CS352H: Computer Systems Architecture

CS352H: Computer Systems ArchitectureLecture 1: What Is Computer Architecture?August 27, 2009University of Texas Computer ScienceCS352HFall 2009Don Fussell1

GoalsUnderstand the “how” and “why” of computer systemorganizationInstruction set architectureSystem organization (processor, memory, I/O)Micro-architectureLearn methods of evaluating performanceMetricsBenchmarksLearn how to make systems go fastPipelining, cachingParallelismLearn how to make systems less power hungry (maybe)University of Texas Computer ScienceCS352HFall 2009Don Fussell22

LogisticsLectures:TTh 3:30-5:00, WEL 3.260Instructor:Don FussellOffice:Office hours:Email:TA:ACES 2.120TTh 2:00-3:00 or by appointmentfussell@cs.utexas.eduDong LiOffice:Office Hours:Email:Grading:Final examMidterm 6 Homework 5 paper critiquesProjectHomework/project late account - 3 days for the semester25%20%30%25%Course webpage:http://www.cs.utexas.edu/ fussell/courses/cs352hUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell33

ResourcesText:Patterson & HennessyComputer Organization & Design:The Hardware/Software Interface, 4th Edition(required)Reference : Hennessy & PattersonComputer Architecture: A Quantitative Approach(not required)Software packagesSPIM for MIPS assemblyVerilog for class projectYou will need a CS department Unix accountUsing Blackboard:Messages about the class (make sure your e-mail address iscorrect)Other resources as needed in addition to the webpageFor your own discussion groupsUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell44

What is expected of youAttend class and participatePublished lecture notes are just “notes”Remember, Powerpoint was designed for content-free industrialpresentations, there’s more here than fits into bulletsDiscussions matter – they’re the best part of the courseBe a hero – ask stupid questions, you won’t be the only one who iswonderingDo the work – the homework and project are more than halfyour grade for a reason, doing them is the heart of the courseThings aren’t always as simple as they look – the devil is in thedetails, so sweat themDon’t procrastinateTake initiative to follow up on your interestsDon’t cheat: follow code of tmlUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell55

Computer Architecture“Computer architecture, like other architecture, is the art of determiningthe needs of the user of a structure and then designing to meet thoseneeds as effectively as possible within economic and technologicalconstraints.” F.P. Brooks, Planning a Computer System, Project Stretch, 1962What does this “design” consist of?A “programming” interface for softwareInstructionsState changesOrganizational principles for processors, memory and I/O devicesProcessors - microarchitectureOverall - system designIn order to:Meet functional and performance targetsWithin constraints, such as cost and powerWhile taking advantage of advances in technologyArchitecture is about making tradeoffsUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell66

What Does a Computer Architect Do?University of Texas Computer ScienceCS352HFall 2009Don Fussell77

CS352H TopicsTechnology TrendsInstruction set architecturesMIPS instruction setDesigning for performancePipeliningInstruction level parallelism (ILP)StaticDynamicMemory hierarchies and cachesVirtual memoryMultiprocessors and multicoreSystem measurementSystem implementationUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell88

Application constraintsApplications drive machine balanceNumerical simulation, scientific computing, 3-d graphics/gamesFloating point performanceMemory bandwidthTransaction processing, databasesI/O throughputInteger CPU performanceDecision supportI/O bandwidthEmbedded controllersI/O timing, interfacesPowerVideo processingLow precision pixel arithmeticUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell99

A Little Perspective: Processor PerformanceLog scale!!!!1Hz 1 cycle per second. 1GHz 109 cycles per secondTaken from: Hennessy & Patterson Computer Architecture: A Quantitative Approach, 4th edUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell1010

Technology scalingCPUsDevice density: 2x every 1.5 years ( 60%per year)Latency: 2x every 5 years ( 15% per year)Memory (DRAM)Capacity: 4x every 3 years ( 60% per year)(2x every two years lately)Latency: 1.5x every 10 yearsCost per bit: decreases about 25% per yearHard drives:Capacity: 4x every 3 years ( 60% per year)Bandwidth: 2.5x every 4 yearsLatency: 2x every 5 yearsBoards:Wire density: 2x every 15 yearsCables:No changeUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell1111

