Re: Poor memory context performance in large hash joins

On 2017-02-27 19:20:56 +0100, Tomas Vondra wrote:
> On 02/27/2017 12:55 PM, Andres Freund wrote:
> > On 2017-02-24 15:18:04 -0800, Andres Freund wrote:
> > > On 2017-02-24 15:12:37 -0800, Andres Freund wrote:
> > > > On 2017-02-24 18:04:18 -0500, Tom Lane wrote:
> > > > > Concretely, something like the attached.  This passes regression tests
> > > > > but I've not pushed on it any harder than that.
> > > >
> > > > Heh, I'd just gotten something that didn't immediately crash anymore ;)
> > > >
> > > > Running your patch against Jeff's test-case, verified before that I
> > > > could easily reproduce the O(N^2) cost.
> > >
> > > Oh, that didn't take as long as I was afraid (optimized/non-assert build):
> > >
> > > postgres[26268][1]=# SET work_mem = '13GB';
> > > SET
> > > Time: 2.591 ms
> > > postgres[26268][1]=# select count(*) from foobar2 where not exists (select 1 from foobar t where t.titleid=foobar2.titleid);
> > > Time: 268043.710 ms (04:28.044)
> >
> > As another datapoint, I measured this patch against the problem from
> > https://www.postgresql.org/message-id/20170227111732.vrx5v72ighehwpkf@alap3.anarazel.de
> > (see top post in thread), and it indeed fixes the runtime issue -
> > there's still considerably higher memory usage and some runtime
> > overhead, but the quadratic behaviour is gone.
> >
> > I think we should go forward with something like this patch in all
> > branches, and only use Tomas' patch in master, because they're
> > considerably larger.
> >
>
> So you've tried to switch hashjoin to the slab allocators? Or what have you
> compared?

No, sorry for not being more explicit about this.  Meant that Tom's
patch addresses the performance issue in the reorderbuffer.c to a good
degree (i.e. gets rid of the quadratic cost, even though constants are
higher than w/ your patches).  As the patch here is a lot smaller, it
seems like a better choice for the back-branches than backporting
slab.c/generation.c.

Andres