[slurm-users] Manage access to specialized nodes: Reservation, Queue, or Features
dragowsky at case.edu
Tue Jun 18 18:30:24 UTC 2019
We're running Slurm 17.02.2.
- We have implemented OnDemand in our cluster, including the Jupyter app
across all the compute nodes. The Interactive Desktop application, however,
is installed on a small set of compute nodes during an extended validation
period. Installation and initial testing were conducted by setting these 5
nodes aside using a reservation. Testing was limited by setting specific
accounts and users.
These nodes are now ready for validation by all cluster account holders.
I had hoped to adjust the reservation to continue for the general
want to prioritize use of these nodes for the desktop app*). I am
finding that the reservation can be modified to set either Users=(null), or
Accounts=(null); but not both.
- [mrd20 at hpc3 reservations]$ scontrol show res
Nodes=compt[162-166] NodeCnt=5 CoreCnt=80 Features=(null)
Accounts=(null) Licenses=(null) State=ACTIVE BurstBuffer=(null) Watts=n/a
Is there a work-around? I had merely tried:
> sudo scontrol update reservation=ood-interactive Users=
an analog of the command used to set 'Accounts=' to null.
*Is there a workaround to open a reservation to all users and accounts?*
I'm realizing that a reservation is not the ideal long-term control method.
We are a bit ambivalent about installing the desktop software as part of
My short-term workaround is to catalog our most active accounts, set
'Accounts=<top-50-active-accounts>' and set Users null, Then promote the
service for evaluation, adding further accounts by request.
Am I being *too* short-sighted in persisting with a reservation to control
access to these nodes? We may implement the service through a separate
queue if we decide to restrict the desktop app software to just a limited
number of nodes. I haven't yet thought of a way to prioritize availability
through setting a feature, since it's my perception that OnDemand is only
successful if requested resources become available with only a short delay
(~5 min or so).
E.M. Dragowsky, Ph.D.
Research Computing -- UTech
Case Western Reserve University
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the slurm-users