On Tue, 13 Jul 2021 at 15:15, David Rowley <dgrowleyml@gmail.com> wrote:
> In theory, we likely could get rid of the small regression by having
> two versions of ExecSort() and setting the correct one during
> ExecInitSort() by setting the function pointer to the version we want
> to use in sortstate->ss.ps.ExecProcNode.
Just to see how it would perform, I tried what I mentioned above. I've
included what I ended up with in the attached POC patch.
I got the following results on my AMD hardware.
Test master v8 patch comparison
Test1 448.0 671.7 149.9%
Test2 316.4 317.5 100.3%
Test3 299.5 381.6 127.4%
Test4 219.7 229.5 104.5%
Test5 226.3 254.6 112.5%
Test6 197.9 217.9 110.1%
Test7 179.2 185.3 103.4%
Test8 389.2 544.8 140.0%
This time I saw no regression on tests 2, 4 and 7.
I looked to see if there was anywhere else in the executor that
conditionally uses a different exec function in this way and found
nothing, so I'm not too sure if it's a good idea to start doing this.
It would be good to get a 2nd opinion about this idea. Also, more
benchmark results with v6 and v8 would be good too.
David