Re: Using POPCNT and other advanced bit manipulation instructions

On Thu, 20 Dec 2018 at 20:17, Gavin Flower
<GavinFlower(at)archidevsys(dot)co(dot)nz> wrote:
> Looking at the normalized standard deviations, the patched results have
> a higher than 5% chance of being better simply by chance. I suspect
> that you have made an improvement, but the statistics are not convincing.

Yeah, I'd hoped that I could have gotten a better signal to noise
ratio by running the test many times, but you're right. That was on
my laptop. I've run the test again on an AWS instance and the results
seem to be a bit more stable. Same table with 1 int column and 100m
rows. statistics set to 10.

Unpatched

postgres=# analyze a;

Time: 38.248 ms
Time: 35.185 ms
Time: 35.067 ms
Time: 34.879 ms
Time: 34.816 ms
Time: 34.558 ms
Time: 34.722 ms
Time: 34.427 ms
Time: 34.214 ms
Time: 34.301 ms
Time: 35.751 ms
Time: 33.993 ms
Time: 33.880 ms
Time: 33.617 ms
Time: 33.381 ms
Time: 33.326 ms

Patched:

postgres=# analyze a;

Time: 34.421 ms
Time: 33.523 ms
Time: 33.230 ms
Time: 33.678 ms
Time: 32.987 ms
Time: 32.914 ms
Time: 33.165 ms
Time: 32.707 ms
Time: 32.645 ms
Time: 32.814 ms
Time: 32.082 ms
Time: 32.143 ms
Time: 32.310 ms
Time: 31.966 ms
Time: 31.702 ms
Time: 32.089 ms

Avg +5.72%, Median +5.29%

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