Open kimchitsigai opened 2 years ago
Hi,
I've added a logging instruction in slurm_drmaa/session.c
:
fsd_log_debug(("job_desc.tres_per_node = %s", job_desc.tres_per_node));
and it produced this output:
job_desc.tres_per_node = gpu:1
And finally, I get the Invalid Trackable (TRES) specification RESource error
.
By changing the input from --gres=gpu:1
to --gres=gres:gpu:1
, the logging instruction outputs:
job_desc.tres_per_node = gres:gpu:1
as expected. The TRES error disappeared and the job was correctly submitted.
By looking at the code in https://github.com/SchedMD/slurm/blob/master/src/common/gres.c and https://github.com/SchedMD/slurm/blob/master/src/common/slurm_opt.c it seems that Slurm is expecting GPU resource requests formatted as gres:gpu:1
and not as gpu:1
as in the previous versions (20.x) of Slurm.
Shouldn't slurm-drmaa be updated to take into account this change in Slurm?
Best, Kimchi
I can confirm that changing --gres=gpu:1
to --gres=gres:gpu:1
solves the TRES error.
Thanks, yes, it sounds like we should probably prepend gres:
to the string in the job template. gres
in the job template was generalized to tres_per_node
in 18.08, I am not sure if this is when it broke.
Hi,
I'm using slurm-drmaa to submit a job and I get the error below:
The same job without
`--gres=gpu:1
works fine. slurm-drmaa version is 1.1.3 and slurm version is 21.08.6. Os is RHEL 8.4.Any hint would be greatly appreciated, Kimchi