Fix job state recovery logic in which a job's accounting frequency was
not set. This would result in a value of 65534 seconds being used (the
equivalent of NO_VAL in uint16_t), which could result in the job being
requeued or aborted.

Problem reported by BYU. slurm.conf included a file one byte in
length. Logic created a buffer one byte long and used fgets()
to read the file. fgets() reads one byte less than the buffer
size to include a trailing '\0', so it fails to read the file.

Basically the system size has to be set up before you call the
priority/multifactor plugin.  If a job is finishing while the slurmctld
is starting then it would fatal on the init if it wasn't set up.

core reservations and reservation prolog/epilog

I'm running Slurm 2.6.0 and MWM 7.2.4 in our test cluster at the moment.  I happened to notice that node load reporting wasn't consistent- periodically you'd see a "sane" load reported in Moab, but most of the time the reported load was zero despite an accurate CPULoad value reported by "scontrol show node".

Finally got to digging into this.  It appears that the only time load was being reported properly was in the Moab scheduling cycle directly after slurmctld did a node ping.  In subsequent scheduling cycles the load (again, as reported by Moab) was back to zero.

The node ping is significant as that is the only time the node is updated- since the wiki2 interface only reports records that change, and the load record isn't changed, it isn't reported in the queries after the node ping.

Judging from this behavior, I'm guessing that Moab does not store the load value- every time it queries resources in Slurm it sets the node's load back to zero.

I've altered src/plugins/sched/wiki2/get_nodes.c slightly- basically moved the section that reports CPULOAD above the check for updated info (update_time > last_node_update).

So I don't know if this is the appropriate way to fix it.  The wiki specification that Adaptive has published doesn't seem to indicate how this should function.  Either MWM should assume the last value reported is still accurate or Slurm needs to report it for every wiki GETNODES command.

Anyway, the patch is attached, it seems to be working for me, and I've rolled it into our debian build directory.  YMMV.

Michael

I don't see how this could happen, but it might explain something
reported by Harvard University. In any case, this could prevent
an infinite loop if the task distribution funciton is passed a
job allocation with zero nodes.

This problem was reported by Harvard University and could be
reproduced with a command line of "srun -N1 --tasks-per-node=2 -O id".
With other job types, the error message could be logged many times
for each job. This change logs the error once per job and only if
the job request does not include the -O/--overcommit option.

was down (slurmctld not running) during that time period.

version of Slurm.

sacct.

Remove documentation about "*cpu" portion of gres specification
in the man pages for salloc, sbatch, and srun. Support for this
specification was never implemented nor does the GRES data structure
include a field for it.

as the job completes.

of at multifactor poll.

Need higher memory limits due to pmdv12 size
pmdv12 fails to recognize immediate application exit, hangs with defunct process