Correct tracking of licenses for suspended jobs on slurmctld reconfigure or
restart. Previously licenses for suspended jobs were not counted, so
the license count could be exceeded with those jobs get resumed.

commit a9dc50d4.

to commit a9dc50d4.

Refine commit 5f89223f based upon
feedback from David Gloe:
* It's not only MPI jobs, but anything that uses PMI. That includes MPI,
shmem, etc, so you may want to reword the error message.
* I added the terminated flag because if multiple tasks on a node exit,
you would get an error message from each of them. That reduces it to one
error message per node. Cray bug 810310 prompted that change.
* Since we're now relying on --kill-on-bad-exit, I think we should update
the Cray slurm.conf template to default to 1 (set KillOnBadExit=1 in
contribs/cray/slurm.conf.template).
bug 1171

Treat Cray MPI job calling exit() without mpi_fini() as fatal error for
that specific task and let srun handle all timeout logic.
Previous logic would cancel the entire job step and srun options
for wait time and kill on exit were ignored. The new logic provides
users with the following type of response:

$ srun -n3 -K0 -N3 --wait=60 ./tmp
Task:0 Cycle:1
Task:2 Cycle:1
Task:1 Cycle:1
Task:0 Cycle:2
Task:2 Cycle:2
slurmstepd: step 14927.0 task 1 exited without calling mpi_fini()
srun: error: tux2: task 1: Killed
Task:0 Cycle:3
Task:2 Cycle:3
Task:0 Cycle:4
...

bug 1171

This fixes a race condition if the slurmctld needed to power up
a node shortly after startup. Previously it would execute the
ResumeProgram twice for effected nodes.

Without this change, a node in the cloud that failed to power up,
would not have its NoResponding flag cleared, which would prevent
its later use. The NoResponding flag is now cleared when manuallly
when the node is modified to PowerDown.

If a batch job launch to the cloud fails, permit an unlimited
number of job requeues. Previously the job would abort on the
second launch failure.

This reverts commit 4d03d0b4.

Make sure the correct Author is attributed here.

This reverts commit 1891936e.

This has apparently been broken from the get go.  This fixes bug 1172.

test21.22 should be updated to test the dump and load of a file that is
generated.

since this represents the user could be exaggerating their system.

using --ntasks-per-node.

This is related to bug 1145.  What was happening is all the cpus were
allocated on one socket instead of a cyclic method.  While this is allowed
it is strange and resulted in this bug.  There appears to be a different
bug as to why the tasks were laid out in a block fashion in the first
place.

with no way to get them out.

This fixes bug 1134.  It is advised the pro/epilog to call xtprocadmin in
the script instead of returning a non-zero exit code.

The job could have been purged from a short MinJobAge and the
trigger would then point to an invalid job.
Bug #1144

Note that PlugStackConfig defaults to plugstack.conf in the same
directory as slurm.conf. The added logic tests if the file actually
exists (using stat) and if not found then do not fork/exec slurmstepd
to invoke the spank prolog/epilog. This saves about 14msec on startup
and 14msec on shutdown if no spank plugins are configured. It also
eliminates some possible failures (e.g. if fork() fails, or the
slurmstepd processes can not exec()). This logic also caches the
PlugStackConfig value and reads it again on reconfigure, but avoid
reading the value for each job.
bug 982

Add "void" argument to a function and rename a local function to have
a prefix of "_"

9b00f12c

Signed-off-by: Danny Auble <da@schedmd.com>

if a node is down, then permit setting its state to power down,
which causes the SuspendProgram to run and set the node state
back to cloud.

If a node is powered down, then do not power it up on slurmctld
restart.

The power up/down request only takes effect after the ResumeTimeout
or SuspendTimeout is reached in order to avoid a race condition.

Bug #1143

This commit fixes a bug we observed when combining select/linear with
gres. If an allocation was requested with a --gres argument an srun
execution within that allocation would stall indefinitely:

-bash-4.1$ salloc -N 1 --gres=gpfs:100
salloc: Granted job allocation 384049
bash-4.1$ srun -w j3c017 -n 1 hostname
srun: Job step creation temporarily disabled, retrying

The slurmctld log showed:

debug3: StepDesc: user_id=10034 job_id=384049 node_count=1-1 cpu_count=1
debug3:    cpu_freq=4294967294 num_tasks=1 relative=65534 task_dist=1 node_list=j3c017
debug3:    host=j3l02 port=33608 name=hostname network=(null) exclusive=0
debug3:    checkpoint-dir=/home/user checkpoint_int=0
debug3:    mem_per_node=62720 resv_port_cnt=65534 immediate=0 no_kill=0
debug3:    overcommit=0 time_limit=0 gres=(null) constraints=(null)
debug:  Configuration for job 384049 complete
_pick_step_nodes: some requested nodes j3c017 still have memory used by other steps
_slurm_rpc_job_step_create for job 384049: Requested nodes are busy

If srun --exclusive would have be used instead everything would work fine.
The reason is that in exclusive mode the code properly checks whether memory
is a reserved resource in the _pick_step_node() function.
This commit modifies the alternate code path to do the same.

(i.e ArchiveJobs PurgeJobs).  This is only a cosmetic change.

on slurmdbd startup.

lots of jobs.