Presentations: Jack Dongarra, University of Tennessee & ORNL. The HPL Benchmark: Past, Present & Future. Mike Heroux, Sandia National Laboratories

Transcription

1 HPC Benchmarking Presentations: Jack Dongarra, University of Tennessee & ORNL The HPL Benchmark: Past, Present & Future Mike Heroux, Sandia National Laboratories The HPCG Benchmark: Challenges It Presents to Current & Future Systems Mark Adams, LBNL HPGMG: A Supercomputer Benchmark & Metric David A. Bader, Georgia Institute of Technology Graph500: A Challenging Benchmark for High Performance Data Analytics 1

2 The HPL Benchmark: Past, Present & Future Jack Dongarra University of Tennessee Oak Ridge National Laboratory University of Manchester 6/21/16 2

3 Confessions of an 3 Accidental Benchmarker Appendix B of the Linpack Users Guide Designed to help users extrapolate execution Linpack software package First benchmark report from 1977; Cray 1 to DEC PDP-10 time for

4 4 Started 37 Years Ago Have seen a Factor of 6x From 14 Mflop/s to 93 Pflop/s In the late 70 s the fastest computer ran LINPACK at 14 Mflop/s Today with HPL we are at 93 Pflop/s Nine orders of magnitude doubling every 14 months About 7 orders of magnitude increase in the number of processors Plus algorithmic improvements Began in late 70 s time when floating point operations were expensive compared to other operations and data movement

5 5

6 6 Linpack Benchmark Over Time In the beginning there was the Linpack 100 Benchmark (1977) n=100 (80KB); size that would fit in all the machines Fortran; 64 bit floating point arithmetic No hand optimization (only compiler options) Linpack 1000 (1986) n=1000 (8MB); wanted to see higher performance levels Any language; 64 bit floating point arithmetic Hand optimization OK Linpack TPP (1991) (Top500; 1993) Any size (n as large as you can; n = 12x10 6 ; 1.2 PB); Any language; 64 bit floating point arithmetic Hand optimization OK Strassen s method not allowed (confuses the op count and rate) Reference implementation available In all cases results are verified by looking at: Operations count for factorization 2 1 n n ; solve Ax b = O(1) A x nε 2 2n

7 7 LINPACK to HPL to TOP500 LINPACK Benchmark report, ANL TM-23, 1984 Performance of Various Computers Using Standard Linear Equations Software, listed about 70 systems. Over time the LINPACK Benchmark when through a number of changes. Began with Fortran code, run the code as is, no changes, N = 100 (Table 1) Later N = 1000 introduced, hand coding to allow for optimization and parallelism (Table 2) Timing harness provided to generate matrix, check the solution The basic algorithm, GE/PP, remained the same started putting together Table 3 (Toward Peak Performance) of the LINPACK benchmark report. N allowed to be any size Timing harness provided to generate matrix, check the solution List R max, N max, R peak In 2000 we put together an optimized implementation of the benchmark, called High Performance LINPACK or HPL. Just needs optimized version of BLAS and MPI.

8 8 TOP500 In 1986 Hans Meuer started a list of supercomputer around the world, they were ranked by peak performance. Hans approached me in 1992 to put together our lists into the TOP500. The first TOP500 list was in June 1993.

9 9 Rules For HPL and TOP500 Algorithm is Gaussian Elimination with partial pivoting. Excludes the use of a fast matrix multiply algorithm like "Strassen's Method Excludes algorithms which compute a solution in a precision lower than full precision (64 bit floating point arithmetic) and refine the solution using an iterative approach. The authors of the TOP500 reserve the right to independently verify submitted LINPACK results, and exclude computer from the list which are not valid or not general purpose in nature. Any computer designed specifically to solve the LINPACK benchmark problem or have as its major purpose the goal of a high TOP500 ranking will be disqualified.

