Chapter1ComputerAbstractionsandTechnologyChapter1ComputerAbstractionsTheComputerRevolutionProgressincomputertechnologyUnderpinnedbyMoore’sLaw

WhatisMoore’sLaw???Moore'slawdescribesalong-termtrendinthehistoryofcomputinghardware.Thequantityoftransistorsthatcanbeplacedinexpensivelyonanintegratedcircuithasdoubledapproximatelyeverytwoyears.§1.1Introductionppt課件TheComputerRevolutionProgresMoore’sLawThetrendhascontinuedformorethanhalfacenturyandisnotexpectedtostopuntil2015orlater.ppt課件Moore’sLawThetrendhascontiTheComputerRevolutionMakesnovelapplicationsfeasibleComputersinautomobiles§1.1Introductionppt課件TheComputerRevolutionMakesnTheComputerRevolutionMakesnovelapplicationsfeasibleCellphones§1.1Introductionppt課件TheComputerRevolutionMakesnTheComputerRevolutionMakesnovelapplicationsfeasibleHumangenomeprojectWorldWideWebSearchEngines§1.1Introductionppt課件TheComputerRevolutionMakesnTheComputerRevolutionComputersarepervasive§1.1Introductionppt課件TheComputerRevolutionComputeClassesofComputersQuestion:Howdoyouclassifycomputers?DesktopComputersServerComputersEmbeddedComputersppt課件ClassesofComputersQuestion:ClassesofComputersDesktopcomputersPCGeneralpurpose,varietyofsoftwareSubjecttocost/performancetradeoffppt課件ClassesofComputersDesktopcoClassesofComputersServercomputersNetworkbasedHighcapacity,performance,reliabilityRangefromsmallserverstobuildingsized

World’ssmallestwebserverppt課件ClassesofComputersServercomClassesofComputersEmbeddedcomputersHiddenascomponentsofsystemsStringentpower/performance/costconstraintsppt課件ClassesofComputersEmbeddedcTheProcessorMarketppt課件TheProcessorMarketppt課件WhatYouWillLearnHowprogramsaretranslatedintothemachinelanguageAndhowthehardwareexecutesthemThehardware/softwareinterfaceWhatdeterminesprogramperformanceAndhowitcanbeimprovedHowhardwaredesignersimproveperformanceWhatisparallelprocessingppt課件WhatYouWillLearnHowprogramLevelsofProgramCodeHigh-levellanguageLevelofabstractionclosertoproblemdomainProvidesforproductivityandportabilityAssemblylanguageTextualrepresentationofinstructionsHardwarerepresentationBinarydigits(bits)Encodedinstructionsanddatappt課件LevelsofProgramCodeHigh-levBelowYourProgramApplicationsoftwareWritteninhigh-levellanguage(HLL)SystemsoftwareCompiler:translatesHLLcodetomachinecodeOperatingSystem:servicecodeHandlinginput/outputManagingmemoryandstorageSchedulingtasks&sharingresourcesHardwareProcessor,memory,I/Ocontrollers§1.2BelowYourProgramppt課件BelowYourProgramApplicationUnderstandingPerformanceAlgorithmDeterminesnumberofoperationsexecutedProgramminglanguage,compiler,architectureDeterminenumberofmachineinstructionsexecutedperoperationProcessorandmemorysystemDeterminehowfastinstructionsareexecutedI/Osystem(includingOS)DetermineshowfastI/Ooperationsareexecutedppt課件UnderstandingPerformanceAlgorComponentsofaComputerSamecomponentsfor

allkindsofcomputerDesktop,server,

embeddedInput/outputincludesUser-interfacedevicesDisplay,keyboard,mouseStoragedevicesHarddisk,CD/DVD,flashNetworkadaptersForcommunicatingwithothercomputers§1.3UndertheCoversTheBIGPictureppt課件ComponentsofaComputerSamecAnatomy(結(jié)構(gòu))ofaComputerOutputdeviceInputdeviceInputdeviceNetworkcableppt課件Anatomy(結(jié)構(gòu))ofaComputerOutpAnatomyofaMouseOpticalmouseLEDilluminatesdesktopSmalllow-rescameraBasicimageprocessorLooksforx,ymovementButtons&wheelSupersedesroller-ballmechanicalmouseppt課件AnatomyofaMouseOpticalmousThroughtheLookingGlassLCDscreen:pictureelements(pixels)Mirrorscontentofframebuffermemoryppt課件ThroughtheLookingGlassLCDsOpeningtheBoxppt課件OpeningtheBoxppt課件InsidetheProcessor(CPU)Datapath:performsoperationsondataControl:sequencesdatapath,memory,...CachememorySmallfastSRAMmemoryforimmediateaccesstodataSRAM–StaticRandomAccessMemoryppt課件InsidetheProcessor(CPU)DataInsidetheProcessorAMDBarcelona:4processorcoresppt課件InsidetheProcessorAMDBarcelASafePlaceforDataVolatile(易變的)mainmemoryLosesinstructionsanddatawhenpoweroffNon-volatilesecondarymemoryMagneticdiskFlashmemoryOpticaldisk(CDROM,DVD)ppt課件ASafePlaceforDataVolatileNetworksCommunicationandresourcesharingLocalareanetwork(LAN):EthernetWithinabuildingWideareanetwork(WAN):theInternetppt課件NetworksCommunicationandresoAbstractionsInstructionSetArchitecture(ISA)Aninterfacebetweenthehardwareandthelowest-level

