[mpich-discuss] new to mpi: performance issues when using mpich
Raffenetti, Ken
raffenet at anl.gov
Wed Feb 16 13:22:15 CST 2022
--cpus-per-task is not equivalent to -ppn when using mpiexec/mpirun. -ppn says launch this many processes per node. --cpus-per-tasks says launch as many processes as you can so that each has access to at least 8 CPU cores. Trying tuning the launch options so you are sure you are launching the same number of procs in either configuration.
Ken
On 2/16/22, 12:36 PM, "Sopena Ballesteros Manuel via discuss" <discuss at mpich.org> wrote:
I am learning how to install and run mpi aplications. My goal is to install mpich and run osu micro benchmark in order to understand how it works.
so far I got it working but I must be doing wrong because the latency tests does not looks right to me.
/scratch/e1000/msopena/mpich-installed/bin/mpirun -f /scratch/e1000/msopena/hosts -ppn 8 /scratch/e1000/msopena/osu-micro-benchmarks-5.8/mpi/collective/osu_alltoall
# OSU MPI All-to-All Personalized Exchange Latency Test v5.8
# Size Avg Latency(us)
1 104.87
2 101.44
4 105.47
8 94.35
16 95.14
32 239.45
64 293.59
128 437.93
256 414.21
512 444.74
1024 655.93
2048 529.20
4096 551.84
8192 574.99
16384 644.90
32768 790.70
65536 4169.19
131072 4756.38
262144 6782.57
524288 12299.24
1048576 22880.86
However, I get notmal results when the osu benchmarks through srun
eg
$ srun -N3 --cpus-per-task=8 /scratch/e1000/msopena/osu-micro-benchmarks-5.8/mpi/collective/osu_alltoall
# OSU MPI All-to-All Personalized Exchange Latency Test v5.8
# Size Avg Latency(us)
1 27.43
2 27.97
4 29.26
8 30.69
16 31.77
32 34.89
64 42.06
128 60.83
256 115.84
512 205.33
1024 282.61
2048 446.41
4096 733.10
8192 1421.99
16384 2832.45
32768 5631.33
65536 11242.08
131072 22645.46
262144 45544.75
524288 90440.93
1048576 186709.32
Could someone please help me to troubleshoot could be wrong when I use mpich stand alone?
thank you very much
More information about the discuss
mailing list