Accelerating the Adaptive Tempering Monte Carlo Method with CUDA Graphics Processing Units

Abstract

Molecular Dynamics (MD) has been and continues to be a popular method of molecular simulation because it is easily parallelizable. Parallel programming has become less burdensome for the science community, and competition in MD algorithm development has given MD avant-garde positions in molecular, bio-systems, materials, and nano-systems simulation. In contrast, inherently serial Monte Carlo (MC) methods have been largely ignored in the recent advancements of parallel computing technology. The trend exists even though MC methods based on statistical mechanics principles are superior for studying thermodynamics properties such as entropy and free energy. In my dissertation I present a means of parallelizing MC molecular simulation such that in time the popularity of MC may be restored to that of MD. The Adaptive Tempering Monte Carlo method (ATMC) employs the Metropolis MC (MMC) sampling criterion; therefore, both ATMC and MMC are inherently serial algorithms. ATMC is a multicanonical ensemble algorithm that optimizes system configuration by searching for the most ordered state. This algorithm was developed by Dong and Blaisten-Barojas in 2006. My algorithm accelerates ATMC and MMC in a novel implementation exploiting state of the art parallel processing technology, namely NVIDIA® Compute Unified Device Architecture (CUDA) Graphics Processing Units (GPUs). My implementation source code is written in CUDA C, NVIDIA's extension to the C programming language for parallel programming, and summarily compiled by NVCC, NVIDIA's CUDA version 4.0 C compiler. My CUDA GPU-accelerated implementation is verified against a 2010 study by Dai and Blaisten-Barojas of pyrrole oligomers (specifically, 12-Py chains), an interesting material for its application in artificial muscles, actuators, chemical remediation, among others. This previous study put forward a partially coarse-grained model potential for reduced pyrrole oligomers at the polypyrrole experimental density. I introduced a revision to this potential model apropos for condensed phases of oligopyrroles. Verification includes comparison of total potential energy, intra-oligomer energy, inter-oligomer energy, end-to-end distance, radius of gyration, and two order parameters that characterize the chain ordering in the condensed phase. Bending and dihedral angles are also examined. In addition, I performed a benchmark of my accelerated algorithms that show a speed-up factor greater than 60 with respect to the implementation in CPU. This extremely fast implementation is reached for systems larger than about 250,000 pyrrole monomers. Speed-ups in this range are unique in the published literature. A journal article is in preparation to report this achievement. My novel accelerated implementation has already been applied in a study of oxidized oligopyrrole. A contributed presentation was presented at the American Physical Society March Meeting in Baltimore, March 2013 and is soon to be published in a physical chemistry journal.