High Performance Computing Techniques for Metropolis Monte Carlo Simulations of Oxidized Polypyrrole

Abstract

The dissertation research aims to develop and improve the high performance computingtechniques related to the parallelization of the Metropolis Monte Carlo Method (MMC). The MMC is inherently serial due to its basic nature that requires sequential evaluations of properties at the current iteration before moving to the next iteration. As a result of this constraint, the MMC algorithm parallelization is challenging. Additionally, the MMC applied to systems that have pairwise and three-body interactions between the N atoms is a computationally intense problem with a processing time growth rate on the order of O (N3). Due to the serial character of the MMC algorithm, effort is usually spent in the optimization of the internals of a single iteration. My research has focused on two major goals: (i) optimization and parallelization of theMMC algorithm and (ii) optimization and parallelization of the implementation as applied to the chemically oxidized polypyrrole (PPy) system in condensed phases modeled with a coarse grained potential for prediction of its structural and thermodynamic properties. The proposed PPy model potential is well suited for parallelization and optimization because most the interactions between different polymer chains and elemental dopants are described by a sum of pairwise additive functions which are good candidates for parallelization. My efforts narrowed down on developing a scalable technique for the parallelization of the MMC across multiple compute nodes utilizing CPU and GPU processing. My implementation provides significant speedup over traditional approaches with the GPU enabled version being 342 times faster than the CPU alone version. Overall, my implementation combines the optimized MMC simulations with a parametricsearch of the constants entering in the coarse grained modeling of polymers. Novel to this field is an automated combination of parameters determination that hops between quantum mechanics analysis done by my colleague and MMC calculation of critical thermodynamic properties of the system that should match experimental results. A result is a workflow that employs delocalized hardware that work as a team producing a compendium of system properties simultaneously. The targeted condensed system properties of the polymer include density, enthalpy, potential energy, cohesive energy, thermal expansion coefficient, Hildebrand parameter, radial distribution functions, bulk modulus, sonic velocity, polymer radius of gyration and end-to-end distances, orientational and vector order parameters, among others. Hence, my implementation brings large scale simulation abilities to desktops, laptops and legacy computing hardware in addition to the a more efficient employment of computer clusters such as Argo and Hopper. The resource pool of available computing hardware to perform research and analysis of polymer systems such as PPy containing hundreds of thousands particles is significantly expanded.