On Fri, Mar 20, 2015 at 10:02:23AM -0700, Linus Torvalds wrote:

On Thu, Mar 19, 2015 at 9:13 PM, Dave Chinner [off-list ref] wrote:

quoted

Testing now. It's a bit faster - three runs gave 7m35s, 7m20s and
7m36s. IOWs's a bit better, but not significantly. page migrations
are pretty much unchanged, too:

           558,632      migrate:mm_migrate_pages ( +-  6.38% )

Ok. That was kind of the expected thing.

I don't really know the NUMA fault rate limiting code, but one thing
that strikes me is that if it tries to balance the NUMA faults against
the *regular* faults, then maybe just the fact that we end up taking
more COW faults after a NUMA fault then means that the NUMA rate
limiting code now gets over-eager (because it sees all those extra
non-numa faults).

Mel, does that sound at all possible? I really have never looked at
the magic automatic rate handling..

It should not be trying to balance against regular faults as it has no
information on it. The trapping of additional faults to mark the PTE
writable will alter timing so it indirectly affects how many migration
faults there but this is only a side-effect IMO.

There is more overhead now due to losing the writable information and
that should be reduced so I tried a few approaches.  Ultimately, the one
that performed the best and was easiest to understand simply preserved
the writable bit across the protection update and page fault. I'll post
it later when I stick a changelog on it.

-- 
Mel Gorman
SUSE Labs