FMapper: Scalable read mapper based on succinct hash index on SunWay TaihuLight

One of the most important application in bioinformatics is read mapping. With the rapidly increasing number of reads produced by next-generation sequencing (NGS) technology, there is a need for fast and efficient high-throughput read mappers. In this paper, we present FMapper - a highly scalable read mapper on the TaihuLight supercomputer optimized for its fourth-generation ShenWei many-core architecture (SW26010). In order to fully exploit the computational power of the SW26010, we employ dynamic scheduling of tasks, asynchronous I/O and data transfers and implement a vectorized version of the banded Myers algorithm tailored to the 256 bit vector registers of the SW26010. Our performance evaluation demonstrates that FMapper using all 4 compute groups of a single SW26010 processor outperforms S-Aligner on the same hardware as well as RazerS3, Hobbes3, Minimap2 and BWA running on a 4-core Xeon W-2123v3 CPU and achieves speedups of 4.7, 24.8, 2.4, 4.6 and 14.7 respectively. Using several optimizations, we achieve a speedup of 6 compared to the naive implementation on one compute group of an SW26010 processor and a strong scaling efficiency of 65% on 512 compute groups. (c) 2021 Published by Elsevier Inc.

Automated summary

This paper introduces FMapper, a highly scalable read mapper optimized for the SW26010 many-core architecture on the TaihuLight supercomputer. By implementing dynamic task scheduling, asynchronous I/O and data transfers, and a vectorized version of the banded Myers algorithm tailored to the SW26010 256 bit vector registers, FMapper outperforms other read mappers in performance evaluation.