[BMT] STREAM

STREAM: Sustainable Memory Bandwidth in High Performance Computers

- 참고 : https://www.cs.virginia.edu/stream

 

 

1. Source code download

$ wget https://www.cs.virginia.edu/stream/FTP/Code/Versions/stream_omp.c

 

2. Compile

$ gcc -O3 -fopenmp -D OPENMP stream_omp.c -o stream_c.x

 

3. test

- test script 작성

#!/bin/sh #SBATCH -J stream #SBATCH -p short #SBATCH -N 1 #SBATCH -n 1 #SBATCH -o %x_%j.out #SBATCH -e %x_%j.err #SBATCH --time=01:00:00 export OMP_NUM_THREADS=32 /home/hpcd/leesangy/stream/stream_c.x exit 0

 

- 작업 제출

$ sbatch job.sh

 

- 결과는 아래와 같이 출력 됩니다.

------------------------------------------------------------- This system uses 8 bytes per DOUBLE PRECISION word. ------------------------------------------------------------- Array size = 2000000, Offset = 0 Total memory required = 45.8 MB. Each test is run 10 times, but only the *best* time for each is used. ------------------------------------------------------------- Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 Number of Threads requested = 32 ------------------------------------------------------------- Your clock granularity/precision appears to be 1 microseconds. Each test below will take on the order of 87 microseconds.    (= 87 clock ticks) Increase the size of the arrays if this shows that you are not getting at least 20 clock ticks per test. ------------------------------------------------------------- WARNING -- The above is only a rough guideline. For best results, please be sure you know the precision of your system timer. ------------------------------------------------------------- Function      Rate (MB/s)   Avg time     Min time     Max time Copy:      681308.2640       0.0001       0.0000       0.0001 Scale:     604584.3604       0.0001       0.0001       0.0001 Add:       696631.8062       0.0001       0.0001       0.0001 Triad:     696631.8062       0.0001       0.0001       0.0001 ------------------------------------------------------------- Solution Validates -------------------------------------------------------------

 

- 테스트#1 : compiler : gcc 4.8.5, option : -O2 -fopenmp

Threads	Copy(MB/s)	Scale(MB/s)	Add(MB/s)	Triad(MB/s)
1	35394.9705	21461.1014	34215.9402	31516.3732
10	200026.4203	110014.5311	202541.8431	129387.2699
24	371794.2604	251815.6248	390167.8140	300039.6304
32	464421.2042	323416.2120	480493.0597	377723.4371
64	464421.2042	438620.0261	479349.0286	475949.3901

 

- 테스트#2 : compiler : gcc 9.2.0, option : -O3 -fopenmp -march=znver1

Threads	Copy(MB/s)	Scale(MB/s)	Add(MB/s)	Triad(MB/s)
1	51286.8659	36611.4915	49236.1438	45579.9393
10	271146.9253	162491.1961	177068.2427	175218.9661
24	593883.7522	532610.0317	608237.4381	608237.4381
32	681308.2640	604584.3604	696631.8062	696631.8062
64	654720.6244	651542.3689	748425.9926	751218.6269

# AMD EPYC™ 7xx1 Series Processors 최적화 옵션 "-march=znver1" 사용을 위하여 gcc 9.2 컴파일러 사용

# 참고 : https://www.amd.com/system/files/documents/compiler-options-guide-amd-epyc-7xx1-series-processors.pdf

 

4. 기타(테스트 환경)

구분	상세 정보	비고
Hostname	aixl645dn04
Platform	Apollo 6500 Gen10+ XL645d (Bantha)
Processor	1 x 7543	hyperthreading enabled
OS	Ubuntu 20.04
Memory	8 x 64GB 3200MHz = 512GB
GPU	4 x A100-SXM4-40GB