softwareTheabstractimageofacomputingsystemthatisseenbyamachine/assemblylanguageprogrammerIncludinginstructions,registers,memoryaccess,I/O,…TheBIGPictureppt課件AbstractionsInstructionSetArISAsSystem/360andupwardscompatiblesuccessorsz/ArchitecturePowerArchitecturePDP-11SPARCSuperHTricoreTransputerUNIVAC1100/2200seriesVAXx86IA-32(32-bitx86,firstimplementedintheIntel80386)x86-64(64-bitsupersetofIA-32,firstimplementedintheAMDOpteron)EISC(AE32K)4004,40406800,6502,6809,68HC11,68HC08.8008,8080,8085,Z80,Z180,eZ80,etc.8048,8051,etc.Z8,eZ8,etc.AlphaARMBurroughsB5000seriesBurroughsB6000/B7000serieseSi-RISCIA-64(Itanium)Mico32MIPSMotorola68kPA-RISCIBM700/7000linesppt課件ISAsSystem/360andupwardscomAbstractionsApplicationBinaryInterface(ABI)thelow-level

interface

betweenanapplicationprogramandOS

ABIscoverdetailssuchasdatatype,size,alignment,callingconvention,binaryformatofobjectfiles,etc.Definesastandardforbinaryportabilityacrosscomputers.ppt課件AbstractionsApplicationBinaryAbstractionsImplementationHardwarethatobeysthearchitectureabstraction

ManyimplementationsforthesameISAExample:IntelPentiumvsAMDAthlon,almostidenticalISA,butradicallydifferentinternaldesigns.ppt課件AbstractionsImplementationppt課PERFORMANCEppt課件PERFORMANCEppt課件DefiningPerformanceWhichairplanehasthebestperformance?§1.4Performanceppt課件DefiningPerformanceWhichairpDefiningPerformanceWhichairplanehasthebestperformance?§1.4Performanceppt課件DefiningPerformanceWhichairpDefiningPerformancePerformance=SpeedE.g.,takingasinglepassengerfromonepointtoanotherWinner:ConcordePerformance=PassengerThroughputPassengersxm.p.hE.g.,transporting450passengersfromonepointtoanotherWinner:Boeing747ppt課件DefiningPerformancePerformancResponseTimeandThroughputResponsetimeHowlongittakestodoataskThroughputTotalworkdoneperunittimee.g.,tasks/transactions/…perhourppt課件ResponseTimeandThroughputReResponseTimeandThroughputHowareresponsetimeandthroughputaffectedbyReplacingtheprocessorwithafasterversion?Addingmoreprocessors?We’llfocusonresponsetimefornow…ppt課件ResponseTimeandThroughputHoRelativePerformanceDefinePerformance=1/ExecutionTime“XisntimefasterthanY”Example:timetakentorunaprogram10sonA,15sonBExecutionTimeB/ExecutionTimeA

=15s/10s=1.5SoAis1.5timesfasterthanBppt課件RelativePerformanceDefinePerMeasuringExecutionTimeElapsedtimeTotalresponsetime,includingallaspectsProcessing,I/O,OSoverhead,idletimeDeterminessystemperformanceCPUtimeTimespentprocessingagivenjobDiscountsI/Otime,otherjobs’sharesComprisesuserCPUtimeandsystemCPUtimeDifferentprogramsareaffecteddifferentlybyCPUandsystemperformanceppt課件MeasuringExecutionTimeElapseCPUClockingOperationofdigitalhardwaregovernedbyaconstant-rateclockClock(cycles)Datatransfer

andcomputationUpdatestateClockperiodClockcycletime:durationofaclockcyclee.g.,250ps=0.25ns=250×10–12sClockfrequency(clockrate):cyclesperseconde.g.,4.0GHz=4000MHz=4.0×109Hzppt課件CPUClockingOperationofdigitCPUTimePerformanceimprovedbyReducingnumberofclockcyclesIncreasingclockrateHardwaredesignermustoftentradeoffclockrateagainstcyclecountppt課件CPUTimeppt課件CPUTimeExampleComputerA:2GHzclock,10sCPUtimeDesigningComputerBAimfor6sCPUtimeCandofasterclock,butcauses1.2×clockcyclesHowfastmustComputerB’sclockratebe?

