If a job has non-responding node, retry job step create rather than
returning with DOWN node error.
bug 734

Also correct related documentation

Completing nodes is removed when calling _try_sched() for a job, which
is the case in select_nodes(). If _try_sched() thinks the job can run
now but select_nodes() returns ESLURM_NODES_BUSY, the backfill loop will
be ended.

Do not resume a job with specialized cores on a node running another job
with specialized cores (only one can run at a time).
bug 792

Fix sinfo -R to print each down/drained node once, rather than once per
partition. This was broken in the sinfo change to process each partition's
information in a separate pthread.

Correct sinfo --sort fields to match documentation: E => Reason,
H -> Reason Time (new), R -> Partition Name, u/U -> Reason user (new)

Without this patch, jobs with an infinite time limit would have
their preemption GraceTime ignored.

related to bug 789

them.

Added ChosLoc configuration parameter in slurm.conf (Chroot OS tool
location).
bug 685

In slurm.spec file, replace "Requires cray-MySQL-devel-enterprise" with
"Requires mysql-devel" per David Gloe.

Permit jobs steps full control over cpu_bind options if specialized cores
are included in the job allocation.
bug 782

Related to bug 771

In version 14.03.2 was using "slurm_<jobid>_4294967294.out" due to
error in job array logic.

cnode counts.

This is for bug 775

regression from 2a674aee

is running.

is running.

together.

Switch/nrt - Properly track usage of CAU and RDMA resources with multiple
tasks per compute node. Previous logic would allocate resources once per
task and then deallocate once per node, leaking CMA and RDMA resources
and preventing their use by future jobs.

If a job is held, then only release it with the "scontrol release <jobid>"
command rather than a simple reset of the job's priority. This is needed to
support job arrays better. Otherwise a priority reset of a job array
would free all requeued/held jobs from that job array rather than
leaving them held.

Modify slurmd to keep track of which jobs have already been launched.
It the launch is complete, then process suspend requests immediately.
Previously the suspend request was always delayed by 1 second, which
adversely impacts gang scheduling performance. If the job can't be
found (say after a slurmd restart), then delay the suspend by up
to 3 seconds, but only once.