Re: Found the commit that causes the OOMs

[Wu Fengguang <fengguang.wu@gmail.com>]

On Sun, Jul 05, 2009 at 07:38:54PM +0900, KOSAKI Motohiro wrote:
> > >> OK. thanks.
> > >> I plan to submit this patch after small more tests. it is useful for OOM analysis.
> > >
> > > It is also useful for throttling page reclaim.
> > >
> > > If more than half of the inactive pages in a zone are
> > > isolated, we are probably beyond the point where adding
> > > additional reclaim processes will do more harm than good.
> > 
> > Maybe we can try limiting the isolation phase of direct reclaims to
> > one per CPU?
> > 
> >         mutex_lock(per_cpu_lock);
> >         isolate_pages();
> >         shrink_page_list();
> >         put_back_pages();
> >         mutex_unlock(per_cpu_lock);
> > 
> > This way the isolated pages as well as major parts of direct reclaims
> > will be bounded by CPU numbers. The added overheads should be trivial
> > comparing to the reclaim costs.
> 
> hm, this idea makes performance degression on few CPU machine, I think.

Yes, this is also my big worry. But one possible workaround is to
allow N direct reclaims per CPU.

> e.g.
> if system have only one cpu and sysmtem makes lumpy reclaim, lumpy reclaim
> makes synchronous pageout and it makes very long waiting time.

We can temporarily drop the lock during the writeback waiting.
0-order reclaims shall not be blocked by ongoing high order reclaims.

> I suspect per-cpu decision is not useful in this area.

Maybe. I'm just proposing one more possible way to choose from :)

Thanks,
Fengguang