Definition Manydefinitions(2)ofsupercomputershavecomeandgone.Somefavoritesofthepastare
anycomputercostingmorethantenmilliondollars, anycomputerwhoseperformanceislimitedbyinput/output(I/O)ratherthanbytheCPU, anycomputerthatisonlyonegenerationbehindwhatyoureallyneed.

Ingeneral,asupercomputerperformsabovenormalcomputing,inthesensethatasuperstarisabovetheother stars.And,yes,youcanbuildyourownpersonalsupercomputer.Thiswillbeexplainedinalaterchapteron"Grid" computing.Notethatsuperanythingdoesnotmeantheabsolutefastestorbest. Thoughallcontemporarypersonalcomputersperforminthetensorhundredsofmegaflops(millionsofcalculationsper second),theystillcannotsolvecertainproblemsfastenough.Itwasonlyinthebeginningof2000thatthe supercomputingarenamovesintothegigaflopsregion.Whatthismeansisthatyoucanhaveacomputercalculate problemsatthespeedofafewgigaflops,butdoingthesamecalculationsat"just"100megaflopsandwithinacceptable time,too,isalmostimpossible.Meaning,withsupercomputersyoucandocalculationswithinatimelimitorsessionthat isacceptabletotheuser.Namely:YOU. Toputitstronger:youcandoanythinginrealtime(meaning:now,rightaway,immediately)withasupercomputerthat cannotbedoneinyourlifetimewithonesinglePC.

screendumpofNEC'sESS So,certaintasksare,insomeways,notpossibletodoinrealtimeonPCs.(2)ForexampleitwilltakeasinglePCmore thanafewdays(weeks)tocalculateaweathermap.Resultinginthepredictionsoftheweatherseveraldaysoldwhen themapisfinished.Thatdoesn'tsoundmuchlikeaprediction,doesit?Asupercomputerdoesthesamejobinafew minutes.That'smorelikewhatwewantasusers:topspeed. Theissueofcostandtime Constructionofsupercomputersisanawesomeandveryexpensivetask.Togetamachinefromthelaboratorytothe marketmaytakeseveralyears.Themostrecentdevelopmentcostsofsupercomputersvariedbetween150to500 milliondollarsormore.Youcanimaginethataprojectlikethatdrawsonalltheresourcesacompanyhas.Thisisoneof themajorreasonsthatthedevelopmentofasupercomputeriskeptveryhushhush.Thelatestsupersareonlypossible

www.thocp.net/hardware/supercomputers.htm

2/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

tocreatewiththehelpofgovernmentsandoneormorelargesizecompanies. Usingasupercomputerisexpensiveaswell.Asauser,youarechargedaccordingtothetimeyouusethesystemwhatis expressedinthenumberofprocessor(CPU)secondsyourprogramruns.Intherecentpast,Cray(oneofthefirst supercomputers)timewas$1,000perhour.Theuseofthis"Craytime"wasaverycommonwaytoexpresscomputer costsintimeanddollars. Theneed Whydoweneedasupercomputer?Well,asanormalpersononthestreet,youdon't.YourcellphoneorPDAhasmore computingpowerthanthefirstmainframesliketheENIACorMark1.Withtheinformationglutfloodingyoursenses,and thebloatedsoftwaretryingtochannelthat,wewillprobablyneedextremecomputingpowerinmaybeafewdecades. Thearchitectureincreatingthatpowerisalreadyonthehorizon:wirelessLANs,infobottechnology,gridcomputingand virtualcomputingcentersthussharedcomputingwillallbepartofourdailyusedequipment.Computerswillevenbe sewnintoourclothing.(SeeMIT'swearablecomputingproject) Whoreallyneedssupercomputingtodayaremostlyscientistsperformingmasscomputingatultrahighspeed.Theyuse suchcomputersinallimaginabledisciplines:spaceexplorationandrelatedimagery(picturinggalaxiesandintergalactic matter),environmentalsimulations(globalwarmingeffects)mathematics,physics(thesearchforthereallysmallestpart ofmatter),genetechnology(whatgeneitisthatmakesusold),andmanyothers. Morerealworldexamplesare:industrialandtechnologicalapplications,worldspanningfinancialandeconomicalsystems inwhichspeedisessential.Also,moreandmoresupercomputersareusedforcreatingsimulationsforbuildingairplanes, creatingnewchemicalsubstances,newmaterials,andtestingcarcrasheswithouthavingtocrashacar.Supercomputers areusedforapplicationswhereitwilltakemorethanafewdaystogettheresultsorwhentheresultsareimpossiblefor aslowercomputertocalculate. Chronology Belowfollowsashortnarrativeonhowsupercomputersevolvedfrommainframesandthelike.Again,theneedto developsupersdidnotcomeoutoftheblue.Governmentandprivatecompaniesalikeactedontheneedofthemarket, theneedtobringdowncostlycomputertime,andtocalculateasfastaspossibletosavetimethusmoney.Thelatteris notalwaystheprimaryreason.

