On 2016-04-05 12:14:35 -0400, Robert Haas wrote:
> On Tue, Apr 5, 2016 at 11:30 AM, Andres Freund <andres@anarazel.de> wrote:
> > On 2016-04-05 20:56:31 +0530, Amit Kapila wrote:
> >> This fluctuation started appearing after commit 6150a1b0 which we have
> >> discussed in another thread [1] and a colleague of mine is working on to
> >> write a patch to try to revert it on current HEAD and then see the results.
> >
> > I don't see what that buys us. That commit is a good win on x86...
>
> Maybe. But I wouldn't be surprised to find out that that is an
> overgeneralization. Based on some results Mithun Cy showed me this
> morning, I think that some of this enormous run-to-run fluctuation
> that we're seeing is due to NUMA effects. So some runs we get two
> things that are frequently accessed together on the same NUMA node and
> other times they get placed on different NUMA nodes and then
> everything sucks. I don't think we fully understand what's going on
> here yet - and I think we're committing changes in this area awfully
> quickly - but I see no reason to believe that x86 is immune to such
> effects. They may just happen in different scenarios than what we see
> on POWER.
I'm not really following - we were talking about 6150a1b0 ("Move buffer
I/O and content LWLocks out of the main tranche.") made four months
ago. Afaics the atomic buffer pin patch is a pretty clear win on both
ppc and x86?
I agree that there's numa effects we don't understand. I think
re-considering Kevin's patch that did explicit numa hinting on linux
would be rather worthwhile. It makes a lot of sense to force shmem to be
force-interleaved, but make backend local memory, uh, local.
Greetings,
Andres Freund
