Re: Deferring some AtStart* allocations?

On 2014-10-09 15:01:19 -0400, Robert Haas wrote:
> On Thu, Oct 9, 2014 at 8:20 AM, Andres Freund <andres(at)2ndquadrant(dot)com> wrote:
> > On 2014-10-09 08:18:18 -0400, Robert Haas wrote:
> >> On Thu, Oct 9, 2014 at 5:34 AM, Andres Freund <andres(at)2ndquadrant(dot)com> wrote:
> >> > Interesting - in my local profile AtStart_Inval() is more pronounced
> >> > than AfterTriggerBeginQuery(). I've quickly and in a ugly fashion hacked
> >> > AtStart_Inval() out of readonly queries ontop of your patch. Together
> >> > that yields a ~3.5% performance improvement in my trivial 'SELECT * FROM
> >> > tbl WHER pkey = xxx' testcase.
> >>
> >> Whoa. Now that's clearly significant.
> >
> > Well, my guess it'll be far less noticeable in less trivial
> > workloads. But it does seem worthwile.
> >
> >> You didn't attach the patch; was that inadvertent, or was it too ugly
> >> for that?
> >
> > Far, far too ugly ;). I just removed the AtStart() call from xact.c and
> > sprinkled it around relevant places instead ;)
>
> OK, here's an attempt at a real patch for that. I haven't perf-tested this.

Neato. With a really trivial SELECT:

before:
tps = 28150.794776 (excluding connections establishing)
after:
tps = 29978.767703 (excluding connections establishing)

slightly more meaningful:

before:
tps = 21272.400039 (including connections establishing)
after:
tps = 22290.703482 (excluding connections establishing)

So that's a noticeable win. Obviously it's going to be less for more
complicated stuff, but still...

I've not really looked at the patches though.

Greetings,

Andres Freund

--
Andres Freund http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services