Preludetosupercomputing19321964
Upto1965thereweresomeexperimentalsupercomputers.ButthefirstsuccessfulonewastheCDC6600. 1939 1940 AtanasoffBerryComputercreatedatIowaState KonradZuseZ2usestelephonerelaysinsteadofmechanicallogicalcircuits

1942

ENIAC1942 1943 ColossusBritishvacuumtubecomputer

www.thocp.net/hardware/supercomputers.htm

3/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

1944

ManchesterMark11944 1946 J.PresperEckert&JohnMauchly,ACM,AEEI,ENIAC,

HarvardMarkII 1947 FirstTransistor,HarvardMarkII(MagneticDrumStorage) Whirlwind

1948 ManchesterMarkI(1ststoredprogramdigitalcomputer),WhirlwindatMIT UNIVACI

1950 AlanTuringTestofMachineIntelligence,UNIVACI(USCensusBureau) 1951 WilliamShockleyinventstheJunctionTransistor 1952 ILLIACI,UNIVACIatLivermorepredicts1952election,MANIACbuiltatLosAlamos,AVIDAC builtatArgonne

www.thocp.net/hardware/supercomputers.htm

4/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

EDVAC

IBM7011953 1953 EDVAC,IBM701

1954

IBM650(firstmassproducedcomputer), FORTRANdevelopedbyJohnBackus ORACLEOakRidgeAutomatedComputerAndLogicalEngine

1955

TexasInstrumentsintroducesthesilicontransistor, UNIVACIIintroduced

1956

IBM7041956

IAS

1956 MANIAC2,DEUCE(fixedheaddrummemory),cloneofIAS

www.thocp.net/hardware/supercomputers.htm

5/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

1958 NipponTelegraph&TelephoneMusasino1:1stparametroncomputer,JackKilbyFirstintegrated circuitprototypeRobertNoyceworksseparatelyonIC's,NEC1101&1102

1960

PaulBaranatRanddevelopspacketswitching,NEAC2201, Whirlwindairtrafficcontrol, LivermoreAdvancedResearchComputer(LARC), ControlDataCorporationCDC1604, Firstmajorinternationalcomputerconference

Stretch 1961 IBMStretchMultiprogramming

IBM709

IBM7090console

1962

ControlDataCorporationopenslabinChippewaFallsheadedbySeymourCray(19251996), Telestarlaunched,

www.thocp.net/hardware/supercomputers.htm

6/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

Atlasvirtualmemoryandpipelinedoperations, TimesharingIBM709and7090

IBM360 1964 IBM360thirdgenerationcomputerfamily. Limitedtestbantreaty IEEEformed 1965 TheSageSystem,

Supercomputerera
Before1965,thereweremightycomputersakintowhatarecalledMainframes.Theywerefastbutnotfastenough. Usualdevelopmentssetinmotioninthe"presuper"eralaidthebasisforwhatfollowed.Industryandgovernment organizationsprimarilyintheUSAfelttheurgentneedforfastercomputing.Sincecalculatingasimplestressmodulefor designingenginestookseveraldaysonacontemporarymainframe,suckingupallavailableresources,theneedfora newclassofcomputingwasalmostpalpable:Supercomputing.ItwasSeymourCraythatbroughtittous. 1965

CDC6600designedbySeymourCray(Firstcommerciallysuccessfulsupercomputerspeedof 9megaflops) 1966 RS232Cstandardfordataexchangebetweencomputers&peripherals, IBM1360Photostoreonlineterabitmassstorage 1967 CMOSintegratedcircuits, TexasInstrumentsAdvancedScientificComputer(ASC) TheAdvancedScientificComputer,orASC,wasasupercomputerarchitecture designedbyTexasInstruments(TI)between1966and1973.KeytotheASC's designwasasinglehighspeedsharedmemory,whichwasaccessedbyanumber ofprocessorsandchannelcontrollers,inafashionsimilartoSeymoreCray's groundbreakingCDC6600.OnekeydifferencewasthattheCPUwassplitintwo,

www.thocp.net/hardware/supercomputers.htm

7/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

theoperatingsystemranonthe"peripheralprocessor",whileapplicationswererun onthededicated,slave,ALU,whichtheyreferredtoas"the"CPU.TheALU/CPU wasoneofthefirsttoincludevectorprocessinginstructions.(sourcewikipedia.org) 1968 RANDdecentralizedcommunicationnetworkconcept, DonaldKnuthalgorithms&datastructuresseparatefromtheirprograms, UNIVAC9400 1969 ARPAnet, SeymourCrayCDC7600(40megaflops) 1970 UnixdevelopedbyDennisRitchie&KennethThomson

