Scalable Self-Tuning Implementation of Smith-Waterman Algorithm for Multicore CPUs

Improved version of the Smith-Waterman algorithm (SWA) is most widely used for local alignment of a pattern (or query) sequence with a Database (DB) sequence. This dynamicprogramming algorithm is computation intensive. To reduce time for computing alignment score matrix, parallel versions have been implemented on GPUs and multicore CPUs. These parallel versions have shown significant speedup when compared with their corresponding sequential versions. Our initial evaluation of an OpenMP parallelization of SWA has shown linear speedup on multicore CPUs, but a closer look at performance data from both sequential and parallel versions have revealed two undesired effects: (i) As the length of the DB sequence increases, the number of elements of the alignment score matrix H computed in per unit time initially increases, then reaches a maximum, and finally decreases continuously; (ii) the length of the DB sequence where decline starts is different for different CPUs. To overcome the computation rate decline we have proposed a run-time self-tuning algorithm. It determines the length, l, of a DB sequence that maximize computation rate during execution time. Then, divide computation of H into computation of a set of submatrices, such that the number of columns in each submatrix is about l. Our study also found that the number of per-core-threads that delivers the highest rate of computation differs from CPU to CPU. Our proposed algorithm determines optimal number of threads during execution time and creates optimal number of threads for highest possible computation rate. Our extensive evaluations of the proposed self-tuning algorithm on three different multicore multi-CPU shared memory machines have shown significant performance improvement.