Don't put all nodes to power save mode

48 views
Skip to first unread message

Maciej Skarysz

unread,
Aug 5, 2020, 2:29:30 PM8/5/20
to google-cloud-slurm-discuss
Hi,

I understand that slurmctld is deciding to put idle nodes to power save mode. It runs SuspendProgram. I want to always keep one node idle (not idle~), is it possible? Note that I can't use SuspendExcNodes as I don't know a priori which node will stay idle, I only know I always what to have one.

I would appreciate any suggestions.

Best,
Maciej

Keith Binder

unread,
Aug 6, 2020, 9:11:10 AM8/6/20
to Maciej Skarysz, google-cloud-slurm-discuss

If you re-deploy using the field "static_node_count" under a partition, you can specify a certain number of nodes that will stay in the "idle" state.



--
You received this message because you are subscribed to the Google Groups "google-cloud-slurm-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to google-cloud-slurm-...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/google-cloud-slurm-discuss/2cfd0419-f52c-45ca-abfe-81ea2c7f4846o%40googlegroups.com.

Maciej Skarysz

unread,
Aug 6, 2020, 9:23:38 AM8/6/20
to google-cloud-slurm-discuss
Hi Keith,

Many thanks for your answer. The "static_node_count" effectively sets SuspendExcNodes in slurm.conf and this indeed keep nodes in state idle. But this option only works for non-preemptible machines, and is not exactly what I want. I want to have (n+1)th node idle where n is the number of allocated nodes. When I set static_node_count=1 and then I allocate a job to this node my number of idle nodes will be 0. I hope I didn't make it harder to understand.

Best,
Maciej

Wyatt Gorman

unread,
Sep 16, 2020, 12:53:38 PM9/16/20
to Maciej Skarysz, google-cloud-slurm-discuss
Hi Maciej,

The static nodes that are created by the YAML field are not preemptible, regardless of the configuration of the burst nodes in the rest of the partition.

Currently there is no way to arbitrarily set a given node in a job as non-preemptible. You would have to split things between a preemptible partition and a non-preemptible one.


Wyatt Gorman

HPC Solutions Manager

https://cloud.google.com/hpc




--
You received this message because you are subscribed to the Google Groups "google-cloud-slurm-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to google-cloud-slurm-...@googlegroups.com.
Reply all
Reply to author
Forward
0 new messages