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

[Mel Gorman <mel@csn.ul.ie>]

On Wed, Jun 10, 2009 at 10:14:40AM +0800, Wu Fengguang wrote:
> On Tue, Jun 09, 2009 at 11:06:19PM +0800, Mel Gorman wrote:
> > On Tue, Jun 09, 2009 at 09:38:04PM +0800, Wu Fengguang wrote:
> > > On Tue, Jun 09, 2009 at 05:40:50PM +0800, Mel Gorman wrote:
> > > > 
> > > > Conceivably though, zone_reclaim_interval could automatically tune
> > > > itself based on a heuristic like this if the administrator does not give
> > > > a specific value. I think that would be an interesting follow on once
> > > > we've brought back zone_reclaim_interval and get a feeling for how often
> > > > it is actually used.
> > > 
> > > Well I don't think that's good practice. There are heuristic
> > > calculations all over the kernel. Shall we exporting parameters to
> > > user space just because we are not absolutely sure? Or shall we ship
> > > the heuristics and do adjustments based on feedbacks and only export
> > > parameters when we find _known cases_ that cannot be covered by pure
> > > heuristics?
> > > 
> > 
> > Good question - I don't have a satisfactory answer but I intuitively find
> > the zone_reclaim_interval easier to deal with than the heuristic.  That said,
> > I would prefer if neither was required.
> 
> Yes - can we rely on the (improved) accounting to make our "failure feedback"
> patches unnecessary? :)
> 

Am awaiting test results to answer that question :)

> Thanks,
> Fengguang
> 
> > In the patchset, I've added a counter for the number of times that the
> > scan-avoidance heuristic fails. If the tmpfs problem has been resolved
> > (patch with bug reporter, am awaiting test), I'll drop zone_reclaim_interval
> > altogether and we'll use the counter to detect if/when this situation
> > occurs again.
> 

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