HPCA - 7meiyang/ecg700/readings/presentation_spec_benchmarks.pdf · HPCA - 7 Performance...
Transcript of HPCA - 7meiyang/ecg700/readings/presentation_spec_benchmarks.pdf · HPCA - 7 Performance...
spec
Standard Performance Evaluation Corporation
Kaivalya M. Dixit
Chief Cook & Bottlewasher, SPECServer Performance, IBM
HPCA - 7
Performance SPECulations - Benchmarks, Friend or Foe
Monterrey, MexicoJanuary 23, 2001
spec
Standard Performance Evaluation Corporation
Technology TrendsPopular Benchmarks -- will they ever die?Applications vs. Benchmarks - one perspectiveTop 500 Super Computers Performance Problem Benchmark viewed by: Engr, Mktg, Standards, Boss & UsersSPEC - infrastructureSPEC benchmarksCPU2K MPG - blind submissionCPU2k MPG - one year laterComing attractionsIssuesSummary
Topics
spec
Standard Performance Evaluation Corporation
Transistors (1971-2000)Technology (1971-2000)
7173
7577
7981
8385
8789
9193
9597
990.1
1
10
1000
10000
100000
1000000
10000000
100000000
4004
P4
486
P II
P III
mic
ron
s#
of
Tra
ns
isto
rs
spec
Standard Performance Evaluation Corporation
Frequency (1971-2000)Technology (1971-2000)
71 73 75 77 79 81 83 85 87 89 91 93 95 97 990.1
1
10
0.1
1
10
100
1000
10000
P III
mic
ron
sM
Hz
4004
486
P4
spec
Standard Performance Evaluation Corporation
Chip #Transistors (Million )
Frequency (MHz)
Alpha 15.2 833AMD 37 1200HP 116 552IBM 44 600Sun 29 900SGI 7.2 400
Other Architectures - Today
IBM POWER 4 has announced 170+ million transistors
spec
Standard Performance Evaluation Corporation
Tomorrow Copper/silver, Silk, SOI, Electron Beam and other technologies may extend Moore's law for next 10 to 12 yearsSmall Transistors ( 3 atoms )High Clockrates ( 10-20 GHz )System on a chipCustom processorsHigh performance and Low power (oxymoron?)GIPs, Gflops and GDB available to everyoneChallenge Amdahl's Law?
spec
Standard Performance Evaluation Corporation
DhrystonesSPECTPCSAP
OracleEEMBCLinpackStreams
Dhrystones
DhampstonesLinpack
PerfectDhampstones
???
Streams
Benchmarks - Yesterday, Today & Tomorrow
More $$$s
Larger
spec
Standard Performance Evaluation Corporation
Linpack poor predictor (2 x Linpack = ~10% App. performance)SPECfp95 (2 x SPECfp95 = 2 x App. performance)BW demand on SPECfp2000 higher than App.Icache misses of SPECfp2000 an order of magnitude smaller than App.Mispredicted branches similar to App.Ratio of MemOPs to FPOPS 1-4 with mean of 2 for Apps. but 0.75 to 2 with mean of 1 for BenchmarkWe need to gather and share a lot more data from Apps. Reference: http://www.cs.varginia.edu/~mccalpin/wwc-keynote.html
Applications vs. Benchmarks
spec
Standard Performance Evaluation Corporation
Highlights from the Top 100:
ASCI White is new #1 with 4.9 TFlops on LinpackDOE ASCI systems hold the first 4 positions7 systems have Linpack performance above 1 TFlops18 systems have peak performance above 1 TFlops including one commercial system (#15 at Charles Schwab)0.89 Tflops is the entry point for the Top 10Super Computer Rating either MIPS or MFLOPS -- why?
