Reactive, polarizable molecular dynamics simulations are a crucial tool for the high-fidelity study of large systems with chemical reactions. In support of this, several approaches have been employed with varying degrees of computational cost and physical accuracy. One of the more successful approaches in recent years, the reactive force field (ReaxFF) model, wasdesigned to fill the gap between traditional classical models and quantum mechanical models by incorporating a dynamic bond order potential term. When coupling ReaxFF with dynamic global charges models for electrostatics,special considerations are necessary for obtaining highly performant implementations, especially on modern high-performance computing architectures.In this work, we detail the performance optimization of the PuReMD (PuReMD Reactive Molecular Dynamics) software package, an open-source, GPLv3-licensed implementation of ReaxFF coupled with dynamic charge models. We begin byexploring the tuning of the iterative Krylov linear solvers underpinning the global charge models in a shared-memory parallel context using OpenMP, with the explicit goal of minimizing the mean combined preconditioner and solver time. We found that with appropriate solver tuning, significant speedups and scalability improvements were observed. Following these successes, we extend these approaches to the solvers in the distributed-memory MPI implementation of PuReMD, as well as broaden the scope of optimization to other portions of the ReaxFF potential such as the bond order computations. Here again we find that sizable performance gains were achieved for large simulations numbering in the hundreds of thousands of atoms.With these performance improvements in hand, we next change focus to another important use of PuReMD -- the development of ReaxFF force fields for new materials. The high fidelity inherent in ReaxFF simulations for different chemistries oftentimes comes at the expense of a steep learning curve for parameter optimization, due in part to complexities in the high dimensional parameter space and due in part to the necessity of deep domain knowledge of how to adequately control the ReaxFF functional forms. To diagnose and combat these issues, a study was undertaken to optimize parameters for Li-O systems using the OGOLEM genetic algorithms framework coupled with a modified shared-memory version of PuReMD. We found that with careful training set design, sufficient optimization control with tuned genetic algorithms, and improved polarizability through enhanced charge model use, higher accuracy was achieved in simulations involving ductile fracture behavior, a difficult phenomena to hereto model correctly.Finally, we return to performance optimization for the GPU-accelerated distributed-memory PuReMD codebase. Modern supercomputers have recently achieved exascale levels of peak arithmetic rates due in large part to the design decision to incorporate massive numbers of GPUs. In order to take advantage of such computing systems, the MPI+CUDA version of PuReMD was re-designed and benchmarked on modern NVIDIA Tesla GPUs. Performance on-par with or exceeding the LAMMPS Kokkos, a ReaxFF implementation developed at Scandia National Laboratories, with PuReMD typically out-performing LAMMPS Kokkos at larger scales.
Read
