Better server selection client code?

[Andrew Harbick]

> Even I had suggested a deterministic way of picking a point, which was
> based on the largest available free arc on the unit circle. But as
> Andrew added, this will distribute load uniform only when the number of
> servers are a power of 2.
> (http://lists.danga.com/pipermail/memcached/2006-October/002940.html )

Right, what my client didn't do at the time was to place 
(deterministically) any one server in multiple places on the unit 
circle.  I was placing each server in only one place on the unit circle.

Imagine using a bunch of different hashes to put one server at different 
spots on the circle, and then repeat that process with other servers. 
You'll get an even distribution of load and still be able to have only a 
few servers.