spec
Standard Performance Evaluation Corporation
performance = f (MHz,Path Length, CPI,..)Perception: High Performance = High MHzCISC, RISC, VLIW, SMT,CMP, ...ISA, pipe-lining, super scalar, branch predictors, Reg. renaming, OOO, caches, load/store reordering, prefetch, memory, ...Technology, compiler, language, ....SMP, NUMA, MPP, Clusters, ...Benchmarks vs. Applications
Performance Equation
spec
Standard Performance Evaluation Corporation
Linpack inspired fma, load / store units, & compiler tricksDhrystone inspired string instructionsStream influenced CPU to memory designsSPEC inspired preprocessors, improved cache, branch predictors, OOO, ISAs, optimizers for graphics, web, compilers, Libs and FDOsTPC drove improvement of OS, database & I/OBoth TPC & SPEC have helped improve H/W and SW scalabilityAll Benchmark Specials are BAD
Benchmark Influence
spec
Standard Performance Evaluation Corporation
Performance
BenchmarkEngineer
Workloads
Analysis
Rules
Data
Metrics
Bias
Debate
#1
Lawyer
Marketing
Benchmark Metric = Complex
spec
Standard Performance Evaluation Corporation
Performance
BenchmarkMarketing
Papers
Value
Customer
Info
Driver
Bias
#1
Benchmark Metric = Single Number
spec
Standard Performance Evaluation Corporation
Performance
BenchmarkStandards
Papers
Rules
More BM
Analysis
Metrics
Bias
Workloads
Benchmark Metric = Full Disclosure
spec
Standard Performance Evaluation Corporation
Performance
BenchmarkBoss
Happy
#1
Value
$$$?
Benchmark Metric = Don't care
spec
Standard Performance Evaluation Corporation
Performance
BenchmarkUser#1
Value
Benchmark Metric =
spec
Standard Performance Evaluation Corporation
Board of DirectorsLegal, HQ, PR
OSG GPC HPG
Committees & People Who Do Real Work!
Infrastructure
60+ Members and Associates
spec
Standard Performance Evaluation Corporation
OSG GPC HPGSPECjbb2000 SPECviewperf 6.11 SPEC HPC96 V1.10
SPEC CPU2000 SPECglperf 3.1.2
SPECweb99 SPECapc for Pro/ENGINEER 2000i
SPECjvm98 SPECapc for SolidWorks99
SPECsfs97 SPECapc for Unigraphics V15
Popular Benchmarks
Retired: SPEC89, 92, 95, SPECweb96 ..
0.64
0.490.46
0.83
0.40
Compaq IBM Intel SGI Sun
SP
EC
int2
000/
MH
zSPECint2000/MHz
667
500
0.67
0.87
Gain due to optimization flags
340 733 400 333
11/99
0.62 0.64
0.35
0.82
0.49
Compaq IBM Intel SGI Sun
SP
EC
int2
000/
MH
zSPECint2000/MHz
833
500
0.70
0.88
Gain due to optimization flags
450 1500 400 900
0.36
0.52
0.65 01/01
One Year Later
0.770.82
0.33
0.90
0.38
Compaq SGI Intel IBM Sun
SP
EC
fp20
00/M
Hz
SPECfp2000 / MHz
0.87Gain due to optimization flags
667 400 733 200 333
0.87
11/99
0.71
0.96
0.35
0.79
0.47
Compaq SGI Intel IBM Sun
SP
EC
fp20
00/M
Hz
SPECfp2000 / MHz
Gain due to optimization flags
833 400 1500 450 900
1.00
0.79
0.37
0.91
0.54
01/01
One Year Later
spec
Standard Performance Evaluation Corporation
Coming Attractions
Released 01/19/2001 - SPECmail2000Server side Java - J2EE Benchmark High Performance Computing - SPEComp2000SPEC Annual Meeting (Jan 24, 2001) will be held in Austin, TXSearching for CPU 200X benchmarksFor latest information for SPEC and TPC go to http://www.spec.org & http://www.tpc.org
spec
Standard Performance Evaluation Corporation
IssuesBenchmarking is getting more and more EXPENSIVE & DIFFICULTBenchmark development cycles are getting longer - technology racing aheadBiggest problem with benchmarks is they represent yesterday's workloadsBenchmarks help to MOVE bottlenecks not REMOVE themBenchmarketing takes over benchmarkingOld benchmarks are good members of NRA (Not Relevant Application) club - vigilance necessary
spec
Standard Performance Evaluation Corporation
Summary
GIPs, GFLOPs & GDB used for simpler MMIGood Benchmarks will improve H/W & S/WBiggest problem with benchmarks is that often times they represent yesterday's workloadsBenchmarks help to MOVE bottlenecks not REMOVE themBest way to help is provide what you think is a good benchmark to standards organizationSPEC has a bounty program and we are working on future benchmarksQ & A