You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This cli filter expects to be able to pull node features based on the name of a partition which means that using the --constraints syntax is null a void.
This means partitions that overlap are unsupported as their feature set will vary based on the nodes features.
Can this please be fixed as it restricts the use of a key design feature of Slurm?
The text was updated successfully, but these errors were encountered:
The filter does operate on the assumption that cpu counts, gpu counts and memory configuration are homogeneous in a partition in order to compute:
memory per CPU (modified by SMT count) to use in the absence of an explicit memory request, as a workaround for a slurm buglimitation;
CPU to GPU ratios to supply a --cpus-per-gpu value for proportional shared access on GPU partitions.
It relies on there being sensible values for DefMemPerCPU and DefMemPerGPU in the partition configuration.
Similarly, it doesn't support allocations with a list of requested partitions.
I think for heterogeneous partitions or partitions with configurations without these assumptions above it would be best to bypass the cli_filter. As far as I know, the only way to do that is to supply an alternative SLURM_CONF, but we could add a check for an environment variable that directs the filter to pass on options unfiltered if that would be useful in these cases.
This cli filter expects to be able to pull node features based on the name of a partition which means that using the --constraints syntax is null a void.
This means partitions that overlap are unsupported as their feature set will vary based on the nodes features.
Can this please be fixed as it restricts the use of a key design feature of Slurm?
The text was updated successfully, but these errors were encountered: