MrBayes for Phylogenetic Inference Using Protein Data on a GPU Cluster.

MrBayes is a widely used software for Bayesian phylogenetic inference: we input biological sequence data from various taxonomic groups, and MrBayes returns its estimate of the phylogenetic tree which gave rise to those taxa. This paper presents ta(MC)(^{3}), based on its predecessor a(MC)(^{3}), which, for protein datasets, improves computational efficiency and overcomes major obstacles in analyzing larger datasets on HPCs with multiple Graphics Processing Units (GPUs). The major improvements are (a) a new task mapping strategy, (b) the use of Kahan summation to resolve non-convergence issues, and (c) the introduction of 64-bit variables. We evaluate ta(MC)(^{3}) on real-world protein datasets both on a desktop server and the Tianhe-1A supercomputer. With a single GPU, ta(MC)(^{3}) is nearly 90 times faster compared with the serial version of MrBayes, up to around 9 times faster than MrBayes utilizing a GPU via the BEAGLE library, and up to 2.5 times faster than a(MC)(^{3}). On larger datasets with 64 nodes (GPUs) on Tianhe-1A, ta(MC)(^{3}) is capable of obtaining (1000+) speedup vs. serial MrBayes.