Remove the SchedulerParameters option of "assoc_limit_continue", making it
    the default value. Add option of "assoc_limit_stop". If "assoc_limit_stop"
    is set and a job cannot start due to association limits, then do not attempt
    to initiate any lower priority jobs in that partition. Setting this can
    decrease system throughput and utlization, but avoid potentially starving
    larger jobs by preventing them from launching indefinitely.

launching a job, instead the job will fail and drain the node if the env
isn't loaded normally.

bug 2546

with the slurmdbd.

make the wait to return data only hit after 500 nodes and configurable
based on the TcpTimeout value.

There was a subtle bug in how tasks were bound to CPUs which could result
in an "infinite loop" error. The problem was various socket/core/threasd
calculations were based upon the resources allocated to a step rather than
all resources on the node and rounding errors could occur. Consider for
example a node with 2 sockets, 6 cores per socket and 2 threads per core.
On the idle node, a job requesting 14 CPUs is submitted. That job would
be allocted 4 cores on the first socket and 3 cores on the second socket.
The old logic would get the number of sockets for the job at 2 and the
number of cores at 7, then calculate the number of cores per socket at
7/2 or 3 (rounding down to an integer). The logic layouting out tasks
would bind the first 3 cores on each socket to the job then not find any
remaining cores, report the "infinite loop" error to the user, and run
the job without one of the expected cores. The problem gets even worse
when there are some allocated cores on a node. In a more extreme case,
a job might be allocated 6 cores on one socket and 1 core on a second
socket. In that case, 3 of that job's cores would be unused.
bug 2502

This is a revision to commit 1ed38f26
The root problem is that a pthread is passed an argument which is
a pointer to a variable on the stack. If that variable is over-written,
the signal number recieved will be garbage, and that bad signal
number will be interpretted by srun to possible abort the request.

The previous commit obviously fixed a problem, but introduced a different
set of problems. This will be pursued later, perhaps in version 16.05.

With some configurations and systems, errors of the following sort were
occuring:
task/cgroup: task[1] infinite loop broken while trying to provision compute elements using block
task/cgroup: task[1] unable to set taskset '0x0'

Bug 1706

burst_buffer/cray - If the pre-run operation fails then don't issue
    duplicate job cancel/requeue unless the job is still in run state. Prevents
    jobs hung in COMPLETING state.
bug 2587

isn't kept up to date in the cache.

Fix gang scheduling resource selection bug which could prevent multiple jobs
    from being allocated the same resources. Bug was introduced in 15.08.6,
    commit 44f491b8

Here's how to reproduce on smd-server with 2 sockets, 6 cores per
socket and 2 threads per core, just run the following command line
3 times in quick succession (all active at the same time):
srun --cpus-per-task=4 -m block sleep 30
What was happening is the first job would be allocated cores 0+1
The second job would be allocated cores 2+3
The thrid job would test use of cores 0-3 then exit because the
 job only needs 4 CPUs. The resulting core binding would include
 NO CPUs. The new logic tests that the core being considered for
 use actually has some resources available to the job before
 updating the counter which is being tested against the needed
 CPU counter.

Specifically add the HWLOC_TOPOLOGY_FLAG_WHOLE_SYSTEM flag when
loading configuration from HWLOC library. Previous logic in
task/cgroup did not do this, which was different behaviour from
how slurmd gets configuration information. Here's the HWLOC
documentation:
HWLOC_TOPOLOGY_FLAG_WHOLE_SYSTEM
Detect the whole system, ignore reservations and offline settings.
Gather all resources, even if some were disabled by the administrator.
For instance, ignore Linux Cpusets and gather all processors and memory
nodes, and ignore the fact that some resources may be offline.

Without this flag, I was rarely observing a bad core count, which
resulted in the logic layout out tasks wrong and generating an error:
task/cgroup: task[0] infinite loop broken while trying to provision compute elements using cyclic

bug 2502

burst_buffer/cray: Set environment variables just before starting job rather
    than at job submission time to reflect persistent buffers created or
    modified while the job is pending.
bug 2545

buffer is found.

Bug 2576

What happened was a function was doing a double read lock which isn't
awesome to begin with, but not really horrible (if all you are doing is
read locks anyway).  The problem was after the first lock was locked a
different thread was going for a write lock and so when the second
read lock came in it created deadlocked.

