[QE-users] To the OneAPI MPI user

[Kazume NISHIDATE]

Sometimes, QE compiled by the OneAPI MPI does not return converged
results in reasonable CPU time, especially in the hybrid functional
calculation.

I encountered this problem last summer and consulted Dell Technologies
Japan, which solved the problem at that time and taught me the
solution. But I forgot about it and re-encountered the problem this
summer again! So, I’m posting my experience here to remember it. :)

This is because the shared memory treatment in a node is changed from
that of the Intel MPI. In this case, you can try,

FI_PROVIDER=psm3

in your JOB script. It forces the QE to use the Intel MPI style shared
memory treatment as before.

Another approach is to recompile the QE using OpenMPI. The QE will
then adopt the Intel MPI style shared memory treatment, similar to the
original setup.



西館数芽
Kazume NISHIDATE Ph.D

Department of Systems Innovation Engineering,
Graduate School of Science and Engineering, Iwate University
4-3-5 Ueda, Morioka, Iwate, 020-8551 JAPAN
Phone:+81-19-621-6391
kazume.nishidate at gmail.com, nisidate at iwate-u.ac.jp
https://sites.google.com/site/nisidatelab/