Re: Terminate process that fails on a constrained allocation

[Andi Kleen <ak@suse.de>]

On Wednesday 08 February 2006 19:57, Paul Jackson wrote:
> > No it only disables the oom killer for constrained allocations.
> 
> But on many big numa systems, the way they are administered,
> that affectively disables the oom killer.

I guess it won't matter because they are administrated with appropiate ulimits
I guess. And to be honest the OOM killer never really worked all that
well, so it's not a big loss.

[still often wish we had that "global virtual memory ulimit for uid"]


> I've yet to be convinced that the oom killer is our friend,
> and half of me (not seriously) is almost wishing it were
> gone.

It's more than half of me near seriously agreeing with you.

> Would another option be to continue to fine tune the heuristics
> that the oom killer uses to pick its next victim?
> 
> What situation did you hit that motivated this change?

It's a long known design bug of the NUMA policy, but it recently 
hit with some test program again.


-Andi