[RFC] [PATCH] Cgroup based OOM killer controller

Nikanth Karthikesan knikanth at suse.de
Thu Jan 22 02:10:07 PST 2009


On Thursday 22 January 2009 15:09:28 David Rientjes wrote:
> On Thu, 22 Jan 2009, Nikanth Karthikesan wrote:
> > > You can't specify different behavior for an oom cgroup depending on
> > > what type of oom it is, which is the problem with this proposal.
> >
> > No. This does not disable any such special selection criteria which is
> > used without this controller.
>
> I didn't say it disabled it; the cpuset preference is actually implemented
> in the badness() score and not specifically excluded in
> select_bad_process().  That's because it's quite possible that a task has
> allocated memory in a cpuset and then either moved to a separate cpuset or
> had it's mems_allowed changed.
>
> Please try it and you'll see.  Create two cpusets, cpuset A and cpuset B.
> Elevate cpuset A's oom.victim value and then trigger an oom in cpuset B.
> Your patch will cause a task from cpuset A to be killed for a cpuset B
> triggered oom which, more often than not, will not lead to future memory
> freeing.
>
> It's quite possible that cpuset A would be preferred to be killed in a
> global unconstrained oom condition, however.  That's the only reason why
> one would elevate its oom.victim score to begin with.  But it doesn't work
> for cpuset-constrained ooms.
>
> It's not going to help if it I explain it further and you don't try it out
> on your own.  Thanks.

Thanks for the clear explanation. Cpuset does it by reducing the badness to 
1/8th for tasks. So using oom-controller could kill some innocent processes on 
some other cpuset! 

But it is possible to have the same effect with oom_adj, having oom_adj=4 for 
a task on a diff cpuset will do the same(assuming they have similar badness).

I think cpusets preference could be improved, not to depend on badness, with 
something similar to what memcg does. With or without adding overhead of 
tracking processes that has memory from a node.

Thanks
Nikanth


More information about the Containers mailing list