GPU Parallelism of Phylogenetic Likelihood Estimates for Protein Data

Abstract

The lack of a systematic and reliable approach for parallelizing the analysis of protein sequences impedes the effectiveness of using a wide range of phylogenetic analysis tools on GPUs. This paper attempts to bridge this gap by proposing a new parallelization approach. We study the impact of workloads, resource utilization, and variance in load-levels when calculating conditional likelihood probabilities. When then propose a more efficient method for parallelizing the phylogenetic analysis of protein sequence data on GPUs. In comparison with the serial version of MrBayes v3.1.2, implemented on a single CPU core, the proposed method, tgpMC3, achieves a peak speedup ratio of 117× by two NVIDIA Tesla K20 GPU cards on the Tianhe-1A supercomputer's GPU nodes. In comparison with the taMC3 method, another state-of-the-art GPU method for the analysis of protein sequences, the proposed method outperforms it by speedup factors ranging from 2.2 to 2.6×. The experimental results suggest that our large-scale phylogenetic analysis has significant implications for future research on high performance phylogenetic analysis.