Hi,
On 2021-04-25 01:27:24 +0300, Yura Sokolov wrote:
> It is quite interesting result. Simplehash being open-addressing with
> linear probing is friendly for cpu cache. I'd recommend to define
> SH_FILLFACTOR with value lower than default (0.9). I believe 0.75 is
> suitable most for such kind of hash table.
It's not a "plain" linear probing hash table (although it is on the lookup
side). During insertions collisions are reordered so that the average distance
from the "optimal" position is ~even ("robin hood hashing"). That allows a
higher load factor than a plain linear probed hash table (for which IIRC
there's data to show 0.75 to be a good default load factor).
There of course may still be a benefit in lowering the load factor, but I'd
not start there.
David's test aren't really suited to benchmarking the load factor, but to me
the stats he showed didn't highlight a need to lower the load factor. Lowering
the fill factor does influence the cache hit ratio...
Greetings,
Andres Freund
