I saw you got some good answers, but a quick note on mpi. For some of
them, you are compiling it yourself, they can be "slurm-aware" (eg:
openmpi). Then when you do 'mpirun' it automatically knows your
inherited hostlist and you need do nothing extra when running.
Brian Andrus
On 5/18/2018 12:11 AM, Mahmood Naderan wrote:
Hi,
Is there any slurm variable to read the node names of a partition?
There is an MPI option --hostfile which we can write the node names.
I want to use something like this in the sbatch script:
#SBATCH --partition=MYPART
... --hostfile $SLURM_NODES_IN_PARTITION ....
I can manually manipulate the output of scontrol to extract node
names. Like this:
[mahmood@rocks7 ~]$ scontrol show partition MYPART | grep -w Nodes |
cut -d '=' -f 2
compute-0-[4-6]
But that basically is not
compute-0-4
compute-0-5
compute-0-6
Which I have to post process more. Any better idea?
Regards,
Mahmood