Dear Thomas, On 01/20/2017 08:01 PM, Dr. Thomas Orgis wrote: > I am using slurm-15.08.13 on the nodes for slurmd and client commands > while the batch server is still on 15.08.12 since I intended to do that ... > > Anyhow, am I observing normal behaviour with the prolog and am I just > misunderstanding some configuration? Is there a bug?
I think this is a bug. We have observed the same behaviour with Slurm version 14.11, I have not tested with any more recent version. We have used a work-around for several years due to Slurms Prolog not behaving as we need it to. Instead of configuring Prolog in slurm.conf, we use PrologSlurmctld and that script uses pdsh to run our own "Prolog" script on allocated nodes. If I recall correctly we initially implemented this on Slurm 2.4. With Slurm 14.11 the combination of PrologFlags=Alloc and SchedulerParameters=nohold_on_prolog_fail should in theory give us the Prolog behaviour we want. When our testing showed that prolog failures was ignored with this configuration we didn't have time to investigate further and simply re-enabled our work-around instead. Regards, Pär Lindfors, NSC
