Updates of existing bluegene advanced reservations did not work at all.
Some multi-core configurations resulting in an abort due to creating
  core_bitmaps for the reservation that only had one bit per node rather
  than one bit per core.
These bugs were introduced in commit 5f258072

job data structure.

Logic was assuming the reservation had a node bitmap which was
  being used to check for overlapping jobs. If there is no node
  bitmap (e.g. a licenses only reservation), an abort would result.

Bug 1739

Bug 1743

Prevent double free.

Bug 1733

In your node_mgr fix to keep rebooted nodes down (commit 9cd15dfe), you
forgot to consider the case of nodes that are powered up but are responding after
ResumeTimeout seconds (the maximum time permitted). Such nodes are
marked DOWN (because they didn't respond within ResumeTimeout seconds) than
should become silently available when ReturnToService=1 (as stated in the slurm.conf manual)

With your modification when such nodes are finally responding, they are seen as
rebooted nodes and remain in the DOWN state (with the new reason: Node
unexpectedly rebooted) even when ReturnToService=1 !

Correction of commit 3c2b46af

In your node_mgr fix to keep rebooted nodes down (commit 9cd15dfe), you
forgot to consider the case of nodes that are powered up but are responding after
ResumeTimeout seconds (the maximum time permitted). Such nodes are
marked DOWN (because they didn't respond within ResumeTimeout seconds) than
should become silently available when ReturnToService=1 (as stated in the slurm.conf manual)

With your modification when such nodes are finally responding, they are seen as
rebooted nodes and remain in the DOWN state (with the new reason: Node
unexpectedly rebooted) even when ReturnToService=1 !

My patch to obtain the correct behaviour:

Equivalent fix as e1a00772
for select/serial rather than select/cons_res

1. I submit a first job that uses 1 GPU:
$ srun --gres gpu:1 --pty bash
$ echo $CUDA_VISIBLE_DEVICES
0

2. while the first one is still running, a 2-GPU job asking for 1 task per node
waits (and I don't really understand why):
$ srun --ntasks-per-node=1 --gres=gpu:2 --pty bash
srun: job 2390816 queued and waiting for resources

3. whereas a 2-GPU job requesting 1 core per socket (so just 1 socket) actually
gets GPUs allocated from two different sockets!
$ srun -n 1  --cores-per-socket=1 --gres=gpu:2 -p testk --pty bash
$ echo $CUDA_VISIBLE_DEVICES
1,2

With this change #2 works the same way as #3.
bug 1725