Performance Engineering of Reactive Molecular Dynamics Simulations

Abstract

Reactive molecular dynamics is the best-performing option for simulating chemical systems on the order of thousands of atoms, but its high computational cost often limits the temporal scale of simulation. In order to observe scientific phenomena of interest, we need implementations of interatomic potentials which are highly efficient and scalable on modern architectures. Parallel computing is now ubiquitous, and today’s supercomputing clusters often consist of multicore nodes with high on-node parallelism. Current implementations of ReaxFF display good scaling across many distributed nodes, but fall short in taking full advantage of compute available on an individual CPU or GPU.

This thesis presents analysis and performance optimization of the widely used LAMMPS ReaxFF implementations. I analyze the performance characteristics of the USER-REAXC, USER-OMP, and Kokkos implementations in LAMMPS, profiling and describing bottlenecks in each. I then provide optimizations to serial and parallel CPU code which increase the efficiency and parallel thread scaling of USER-OMP. Using an Intel Xeon Platinum 8260, the resulting code obtains a speedup of 1.5-3x and shows scaling with twice as many OpenMP threads on a 1152-atom Hafnium Diboride simulation. I show performance improvements on various simulation sizes up to 44K atoms, and present independently verified correctness on an AMD Ryzen Threadripper 3970X.