ppt課件CPUTimeExampleComputerA:2GInstructionCountandCPIInstructionCountforaprogramDeterminedbyprogram,ISAandcompilerAveragecyclesperinstructionDeterminedbyCPUhardwareIfdifferentinstructionshavedifferentCPIAverageCPIaffectedbyinstructionmixCPIppt課件InstructionCountandCPIInstrCPIExampleComputerA:CycleTime=250ps,CPI=2.0ComputerB:CycleTime=500ps,CPI=1.2SameISAWhichisfaster,andbyhowmuch?Aisfaster……bythismuchppt課件CPIExampleComputerA:CycleTCPIinMoreDetailIfdifferentinstructionclassestakedifferentnumbersofcyclesWeightedaverageCPIRelativefrequencyppt課件CPIinMoreDetailIfdifferentCPIExampleAlternativecompiledcodesequencesusinginstructionsinclassesA,B,CSequence1:IC=5ClockCycles

=2×1+1×2+2×3

=10Avg.CPI=10/5=2.0Sequence2:IC=6ClockCycles

=4×1+1×2+1×3

=9Avg.CPI=9/6=1.5Whatisavg.CPI?IC=InstructionCountppt課件CPIExampleAlternativecompilePerformanceSummaryPerformancedependsonAlgorithm:affectsIC,possiblyCPIProgramminglanguage:affectsIC,CPICompiler:affectsIC,CPIInstructionsetarchitecture:affectsIC,CPI,TcTheBIGPictureppt課件PerformanceSummaryPerformancePOWERppt課件POWERppt課件PowerTrendsInCMOSICtechnology§1.5ThePowerWall×1000×305V→1Vppt課件PowerTrendsInCMOSICtechnolReducingPowerSupposeanewCPUhas85%ofcapacitiveloadofoldCPU15%voltageand15%frequencyreductionThepowerwallWecan’treducevoltagefurtherWecan’tremovemoreheatHowelsecanweimproveperformance?ppt課件ReducingPowerSupposeanewCPUniprocessorPerformance§1.6TheSeaChange:TheSwitchtoMultiprocessorsConstrainedbypower,instruction-levelparallelism,memorylatencyppt課件UniprocessorPerformance§1.6TMultiprocessorsMulticoremicroprocessorsMorethanoneprocessorperchipRequiresexplicitlyparallelprogrammingComparewithinstructionlevelparallelismHardwareexecutesmultipleinstructionsatonceHiddenfromtheprogrammerHardtodoProgrammingforperformanceLoadbalancingOptimizingcommunicationandsynchronizationppt課件MultiprocessorsMulticoremicroManufacturingppt課件Manufacturingppt課件ManufacturingICsYield:proportionofworkingdiesperwafer§1.7RealStuff:TheAMDOpteronX4ppt課件ManufacturingICsYield:proporAMDOpteronX2WaferX2:300mmwafer,117

chips,90nmtechnologyX4:45nmtechnologyppt課件AMDOpteronX2WaferX2:300mmIntegratedCircuitCostNonlinearrelationtoareaanddefectrateWafercostandareaarefixedDefectratedeterminedbymanufacturingprocessDieareadeterminedbyarchitectureandcircuitdesignppt課件IntegratedCircuitCostNonlineBenchmarkingppt課件Benchmarkingppt課件SPECCPUBenchmarkProgramsusedtomeasureperformanceSupposedlytypicalofactualworkloadStandardPerformanceEvaluationCorp(SPEC)DevelopsbenchmarksforCPU,I/O,Web,…SPECCPU2006ElapsedtimetoexecuteaselectionofprogramsNegligibleI/O,sofocusesonCPUperformanceNormalize

relativetoreferencemachineSummarizeasgeometricmeanofperformanceratiosCINT2006(integer)andCFP2006(floating-point)ppt課件SPECCPUBenchmarkProgramsuseCINT2006forOpteronX42356ppt課件CINT2006forOpteronX42356ppSPECPowerBenchmarkPowerconsumptionofserveratdifferentworkloadlevelsPerformance:ssj_ops/secPower:Watts(Joules/sec)ppt課件SPECPowerBenchmarkPower