Re: [PATCH 1/3] Reintroduce zone_reclaim_interval for whenzone_reclaim() scans and fails to avoid CPU spinning at 100% on NUMA

[Wu Fengguang]

On Tue, Jun 09, 2009 at 04:14:25PM +0800, Mel Gorman wrote:
> On Tue, Jun 09, 2009 at 09:58:22AM +0800, Wu Fengguang wrote:
> > On Mon, Jun 08, 2009 at 09:01:28PM +0800, Mel Gorman 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() can be in a situation where it scans excessively without
> > > making progress.
> > > 
> > > One such situation is where a large tmpfs mount is occupying a large
> > > percentage of memory overall. The pages do not get cleaned or reclaimed by
> > > zone_reclaim(), but the lists are uselessly scanned frequencly making the
> > > CPU spin at 100%. The scanning occurs because zone_reclaim() cannot tell
> > > in advance the scan is pointless because the counters do not distinguish
> > > between pagecache pages backed by disk and by RAM.  The observation in
> > > the field is that malloc() stalls for a long time (minutes in some cases)
> > > when this situation occurs.
> > > 
> > > Accounting for ram-backed file pages was considered but not implemented on
> > > the grounds it would be introducing new branches and expensive checks into
> > > the page cache add/remove patches and increase the number of statistics
> > > needed in the zone. As zone_reclaim() failing is currently considered a
> > > corner case, this seemed like overkill. Note, if there are a large number
> > > of reports about CPU spinning at 100% on NUMA that is fixed by disabling
> > > zone_reclaim, then this assumption is false and zone_reclaim() scanning
> > > and failing is not a corner case but a common occurance
> > > 
> > > This patch reintroduces zone_reclaim_interval which was removed by commit
> > > 34aa1330f9b3c5783d269851d467326525207422 [zoned vm counters: zone_reclaim:
> > > remove /proc/sys/vm/zone_reclaim_interval] because the zone counters were
> > > considered sufficient to determine in advance if the scan would succeed.
> > > As unsuccessful scans can still occur, zone_reclaim_interval is still
> > > required.
> > 
> > Can we avoid the user visible parameter zone_reclaim_interval?
> > 
> 
> You could, but then there is no way of disabling it by setting it to 0
> either. I can't imagine why but the desired behaviour might really be to
> spin and never go off-node unless there is no other option. They might
> want to set it to 0 for example when determining what the right value for
> zone_reclaim_mode is for their workloads.
> 
> > That means to introduce some heuristics for it.
> 
> I suspect the vast majority of users will ignore it unless they are runing
> zone_reclaim_mode at the same time and even then will probably just leave
> it as 30 as a LRU scan every 30 seconds worst case is not going to show up
> on many profiles.
> 
> > Since the whole point
> > is to avoid 100% CPU usage, we can take down the time used for this
> > failed zone reclaim (T) and forbid zone reclaim until (NOW + 100*T).
> > 
> 
> i.e. just fix it internally at 100 seconds? How is that better than
> having an obscure tunable? I think if this heuristic exists at all, it's
> important that an administrator be able to turn it off if absolutly
> necessary and so something must be user-visible.

That 100*T don't mean 100 seconds. It means to keep CPU usage under 1%:
after busy scanning for time T, let's go relax for 100*T.
--
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/