Efficient implementation of MrBayes on multi-GPU.

Area of Science:

• Computational Biology
• Bioinformatics
• High-Performance Computing

Background:

• Bayesian inference using Metropolis-coupled Markov chain Monte Carlo (MCMCMC) is crucial for phylogenetic analysis.
• Existing MCMCMC algorithms struggle with the computational demands of massive real-world DNA datasets.
• Graphics Processing Units (GPUs) offer significant acceleration potential for complex computations.

Purpose of the Study:

• To develop an efficient GPU-accelerated implementation of MCMCMC for the MrBayes program.
• To enhance the speed and scalability of Bayesian phylogenetic inference for large DNA sequence data.
• To leverage the parallel processing capabilities of GPUs for faster analysis.

Main Methods:

• Implementation of an adaptive MCMCMC (aMCMCMC) algorithm on Compute Unified Device Architecture (CUDA).
• Dynamic adjustment of task granularity to optimize GPU core utilization based on data size and hardware.
• Development of adaptive sequence splitting/combining and node-by-node task scheduling for improved concurrency and reduced overhead.

Main Results:

• Achieved up to 63x speedup on a single GPU, 170x with four GPUs, and 478x with a 32-node GPU cluster compared to serial MrBayes.
• Demonstrated significant performance gains over previous MCMCMC algorithms.
• Showcased excellent scalability on large GPU clusters.

Conclusions:

• The aMCMCMC implementation dramatically accelerates Bayesian phylogenetic inference.
• This GPU-based approach effectively handles massive DNA datasets, overcoming previous computational limitations.
• The method offers a scalable and efficient solution for modern phylogenetic research.