Fast Solvers for Charge Distribution Models on Shared Memory Platforms

Abstract

Including atom polarizability in molecular dynamics (MD) simulations is important for high-fidelity simulations. Linear solvers for charge models that are used to dynamically determine atom polarizations constitute significant bottlenecks in terms of time-to-solution and the overall scalability of polarizable and reactive force fields. We present properly customized preconditioning techniques to accelerate the iterative solvers used for several charge models and develop their efficient shared memory parallel implementations in the open source PuReMD (Purdue Reactive Molecular Dynamics) software package. With these goals in mind, special attention has been paid to minimizing the mean combined preconditioner construction and solver time. Detailed analysis of how different preconditioning techniques affect solver convergence rate and the overall performance is presented. Incomplete LU/Cholesky and sparse approximate inverse (SAI) based schemes that produce good quality factors with a relatively low number of nonzeros have been observed to yield significant speedups over a baseline Jacobi preconditioner. These results are significant as they can enable efficient simulations of small to moderate-sized systems on multicore computers, but, more importantly, they serve as a basis for distributed memory solvers.