slurm generic resources: https://slurm.schedmd.com/gres.html
GPUs are tracked by Slurm as generic resources. Selecting GPUs in the OOD form can be error prone because the user must also select an account, partition, and QoS such that their job can be scheduled on a node with their requested GPU type. When this happens, the user gets this message:
Which is not very helpful.
The auto_queues and auto_accounts widgets already work together to prevent the user from submitting a job with incompatible parameters, so this feature would be consistent with what's already there.
notes:
auto_gres would take an option gres_type, which will usually be "gpu".- The
auto_gres widget would have to be a select and an input type=number, which might not be straightforward to implement.
- A new attribute would have to be added to the
NodeInfo class in ood_core.
- Ideally,
PartitionInfo could be extended to document which nodes it includes, so that auto_gres can integrate with auto_queues to prevent the error message above.
slurm generic resources: https://slurm.schedmd.com/gres.html
GPUs are tracked by Slurm as generic resources. Selecting GPUs in the OOD form can be error prone because the user must also select an account, partition, and QoS such that their job can be scheduled on a node with their requested GPU type. When this happens, the user gets this message:
Which is not very helpful.
The
auto_queuesandauto_accountswidgets already work together to prevent the user from submitting a job with incompatible parameters, so this feature would be consistent with what's already there.notes:
auto_greswould take an optiongres_type, which will usually be"gpu".auto_greswidget would have to be aselectand aninput type=number, which might not be straightforward to implement.NodeInfoclass in ood_core.PartitionInfocould be extended to document which nodes it includes, so thatauto_grescan integrate withauto_queuesto prevent the error message above.