Jobs below the specified threshold will not have resources reserved for them.
bug 2565

Avoid possibly aborting srun that gets simultaneous SIGSTOP+SIGCONT while
    creating the job step. The result is that the signal hanlder gets a
    argument (the signal received) of zero.

Here's a log, window 1:
$ srun hostname
srun: Job step creation temporarily disabled, retrying
srun: I Got signal 18
srun: I Got signal 18
srun: I Got signal 18
srun: I Got signal 18
srun: I Got signal 18
srun: I Got signal 18
srun: I Got signal 18
srun: I Got signal 18
srun: I Got signal 18
srun: I Got signal 18
srun: I Got signal 18
srun: I Got signal 18
srun: I Got signal 0
srun: Cancelled pending job step

Window 2:
$  kill -STOP 18696 ; kill -CONT 18696
$  kill -STOP 18696 ; kill -CONT 18696
$  kill -STOP 18696 ; kill -CONT 18696
....

bug 2494

Change how a node's allocated CPU count is calculated to avoid double
    counting CPUs allocated to multiple jobs at the same time.
    Previous logic would sum the maximum number of CPUs allocated by each
    partition for any time slice, which could double count CPUs allocated
    to multiple jobs. New logic ORs bitmap of allocated CPUs for every
    partition and time slice, then counts the total for a given node.
    This avoids double counting CPUs allocated to multiple jobs, but
    does not remove from the count CPUs which have been allocated to
    jobs which might be suspended by the gang scheduler (either for
    time slicing or preemption).

The uid is used as part of the hash function, must remove old reference
and recalculate if it may change, otherwise _delete_assoc_hash
will not find it again when the association is removed, causing
slurmctld to segfault.

Bug 2560.

Add --gres-flags=enforce-binding option to salloc, sbatch and srun commands.
    If set, the only CPUs available to the job will be those bound to the
    selected GRES (i.e. the CPUs identifed in the gres.conf file will be
    strictly enforced rather than advisory).
bug 1725

Previous gang scheduling logic maintained information about resources
  originally allocated to the job and made scheduling decisions on
  that basis.
bug 2494

Update gang scheduling table when job manually suspended or resumed. Prior
    logic could mess up job suspend/resume sequencing.
bug 2494

time.

https://bugs.schedmd.com/show_bug.cgi?id=2547

The code just wasn't fully baked before and was probably written before
a lot of the other supporting code was done i.e
assoc_mgr_set_assoc|qos_tres_cnt were done specifically for this kind of
thing.  Many of the usage structures weren't realloced either as well as
the tres_cnt local to each qos and assoc wasn't updated.  So all in all
pretty bad code - bad Danny.  This makes sure all this sets up and no
memory corruption happens.

Generate burst buffer use completion email immediately afer teardown
    completes rather than at job purge time (likely minutes later).
bug 2539

Change burst buffer use completion message from
    "SLURM Job_id=1360353 Name=tmp Staged Out, StageOut time 00:01:47" to
    "SLURM Job_id=1360353 Name=tmp StageOut/Teardown time 00:01:47"

Bug 2396

Bug 2543.

on only one port like TopologyParam=NoInAddrAny does for everything else.

There's no /proc on *BSD, and BSD handles OOM in a completely different way.

Return [0-100:2] formatting, rather than [0,2,4,6,8,...] when using
a step function.

Was inadvertantly broken in 14.11 with commit 5ffdca92.

Bug 2535.

Need higher count for KNL processor.

Fix Cray NHC spawning on job requeue. Previous logic would leave nodes
allocated to a requeued job as non-usable on job termination.

Specifically, each job has a "cleaning/cleaned" flag. Once a job
terminates, the cleaning flag is set, then after the job node health
check completes, the value gets set to cleaned. If the job is requeued,
on its second (or subsequent) termination, the select/cray plugin
is called to launch the NHC. The plugin sees the "cleaned" flag
already set, it then logs:
error: select_p_job_fini: Cleaned flag already set for job 1283858, this should never happen
and returns, never launching the NHC. Since the termination of the
job NHC triggers releasing job resources (CPUs, memory, and GRES),
those resources are never released for use by other jobs.

Bug 2384