This problem is reproducible by launching a job then killing the
  slurmstepd process. Under those conditions, requeue the job if
  possible (i.e. batch job with requeue option/configuration).
  This patch also improves the slurmctld logging when this happens.
bug 1889

Correct RebootProgram logic when executed outside of a maintenance
  reservation. Previous logic would mark the node up upon response
  to the reboot RPC (from slurmctld to slurmc) and when the node
  actually rebooted, flag that as an unexpected reboot. This new
  logic checks the node's up time to not mark the compute node as
  being usable until the reboot actually takes place.
but 1866

Add environment variables SLURM_ARRAY_TASK_MAX, SLURM_ARRAY_TASK_MIN,
SLURM_ARRAY_TASK_STEP for job arrays.
bug 1600

to free it otherwise.

Bug 1826

association manager.

Prevent job array task ID from being reported as NO_VAL if last task in the
  array gets requeued. The problem is that when that task starts, the task
  bitmap entry for it stays set, but the task counter gets decremented.
  If that job then gets requeued, under some conditions a failure to schedule
  it results in the array_task_id in the job record getting set to NO_VAL.
  Then when building the job info to report for squeue/scontrol, the string
  showing the pending task ID's is not rebuilt due to that counter being
  zero. All indications are that the job runs fine, only the information
  reported to squeue/scontrol is wrong.
bug 1790

accounting correctly to avoid meaningless errors about overflow.

Bug 1873

intended for the head of the tree.

or binary.

wouldn't get the correct protocol version to launch a step.

Bug 1831

Bug 1869

Fix gang scheduling/preemption issue that could cancel job at startup.
I have not been able to reproduce the reported problem, but this
should prevent the reported problem.
bug 1880

This reverts commit 406cb529.

Conflicts:
	NEWS

This works ok up until 1048576M then it will make the divisor
1048576 and when it goes to divide it only does it once instead of 2 times
returning 1G instead of 1T.

We have decided to wait until 15.08 for a no convert option to be added.

This reverts commit 406cb529.

Conflicts:
	NEWS

This works ok up until 1048576M then it will make the divisor
1048576 and when it goes to divide it only does it once instead of 2 times
returning 1G instead of 1T.

We have decided to wait until 15.08 for a no convert option to be added.

Bug 858

Fix case where job would get the wrong cpu count when using --ntasks-per-core and --cpus-per-task together.

Bug 1370

We are using a non-default AuthInfo configuration and based on
log-messages we see I believe this is not properly handled in certain
parts of the code.

Typical log message:
Aug 12 17:06:15 t02n20 slurmd[27001]: error: Munge encode failed:
Failed to access "/var/run/munge/munge.socket.2": No such file or
directory
Aug 12 17:06:15 t02n20 slurmd[27001]: error: Creating authentication
credential: Socket communication error
Aug 12 17:06:15 t02n20 slurmd[27001]: error: stepd_connect to 3165.0
failed: Protocol authentication error
Aug 12 17:06:15 t02n20 slurmd[27001]: error: If munged is up, restart
with --num-threads=10

Unfortunately the reservation's core_bitmap is a global bitmap and the nodes
in the system may have changed in terms of their node count or nodes have been
added or removed. Make best effort to rebuild the reservation's core_bitmap
on the limited information currently available. The specific cores might
change, but this logic at least leaves their count constant and uses the
same nodes
bug 1850

Bug 1770

thing (-F exclusive) instead of -F share.

bug 1849

This reverts part of commit c2fbf88f
see bug 1846

Fix scheduling anomaly with job arrays submitted to multiple partitions,
    jobs could be started out of priority order.
bug 1892

portion of a job array is pending after it's initial start in the
database it could create another row for the remain array instead
of using the already existing row.

bug 1840

This commit adds a new field free_mem to the node_info struct
and augments the 15.08 RPC protocol to exchange this information
as part of the ping and registration status messages. The
free memory is taken from the sysinfo() information on systems
that support this function. Overall, the implementation sticks
to the way cpu_load (the 5 min load average) is handled.