We’re running slurm 17.11.5 on RHEL 7 and have been having issues with jobs escaping their cgroup controls on GPU devices.


For example we have the following steps running on a single node:

# ps auxn | grep [s]lurmstepd
       0  2380  0.0  0.0 538436  3700 ?        Sl   07:22   0:02 slurmstepd: [46609.0]
       0  5714  0.0  0.0 472136  3952 ?        Sl   Apr11   0:03 slurmstepd: [46603.0]
       0 17202  0.0  0.0 538448  3724 ?        Sl   Apr11   0:03 slurmstepd: [46596.0]
       0 28673  0.0  0.0 538380  3696 ?        Sl   Apr10   0:39 slurmstepd: [46262.0]
       0 44832  0.0  0.0 538640  3964 ?        Sl   Apr11   1:12 slurmstepd: [46361.0]


But not all of those are reflected in the cgroup device hierarchy:

# lscgroup | grep devices | grep slurm
devices:/slurm
devices:/slurm/uid_2093
devices:/slurm/uid_2093/job_46609
devices:/slurm/uid_2093/job_46609/step_0
devices:/slurm/uid_11477
devices:/slurm/uid_11477/job_46603
devices:/slurm/uid_11477/job_46603/step_0
devices:/slurm/uid_11184
devices:/slurm/uid_11184/job_46596
devices:/slurm/uid_11184/job_46596/step_0


This issue only seems to happen after a job has been running for a while, as when it is first started the cgroup controls work as expected.  In this example, the jobs that have escaped the controls (46361,46362) have been running for over a day:

# squeue -j 46609,46603,46596,46262,46361
             JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
             46596     dpart     bash     yhng  R   10:56:00      1 vulcan14
             46609 scavenger     bash    yaser  R    1:52:37      1 vulcan14
             46603 scavenger     bash  jxzheng  R    9:47:26      1 vulcan14
             46361     dpart     bash  jxzheng  R 1-08:31:14      1 vulcan14
             46262     dpart Weighted  umahbub  R 1-18:07:07      1 vulcan14


So it seems that at some point slurm, or something else, comes in and modifies the cgroup hierarchy, but we haven’t had much luck in tracking down what.

We're seeing this happen on multiple different clusters, and it has been occurring since at least 17.02.