search

aiidateam / aiida-core

The official repository for the AiiDA code
https://aiida-core.readthedocs.io
Other
434 stars 188 forks source link

SLURM Scheduler throws error if underutilizing CPU resources #3439

Open zooks97 opened 5 years ago

zooks97 commented 5 years ago

Currently, the SLURM scheduler checks that num_cores_per_machine == num_cores_per_mpiproc * num_mpiprocs_per_machine. However, a user may like to use a number of cores less than the number available because of parallelization restrictions (e.g. in pw.x, if you are running a calculation with 11 kpoints on a 12-core node, it may be best to use 11 k-pools on 11 mpiprocs with 1 core per mpiproc for the best performance).

It would be better to check that num_cores_per_machine >= num_cores_per_mpiproc * num_mpiprocs_per_machine.

sphuber commented 5 years ago

@giovannipizzi I labeled it as a feature request for now. If you also agree we can turn it into accepted feature and implement it.