Machines Adapt to Changes in Technology1970s1990sMulti-chip CPUsSemiconductor memory is veryexpensiveMicro-coded controlComplex instruction setsLots of transistorsComplex control to exploit ILPLots of on-chip memoryMulti-level caches1980s2000sSingle-chip CPUsSome on-chip RAMSimple, hard-wired controlSimple instruction setsSmall on-chip cachesUniversity of Texas Computer ScienceApproaching 1B transistors!Slow wiresPower becoming expensiveMulticoresCS352HFall 2009Don Fussell12

The Times They Are A-Changin'19622001University of Texas Computer ScienceCS352HFall 2009Don Fussell1313

What is a Computer?Supercomputer?Web server?Game console?Desktop?Notebook?Microcontroller?Each of these “lives” in a different space with its ownrequirements and constraintsWe need to pick one for this courseDesktopUsing the MIPS ISAUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell1414

Interface DesignInterface vs. ImplementationVisibilityThree types of interfaces:Between layersAPI, ISABetween modulesSCSI, PCIStandard representationsIEEE floating point, ASCIIA good interfaceLasts through several generations of implementationsIs simple – ‘economy of mechanism’ (KISS)University of Texas Computer ScienceCS352HFall 2009Don Fussell1515

Instruction-Set Architecture (ISA)HW/SW interfaceSW ImpactSupport OS functionsRestartable instructionsMemory managementA good compiler targetSimpleOrthogonalDenseHW ImpactAmenable to efficientimplementation over timeAmenable to parallelizationUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell1616

System-Level OrganizationDesign at the level ofprocessors, memories, More important toapplication performancethan CPU designFeeds and speedsConstrained by pincounts and signalingratesSystem balanceSun SparcStation20MBus ModuleCPUL2 CCL64852MBusMBus controlM-S AdapterSTDIOSBusApplication-specificSBusDMADriven byPerformance/cost goalsAvailable componentsTechnology constraintsUniversity of Texas Computer ScienceDRAMControllerserialkbdmouseaudioRTCBoot PROMFloppySCSIEthernetSBusCardsCS352HFall 2009Don Fussell1717

Micro-architectureRegister-transfer-level (RTL)Implement instruction setExploit technology capabilitiesLocality & concurrencyFloating-point UnitInteger UnitIterative processGenerate architectureEstimate costEvaluate performanceOvercoming sequential natureof programsPipeliningMultiple issueDynamic schedulingBranch prediction/speculationUniversity of Texas Computer ScienceTI SuperSPARC Bus InterfaceFall 2009Don Fussell1818

Performance EvaluationWhat’s the right measure?Instruction count? cache b/w? I/O throughput?Application response time? Application throughput?Measure what?Benchmarks? Real applications? Traces?Measure how?Simulation? Math models? Real tests?University of Texas Computer ScienceCS352HFall 2009Don Fussell1919

Major Components of a ComputerDesktop design targetProcessor: 25% of costMemory: 25% of costRest (I/O devices, power supply, enclosure): versity of Texas Computer ScienceInputOutputCS352HFall 2009Don Fussell2020

PC ChassisUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell2121

Motherboard organizationMac miniUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell2222

PC MotherboardUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell2323

Pentium 4 ChipUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell24

SummaryArchitecture: Structure of “things”Interplay between technology and architectureComputer architect uses quantitative methods to:Design interfacesDevelop machine organization: system-level & micro-architectureThis course:Instruction-Set Architectures (ISAs)Performance evaluation (a tiny bit)PipeliningMemorySystem organizationParallelismPowerUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell2525

First AssignmentRead the Moore paper (see webpage)Write a one page critiqueTypewritten, font size 10-12Critique formatHeadline message: a one sentence tag line for the paperElevator pitch: a one paragraph summary of the paper’s keymessagesA couple of paragraphs to support and analyze the elevator pitchHand in your one page review at the beginning of next classYou should use this format for all other papers as wellUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell2626

Next TimeHand in a hard copy of your one-page typewritten review of theMoore paper at the beginning of the classWe’ll discuss the paperThen, instruction set architectures, specifically the MIPS ISARead Chap 1 for backgroundRead Chap 2.1-2.10 for class prepUniversity of Texas Computer ScienceCS352HFall 2009Don Fussell2727

Computer Architecture “Computer architecture, like other architecture, is the art of determining the needs of the user of a structure and then designing to meet those needs as effectively as possible within economic and technological constraints.” F.P. Brooks, Planning a Computer System, Project Stretch, 1962 What does this “design .