Dear Slurm users,
after opening our new cluster (62 nodes - 250 GB RAM, 64 cores each -
Rocky Linux 8.6 - Kernel 4.18.0-372.16.1.el8_6.0.1 - Slurm 22.05) for
"friendly user" test operation about 6 weeks ago we were soon facing
serious problems with nodes that suddenly become unresponsive (so much
so that only a hard reboot via IPMI gets them back).
We were able to narrow the problem down to one similar to this one:
https://github.com/apptainer/singularity/issues/5850
Although in our case it's not related to Singularity but generally to
cgroups.
We are using cgroups in our Slurm configuration to limit RAM, CPUs and
devices. In the beginning we did *not* limit swap space (we are doing so
now to work around the problem but would like to allow at least some
swap space).
We are able to reproduce the problem *outside* Slurm as well by using
the small test program mentioned in the above Singularity GitHub-issue
(https://gist.github.com/pja237/b0e9a49be64a20ad1af905305487d41a) with
these steps (for cgroups/v1):
cd /sys/fs/cgroup/memory
mkdir test
cd test
echo $((5*1024*1024*1024)) > memory.limit_in_bytes
echo $$ > cgroup.procs
/path/to/mempoc 2 10
After about 10 to 30 minutes the problem occurs.
We tried to switch to cgroups/v2. Which does solve the problem for the
manual case outside Slurm:
cd /sys/fs/cgroup
mkdir test
cd test
echo "+memory" > cgroup.subtree_control
mkdir test2
echo $((5*1024*1024*1024)) > test2/memory.high
echo $$ > test2/cgroup.procs
/path/to/mempoc 2 10
Now it runs for days and weeks without any issues!
But when we run the same thing in Slurm (with cgroups/v2 configured to
*not limit* swapping) by using
sbatch --mem=5G --cpus-per-task=10 \
--wrap "/path/to/mempoc 2 10"
the nodes still become unusable after some time (1 to 5 hours) with the
usual symptoms.
Did anyone of you face similar issues?
Are we missing something?
Is it unreasonable to think our systems should stay stable even when
there is cgroup-based swapping?
Kind regards,
Hermann