Cray11976(courtesyCrayInc.) 1976 CrayResearchCRAYIvectorarchitecture(designedbySeymourCray,shapedthecomputer industryforyearstocome), DatapointintroducesARC(firstlocalareanetwork)

VAX11780(5)

1978 1979 1980

DECintroducesVAX11/780(32bitsuper/minicomputer) Xerox,DEC,Intelethernetsupport DavidA.PattersonandJohnHennessy"reducedinstructionset", CDCCyber205

CrayXMP(courtesyCrayInc.)

www.thocp.net/hardware/supercomputers.htm

8/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

1982

CrayXMP, JapanfifthgenerationcomputerprojectisintendedforresearchintoArtificialIntelligence. Theprojectgoalsmayneverbereachedbutthespinoffofthisprojectisveryrewarding towardothersupercomputingprojects.

NECSX2(courtesyofNEC)seehpcslide 1984 ThinkingMachinesandNcubearefounded:parallelprocessingisintroduced. Mainstreamsupercomputers: HitachiS810/20, FujitsuFACOMVP200, ConvexC1, NECSX2 1985

www.thocp.net/hardware/supercomputers.htm

9/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

ThinkingMachines,TheConnectionMachine,asthismodeliscalled,contains64,000CPUs 1986 1987 1988 Firstdistributedmemoryparallelcomputer(InteliPSC/1,32CPUs)deliveredtoORNL IBM3090VPF. EvansandSutherlandES1,FujitsuVP400E,NSFnetestablished Apollo,Ardent,andStellarGraphicsSupercomputers, HitachiS820/80,HypercubesimulationonaLANatORNL, 1989 CRAYYMP, TimBernersLee:WorldWideWebprojectatCERN, SeymourCray:FoundsCrayComputerCorp.BeginsCRAY3usinggalliumarsenidechips, 1990 BellLabs:allopticalprocessor, IntellaunchesparallelsupercomputerwithRISCmicroprocessors

NECSX3(courtesyNEC)seehpcslide 1991

www.thocp.net/hardware/supercomputers.htm

10/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

GordonBellprojectsthespeedofthenextfewCraysupercomputers Japanannouncesplansforthesixthgenerationcomputerbasedonneuralnetworks FirstMboneaudiomulticasttransmittedontheNet Mainstreamsupercomputers: NECSX3, HewlettPackardandIBMRISCbasedcomputers FujitsuVP2600 1992 ThinkingMachinesCM5

CrayT3D(courtesyCrayInc.)

1993 1994

CRAYT3D NetscapeandNCSAMosaicwebbrowsersopenuptheInternettononcomputerlitterates. LeonardAdlemanDNAascomputingmedium.

www.thocp.net/hardware/supercomputers.htm

11/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

IntelASCIred(AdvancedSuperComputerInitiative)

1997 ASCIRedfirstteraflopcomputerdelivered CalTech/JPLsimulates50,000syntheticforce 2000

BlueGeneisIBM'sfirstspecialpurposesupercomputerandwillbeusedformodelinghuman proteins.Unravelingthehumangeneticcodecanhelppharmaceuticallaboratoriestodevelop medicines. 2002

NEC'sEarthSimulatorrunningat35.61TFlops.(trillionoperationspersecond)TheEarth Simulatorconsistsof640supercomputers.Itisprimarilydesignedforenvironmental simulations. 2004 Inafewyears,architecturesofsuperschangeddramatically.Comparebothgraphsbelowfrom 2001(top)and2004(bottom)

www.thocp.net/hardware/supercomputers.htm

12/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

Theshareof clustercomputersgrewfromamere10%toalmost60%,SMP(singelmultipleprocessor) seemstohavedisappeared,andMPPdownto25%.ItseemsthatClustercomputerswonthe raceinsupercomputingland.Bothgraphsarecourtesyofwww.top500.org.

