Re: [RFC-PATCH 0/7] Memory Compaction v1

From: Mel Gorman
Date: Thu Jan 21 2010 - 05:11:31 EST

On Thu, Jan 21, 2010 at 12:12:11PM +0900, KOSAKI Motohiro wrote:
> Hi Mel,
> Sorry, I haven't read this patch at all.
> > The time differences are marginal but bear in mind that this is an ideal
> > case of mostly unmapped buffer pages. On nice set of results is between
> > allocations 13-18 where no pages were reclaimed, some compaction occured
> > and 300 huge pages were allocated in 0.16 seconds. Furthermore, compaction
> > allocated a high higher percentage of memory (91% of RAM as huge pages).
> >
> > The downside appears to be that the compaction kernel reclaimed even more
> > pages than the vanilla kernel. However, take the cut-off point of 880 pages
> > that both kernels succeeded. The vanilla kernel had reclaimed 105132 pages
> > at that point. The kernel with compaction had reclaimed 59071, less than
> > half of what the vanilla kernel reclaimed. i.e. the bulk of pages reclaimed
> > with the compaction kernel were to get from 87% of memory allocated to 91%
> > as huge pages.
> >
> > These results would appear to be an encouraging enough start.
> >
> > Comments?
> I think "Total pages reclaimed" increasing is not good thing ;)

First, I made a mistake in the patch. With the bug fixed, they're
reduced. See the post later in the thread

> Honestly, I haven't understand why your patch increase reclaimed and
> the exactly meaning of the your tool's rclm field.
> Can you share your mesurement script? May I run the same test?

Unfortunately at the moment it's part of a mini-testgrid setup I run out
of the house. It doesn't lend itself to being stand-alone. I'll break it
out as part of the next release.

> I like this patch, but I don't like increasing reclaim. I'd like to know
> this patch require any vmscan change and/or its change mitigate the issue.

With the bug repaired, reclaims go from 105132 to 45935 with more huge
pages allocated so right now, no special action is required.

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
Please read the FAQ at