[Mvapich-discuss] Fix for collective validation with ROCm
McDougall, Damon
Damon.Mcdougall at amd.com
Wed Jan 11 19:33:23 EST 2023
[AMD Official Use Only - General]
Hi,
I think I fixed a bug in OSU MPI Microbenchmarks 7.0.1. I've attached a patch that I think fixes it. I hope it is helpful.
Before the patch:
$ srun -N 1 -n 2 ./get_local_rank ./mpi/collective/osu_alltoall -c -d rocm
# OSU MPI-ROCM All-to-All Personalized Exchange Latency Test v7.0
# Size Avg Latency(us) Validation
1 0.74 Fail
DATA VALIDATION ERROR: ./mpi/collective/osu_alltoall exited with status 1 on message size 1.
After the patch:
$ srun -N 1 -n 2 ./get_local_rank ./mpi/collective/osu_alltoall -c -d rocm
# OSU MPI-ROCM All-to-All Personalized Exchange Latency Test v7.0
# Size Avg Latency(us) Validation
1 0.74 Pass
2 0.80 Pass
4 0.67 Pass
8 0.81 Pass
16 0.75 Pass
32 0.80 Pass
64 0.85 Pass
128 0.90 Pass
256 1.36 Pass
512 1.49 Pass
1024 2.67 Pass
2048 2.72 Pass
4096 2.80 Pass
8192 11.93 Pass
16384 12.19 Pass
32768 13.14 Pass
65536 14.71 Pass
131072 17.30 Pass
262144 22.54 Pass
524288 32.67 Pass
1048576 44.61 Pass
Thanks,
Damon
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osu.edu/pipermail/mvapich-discuss/attachments/20230112/1128173a/attachment-0009.html>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: rocm_collective_validation.patch.txt
URL: <http://lists.osu.edu/pipermail/mvapich-discuss/attachments/20230112/1128173a/attachment-0009.txt>
More information about the Mvapich-discuss
mailing list