A parallel BSP algorithm for irregular dynamic programming

Dynamic programming is a widely applied algorithm design technique in many areas such as computational biology and scientific computing. Typical applications using this technique are compute-intensive and suffer from long runtimes on sequential architectures. Therefore, several parallel algorithms for both fine-grained and coarse-grained architectures have been introduced. However, the commonly used data partitioning scheme can not be efficiently applied to irregular dynamic programming algorithms, i.e. dynamic programming algorithms with an uneven load density pattern. In this paper we present a tunable parallel Bulk Synchronous Parallel (BSP) algorithm for such kind of applications. This new algorithm can balance the workload among processors using a tunable block-cyclic data partitioning method and thus is capable of getting almost linear performance gains. We present a theoretical analysis and experimentally show that it leads to significant runtime savings for pairwise sequence alignment with general gap penalties using BSPonMPI on a PC cluster.