Architecture
Thereare,inprinciple,severalwaysyoucandesignasupercomputerandthereisonlyonejudge:speed,speedand speed. Thefollowingexcerptisfroma1989lecturedeliveredbyGordonBellandillustratesnicelythecontemporarydesign

www.thocp.net/hardware/supercomputers.htm

13/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

pitfalls. TheElevenRulesofSupercomputerDesign(7)
1)Performance,performance,performance.Peoplearebuyingsupercomputersforperformance.Performance,withina broadpricerange,iseverything.Thus,performancegoalsforTitanwereincreasedduringtheinitialdesignphaseeven thoughitincreasedthetargetsellingprice.Furthermore,thefocusonthesecondgenerationTitanwasonincreasing performanceaboveallelse. 2)Everythingmatters.TheuseoftheharmonicmeanforreportingperformanceontheLivermoreLoopsseverely penalizesmachinesthatrunpoorlyonevenoneloop.Italsobringslittlebenefitforthoseloopsthatrunsignificantly fasterthanotherloops.SincetheLivermoreLoopswasdesignedtosimulatetherealcomputationalloadmixat LivermoreLabs,therecanbenoholesinperformancewhenstrivingtoachievehighperformanceonthisrealisticmixof computationalloads. 3)Scalarsmatterthemost.Awelldesignedvectorunitwillprobablybefastenoughtomakescalarsthelimitingfactor. Evenifscalaroperationscanbeissuedefficiently,highlatencythroughapipelinedfloatingpointunitsuchastheVPU canbedeadlyinsomeapplications.TheP3TitanimprovedscalarperformancebyusingtheMIPSR3010scalarfloating pointcoprocessingchip.Thissignificantlyreducedoverheadandlatencyforscalaroperations. 4)Provideasmuchvectorperformanceaspriceallows.Peakvectorperformanceisprimarilydeterminedbybus bandwidthinsomecircumstances,andtheuseofvectorregistersinothers.Thusthebuswasdesignedtobeasfastas practicalusingacosteffectivemixofTTLandECLlogic,andtheVRFwasdesignedtobeaslargeandflexibleas possiblewithincostlimitations.GordonBell'sruleofthumbisthateachvectorunitmustbeabletoproduceatleasttwo resultsperclockticktohaveacceptablyhighperformance. 5)Avoidholesintheperformancespace.Thisisanamplificationofrule2.Certainspecificoperationsmaynotoccur ofteninan"average"application.Butinthoseapplicationswheretheyoccur,lackofhighspeedsupportcansignificantly degradeperformance.AnexampleofthisinTitanistheslowdivideunitonthefirstversion.Apipelineddivideunitwas addedtotheP3versionofTitanbecauseoneparticularbenchmarkcode(Flo82)madeextensiveuseofdivision. 6)Placepeaksinperformance.Marketingsellsmachinesasmuchasormoresothantechnicalexcellence.Benchmark andspecificationwarsareinevitable.Thereforethemostimportantinnerloopsorbenchmarksforthetargetedmarket shouldbeidentified,andinexpensivemethodsshouldbeusedtoincreaseperformance.Itisvitalthatthesystemcanbe calledthe"World'sFastest",eventhoughonlyonasingleprogram.Atypicalwaythatthisisdoneistobuildspecial optimizationsintothecompilertorecognizespecificbenchmarkprograms.Titanisabletodowellonprogramsthatcan makerepeateduseofalongvectorstoredinoneofitsvectorregisterfiles. 7)Provideadecadeofaddressing.Computersneverhaveenoughaddressspace.Historyisfullofexamplesof computersthathaverunoutofmemoryaddressingspaceforimportantapplicationswhilestillrelativelyearlyintheir life(e.g.,thePDP8,theIBMSystem360,andtheIBMPC).Ideally,asystemshouldbedesignedtolastfor10years withoutrunningoutofmemoryaddressspaceforthemaximumamountofmemorythatcanbeinstalled.Sincedynamic RAMchipstendtoquadrupleinsizeeverythreeyears,thismeansthattheaddressspaceshouldcontain7bitsmore thanrequiredtoaddressinstalledmemoryontheinitialsystem.AfirstgenerationTitanwithfullyloadedmemorycards uses27bitsofaddressspace,whileonly29bitsofaddresslinesareavailableonthesystembus.When16MbitDRAM chipsbecomeavailable,Titanwillbelimitedbyitsbusdesign,andnotbyrealestateonitsmemoryboards. 8)Makeiteasytouse.The"dustydeck"syndrome,inwhichuserswanttoreuseFORTRANcodewrittentwoorthree decadesearly,isrampantinthesupercomputerworld.Supercomputerswithparallelprocessorsandvectorunitsare expectedtorunthiscodeefficientlywithoutanyworkonthepartoftheprogrammer.Whilethismaynotbeentirely realistic,itpointsouttheissueofmakingacomplexsystemeasytouse.Technologychangestooquicklyforcustomers

