Re: [PATCH] Use optimized single-datum tuplesort in ExecSort

Em qua., 14 de jul. de 2021 às 21:21, David Rowley <dgrowleyml@gmail.com>
escreveu:

> On Thu, 15 Jul 2021 at 12:10, Ranier Vilela <ranier.vf@gmail.com> wrote:
> >
> > Em qua., 14 de jul. de 2021 às 20:43, David Rowley <dgrowleyml@gmail.com>
> escreveu:
> >>
> >> On Thu, 15 Jul 2021 at 05:55, Ranier Vilela <ranier.vf@gmail.com>
> wrote:
> >> > I repeated (3 times) the benchmark with v8 here,
> >> > and the results were not good.
> >>
> >> Do you have any good theories on why the additional branching that's
> >> done in v7b vs v8 might cause it to run faster?
> >
> >
> > Branch Predictions works with *more* probable path,
> > otherwise a penalty occurs and the cpu must revert the results.
>
> But, in v8 there is no additional branch, so no branch to mispredict.
> I don't really see how your explanation fits.
>
In v8 the branch occurs at :
+ if (ExecGetResultType(outerPlanState(sortstate))->natts == 1)

datumSort is tested first.

Cpu time is a more expensive resource.
Always is executed two branches, if it is right path, win,
otherwise occurs a penalty time.


> It seems much more likely to me that the results were just noisy.  It
> would be good to see if you can recreate them consistently.
>
I do.
Can you please share results with v7b?

regards,
Ranier Vilela