[slurm-users] [External] Another question about partition and node allocation

Michael Robbert mrobbert at mines.edu
Wed Apr 15 13:42:49 UTC 2020

The more flexible way to do this is with QoS. (PreemptType=preempt/qos) You'll need to have Accounting enabled and you'll probably want qos listed in AccountingStorageEnforce. Once you do that you create a "shared" for the scavenger jobs, a QoS for each group that buys into resources. Assign the correct number of resources to each group's QoS as GrpTRES limits, add the "shared" qos as preemptable from that qos and give the qos a higher priority. 

Mike Robbert

On 4/10/20, 15:34, "slurm-users on behalf of Renata Maria Dart" <slurm-users-bounces at lists.schedmd.com on behalf of renata at slac.stanford.edu> wrote:

    CAUTION: This email originated from outside of the Colorado School of Mines organization. Do not click on links or open attachments unless you recognize the sender and know the content is safe.
    Hi, we have 40 nodes (all the same, amd nodes with 128 cores) which
    have all been purchased by different groups at our lab and each group
    would like to have immediate access of course to what they have paid
    for.  The stakeholder groups are also fine with allowing the general
    public to use their hosts/cores provided they can preempt the general
    public's jobs.  One way I can see to do that is to assign specific
    nodes to each stakeholder group defined as a partition, something like this:
    PartitionName=shared     Default=yes   Priority=10   MaxTime=5-00:00:00     DefaultTime=30            PreemptMode=CANCEL   State=UP   Nodes=amd[0001-0040]
    PartitionName=exp1       Default=no    Priority=50   MaxTime=5-00:00:00     DefaultTime=1-00:00:00    PreemptMode=OFF      State=UP   Nodes=amd[0001-0003]
    PartitionName=exp2       Default=no    Priority=50   MaxTime=5-00:00:00     DefaultTime=1-00:00:00    PreemptMode=OFF      State=UP   Nodes=amd[0004-0019]
    PartitionName=exp3       Default=no    Priority=50   MaxTime=5-00:00:00     DefaultTime=1-00:00:00    PreemptMode=OFF      State=UP   Nodes=amd[0020-0040]
    Is this the most efficient and best use of resources?  In the above
    scenario if scavenger jobs are running on a given experiment's hosts
    and the experiment needs to run jobs, then scavenger jobs get
    preempted, even if there are idle hosts in the other stakeholder
    partitions.  Is there a way to guarantee say exp1 that they will have
    priority on 386 cores but not necessarily tie them to 3 specific hosts?
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 5173 bytes
Desc: not available
URL: <http://lists.schedmd.com/pipermail/slurm-users/attachments/20200415/810fc2de/attachment.bin>

More information about the slurm-users mailing list