Parallel Suffix Array Construction for Shared Memory Architectures

We present the design of the algorithm for constructing the suffix array of a string using manycore GPUs. Despite of the wide usage in text processing and extensive research over two decades there was a lack of efficient algorithms that were able to exploit shared memory parallelism (as multicore CPUs as manycore GPUs) in practice. To the best of our knowledge we developed the first approach exposing shared memory parallelism that significantly outperforms the state-of-the-art existing implementations for sufficiently large inputs. We reduced the suffix array construction problem to a number of parallel primitives such as prefix-sum, radix sorting, random gather and scatter from/to the memory. Thus, the performance of the algorithm merely depends on the performance of these primitives on the particular shared memory architecture. We demonstrate its performance on manycore GPUs, but the method can also be applied for other parallel architectures, such as multicores, CELL or Intel MIC.