[slurm-users] Wedged nodes from cgroups, OOM killer, and D state process

[John Hearns]

Chris, I have delved deep into the OOM killer code and interaction with
cpusets in the past (*).
That experience is not really relevant!
However I always recommend looking at this sysctl parameter
 min_free_kbytes
https://access.redhat.com/documentation/en-us/red_hat_enterprise_linux/6/html/performance_tuning_guide/s-memory-tunables
I think of this as the 'wriggle room' the system has when it is playing
Tetris with memory pages.
In there past the min_free_kbytes value was by default ridiculously low.
Look at what value you have and consider increasing by a big factor.


(*)A long time ago, when managing a large NUMA machine, I had an
application which would have a memory leak and trigger the OOM killer.
The application was being run in a cgroup with cpuset and memory location
limits. Once the OOM killer kicked in I saw that the system went 'off the
air' for two minutes, which was quite serious. Finally traced to the OOM
killer inspecting every page in memory on the system, independent of the
cgroup, to see if it had been touched by this process.












On Fri, 30 Nov 2018 at 09:31, Ole Holm Nielsen <Ole.H.Nielsen at fysik.dtu.dk>
wrote:

> On 29-11-2018 19:27, Christopher Benjamin Coffey wrote:
> > We've been noticing an issue with nodes from time to time that become
> "wedged", or unusable. This is a state where ps, and w hang. We've been
> looking into this for a while when we get time and finally put some more
> effort into it yesterday. We came across this blog which describes almost
> the exact scenario:
> >
> > https://rachelbythebay.com/w/2014/10/27/ps/
> >
> > It has nothing to do with Slurm, but it does have to do with cgroups
> which we have enabled. It appears that processes that have hit their
> ceiling for memory and should be killed by oom-killer, and are in D state
> at the same time, cause the system to become wedged. For each node wedged,
> I've found a job out in:
> >
> > /cgroup/memory/slurm/uid_3665/job_15363106/step_batch
> > - memory.max_usage_in_bytes
> > - memory.limit_in_bytes
> >
> > The two files are the same bytes, which I'd think would be a candidate
> for oom-killer. But memory.oom_control says:
> >
> > oom_kill_disable 0
> > under_oom 0
> >
> > My feeling is that the process was in D state, the oom-killer tried to
> be invoked, but then didn't and the system became wedged.
> >
> > Has anyone run into this? If so, whats the fix? Apologies if this has
> been discussed before, I haven't noticed it on the group.
> >
> > I wonder if it’s a bug in the oom-killer? Maybe it's been patched in a
> more recent kernel but looking at the kernels in the 6.10 series it doesn't
> look like a newer one would have a patch for a oom-killer bug.
> >
> > Our setup is:
> >
> > Centos 6.10
> > 2.6.32-642.6.2.el6.x86_64
> > Slurm 17.11.12
>
> As far as I remember, Cgroups underwent a major upgrade with RHEL/CentOS
> 7.  Maybe you should try to upgrade to CentOS 7.5 :-)
>
> /Ole
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.schedmd.com/pipermail/slurm-users/attachments/20181130/cfcbc8e8/attachment.html>