Bug 7517

Bug 7517

"NVML Library" is redundant.
Add missing env var styling.
Remove awkwardly-placed, duplicated paragraph on MPS in gres.conf man
page. That information is already in gres.html. Also, saying that MPS
is required to be specified in gres.conf is wrong.
Add a reference to gres.html for a more in-depth explanation of GPU
device IDs with NVML.

Bug 7517

Bug 7517

Bug 7517

Add some slurm.conf validation to make sure that typed GRES and
untyped GRES don't mix (e.g. do not allow gres=gpu:1,gpu:tesla:1).

Bug 7517

Bug 7517

Bug 7517

Remove dependence on fast_schedule param.
Add slurm.conf gres specifications for existing tests and make other
adaptations where possible to work with new changes.
Remove some obsolete tests.
Change test descriptions where appropriate.

Bug 7517

Bug 7517

If a configured GPU matches a GPU on the system, match them together.
If a configured GPU has mismatched Cores or Links with the system GPU,
then omit that configured GPU from the final list.

Bug 7517

Move gres_plugin_init_node_config() before
gres_plugin_node_config_load(), so it initializes node_gres_list.
Pass node_gres_list into gres_plugin_node_config_load(), so it gets
the GRES defined in slurm.conf and merges with gres.conf properly.

Bug 7515

Before, data from slurm.conf was only partially merged if there was
no corresponding entry in gres.conf.
Now, Slurm tries to match up gres.conf records to what is defined in
slurm.conf.

Bug 7517

Bug 7517

Bug 7517

Because the output is alphabetically sorted, "error:" always comes
before "GRES_PARSABLE." Thus, when looking for "WARNING:" instead,
that comes *after* GRES_PARSABLE and can royally mess things up.

The fix is to insulate the output and error regexes from each other by
running them in separate expect loops. This is way more robust and
will guarantee that they never again interfere with each other.

Bug 7517

Create a consistent place for the argument to be specified in tests.

Bug 7517

Add an option to manually disable output sorting and allow extra
logging output.

The problem is that the output is sorted, and is no longer in the
order it printed out. This is confusing when trying to debug the
the test program with an elevated log level.

Bug 7517

This is needed to get rid of some "possibly lost" blocks in Valgrind.

Bug 7517

Bug 7517

Bug 7169

Continuation of 2c44fcf6.

Bug 7842.

Increase the maximum array len large to be packed/unpacked with one order of
magnitude, since the current value demonstrated it is not enough when an MPI
program spawns a considerable amount of tasks over a big set of nodes.

This limit was introduced in 627928f4.

Bug 7495

Previous version relayed in the command "time".
It sent STOP and CONT signals and counted time elapsed.
The problem is that defuct childern of stopped parents are not fully
killed, and "time" keep counting time until the actual parent
continue. So, the values were wrong.

The new version uses signal handlers in a .prog and trap in the
shell to print and check if a signal is received.

NOTE: cgroups has different signaling than linuxproc and pgid.

Bug 7282

In previous commit we have double signaling for normal steps
when using --full.

Bug 7282

Bug 7765

There was never any security to allow for this,
so we are just removing it.

Bug 7765

Admin/Operator users were not able to skip MaxQueryTimeRange when trying
to show/fix runaway jobs.
This commit uses _validate_operator() instead of
_validate_slurm_user() in _get_jobs_cond() as well as check for operators
in _fix_runaway_jobs().

Bug 7765

Bug 7708

Don't remove jobs from preemptee_candidates List.

Bug 7708

Incorrect since 18.08 / commit c1a537db when control_machine
became an array.

Bug 7790.

Since this happens inside a the user process, this can inadvertently
cause the user's job to die by running out of file descriptors.

Bug 7814.

Co-authored-by: William Arndt <warndt@lbl.gov>

Bug 7499

Regression introduced in fb26b706.

Bug 7675

Docs - note that #SBATCH directives stop processing after the first non-comment non-whitespace line.

Bug 7763.

Now the signaling of the batch step and the handeling of the flags is totally
handled in _kill_all_active_steps() in slurmd, and _handle_signal_container()
in stepd to ensure that:
- if KILL_JOB_BATCH then only batch container is signaled
- if KILL_FULL_JOB then batch script and its children are also signaled
- if both of the above then only the batch script and its children are signaled

We do not relay anymore on proctrack_g_signal() to handle the batch step
signaling anymore, therefore it works the same for all proctrack plugins.

This commit also includes minor related fixes in other code handling such
signaling flags, and documentation improvement.

Bug 7282

Bug 7789