Improving Performance of Genomic Aligners on Intel Xeon Phi-Based Architectures

Abstract

Due to the drastic increase in genome sequence data from the GB level to the TB level, a series of powerful computing accelerators depending on manycore architecture have been proposed to cope with this data challenge. GPU and Xeon Phi are the most prevalent accelerators in high-performance computing as indicated by the top 500 supercomputer list. Intel Xeon Phi is a convenient and an attractive accelerator as it is based on the x86 architecture and can readily execute applications that are originally written for x86-based processors without much coding effort. In this paper, we focus on the sequence alignment problem which is one of the most time-consuming steps in the analysis of genomic variants. Our work is based on one of the most widespread alignment tools, BWA aligner. We studied different solutions to efficiently execute BWA aligner on a heterogeneous system that consists of multicore processors and manycore accelerators. We proposed a multi-level strategy (MDPR) based on data parallelization and data replication. The MDPR strategy was executed using several hierarchical levels of parallelism in systems that contain multicore processors (Xeon) and manycore accelerators (Xeon Phi). Through a multi-process and multi-thread scheme which comprises replication and parallelization of certain data structures, we had obtained substantial improvements in BWA execution time. The strategy runs in symmetric mode using a hierarchical scheme that improves memory locality and reduces memory congestion on NUMA architectures as well as on Xeon Phi processors. Although our proposal was applied to BWA, this strategy can be easily extrapolated to other sequence alignment tools that have similar operating principles with those of BWA aligner. Significant improvements were obtained in our experiments, comparing our proposal to other strategies that are executed in native, offload and symmetric mode.