[Python-3000] Performance Notes - new hash algorithm
Larry Hastings
larry at hastings.org
Sun Sep 9 04:24:47 CEST 2007
If the Python community is just noticing the Hsieh hash, that implies
that the Bob Jenkins hashes are probably unknown as well. Behold:
http://burtleburtle.net/bob/hash/doobs.html
To save you a little head-scratching, the functions you want to play
with are hashlittle()/hashlittle2() in "lookup3.c":
http://burtleburtle.net/bob/c/lookup3.c
hashlittle() returns a 32-bit hash; hashlittle2() returns two 32-bit
hashes on the same input (in effect a 64-bit hash). The "little"
implies that the function is better on little-endian machines. (There
is a hashbig(); no hashbig2(), it is left as an exercise for the reader.)
In our testing (at Facebook, for memcached) hashlittle2 was faster than
the Hsieh hash; that was done a year ago (and before I joined) so I
don't have numbers for you.
One goal of Jenkin's hashes is uniform distribution, so these functions
presumably lack the serendipitous "similar inputs hash to similar
values" behavior of Python's current hash function. But why is that a
feature? (Not that I doubt Tim Peters!)
Oh, and, all the Jenkins code is public domain.
Cheers,
/larry/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.python.org/pipermail/python-3000/attachments/20070908/a49272fd/attachment.htm
More information about the Python-3000
mailing list