Inhalt | ||
---|---|---|
|
...
CentOS 7 | Rocky Linux 9 | |
---|---|---|
old partition name | new partition name | current job limits |
● | ● | 512 nodes, 12h wall time |
● | ● | 32 nodes, 1 h wall time |
● | ● | |
● | ● | |
● | ||
● | ||
● | ● |
( ● available ● closed/not available yet )
...
For users of SLURM’s
srun
job launcher:
Open MPI 5.x has dropped support for the PMI-2 API, it solely depends on PMIx to bootstrap MPI processes. For this reason the environment setting was changed fromSLURM_MPI_TYPE=pmi2
toSLURM_MPI_TYPE=pmix
, so binaries linked against Open MPI can be started as usual “out of the box” usingsrun mybinary
. For the case of a binary linked against Intel-MPI, this works too when a recent version (≥2021.11) of Intel-MPI has been used. If an older version of Intel-MPI has been used, and relinking/recompiling is not possible, one can follow the workaround for PMI-2 withsrun
as described in the Q&A section below. Switching fromsrun
tompirun
instead should also be considered.Using more processes per node than available physical cores (PPN > 96; hyperthreads) with the OPX providerwhen defining
FI_PROVIDER=opx
:
The OPX provider currently does not support using hyperthreads/PPN > 96 on the clx partitions. Doing so may result in segmentation faults in libfabric during process startup. If a high number of PPN is really required, the libfabric provider has to be changed back to PSM2 by setting re-definingFI_PROVIDER=psm2
(which is the default setting). Note that the usage of hyperthreads may not be advisable. We encourage users to test performance before using more threads than available physical cores.
Note that Open MPI’smpirun/exec
defaults to use all hyperthreads if a Slurm job/allocation is used that does not explicitely sets explicitly set--ntasks-per-node
(or similar options).
...