On Mon, 2012-01-30 at 09:59 +1100, Bojan Smojver wrote: > With my latest patch, I get that fixed (i.e. the final hash is > perturbed a lot better).
Hmm, different problems emerge. The hash then has a tendency to produce either all odd or even indexes (i.e. lower bits used to address buckets). I think this whole idea of hashing the hash is costing us in loss of randomness. Which makes sense. We just hashed a string and reduced the number of its bits significantly. Then we took that (which is also always the same length) and reduced it again. In the process, we overcooked. I think I'll need to work on a different approach. -- Bojan
