that ended in the OverTimeLimit interval.

Caused by allocating single adapter per node of specific adapter type.

Previous logic would pick CPUs then
reject jobs that can not match GRES to the allocated CPUs. New logic first
filters out CPUs that can not use the GRES, next picks CPUs for the job,
and finally picks the GRES that best match those CPUs.
bug 410

Report anything that is world writable.

/* Current code (<= 2.1) has it so we start the new
 * job with the next step id.  This could be used
 * when restarting to figure out which step the
 * previous run of this job stopped on. */

See
https://groups.google.com/forum/#!topic/slurm-devel/j4izr0L4w8w

due to multiple free calls caused by job arrays submitted to
multiple partitions. The root cause is the job priority array
of the original job being re-used by the subsequent job array
entries. A similar problem that could be induced by the user
specifying a job accounting frequency when submitting a job
array is also fixed.
bug 401

sacctmgr.

If reservation create request included a CoreCnt value and more
nodes are required than configured, the logic in select/cons_res
could go off the end of the core_cnt array. This patch adds a
check for a zero value in the core_cnt array, which terminates
the user-specified array.
Back-port from master of commit 211c224b

This is a correction of a bug introduced in commit
https://github.com/SchedMD/slurm/commit/ac44db862c8d1f460e55ad09017d058942ff6499
That commit eliminated the need of reading the node state information
from squeue for performance reasons (mostly for large parallel systems
in which the Prolog ran squeue, which generates a lot of simultaneous
RPCs, slowing down the job launch process). It also assumed 1 CPU per
node. If a pending job specified a node count of 1 and a task count
larger than one, squeue was reporting the node count of the job as
the same as the task count. This patch moves that same calculation
of a pending job's minimum node count into slurmctld, so the squeue
still does not need to read the node information, but can report the
correct node count for pending jobs with minimal overhead.

to avoid it thinking we don't have a cluster name.

If there are completing jobs, a reconfigure will set wrong job/node
state: all nodes of the completing job will be set allocated, and the
job will not be removed even if the completing nodes are released. The
state can only be restored by restarting slurmctld after the completing
nodes released.

could end up before the job started. Bug 371

This avoids waiting for the job's initiation to fail.

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.

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.

of at multifactor poll.

to drain the node and log error slurmd log file.