lkml.org 
[lkml]   [2009]   [Jun]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 4/4] Reintroduce zone_reclaim_interval for when zone_reclaim() scans and fails to avoid CPU spinning at 100% on NUMA
On Tue, Jun 09, 2009 at 10:54:25PM -0700, Andrew Morton wrote:
> On Tue, 9 Jun 2009 18:01:44 +0100 Mel Gorman <mel@csn.ul.ie> wrote:
>
> > On NUMA machines, the administrator can configure zone_reclaim_mode that is a
> > more targetted form of direct reclaim. On machines with large NUMA distances,
> > zone_reclaim_mode defaults to 1 meaning that clean unmapped pages will be
> > reclaimed if the zone watermarks are not being met. The problem is that
> > zone_reclaim() may get into a situation where it scans excessively without
> > making progress.
> >
> > One such situation occured where a large tmpfs mount occupied a
> > large percentage of memory overall. The pages did not get reclaimed by
> > zone_reclaim(), but the lists are uselessly scanned frequencly making the
> > CPU spin at 100%. The observation in the field was that malloc() stalled
> > for a long time (minutes in some cases) when this situation occurs. This
> > situation should be resolved now and there are counters in place that
> > detect when the scan-avoidance heuristics break but the heuristics might
> > still not be bullet proof. If they fail again, the kernel should respond
> > in some fashion other than scanning uselessly chewing up CPU time.
> >
> > This patch reintroduces zone_reclaim_interval which was removed by commit
> > 34aa1330f9b3c5783d269851d467326525207422 [zoned vm counters: zone_reclaim:
> > remove /proc/sys/vm/zone_reclaim_interval. In the event the scan-avoidance
> > heuristics fail, the event is counted and zone_reclaim_interval avoids
> > excessive scanning.
>
> More distressed fretting!
>

Not at all. One day I'll get a significant patch completed without any
eyebrows raised and the world will end :).

> Pages can be allocated and freed and reclaimed at rates anywhere
> between zero per second to one million per second or more. So what
> sense does it make to pace MM activity by wall-time??
>

None - this is a brute force workaround if the scan heuristic breaks and
was lifted directly from an old patch by Christoph. It could be much
better.

> A better clock for pacing MM activity is page-allocation-attempts, or
> pages-scanned, etc.
>

Agreed. Wu convinced me of that and had some good suggestions. I'm waiting
to hear back from the testers on the new scan heuristics to see if they
are working or not. If they are working now, I'll drop this patch. If we
decide we need it, I'll update with Wu's work. I need to send the tests a
new version now though because of the underflow problem.

--
Mel Gorman
Part-time Phd Student Linux Technology Center
University of Limerick IBM Dublin Software Lab


\
 
 \ /
  Last update: 2009-06-10 12:51    [from the cache]
©2003-2011 Jasper Spaans