At our site we have daily occurring reservation like:

ReservationName=devel StartTime=2011-02-24T08:00:00 EndTime=2011-02-24T18:00:00 
Duration=10:00:00
   Nodes=n[1-4] NodeCnt=4 Features=(null) PartitionName=(null) 
Flags=IGNORE_JOBS,DAILY,SPEC_NODES
   Users=user1,...

I.e. every day between 0800-1800 on nodes n[1-4]. To only allow jobs
below a certain walltime within the reservation we have a partition
called devel on those nodes n[1-4] with this walltime limit.

In order to make use of n[1-4] when the reservation is not in effect
(all days 18:00:01 - 07:59:59), we add them to another partition
during that time and remove them again from the latter partition when
the reservation is in effect, and so on.

When the reservation is not in effect, jobs with an expected endtime
including and after the next reservation start shouldn't be able to
start on n[1-4] - but this might actually happen.

Here is such a case where job 130973 with TimeLimit=1-16:00:00 starts
on these nodes on the other partition (after the nodes have been added
to it, after the reservation has ended and before the reservation has
been updated):

   [2011-02-19T09:34:54] _slurm_rpc_submit_batch_job JobId=130973 usec=1503
   [2011-02-23T18:00:01] update_part: setting nodes to n[1-8,93-364] for 
partition kappa
   [2011-02-23T18:00:11] sched: Allocate JobId=130973 
NodeList=n[1-4,117,119,167,184] #CPUs=64
   [2011-02-23T18:00:15] sched: _slurm_rpc_job_step_create: StepId=130973.0 
n[1-4,117,119,167,184] usec=248
   [2011-02-23T18:00:27] Advance reservation devel one day

All reservations should ideally advance atomically when they end I
guess, before any jobs could be scheduled on the nodes (in this case
at 18:00:01)?

regards,
Per Lundqvist

-- 
Per Lundqvist

National Supercomputer Centre
Linköping University, Sweden

http://www.nsc.liu.se

Reply via email to