Moleclar-dynamics simulations using spatial decomposition and task-based parallelism

Abstract

Molecular Dynamics (MD) simulations are an integral method in the computational studies of materials. This thesis discusses an algorithm for large-scale MD simulations using modern multiand many-core systems on distributed computing networks. In order to utilize the full processing power of these systems, algorithms must be updated to account for newer hardware, such as the many-core Intel Xeon Phi co-processor. The hybrid method is a data-parallel method of parallelization which combines spatial decomposition using the Message Passing Interface (MPI) to distribute the system onto multiple nodes, along with the cell-task method used for task based parallelism on each node. This allows for the improved performance of task based parallelism on single compute nodes in addition to the benefit of distributed computing allowed by MPI. Results from benchmark simulations on Intel Xeon multi-core processors, and Intel Xeon Phi coprocessors are presented. Results show that the hybrid method provides better performance than either spatial decomposition or cell-task methods alone on single nodes, and that the hybrid method outperforms the spatial decomposition method on multiple nodes, on a variety of system configurations.