On Thu, Jan 21, 2021 at 5:14 AM Heikki Linnakangas <hlinnaka@iki.fi> wrote:
> Here you can see that as numsnaps increases, the test becomes slower,
> but then it becomes faster again at 64-66, when it switches to the hash
> table. So 64 seems too much. 32 seems to be the sweet spot here, that's
> where scanning the hash and scanning the array are about the same speed.
That sounds good. I mean, it could be that not all hardware behaves
the same here. But trying to get it into the right ballpark makes
sense.
I also like the fact that this now has some cases where it wins by a
significant margin. That's pretty cool; thanks for working on it!
--
Robert Haas
EDB: http://www.enterprisedb.com