计算化学公社

标题: 求助AMBER-18测试GPU加速问题 [打印本页]

作者
Author: hhhnano 时间: 2020-6-17 11:20
标题: 求助AMBER-18测试GPU加速问题
本帖最后由 hhhnano 于 2020-6-17 11:23 编辑

按照AMBER官方提供的测试软件包测试GPU加速运算速度，两块2080TI计算速度与官方数据接近，但也出现了4个GPU，还有其他问题，仔细检测了AMBER-18安装，并作了测试，发现没有问题，是否官方提供的测试脚本需要结合自己的机器进行修改？官方测试数据。
http://ambermd.org/GPUPerformance.php#RCWBenchmarks

附上计算脚本SH文件

JAC_PRODUCTION_NVE - 23,558 atoms PME 4fs
-----------------------------------------

CPU code 16 cores: |       ns/day =    62.73 seconds/ns = 1377.23
   [0] 1 x GPU: |       ns/day =    990.84 seconds/ns =    87.20
   [1] 1 x GPU: |       ns/day =    974.12 seconds/ns =    88.70
   [2] 1 x GPU: cudaGetDeviceCount failed no CUDA-capable device is detected
|       ns/day =    974.12 seconds/ns =    88.70
   [3] 1 x GPU: cudaGetDeviceCount failed no CUDA-capable device is detected
|       ns/day =    974.12 seconds/ns =    88.70
cudaGetDeviceCount failed no CUDA-capable device is detected
cudaGetDeviceCount failed no CUDA-capable device is detected
Multiple Single GPU Run Performance
   [0] 1 x GPU: |       ns/day =    972.28 seconds/ns =    88.86
   [1] 1 x GPU: |       ns/day =    967.99 seconds/ns =    89.26
   [2] 1 x GPU: grep: mdinfo.1GPU_2: No such file or directory
   [3] 1 x GPU: grep: mdinfo.1GPU_3: No such file or directory
cudaMemcpy GpuBuffer::Download failed an illegal memory access was encountered
      2 x GPU: |       ns/day =    364.00 seconds/ns =    237.36
      4 x GPU: |       ns/day =    192.91 seconds/ns =    447.87
cudaGetDeviceCount failed no CUDA-capable device is detected
cudaMemcpy GpuBuffer::Download failed an illegal memory access was encountered
Multiple 2xGPU Run Performance
   [0,1] 2 x GPU: grep: mdinfo.2GPU_0: No such file or directory
   [2,3] 2 x GPU: grep: mdinfo.2GPU_2: No such file or directory

作者
Author: liuyuje714 时间: 2020-6-17 13:01
修改脚本前面的GPU数和CPU核数啊。

作者
Author: hhhnano 时间: 2020-6-17 15:10
谢谢liuyuje714，我是双CPU（每个CPU28核），两个20280TI显卡，修改了参数如下：
GPU_COUNT=2
CPU_COUNT=28（这里是否修改为56，两颗CPU，56核心），
但是还是有错误出现，而且两个GPU加速计算远比单个GPU加速计算慢，这是什么原因？

JAC_PRODUCTION_NPT - 23,558 atoms PME 4fs
-----------------------------------------

CPU code 28 cores: |       ns/day =    68.90 seconds/ns = 1253.92
   [0] 1 x GPU: |       ns/day =    838.67 seconds/ns =    103.02
   [1] 1 x GPU: |       ns/day =    886.49 seconds/ns =    97.46
Multiple Single GPU Run Performance
   [0] 1 x GPU: |       ns/day =    857.28 seconds/ns =    100.78
   [1] 1 x GPU: |       ns/day =    886.67 seconds/ns =    97.44
      2 x GPU: |       ns/day =    285.90 seconds/ns =    302.21
Multiple 2xGPU Run Performance
   [0,1] 2 x GPU: |       ns/day =    297.59 seconds/ns =    290.33

JAC_PRODUCTION_NVE - 23,558 atoms PME 2fs
-----------------------------------------

CPU code 28 cores: |       ns/day =    38.30 seconds/ns = 2256.10
   [0] 1 x GPU: |       ns/day =    510.40 seconds/ns =    169.28
   [1] 1 x GPU: |       ns/day =    513.35 seconds/ns =    168.31
Multiple Single GPU Run Performance
   [0] 1 x GPU: |       ns/day =    508.17 seconds/ns =    170.02
   [1] 1 x GPU: |       ns/day =    509.15 seconds/ns =    169.69
cudaMemcpy GpuBuffer::Download failed an illegal memory access was encountered
      2 x GPU: |       ns/day =    153.21 seconds/ns =    563.94
Multiple 2xGPU Run Performance
   [0,1] 2 x GPU: |       ns/day =    148.43 seconds/ns =    582.08

JAC_PRODUCTION_NPT - 23,558 atoms PME 2fs
-----------------------------------------

CPU code 28 cores: |       ns/day =    38.49 seconds/ns = 2244.69
   [0] 1 x GPU: |       ns/day =    460.04 seconds/ns =    187.81
   [1] 1 x GPU: |       ns/day =    462.64 seconds/ns =    186.76
Multiple Single GPU Run Performance
   [0] 1 x GPU: |       ns/day =    459.35 seconds/ns =    188.09
   [1] 1 x GPU: |       ns/day =    453.62 seconds/ns =    190.47
      2 x GPU: |       ns/day =    139.81 seconds/ns =    617.96
Multiple 2xGPU Run Performance
   [0,1] 2 x GPU: |       ns/day =    134.95 seconds/ns =    640.26

FACTOR_IX_PRODUCTION_NVE - 90,906 atoms PME

作者
Author: abdoman 时间: 2020-6-18 10:14

hhhnano 发表于 2020-6-17 15:10
谢谢liuyuje714，我是双CPU（每个CPU28核），两个20280TI显卡，修改了参数如下：
GPU_COUNT=2
CPU_COUNT= ...

2个gpu间的数据交换要通过cpu来完成。
如果没有NCCL，建议就是单任务，单显卡（也是amber推荐的）。

作者
Author: hhhnano 时间: 2020-6-18 10:51
谢谢abdoman, NCCL安装需要配NVlink吗？还是直接装NCCL就可以？

欢迎光临计算化学公社 (http://bbs.keinsci.com/)