On Sat, 2009-01-10 at 13:36 -0500, Tom Lane wrote:
> Jeff Davis <pgsql@j-davis.com> writes:
> > I ran 5 times on both old and new code, eliminating the top and bottom
> > and taking the average of the remaining 3, and I got a 6.9% performance
> > improvement with the new code.
>
> The question that has been carefully evaded throughout the discussion
> of this patch is whether the randomness of the hash result is decreased,
> and if so what is that likely to cost us in performance of the various
> hash-dependent algorithms. I would still like to have an answer to that
> before we make a change to gain marginal performance improvement in
> the hash function itself (which is already shown to be barely measurable
> in the total context of a hash-dependent operation...)
>
In:
http://archives.postgresql.org/message-id/20081104202655.GP18362@it.is.rice.edu
Ken showed that the number of hash collisions is the same in the old and
the new code for his test data. Not a direct measurement of randomness,
but it's some kind of indication.
I'm not an authority on either hash functions or statistics, so I'll
have to defer to Ken or someone else to actually prove that the
randomness is maintained.
Regards,Jeff Davis