[RFC] [PATCH] Cgroup based OOM killer controller

[Paul Menage]

On Thu, Jan 22, 2009 at 5:21 AM, Evgeniy Polyakov <zbr at ioremap.net> wrote:
> Having userspace to decide which task to kill may not work in some cases
> at all (when task is swapped and we need to kill someone to get the mem
> to swap out the task, which will make that decision).

That's true in the case of a global OOM. In the case of a local OOM
(caused by memory limits applied via the cgroup memory controller, or
NUMA affinity enforcement applied by cpusets) the userspace handler
can be in a different domain which isn't OOM, and be quite capable of
figuring out who to kill. In our particular use case, it can happen
that a high-priority job hits its memory limits and triggers an OOM,
which causes the system controller daemon to kill some lower-priority
job and reassign some memory from that now-dead low-priority job (and
thus prevent the OOM from killing any process in the original cgroup).
This is something that would be very hard to express via kernel
policy.

Paul