Closed pantaray closed 3 years ago
I tested it, currently SLURMcluster.adapt() doesn't work, it just uses the minimum_jobs and doesn't change at all :(
It's quite easy to check though, I just replaced cluster.scale(total_workers)
in esi_cluster_setup with cluster.adapt(minimum_jobs=1, maximum_jobs=total_workers)
.
It looks like someone else had the same problem https://github.com/dask/dask-jobqueue/issues/463 and fixed it in this pull request https://github.com/dask/distributed/pull/4155/files. I'll keep an eye on it and see if someone adds it to dask then we can test again.
Thank you for looking into this! Nice detective work. The PR looks exactly like what we need - I suggest I'll go ahead and incorporate the necessary code (particularly #12, #13) , comment out the adapt
part and we enable it, once the PR is merged in a new dask release.
in slurmfun jobs would all go in the cue and run when resources became available. Maybe we should be using
cluster.adapt()
instead?