Performance Analysis and Comparison of Three Mrbayes Computational Biology Code on Tianhe-1A Supercomputer

As a popular computational biology program for Bayesian phylogenetic inference, Mr Bayes is extremely computationally expensive with large scale datasets, which has to be implemented on a lot of parallel computers, especially HPCs to meet the computational requirement. Currently, there are three representative MrBayes code using MPI, MPI+OpenMP and MPI+CUDA programming language respectively. However, these three versions have not yet been analyzed on a computer in the order of thousands of nodes with tens of thousands of CPU cores and thousands of GPUs. In this paper, in order to compare and analyze the performance of the three MrBayes code on a HPC, we choose Tian He-1A heterogeneous multi-core supercomputer as our experimental platform. Given large scale input dataset, we use execution time and speedup as criterions to analyze the results from the aspects of performance on single node, strong scalability and weak scalability. The results of this research show that the three versions all have good parallelism and the MPI+OpenMP version achieves the best performance using the same number of compute nodes as the other two versions. Due to exploiting multiple CPU cores not sufficiently in heterogeneous system, the performance of current MPI+CUDA version is weaker than MPI+OpenMP version. In terms of scalability, performance improved a little on a system up on thousands of nodes, attributing to the increasing of communication overhead when it is scaled to large scale system.