A distributed algorithm for the parsimony bootstrap approximation

Accelerating phylogenetic tree reconstruction and bootstrapping is critical, especially to support the study of the evolution of dangerous viruses. In this paper, we propose the MPBoot-MPI, a distributed algorithm efficiently implementing the idea of bootstrap approximation in MPBoot for a parallel computing environment of multiple computational nodes. MPBoot-MPI employs the master-worker paradigm and divides the work in MPBoot into three phases, each with a separate strategy to distribute computing among processes. Since the bootstrap trees are not calculated independently, processes must share results throughout task execution. We propose that when arriving at a checkpoint to report their results to the master process, worker processes apply a stochastic strategy to determine whether to perform the sending, thereby reducing the effect of latency caused by the large size of the message sent. Experiments on simulation and real benchmark datasets showed that MPBoot- MPI on multiple processes obtained MP scores and bootstrap accuracy comparable to MPBoot while achieving a promising speedup ratio. We implemented the proposed method in the MPBoot-MPI program, which is publicly accessible at https://github.com/diepthihoang/mpboot/tree/mpboot-mpi-dev.