Using GPU to accelerate suffix array construction

Suffix array is a simple and powerful alternative to suffix trees in genome analysis, string matching and text compression. Many efforts have been made to construct suffix array in a faster way. We propose a new suffix array construction algorithm fit for data parallel computing by replacing group with sort and describe the implementation issues on CUDA platform. The experiments show that the parallel implementation is simple, fast and scalable for real-world data processing, especially for small alphabet text. The most interesting trait of the parallel code is that the intermediate results generated by GPU can be processed further by serial prefix-doubling algorithms on CPU, which significantly improves the performance of the serial algorithms. The hybrid of parallel and serial codes reaches the peak performance, which reveals the great potential of GPU for high-performance computing in practical applications.