This is due to newer versions of SLURM using /sys/fs/cgroup (as in RHEL7)
rather than /cgroup (as in RHEL6). It would have been a no-brainer to check
and do the right thing, but as it is, you have to specify the cgroup mount
point in cgroup.conf if you're on the wrong system for your slurm
So GRES is a supported TRES type, but it appears to not be supported in the
GrpTRESRunMins limits, which list as :
GrpTRESRunMins=cpu=N,mem=N,energy=N,node=N
We use the cpu limit to reduce long-term monopolization of CPU resources,
and recently could really use the same functionality with
"sacctmgr delete user where defaultqos=normal" results in the following
(looping) strace output from slurmdbd:
---
[pid 5323] read(11,
[pid 5052] <... nanosleep resumed> 0x7f6dba013d00) = 0
[pid 5052] rt_sigprocmask(SIG_BLOCK, [CHLD], [HUP INT QUIT ABRT