Accelerating Phylogenetic Inference on GPUs: an OpenACC and CUDA comparison

Phylogenetic inference is used to derive a "tree of life" for a collection of species whose DNA sequences are known. While several software packages have already been developed to take advantage of GPUs to accelerate phylogenetic inference, they typically require significant changes to the original code, constraining code maintenance. Recently, the OpenACC API was proposed to minimize the programming efforts on accelerator devices. In this work we evaluate the applicability of the OpenACC API for phylogenetic inference using the most recent MrBayes program (version 3.2.2). A new parallelization strategy is proposed that is specifically adapted to the latest version of MrBayes and minimizes the data transfers between the host (CPU) and the accelerating device (GPU). We further implement the proposed strategy using both the OpenACC and CUDA programming frameworks. Experimental results demonstrate that significant performance gains can be achieved using OpenACC with a reduced amount of programming effort. Comparing with state-of-art GPU's implementations, the proposed OpenACC and CUDA implementations achieve a performance gain of up to 5.2x and 5.7x, respectively. Experimental results indicate that with a reduced amount of programming effort, we achieve a performance that is only 10% inferior to one obtained with CUDA, which uses device specific optimizations.