One of our testers discovered a regression in version 2.3.1.  If a job is
pending due to PartitionNodeLimit and the limit is relieved with a
'sacctmgr modify qos name=<qos name> set flags=partitionmaxnodes' new jobs
exceeding the partition limit (but not the QOS limit) are allowed to run.
However, the pending job is never allowed to run.  Attached is a patch to
address this problem.  FYI, this problem doesn't exist in version 2.4.
Patch from Bill Brophy, Bull.

Patch by Alexander Bersenev (Institute of Mathematics and Mechanics, Russia).

There was also some bad code that would reset the conn_type of a block
to SMALL no matter what type of SMALL it was.

Patsh from Rod Schulz, Bull.

This was due to a bug in select/cons_res with some configuration
optiions and job options, especially if there is more than one
thread per core and the job option includes "--threads-per-core=1".
Fixes problem reported by CSCS.

lock was deemed not necessary because the information (db_index) was only
internal and was only modified in the same function later which is
protected by the write lock.

qos on the cluster to fill that in correctly.

incorrectly when getting fairshare from the parent.

wasn't started at the time the eligible message was sent but started
before the db_index was returned information like start time would be lost.

Fix Lua link order

Patch from Yuri D'Ella

Note that if a job is requeued, it's submit time is changed and the record appears
as a duplicate job with a different submit time. Patch by Bill Brophy, Bull.

Put -llua* in LIBS rather than LDFLAGS to get correct link order.
Without this the configure test for Lua fails when using GCC 4.6,
the default compiler on recent Linux distributions like Ubuntu
11.10.

Note this is an old bug. The new code keeps slurmstepd alive and it
keeps trying to send step completion message to slurmctld.

This make the same patch to select/linear as Carles Fenoy's patch to
select/cons_res plugin.

Added gres_cpus test. Without this test it could lead to the error "fatal: cons_res: sync loop not progressing" With this patch a job will be rejected if asking for unavailable configuration.

Print an error rather than setting CUDA_VISIBLE_DEVICES environment
variable to  "NoDevFiles" if no device files defined.

Patch 4f68cde5bd6b4fcf839f6694457373c81d9548ba from chaos/slurm by Don Lipari, LLNL

The change in function call order of commit e60abe43
resulted in slurmd daemons on front-end systems not registering with the
proper node name.

Backfill scheduling - Add SchedulerParameters configuration parameter of
"bf_res" to control the resolution in the backfill scheduler's data about
when jobs begin and end. Default value is 60 seconds (used to be 1 second).

Do not drain the compute or front-end node when trying to start a job
for which the UID is not found

Release locks on cray system after inventory and before backfill scheduling loop
in order to not process more jobs and avoid blocking pending RPCs for so long