Describe how jobs could be lost on slurmctld crash
Backport MPI performance FAQ
Fix bad href tag

renamed the patch for make 3.81 and added a patch for gnu make 4.0

Previous test could fail for node names numeric suffix with more
than one digit. This captures all digits plus checks for names
without a numeric suffix.

Introduced in commit dc7a1fca

Fix error handling for job array create failure due to inability to copy
job files (script and environment).
See bug 1077

are hard links to the first element specification files. If the
controller fails to make the links the files are copied instead.

reserved for higher priority jobs.

I just ran the test suite for slurm 14.04.7, and have a few suggestions
and bugfixes:

Test 1.35 fails on our system (probably because we limit memory with
cgroups).  Changing job_mem_opt from "--mem-per-cpu=64" to
"--mem-per-cpu=192" in line 61 fixes the problem for us.

Test 1.84 fails to recognise node names like "something1-2", ending up
with node names "something1" instead.  Changing NodeName=(\w+) to
NodeName=([^\s]+) fixes the problem.

Test 1.97 reports FAILURE when it discovers that SelectTypeParameters is
not CR_PACK_NODES.  Having "exit 0" instead of "exit 1" in line 50 is
perhaps preferable.

Test 2.18 fails because the variable $partition never gets set, so no
idle nodes are found in line 215.  Setting $partition in globals.local
helps, but should not be needed, IMO.  There is a function
"default_partition" in globals that could perhaps be used.  The same
applies to test 2.19.

Test 12.2 fails on our system because the jobs get killed due to memory
limit.  Increasing the "slack" in job_mem_limit from 4 to 10 in line 269
fixes the problem for us.

Tests 21.30, 21.31 and 21.32 fails when run as a non-privileged user.
Perhaps they should test for it and exit with a warning instead, like
many other tests.

Test 22.1 fails on our system because the time zone is different from
where the test was written.  The problem is that

set midnight 1201766400

is only correct in one time zone (and unfortunately for us, not in
our :).  Perhaps one could use the GNU date command to get the correct
seconds-since-epoch regardless of time zone.  Something like

date +%s --date=2008-01-31

should do it.  Unfortunately, I don't know enough Expect (tcl?) to
suggest how to implement that.

--
Regards,
Bjørn-Helge Mevik, dr. scient,
Department for Research Computing, University of Oslo

Which in this case is what has happened since the launch failed.

since the slurmd will send the same message.

correctly.

ran inside the allocation can read the environment correctly.

Wait up to 20 seconds for gres.conf file before exiting.

This fixes some problems creating advanced reservations on heterogeneous systems, especially when core counts are specified in the reservation.
bug 1068

Make "srun --gres=none ..." work when executed without a job allocation
(i.e. srun creates the allocation plus the step). Previous logic would
try to create the job with a gres value of "none".

Fix for possible error if job has GRES, but the step explicitly requests a
GRES count of zero.

BlueGene/Q's runjob command requires a fully qualified pathname,
so only for that machine type, resolve the name in srun. This
partitially reverts commit 57efc873
but only for BGQ systems.

and only caused confusion since the cpu_bind options mostly refer to a
step we opted to only allow srun to set them in future versions.

Defer job step initiation of required GRES are in use by other steps rather
than immediately returning an error.
bug 1056

had --network= specified.

ProfileHDF5Dir directory as well as all it's sub-directories and files.

This isn't the case in the current code so this isn't as big of a deal.

This logic will also be removed in 14.11 so it become less of a deal as
well, but just to be safe we cover the base.