Discussion:
slurm and grid capabilities
Edsall, William (WJ)
2014-10-03 15:12:43 UTC
Permalink
Hello list,
Could someone please describe the current grid capabilities within Slurm?
Can slurm act as a grid scheduler (like moab) and route jobs between multiple Slurm clusters or do you need Moab+Slurm to achieve this?


Thanks
William
Uwe Sauter
2014-10-03 15:19:40 UTC
Permalink
Hi,

there is some limited capabilities for the tools to send/query clusters
they don't "belong" to (search for the -M option).
Tools "belong" to the cluster that is configured in the slurm.conf they use.

And there is some work taking place at CSCS (Switzerland) that was
presented last week on the SLURM user group meeting, see
http://slurm.schedmd.com/SUG14/inter_cluster.pdf


Regards,

Uwe
Post by Edsall, William (WJ)
Hello list,
Could someone please describe the current grid capabilities within Slurm?
Can slurm act as a grid scheduler (like moab) and route jobs between
multiple Slurm clusters or do you need Moab+Slurm to achieve this?
Thanks
William**
j***@public.gmane.org
2014-10-03 15:29:40 UTC
Permalink
Another option some sites use is to have a single slurm instance
managing multiple clusters (essentially treating it as a single larger
cluster, but with disjoint networks to avoid allocating jobs that span
clusters).
Post by Edsall, William (WJ)
Hello list,
Could someone please describe the current grid capabilities within Slurm?
Can slurm act as a grid scheduler (like moab) and route jobs between
multiple Slurm clusters or do you need Moab+Slurm to achieve this?
Thanks
William
--
Morris "Moe" Jette
CTO, SchedMD LLC
Lipari, Don
2014-10-03 15:46:23 UTC
Permalink
Slurm has a multi-cluster (grid) functionality when Slurm’s database is installed and active. This allows one cluster to submit jobs, and invoke status commands, to another cluster. Jobs submitted that specify multiple cluster candidates will be submitted to the cluster that is projected to run the job soonest. It will remain in that cluster’s pending queue even if workload changes at a later time make it possible for another cluster to run the job sooner.

There are a couple more features in the Moab/Slurm arrangement that Slurm alone does not provide. One is a unique job ID. Jobs submitted to a Moab grid will receive a job ID that is unique across all of the clusters in the grid. This enables the second important feature which is cross-cluster dependencies. A user who submits a job to multiple clusters can specify a single job ID dependency no matter where the dependee job is in the grid or where the dependent job ultimately runs.

As far as I know, Stephen Trofinoff is the only person working on a design which will add the unique job ID and cross-cluster dependency to Slurm (his presentation, http://slurm.schedmd.com/SUG14/inter_cluster.pdf, was cited in a previous response).

Don

From: Edsall, William (WJ) [mailto:***@dow.com]
Sent: Friday, October 03, 2014 8:13 AM
To: slurm-dev
Subject: [slurm-dev] slurm and grid capabilities

Hello list,
Could someone please describe the current grid capabilities within Slurm?
Can slurm act as a grid scheduler (like moab) and route jobs between multiple Slurm clusters or do you need Moab+Slurm to achieve this?


Thanks
William

Loading...