A tile-based parallel Viterbi algorithm for biological sequence alignment on GPU with CUDA

The Viterbi algorithm is the compute-intensive kernel in Hidden Markov Model (HMM) based sequence alignment applications. In this paper, we investigate extending several parallel methods, such as the wave-front and streaming methods for the Smith-Waterman algorithm, to achieve a significant speed-up on a GPU. The wave-front method can take advantage of the computing power of the GPU but it cannot handle long sequences because of the physical GPU memory limit. On the other hand, the streaming method can process long sequences but with increased overhead due to the increased data transmission between CPU and GPU. To further improve the performance on GPU, we propose a new tile-based parallel algorithm. We take advantage of the homological segments to divide long sequences into many short pieces and each piece pair (tile) can be fully held in the GPU's memory. By reorganizing the computational kernel of the Viterbi algorithm, the basic computing unit can be divided into two parts: independent and dependent parts. All of the independent parts are executed with a balanced load in an optimized coalesced memory-accessing manner, which significantly improves the Viterbi algorithm's performance on GPU. The experimental results show that our new tile-based parallel Viterbi algorithm can outperform the wave-front and the streaming methods. Especially for the long sequence alignment problem, the best performance of tile-based algorithm is on average about an order magnitude faster than the serial Viterbi algorithm.