Re: [PATCH] oom: skip frozen tasks

[Michal Hocko]

On Fri 26-08-11 11:53:56, Michal Hocko wrote:
> On Fri 26-08-11 02:21:42, David Rientjes wrote:
> > On Fri, 26 Aug 2011, Michal Hocko wrote:
> > 
> > > Let's give all frozen tasks a bonus (OOM_SCORE_ADJ_MAX/2) so that we do
> > > not consider them unless really necessary and if we really pick up one
> > > then thaw its threads before we try to kill it.
> > > 
> > 
> > I don't like arbitrary heuristics like this because they polluted the old 
> > oom killer before it was rewritten and made it much more unpredictable.  
> > The only heuristic it includes right now is a bonus for root tasks so that 
> > when two processes have nearly the same amount of memory usage (within 3% 
> > of available memory), the non-root task is chosen instead.
> > 
> > This bonus is actually saying that a single frozen task can use up to 50% 
> > more of the machine's capacity in a system-wide oom condition than the 
> > task that will now be killed instead.  That seems excessive.
> 
> Yes, the number is probably too high. I just wanted to start up with
> something. Maybe we can give it another root bonus. But I agree whatever
> we use it will be just a random value...
> 
> > 
> > I do like the idea of automatically thawing the task though and if that's 
> > possible then I don't think we need to manipulate the badness heuristic at 
> > all.  I know that wouldn't be feasible when we've frozen _all_ threads and 
> 
> Why it wouldn't be feasible for all threads? If you have all tasks
> frozen (suspend going on, whole cgroup or all tasks in a cpuset/nodemask
> are frozen) then the selection is more natural because all of them are
> equal (with or without a bonus). The bonus tries to reduce thawing if
> not all of them are frozen.
> I am not saying the bonus is necessary, though. It depends on what
> the freezer is used for (e.g. freeze a process which went wild and
> debug what went wrong wouldn't welcome that somebody killed it or other
> (mis)use which relies on D state).

Anyway, I do agree, the two things (bonus and thaw during oom_kill)
should be handled separately.

-- 
Michal Hocko
SUSE Labs
SUSE LINUX s.r.o.
Lihovarska 1060/12
190 00 Praha 9    
Czech Republic
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/