Re:Short term Redis plans
by pork on Monday November 07, 2011 @05:52PM
(#3208551)
Just curious about the key distribution algorithm -- why use CRC-16(key) % 4096 instead of a fast hash function like djb2, which almost certainly has better key equidistribution?
Re:Short term Redis plans
by antirez on Tuesday November 08, 2011 @03:13PM
(#3212574)
Actually CRC-16 has a much better distribution than djb2 and other algorithms for this use case.
CRC-16 is a remarkably good algorithm for certain kind of hashing, and has a few very interesting qualities being the reminder of a polynominal division. For instance if you hash two bytes data you are guaranteed to have different hashes for every different pair of bytes without collisions.
CRC-16 performed very well in tests also with the very common case of keys with the same prefix, like object:0, object:1, and so forth.
So CRC-16 is very good for our use case and has the advantage of being simple and fast.
http://www.cse.yorku.ca/~oz/hash.html
CRC-16 performed very well in tests also with the very common case of keys with the same prefix, like object:0, object:1, and so forth.
So CRC-16 is very good for our use case and has the advantage of being simple and fast.