[petsc-users] Reaching limit number of communicator with Spectrum MPI

[Victor Eijkhout]

On , 2021Aug21, at 14:08, Victor Eijkhout <eijkhout at tacc.utexas.edu<mailto:eijkhout at tacc.utexas.edu>> wrote:

3b. Dup and free million times takes more time than writing this email. Why do I keep hearing that OpenMPI is so great? Everything slightly non-standard I try is hopelessly slow and broken.

Partial rehabilitation for OpenMPI:

Finished dup/free'ing 1000000 communicators

real    0m9.301s
user    0m9.120s
sys     0m0.114s

So that’s only 4 times slower than Mvapich, but the ludicrously slow performance is only for Spectrum MPI, not OpenMPI per se.

Victor.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20210822/0a1be2ab/attachment.html>