Slurm: Gres vs --gpus configuration, syntax preference on A100

Discussion Zone
, , , , , , ,
1

Previously many sites used the GRES syntax for user facing commands to request GPU using slurm. Now with the newer --gpus syntax is there a preference? Can the two co-exist at the same time for users bringing older codes into new systems or does it have to be one or the other?

For those implementing A100 in slurm, what have you used? Do you rely on the nvml library or utility to find the GPUs on a node or do you explicitly configure the GPU into slurm.conf or gres.conf?

2

I see a few sites with more than one cluster and “old” vs “new” syntax such as https://researchcomputing.princeton.edu/pytorch