Re: [RFC] [PATCH] Cgroup based OOM killer controller

[Nikanth Karthikesan <knikanth@suse.de>]

On Thursday 22 January 2009 02:19:50 David Rientjes wrote:
> On Wed, 21 Jan 2009, Nikanth Karthikesan wrote:
> > This is a container group based approach to override the oom killer
> > selection without losing all the benefits of the current oom killer
> > heuristics and oom_adj interface.
> >
> > It adds a tunable oom.victim to the oom cgroup. The oom killer will kill
> > the process using the usual badness value but only within the cgroup with
> > the maximum value for oom.victim before killing any process from a cgroup
> > with a lesser oom.victim number. Oom killing could be disabled by setting
> > oom.victim=0.
>
> This doesn't help in memcg or cpuset constrained oom conditions, which
> still go through select_bad_process().
>
> If the oom.victim value is high for a specific cgroup and a memory
> controller oom occurs in a disjoint cgroup, for example, it's possible to
> needlessly kill tasks.  Obviously that is up to the administrator to
> configure, but may not be his or her desire for system-wide oom
> conditions.
>
> It may be preferred to kill tasks in a specific cgroup first when the
> entire system is out of memory or kill tasks within a cgroup attached to a
> memory controller when it is oom.
>
> The same scenario applies for cpuset-constrained ooms.  Since oom.victim
> is given higher preference than all tasks' oom_adj values, it is possible
> to needlessly kill tasks that do not lead to future memory freeing for the
> nodes attached to that cpuset.
>
> It also requires that you synchronize the oom.victim values amongst your
> cgroups.

No, this is not specific to memcg or cpuset cases alone. The same needless 
kills will take place even without memcg or cpuset when an administrator 
specifies a light memory consumer to be killed before a heavy memory user. But 
it is up to the administrator to use it wisely. We also provide a panic_on_oom 
option that an administrator could use, not just to kill few more tasks but 
all tasks in the system ;)

Thanks
Nikanth