Performance optimization of reactive molecular dynamics simulations with dynamic charge distribution models on distributed memory platforms

Reactive molecular dynamics (MD) simulations are important for high-fidelity simulations of large systems with chemical reactions. Iterative linear solvers used to dynamically determine atom polarizations in reactive MD models and redundancies related to bond order calculations constitute significant bottlenecks in terms of time-to-solution and the overall scalability of reactive force fields. The objective of this work is to address these bottlenecks. To accomplish this goal, several optimizations are explored including acceleration of the charge model solver through an effective preconditioning technique and a numerical method with reduced communication overheads, as well as initialization and data structure changes for bond order calculations. Detailed scalability analysis of these optimizations and their overall impact is presented. A single-allreduce pipelined non-blocking conjugate gradient (PIPECG) solver coupled with a sparse approximate inverse (SAI) based preconditioner has been observed to yield significant speedups over the baseline standard CG solver with Jacobi preconditioner. These results are significant as they can facilitate scalable simulations of large reactive systems, and presented techniques can be used in other polarizable MD models.