www.thocp.net/hardware/supercomputers.htm

14/21

1/6/13

www.thocp.net/hardware/supercomputers.htm
tohavetimetobecomeanexpertoneachandeverymachineversion. 9)Buildonother'swork.OnemistakeonthefirstversionofTitanwasto"reinventthewheel"inthecaseoftheIPU compiler.StardentshouldhavereliedmoreheavilyontheexistingMIPScompilertechnology,anduseditsresourcesin areaswhereitcouldaddvaluetoexistingMIPSwork(suchasintheareaofmultiprocessing). 10)Designforthenextone,andthendoitagain.Inasmallstartupcompany,resourcesarealwaysscarce,andsurvival dependsonshippingthenextproductonschedule.Itisoftendifficulttolookbeyondthecurrentdesign,yetthisisvital forlongtermsuccess.Extracaremustbetakeninthedesignprocesstoplanaheadforfutureupgrades.Thebestway todothisistostartdesigningthenextgenerationbeforethecurrentgenerationiscomplete,usingapipelinedhardware designprocess.Also,beresignedtothrowingawaythefirstdesignasquicklyaspossible. 11)Haveslackresources.Expecttheunexpected.Nomatterhowgoodtheschedule,unscheduledeventswilloccur.Itis vitaltohavespareresourcesavailabletodealwiththem,eveninastartupcompanywithlittleextramanpoweror capital.

Supercomputerstooktwomajorevolutionarypaths:vectorandscalarprocessing.Asyoucanseefromthegraphbelow thereisabigdifference.

Vectorcalculationsgo7timesfasterbytheclickgraphG.Bell1989

Thewayscalarorvectorcalculationsareexecutedisprincipallydifferent.Onethingissurevectorbasedprocessorsare moreexpensive.Notbecauseofeconomics,butfortechnical/architecturalreasons,vectorCPUsaresimplymuchmore complex.TheESSfromNECin2002isasuperbasedonvectorprocessorsandalsothefastestforthenextfewyears. Thearchitectureissometimesbasedonembeddedtechnologyinthecentralprocessorsandsometimesitisentirelyleft totheoperatingsystemcssoftware.Almostalways,specialtyprocessorsandalotofperipheralsupportarebuilt specificallyforthe(mostly)uniquemachine. Sincethelate1990s,anewarchitecturecomesintogeneralinterest:Gridcomputing.Thisallowsarchitectstodesigna superaroundofftheshelfcomputerpartsliketheonebuiltinSeptember2003byVirginiaTechfrom1100AppleG5 boxesand2200MacG5processors,Thissupercost"only"5.2millionUSDandranat10+Terraflopsreachingthetop3 ofthetop500rankinglist.
(Toseeapromotionalofthissuper:http://images.apple.com/movies/us/apple/g5_va_tech_mwsf_2004/va_tech_480.mov)

www.thocp.net/hardware/supercomputers.htm

15/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

Construction
Theconstructionofsupercomputersissomethingyouhavetoplanverycarefulbecause,onceunderway,thereisno goingbacktodosomemajorrevision.Ifthathappens,thecompanylosesmillionsofdollarsandthatcanresultineither cancelingtheprojectandtryingtomakesomemoneyfromthetechnologydevelopedorcausingthecompanytogo brokeoralmostbroke.AnexampleisCray,since2000anindependentcompanyagain,buttheyhadafewdifficultyears. Mismanagementisanotherfactorthatcausessupercomputerprojectstogobust.Oneexampleisthefifthgeneration projectinJapan.Alotofspinoffcamefromthatproject,verytrue.ButimaginethepossibilitiesiftheJapanesehad succeeded. Thirdis,ofcourse,periodsofeconomicmalaise.Projectsgetstalled.AfineexampleofthatisIntel.In2002,that companyscrappeditssupercomputerprojectandtookthelosses. Allthisdoesnottellusmuchabouthowsupercomputersarebuilt,butitgivesapicturethatnotonlysciencedictates whatisbuildorsuccessful. Surprisinglyenough,supersareoftenbuiltfromexistingCPUsbutthereendsalllikelinesswithexistinghardware. Termslikesuperscalar,vectororientedcomputing,andparallel computingarejustsomeofthetermsusedinthisarena. Since1995,supershavebeenbuiltupfromaGRID,meaningan arrayorclusterofCPUs(evenordinaryPC's)connectedbyaspecial versionof,forexample,Linuxthusactinglikeonebigmachine. Thecostofthistypeofsuperisdramaticallylowercomparedtothe millionsneededtobuild"conventional"typesofsupercomputers.As ifwecansay"conventional"withoutraisinganeyebrowortwo.
Themostobviousisthatourtechnologyisnotcapable

NizamAliaskedus:Whyarethousandsof processorsusedinsupercomputers,whyisthere notonebigprocessorused?(April2003)

Therearevariousreasonsthatsupercomputersdonot
useonesingleCPU.

