Re: [PATCH 4/7] vmscan: narrowing synchrounous lumply reclaimcondition

[Mel Gorman]

On Wed, Oct 27, 2010 at 06:41:38PM +0200, Andrea Arcangeli wrote:
> Hi,
> 
> > <SNIP>
> 
> [...]
> 
> this rejects on THP code, lumpy is unusable with hugepages, it grinds
> the system to an halt, and there's no reason to let it survive. Lumpy
> is like compaction done with an hammer while blindfolded.
> 

The series drastically limits the level of hammering lumpy does to the
system. I'm currently keeping it alive because lumpy reclaim has received a lot
more testing than compaction has. While I ultimately see it going away, I am
resisting it being deleted until compaction has been around for a few releases.

> I don't know why community insists on improving lumpy when it has to
> be removed completely, especially now that we have memory compaction.
> 

Simply because it has been tested and even with compaction there were cases
envisoned where it would be used - low memory or when compaction is not
configured in for example. The ideal is that compaction is used until lumpy
is necessary although this applies more to the static resizing of the huge
page pool than THP which I'd expect to backoff without using lumpy reclaim
i.e. fail the allocation rather than using lumpy reclaim.

> I'll keep deleting on my tree...
> 
> I hope lumpy work stops here and that it goes away whenever THP is
> merged.
> 

Uhhh, I have one more modification in mind when lumpy is involved and
it's to relax the zone watermark slightly to only obey up to
PAGE_ALLOC_COSTLY_ORDER. At the moment, it is freeing more pages than
are necessary to satisfy an allocation request and hits the system
harder than it should. Similar logic should apply to compaction.

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