Nvidia pyxis

[staeglis]

Hi,

what do you think about using the SPANK plugin nvidia pyxis instead of calling enroot directly?

Best,
Stefan

EDIT: Obviously this can't be used for running jobs directly on the frontend (login submit whatever) node. But anyway it's probably better to run such jobs on a dedicated node with probably less resources instead.

[tcojean]

Hi,

That's a good suggestion and generally, it should be doable. I think it could be added as an extra mode on top of the existing code, rather than as a replacement. I think it would require:

• prepare.sh would probably do almost nothing. Since pyxis would manage all the container aspects.
• run.sh would need to add the relevant SLURM, like --container-image.

The one thing I am not entirely sure about yet is how to deal with pipelines that split build and test jobs. Currently, we just reuse the same container instance between the two, we would some coordination to make sure that happens.

For now, I cannot work on this, do you have a specific need?

[staeglis]

Hi,

for now I'm just interested in bundling resources for compute and CI pipelines. pyxis was just an idea to keep the code simple. In the next weeks I hopefully can do some test with current version.