Server selection with arbitrary distribution

Many applications need to pick servers with some desired distribution. For example, in probabilistic quorum systems, one method to generate quorums with high probability of intersection is to randomly pick k/spl radic/n nodes with a fixed probability distribution. Load balancing applications may need to take several samples of the servers with some desired distribution. Existing approaches realize a fixed stationary distribution by controlling the topology of the overlay graph and conducting random walks on it. In particular, existing approaches focus on achieving a uniform distribution. This paper proposes using the distributed Hastings-Metropolis algorithm to achieve any desired stationary distribution without control or global knowledge of the overlay graph. The new method facilitates good load balancing, since heterogeneous server capacity or other factors can be considered in deciding the appropriate distribution by which to pick servers.