> I've now pushed that bug fix so it's fine to remove the change to
> tuplesort.c now.
Thanks, I've rebased the patch, please find attached the v6.
>
> I also did a round of benchmarking on this patch using the attached
> script. Anyone wanting to run it will need to run make installcheck
> first to create the required tables.
I've run your benchmark, keeping the best of three runs each time.
This is an intel laptop, so as many things are running on it there is a lot of
noise...
Both standard and patched run come from a compilation with gcc -O2. No changes
have been done to the default settings.
Query # Master Patched Variation
1 884 1627 184.05%
2 364 375 103.02%
3 568 783 137.85%
4 296 297 100.34%
5 421 484 114.96%
6 359 408 113.65%
7 237 251 105.91%
8 806 1271 157.69%
Since I didn't reproduce your slowdown at all on the first run, I tried to
rerun the benchmark several times and for the "dubious cases" (2, 4 and 7),
the results are too jittery to conclude one way or another in my case. I
don't have access to proper hardware, so not sure if that would be useful in
any way to just run the bench for thousands of xacts instead. I would be
surprised the check adds that much to the whole execution though.
I attach a graph similar to yours for reference.
--
Ronan Dunklau