Afactisthatsupersarethefastestmachinesintheirtime.Yes,we smilelookingbackattheENIACbut,backin1942,itwasamiracle machineandthefastestaround. Therearethreeprimarylimitstoperformanceatthe supercomputerlevel:(1)


individualprocessorspeed, theoverheadinvolvedinmakinglargenumbersofprocessorswork togetheronasingletask, theinput/outputspeedbetweenprocessorsandbetweenprocessorsand memory.

ofproducingachipthatpowerfulasyet.

Lookatitthisway:thereisnowasmuchpowerinone singlechipas1000 ENIAC'sbuiltinthe1940's.Soas technologyadvances,therewillbeasinglechipas powerfulas1000'super'chipswehavetoday.Itisalla matteroftime.

Theotherreasonisthatsinceourtechnologyisnot
capableofmakingthatsuperchip,YET,wecombine manychipstoformone'virtual'processor,operatingas onebigsingleCPU.(seealso gridcomputing) ItisliketheVLAradiotelescope,manysmalldishes makeuponegiantdish,thecombineddishesaremuch morepowerfulandtakelessspacethanagiantsingle onewould. Inotherwords,whencombiningmanyCPUsanda smartoperatingsystem,weavoidproblemslike overheatingandsignalnoisecausedbyelectro tunnelingeffects,alsocalledquantumeffect,thatwill

Input/outputspeedbetweenthedatastoragemediumandmemory isalsoaproblem,butnomoresothaninanyothertypesof computersand,sincesupercomputersallhaveamazinglyhighRAM capacities,thisproblemcanbelargelysolvedwiththeliberal applicationoflargeamountsofmoney.(1) Thespeedofindividualprocessorsisincreasingallthetime,butat greatcostsinresearchanddevelopment.Therealityisthatweare beginningtoreachthelimitsofsiliconbasedprocessors.Seymour Crayshowedthatgalliumarsenidetechnologycouldbemadeto work,butitisverydifficulttoworkwithandveryfewcompanies nowareabletomakeusableprocessorsbasedonGeAs.Itwas

causetheprocessortobehaveabnormally.

Yetanotherreasonismoney.Toputmanysingle

www.thocp.net/hardware/supercomputers.htm

16/21

1/6/13

www.thocp.net/hardware/supercomputers.htm
reasonablycheapCPUsinalargearrayismostly cheaperthantoproduceagiantone.Withthe contemporarytechnologyavailabletousnow,wewould havetoproduceaCPUthesizeofseveralsquare meters.Thiswouldcostseveralnationstheirentire annualbudgetsjusttocreateit.Butthatwill,ineffect, notsolvetheproblemofcreatingamorepowerfulCPU becauseitwillonlybearepetitionandrecombinationof whatwealreadyhave.

suchaproblembackinthoseyearsthatCrayComputerwasforced toacquiretheirownGeAsfoundrysothattheycoulddothework themselves.(1) Thesolutiontheindustryhasbeenturningto,ofcourse,istoadd everlargernumbersofprocessorstotheirsystems,givingthem speedthroughparallelprocessing.Thisapproachallowsthemtouse relativelyinexpensivethirdpartyprocessors,orprocessorsthat weredevelopedforother,highervolumeapplicationssuchas personalorworkstationlevelcomputing.Thus,thedevelopment costsfortheprocessorarespreadoutoverafarlargernumberof processorsthanthesupercomputingindustrycouldaccountforon itsown.(1) However,parallelismbringsproblemsofhighoverheadandthe difficultyofwritingprogramsthatcanutilizemultipleprocessorsat onceinanefficientmanner.Bothproblemshadexistedbeforeas mostsupercomputershadfromtwotosixteenprocessors,butthey weremucheasiertodealwithonthatlevelthanonthelevelof complexityarisingfromtheuseofhundredsoreventhousandsof processors.Ifthesemachinesweretobeusedthewaymainframes hadbeenusedinthepast,thenrelativelylittleworkwasneededas amachinewithhundredsofprocessorscouldhandlehundredsof jobsatatimefairlyefficiently.Distributedcomputingsystems, however,are(orarebecoming,dependingonwhoyouask)more efficientsolutionstotheproblemofmanyuserswithmanysmall