10 #1 System on the Top500 Over the Past 24 Years (18 machines in that club) Top500 List Computer r_max (Tflop/s) n_max Hours MW 6/93 (1) TMC CM-5/ /93 (1) Fujitsu Numerical Wind Tunnel /94 (1) Intel XP/S /94-11/95 (3) Fujitsu Numerical Wind Tunnel /96 (1) Hitachi SR2201/ , /96 (1) Hitachi CP-PACS/ , /97-6/00 (7) Intel ASCI Red , /00-11/01 (3) IBM ASCI White, SP Power3 375 MHz , /02-6/04 (5) NEC Earth-Simulator ,000, /04-11/07 (7) IBM BlueGene/L ,000, /08-6/09 (3) IBM Roadrunner PowerXCell 8i 3.2 Ghz 1,105. 2,329, /09-6/10 (2) Cray Jaguar - XT5-HE 2.6 GHz 1,759. 5,474, /10 (1) NUDT Tianhe-1A, X Ghz NVIDIA 2,566. 3,600, /11-11/11 (2) Fujitsu K computer, SPARC64 VIIIfx 10, ,870, /12 (1) IBM Sequoia BlueGene/Q 16, ,681, /12 (1) Cray XK7 Titan AMD + NVIDIA Kepler 17,590. 4,423, /13 11/15(6) NUDT Tianhe-2 Intel IvyBridge & Xeon Phi 33,862. 9,960, /16 Sunway TaihuLight System 93, ,288,

11 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 11 Run Times for HPL on Top500 Systems 61 hours 30 hours 20 hours 12 hours 11 hours 10 hours 9 hours 8 hours 7 hours 6 hours 5 hours 4 hours 3 hours 2 hours 1 hour 0% 6/1/93 11/1/93 4/1/94 9/1/94 2/1/95 7/1/95 12/1/95 5/1/96 10/1/96 3/1/97 8/1/97 1/1/98 6/1/98 11/1/98 4/1/99 9/1/99 2/1/00 7/1/00 12/1/00 5/1/01 10/1/01 3/1/02 8/1/02 1/1/03 6/1/03 11/1/03 4/1/04 9/1/04 2/1/05 7/1/05 12/1/05 5/1/06 10/1/06 3/1/07 8/1/07 1/1/08 6/1/08 11/1/08 4/1/09 9/1/09 2/1/10 7/1/10 12/1/10 5/1/11 10/1/11 3/1/12 8/1/12 1/1/13 6/1/13

12 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Run Times for HPL on Top500 Systems 12 Hours

13 13 Over the Course of the Run Can t just start the run and stop it. The performance will vary over the course of the run Pflop/s Time in Hours

14 HPL: Reducing Execution Time R max = 1 T 2 T 1 T 2 T 1 G(t)dt Gflop/s Initial section Overestimating performance Rough-estimating performance Middle section Underestimate performance End section time

15 15 How to Capture Performance? Determine the section where the computation and communications for the execution reflect a completed run Pflop/s Time in Hours

16 16 LINPACK Benchmark Still Learning Things We use a backwards error residual to check the correctness of the solution. This is the classical Wilkinson error bound. If the residual is small O(1) then the software is doing the best it can independent of the conditioning of the matrix. We say O(1) is OK, the code allows the residual to be less than O(10). For large problems we noticed the residual was getting smaller.

17 17 LINPACK Benchmark Still Learning Things We use a backwards error residual to check the correctness of the solution. This is the classical Wilkinson error bound. If the residual is small O(1) then the software is doing the best it can independent of the conditioning of the matrix. We say O(1) is OK, the code allows the residual to be less than O(10). For large problems we noticed the residual was getting smaller.

18 18 LINPACK Benchmark Still Learning Things The current criteria might be about O(10 3 ) too lax which allows error for the last bits of the mantissa to go undetected. We believe this has to do with the rounding errors for collective ops when done in parallel, i.e. MatVec and norms A better formulation of the residual might be:

19 19 HPL - Bad Things LINPACK Benchmark is 37 years old TOP500 (HPL) is 23 years old Floating point-intensive performs O(n 3 ) floating point operations and moves O(n 2 ) data. No longer so strongly correlated to real apps. Reports Peak Flops (although hybrid systems see only 1/2 to 2/3 of Peak) Encourages poor choices in architectural features Overall usability of a system is not measured Used as a marketing tool Decisions on acquisition made on one number Benchmarking for days wastes a valuable resource