21/12/2017 18:26, Fabien COELHO пишет:
>
>> I think it is not commitfest ready yet -- I need to add some
>> documentation and tests first.
>
> Yes, doc & test are missing.
>
> From your figures, the murmur2 algorithm output looks way better. I'm
> wondering whether it makes sense to provide a bad hash function if a
> good/better one is available, unless the bad one actually appears in
> some benchmark... So I would suggest to remove fnv1a.
Actually the "bad" one appears in YCSB. But if we should choose the only
one I would stick to murmur too given it provides better results while
having similar computational complexity.
>
> One implementation put constants in defines, the other one uses "const
> int". The practice in pgbench seems to use defines (eg
> MIN_GAUSSIAN_PARAM...), so I would suggest to stick to this style.
That had been my intention from the start until I coded it that way and
it looked ugly and hard to read (IMHO), like:
k *= MURMUR2_M;
k ^= k >> MURMUR2_R;
k *= MURMUR2_M;
result ^= k;
result *= MURMUR2_M;
...etc. And besides it is not a parameter to be tuned and not something
someone would ever want to change; it appears in just a single function.
So I'd better leave it the way it is. Actually I was thinking to do the
same to fnv1a too : )
--
Ildar Musin
Postgres Professional: http://www.postgrespro.com
Russian Postgres Company