Performance Engineering of Directional Message-Passing Algorithms Through a Stencil-Based Approach for Applications in Molecular Dynamics

Abstract

The molecular dynamics method, used by scientists across the fields of physics, materials science, and biology, is an increasingly popular way to simulate particle interactions. Current implementations of molecular dynamics simulators can verify macromolecular structures, examine atomic-level phenomena that cannot be observed directly, and predict the behavior of unstudied proteins. The existing implementations, however, rely on inefficient directional message-passing algorithms on graph neural networks. This thesis presents a novel approach for the optimization of these algorithms using a stencil-like technique. The stencil-based algorithm, called StencilMD, provides both the benefits of parallelization and improved cache locality. The results show that StencilMD successfully reduces the amount of time required to run a molecular dynamics simulation by as much as 28.57% with a corresponding 26.92% decrease in cache misses.