Bug 17176

Bug 17274

Triggered when a node registers with a configured CpuSpecList while
slurmctld configuration has the node without CpuSpecList.

The patch protects against a NULL dereference in bit_equal().

Bug 17322

Extra 'a' in "...and not a reinitialize the..."

Bug 17371

Bug 16731

If using a backup slurmctld and the primary goes down temporarily,
slurmstepd uses the slurmctld timeout to determine when to retry
contacting the primary. Send the value to slurmstepd so it waits the
correct amount of time, rather than NO_VAL16/2.

The slurmctld timeout is initialized to NO_VAL16 in init_slurm_conf()
and the sleep time of slurmctld_timeout / 2 happens in
slurm_send_recv_controller_msg().

Bug 17232

Bug 17360

This partially reverts commit dfb07974. Commit dfb07974 removed code
that limits avail_cpus on a node. This commit restores that code only for
jobs that do not request gres. For jobs that request gres, avail_cpus is
limited later by gres_select_filter_sock_core.

Prior to this commit, for jobs that do not request gres, avail_cpus is not
limited based on the maximum number of tasks that can run on the node.
Therefore, fewer nodes could be assigned to the job than required by tasks.

For example, given nodes with 16 cpus, the following job needs 4 nodes.
However, after commit dfb07974 and before this commit, the job would be
assigned 3 nodes. Three nodes have sufficient cpus to fulfill the job
request, although it does not fulfill the cpus_per_task request.

    srun --mem=0 -n4 --cpus-per-task=12 --exclusive hostname

Bug 17185

Signed-off-by: Marcin Stolarek <cinek@schedmd.com>

Bug 15867

Also clarify that PriorityJobFactor is what's used for the
PriorityWeightPartition calculation.

Bug 17236

Signed-off-by: Nathan Rini <nate@schedmd.com>

In case of CR_CPU_MEMORY can_job_run_on_node can set number of CPUs
available to the job based on --mem-per-cpu to a value that is not a
multiple of ThreadsPerCore. In this case we can't assume that removal
of core always removes ThreadsPerCore. Instead of that assumption just
reset the value number of available cores * ThreadsPerCore.

Bug 17229

This is an update of commit 49bcc0d4. Originally the message was at
debug(), then changed to verbose(), and is now being upgraded to info().

Bug 16664, 17341.

Edit failing condition to not fail because of debug messages.

Bug 14351

Add subtest to check namespaces don't interfere with scheduling and
uid of job

Bug 15441

Update test_123_1.py to work with new slurm-23.11 changes.
The output from reservation errors was changed slightly.

Bug 17203

Update test_116_18.py to work with slurm 23.11.
Specify SelectType=select/linear in slurm.conf

Bug 16860

Bug 17113

slurmctld only considers sockets when allocating resources for a job.
NUMA nodes are not considered sockets unless numa_node_as_socket is
configured.

Bug 17113

In Slurm 23.02, options['uid'] in cli_filter.lua does not return the UID of
the user but returns the result of --uid. This change was made to allow
Linux user namespaces to work. In addition, because a cli_filter is not
secure, an example of rejecting a job due to an invalid constraint is better
done in a job_submit plugin.

For these reasons, move and adapt the example code checking UID and
interactive jobs to job_submit.lua.example.

Bug 17075

Bug 16133

Bug 17330

Since container creation happens in the extern step, it is expected that
_create_ns() will run only once per job per node.  If the job namespace
directory already exists, it is better to assume that BasePath is
configured incorrectly and fail to launch the step than to continue and
potentially break the container.

Bug 16164

Previous logic was only checking a week out for overlapping reoccurring
reservations. Although in some cases those reservation may replace their
nodes and won't overlap, we cannot assume that. The new logic checks the
relevant reoccurring cases advancing the necessary reoccurring periods in
both reservations. It also handles weekdays and weekends properly.

Although commit cd750f97 was not really introducing this issue, it
increased the chance of overlapping reservations in 22.05.0.

Bug 16731

Bug 17192

Bug 17192

Bug 17192

Bug 17192

Bug 17192

As of 23.02, SLURM_NTASKS is only set as an output environment variable for
salloc and sbatch when --ntasks is explicitly requested.
SLURM_NTASKS is no longer set based on calculation of other options.
See commit ef513023 for sbatch and commit 57ec3eb5 for salloc.

Bug 17108

Modify test39.21 to use manual node selection instead of --gres, which
doesn't work with select/linear, to ensure resources are present for test.

Bug 16845

test_gpus_per_node uses --gpus-per-node, which isn't compatible with
select/linear.

Bug 16844

Adjust resource limits (CPUs) to reflect actual node setup instead of
minimal hard-coded limits when on select/linear.

Bug 16836

Test originally checks for exclusive node allocation via oversubscribtion.
Extend these to check for select/linear also to determine exclusive
node allocation.

Bug 16828

Test measures performance with large amount of jobs. Jobs can't be
scheduled on the same node with select/linear, so this makes the
performance of this test much more dependant on the actual configuration.

Bug 16805

Just call net_stream_listen_ports() instead.

Bug 16161.

Signed-off-by: Nathan Rini <nate@schedmd.com>

It is possible to swap between cgroup and linux plugins. Also linux plugin
has a performance penalty on systems with many tasks.

Bug 17099