A lightweight BLASTP and its implementation on CUDA GPUs

The BLAST server in the National Center for Biotechnology Information in the USA receives tens of thousands of queries per day on average. However, the service is always the same for every query even though query lengths vary significantly. In fact, the lengths of a large portion of protein sequences are less than 500. On the other hand, the hit detection process consumes the most of the execution time of BLAST and its core architecture is a lookup table. Following the above reasons, we propose a lightweight BLASTP for servicing not-too-long queries, where a hybrid query-index table is proposed accordingly. Each table entry consists of four bytes that can store up to three query positions. Therefore, a sequence word usually requires only one memory fetch to retrieve its hit information. Furthermore, additional dummy entries are embedded into the table and interleaved with original entries. The entries without any hits and dummy entries both can be used to buffer spilled query positions. The above features result in a much smaller lookup table with a higher utilization rate and a lower cache miss ratio. Experimental results show that the lightweight BLASTP outperforms CUDA-BLASTP with speedups ranging from 1.82 to 3.37 based on the first two critical phases.