Sothetechnologytomakeoneisnotinplaceyet.
Whatisneededistheimplementationofnew technology,somethinglikebioneuralprocessors, quantumbasedCPU's,basedornotbasedon(sub)nano technologyand,alsoanupcomingfield,CPUsbasedon light. Butbythetimeasuperchipwillbethere,Iam convincedthatscientistswillcombinethesenewchips againtocreateanevenmorepowerful"CPUarray." Anddon'tbesurprisedwhenthisprocesswillrepeat itselfuntileternity.

c.robat,editor

tasks.Supercomputers,ontheotherhand,weredesigned,built,andboughttoworkonextremelylargejobsthatcould notbehandledbyanothertypeofcomputingsystem.Wayshadtobefoundtomakemanyprocessorsworktogetheras efficientlyaspossible.Partofthejobishandledbythemanufacturer:extremelyhighendI/Osubsystemsarrangedin topologiesthatminimizetheeffectivedistancesbetweenprocessorswhilealsominimizingtheamountof intercommunicationrequiredfortheprocessorstogettheirjobsdone.(1)Forexample,theESSprojecthasconnected theirclusterswithglasfibercablesandusesspecialvectorprocessors. OperatingSystems Operatingsystemsarespecificallydesignedforeachtypeofsupercomputer.SometimesaredesignofanexistingOSwill dothejobifthesuper'sCPUsbelongtothesamefamily.Butmostly,sincethenewcomputerreallyissomethingnew,a newOSisneeded. Inothercases,onemakesuseofexistingcomponents,withonlyminorchangesneededtomakeprocessorsandIO peripheralstalktoeachotherandthemachineisupandrunning.Buildingsuperslikethatshortensthetimetomarket sometimestoonlyonesingleyear. OperatingsystemsofsupersarenotmuchdifferentfromotherOSofotherplatforms.Theydotheirbusinessona differentscaleandinternalcommunicationtoCPU'sandI/Oiscrucial.SometimesOS'saresplittingupina"parallelway" totakecareforthingsindependentlyfromeachother.Mostcomputershaveainternalclockthattakescareof synchronizingtheIO,somesupersdonotevenhaveaclocklikethatanymore. Thesecretofwhetherornotasuperisatitspeakperformanceoftenliesinthefinetuning.Justaswithcars,amal tunedcarusesalotmoregasandrunsslowerthanawelltunedcarofthesametypeandconfiguration.Optimizingcode (tuning)fortheoperatingsystemrequireshighlyspecializedprogrammersandscientists. Forexample,passingastringofbitsfromonememorymodule/locationtoanotherandbyoptimizingthismovingaround

www.thocp.net/hardware/supercomputers.htm

17/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

ofbitsbyjustamillionthofasecondwillmakeabigdifference!Reachingthatpeakperformancemaytakeasmuchasa fewyears.Howeveracompanymaydecidetoleavepartofthisprocesstotheendusers,mostlyscientists,becauseof economicreasons.Rememberthe80/20rule?With20%ofyoureffortyoucanrealize80%ofyourproject'sgoal,but with80%extratimeandmoneyyoumayonlyimprove20%.Somanufacturerstendtonegotiateabalancebetween effortandresults.Inthescientificworld,thingsaredifferent:onlytheresultcounts. Programmingsupers Programmingsupercomputersisnotathingyouwouldtakeupasanafternoonprogrammingjob.Itisverycomplicated toprogramasupercomputer.Sometimesanentirelynewoperatingsystemisneeded,orevenanewprogramming language.Butagain,asareconstructors,programmersarealsoreluctanttoleavetheproventechnologybehind. Thereareseveralstrategiesthatcanbefollowedinprogrammingsupercomputers. Oneofthemostpopulariscalledmassiveparallelism.Thereareadvantagesbutalsodisadvantagesinparallel computing.Itaddstothecomplexityofaprogramandthusincreasesthenumberofbugsinthecode.Andthatpushes thedeliverydaywaypasttheprojecteddateofoperation. Anotheroneisdistributedcomputingsuchashasbeenavailablesince1998evenviatheInternet(Seti@home,RSA keytesting,etc.). Thefactisthatprogrammingsupersisthemajoroverheadinsupercomputingprojects.Itcantakeanotheryearafter installingasuperbeforethesystembecomesfullyoperational. Pioneers Themostfamousnamesare:Atanasoff,Eckert&Mauchly,SeymourCray. Itishardtotellwhichindividualswereplayingkeyroles(wewillfindoutthough).Especiallyinthisarena,constructinga superisseldomanindividualundertakingbutateameffortledbyoneormorecompanies. Companies
Amdahl, Burroughs, CDC Cray, Fujitsu, Hitachi, HewlettPackard, IBM, Intel, NEC, SGI, Sun, TexasInstruments, ThinkingMachines Univac

