EECS 252 Graduate Computer Architecture Lec 3 – Performance + Pipeline ReviewReview from last lectureOutlineDefinition: PerformancePerformance: What to measureHow Summarize Suite Performance (1/5)How Summarize Suite Performance (2/5)How Summarize Suite Performance (3/5)How Summarize Suite Performance (4/5)How Summarize Suite Performance (5/5)Example Standard Deviation (1/2)Example Standard Deviation (2/2)Comments on Itanium 2 and AthlonFallacies and Pitfalls (1/2)Fallacies and Pitfalls (2/2)CS252: AdministriviaSlide 17A "Typical" RISC ISAExample: MIPS ( MIPS)Datapath vs ControlApproaching an ISA5 Steps of MIPS Datapath Figure A.2, Page A-85 Steps of MIPS Datapath Figure A.3, Page A-9Inst. Set Processor ControllerSlide 25Visualizing Pipelining Figure A.2, Page A-8Pipelining is not quite that easy!One Memory Port/Structural Hazards Figure A.4, Page A-14One Memory Port/Structural Hazards (Similar to Figure A.5, Page A-15)Speed Up Equation for PipeliningExample: Dual-port vs. Single-portData Hazard on R1 Figure A.6, Page A-17Three Generic Data HazardsSlide 34Slide 35Forwarding to Avoid Data Hazard Figure A.7, Page A-19HW Change for Forwarding Figure A.23, Page A-37Forwarding to Avoid LW-SW Data Hazard Figure A.8, Page A-20Data Hazard Even with Forwarding Figure A.9, Page A-21Data Hazard Even with Forwarding (Similar to Figure A.10, Page A-21)Software Scheduling to Avoid Load HazardsSlide 42Control Hazard on Branches Three Stage StallBranch Stall ImpactPipelined MIPS Datapath Figure A.24, page A-38Four Branch Hazard AlternativesSlide 47Scheduling Branch Delay Slots (Fig A.14)Delayed BranchEvaluating Branch AlternativesProblems with PipeliningPrecise Exceptions in Static PipelinesAnd In Conclusion: Control and PipeliningEECS 252 Graduate Computer Architecture Lec 3 – Performance + Pipeline Review David PattersonElectrical Engineering and Computer SciencesUniversity of California, Berkeleyhttp://www.eecs.berkeley.edu/~pattrsnhttp://www-inst.eecs.berkeley.edu/~cs25201/14/19CS252-s06, Lec 02-intro2Review from last lecture•Tracking and extrapolating technology part of architect’s responsibility•Expect Bandwidth in disks, DRAM, network, and processors to improve by at least as much as the square of the improvement in Latency•Quantify Cost (vs. Price)–IC f(Area2) + Learning curve, volume, commodity, margins•Quantify dynamic and static power–Capacitance x Voltage2 x frequency, Energy vs. power•Quantify dependability–Reliability (MTTF vs. FIT), Availability (MTTF/(MTTF+MTTR)01/14/19CS252-s06, Lec 02-intro3Outline•Review•Quantify and summarize performance–Ratios, Geometric Mean, Multiplicative Standard Deviation•F&P: Benchmarks age, disks fail,1 point fail danger•252 Administrivia•MIPS – An ISA for Pipelining•5 stage pipelining•Structural and Data Hazards•Forwarding•Branch Schemes•Exceptions and Interrupts•Conclusion01/14/19CS252-s06, Lec 02-intro4Performance(X) Execution_time(Y) n = =Performance(Y) Execution_time(X) Definition: Performance•Performance is in units of things per sec–bigger is better•If we are primarily concerned with response timeperformance(x) = 1 execution_time(x)" X is n times faster than Y" means01/14/19CS252-s06, Lec 02-intro5Performance: What to measure•Usually rely on benchmarks vs. real workloads•To increase predictability, collections of benchmark applications-- benchmark suites -- are popular•SPECCPU: popular desktop benchmark suite–CPU only, split between integer and floating point programs–SPECint2000 has 12 integer, SPECfp2000 has 14 integer pgms–SPECCPU2006 to be announced Spring 2006–SPECSFS (NFS file server) and SPECWeb (WebServer) added as server benchmarks•Transaction Processing Council measures server performance and cost-performance for databases–TPC-C Complex query for Online Transaction Processing–TPC-H models ad hoc decision support–TPC-W a transactional web benchmark–TPC-App application server and web services benchmark01/14/19CS252-s06, Lec 02-intro6How Summarize Suite Performance (1/5)•Arithmetic average of execution time of all pgms?–But they vary by 4X in speed, so some would be more important than others in arithmetic average•Could add a weights per program, but how pick weight? –Different companies want different weights for their products•SPECRatio: Normalize execution times to reference computer, yielding a ratio proportional to performance =time on reference computer time on computer being rated01/14/19CS252-s06, Lec 02-intro7How Summarize Suite Performance (2/5)•If program SPECRatio on Computer A is 1.25 times bigger than Computer B, thenBAABBre ferenc eAre ferenc eBAePerformancePerformancimeExecutionTimeExecutionTimeExecutionTimeExecutio nTimeExecutionTimeExecutio nTSPECRatioSPECRatio25.1•Note that when comparing 2 computers as a ratio, execution times on the reference computer drop out, so choice of reference computer is irrelevant01/14/19CS252-s06, Lec 02-intro8How Summarize Suite Performance (3/5)•Since ratios, proper mean is geometric mean (SPECRatio unitless, so arithmetic mean meaningless)nniiSPECRatioeanGeometricM1•2 points make geometric mean of ratios attractive to summarize performance:1. Geometric mean of the ratios is the same as the ratio of the geometric means2. Ratio of geometric means = Geometric mean of performance ratios choice of reference computer is irrelevant!01/14/19CS252-s06, Lec 02-intro9How Summarize Suite Performance (4/5)•Does a single mean well summarize performance of programs in benchmark suite?•Can decide if mean a good predictor by characterizing variability of distribution using standard deviation•Like geometric mean, geometric standard deviation is multiplicative rather than arithmetic•Can simply take the logarithm of SPECRatios, compute the standard mean and standard deviation, and then take the exponent to convert back: iniiSPECRatioStDevtDevGeometricSSPECRationea nGeometricMlnexpln1exp101/14/19CS252-s06, Lec 02-intro10How Summarize Suite Performance (5/5)•Standard deviation is more informative if know distribution has a standard form–bell-shaped normal distribution, whose data are symmetric around mean –lognormal distribution, where logarithms of data--not data itself--are normally distributed (symmetric) on a logarithmic scale•For a lognormal distribution, we expect that 68%
View Full Document
Unlocking...