Data dependency reduction in Dynamic Programming matrix

Dynamic Programming (DP) plays an important role in solving a large number of computational problems. As the number of cores per processor is increasing rapidly, new software must be capable of exploiting the advantages of multicore architectures. A typical DP begins with constructing a matrix, and then calculating each element one by one. The standard parallelization spawns multiple threads, one for each row, while maintains the data dependency via thread synchronization. However, as the number of cores increase, the performance degrades due to data dependency. Herein, we propose a novel method for calculating a DP matrix in parallel. In contrast to the standard method that always calculates from up to down and left to right, our method performs the calculation in multiple directions. Therefore, the wait time for data dependency is remarkably reduced. To demonstrate our method, a local sequence alignment algorithm called Smith-Waterman (SW) was chosen as an instance of DP. However, our method is not only limited to SW algorithm, but it is applicable to other DP problems that have the similar patterns of data dependency. A comparison with the standard method was conducted on HP Z800 Workstation equipped with two quad-core processors. The results show that our method performs significantly faster than the classical approach.