Inthemainframeareaintheearly1960stheBUNCHwerethemajorplayersinthatfield.Othersenteredthearenain supercomputingafterCDCandCrayshoweditwaspossibletocreatesupers.

www.thocp.net/hardware/supercomputers.htm

18/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

In40years(2004)therearebutafewplayersleftinthesupercomputingarena:
Cray Dell HP(absorbed:Amdahl,CompaqwhoateDigital) IBM NEC SGI Sun

Thentherearesupergridsmadebyuniversitiesandthemilitary(VirginiaTechwith1100AppleG5dualprocessors,for example)andotherclustersoflesspowerfulmachineswithLinuxspecialOSsforparallelprocessing. Forthemostpart,USAcompaniesareholdingthebatoninsupercomputing.But,fromtimetotime,Japanesefirmsare takingover,justlikeatugofwar. Conclusion Supercomputersevolvedfromasingletaskoriented"CPU"(ENIAC)toamultipurposeCPUclusteredmachine(Deep ThoughtSGI).Thechartshownbelow,takenfromthetop500site(4),illustratesthisnicely.

(4)
Thegraphicshowswhattypeofarchitectureisinvolvedwithcomputersonthetop500rankinglist(verticalbar)

Thenextgenerationofsupercomputingisalreadyannouncingitself: Calculationswiththe(sub)speedoflightviaopticaldevicesandconnections. Opticalcomputing Transistors,theopticalequivalentofthem,arenolongerswitchedbyelectronsbut photons. CalculationsbyrecombiningDNAinaparallelenvironment.

www.thocp.net/hardware/supercomputers.htm

19/21

1/6/13

www.thocp.net/hardware/supercomputers.htm

DNAcomputing

Numerouspossibilitiesaretriedatthesametimethemostoptimalsolutionwillbe "thestrongesttosurvive." Difficulttoimagine,butthinkofitascalculationsbeingdonebeforeyouhavethought ofthem.

Quantumcomputing

Workisdoneintheblinkofaneyesincetimeisofnoessencehere.Massive computingbycalculationsdoneindifferenteventsofprobability.Don'tsayyouweren't told=)

Allnewtechnologieshaveonethingincommon: Things(calculations)arehappeningatthesametimeindifferentplaces. Inotherwords:youcanbuildyourroofatthesametimeasyoubuildyourwallanditallfitstogether!Likeaprefab homethatisputtogetherononespotwithintheblinkofaneye.Theprefabricationhoweverisdoneinotherworkshops ondifferentlocations,allinthesametime. Newmaterialsaretestedandnewinsightsinprogrammingwillslowlygetintroduced,demandinganewbreedof engineersandprogrammers. Inallthefields,especiallythehardwareside,developmentsaretakingplaceattopspeed. Ofcourse,itwilltakeatleastanotherdecadebeforethenewtechnologieswillhittheworkfloor.Severaldecadeslater, thatsametechnologywillbeinplaceinourdesktopmachines,ifwemaystillthinkthatwayaboutcomputerhardware anyway. Eventually,economicswilldecidewhatweasuserscanaffordtobuy. Asalways,manufacturerswillpromisetheneweraofcomputingisjustaroundthecornerespeciallyinthisarenaof computing:supercomputing. Don'tforgetthetimespanittook(1015years)togetanew"hightech"producttothemarketgetsshorter(510 years),ironically,becausemoresupercomputingpowerbecomesavailable. Computersareconstructingandtestingnewcomputers,whichintheirturnareevenfaster,andsoon. Thisconcludestheintroductioninsupercomputing.Ifyouhaveanyquestions,additions,betterwordingoranyother commentpleasefeelfreetoemailus. Introduction|Chronology|Construction|Operatingsystems|Programming|Pioneers|Companies

LastUNovember19,2007>>>>>>>>>>>

Forsuggestionspleasemailthe

editors

Footnotes&References 1 2
3 4 DanCalleei.cs.vt.edu/~history/supercom.calle.html Apple.com www.epm.oml.gov/ssiexpo/histext.html www.top500.org

www.thocp.net/hardware/supercomputers.htm

20/21

1/6/13
5 6 7 http://histoire.info.online.fr/vax.html

www.thocp.net/hardware/supercomputers.htm

NEC'sstrategyinsupercomputing( seehpcslide) Excerptedfrom:TheArchitectureofSupercomputers:Titan,aCaseStudy,Siewiorek&Koopman,AcademicPress,1991,ISBN:0126430608

www.thocp.net/hardware/supercomputers.htm

21/21

You might also like