Section: New Results

AMR-Based Dynamic Load Balancing for Molecular Dynamics Simulations

Modern parallel architectures require applications to generate enough parallelism to feed many cores, which require in turn regular data-parallel instructions to exploit large vector units. We revisit the extensively-studied Classical Molecular Dynamics N-body problem in the light of these hardware constraints. A new data layout is proposed with efficient force computation methods focusing on adaptive mesh refinement techniques, multi-threading, vectorization-friendly, using low memory footprint. Our design is guided by the need for load balancing and adaptivity raised by highly dynamic particle sets, as typically observed in simulations of strong shocks resulting in material micro-jetting. We analyze performance results on several simulation scenarios, over clusters equipped with Intel Xeon Phi knl processors. Performance obtained with our implementation using OpenMP is close to state-of-the-art implementations (LAMMPS) using MPI on steady particles simulations, and outperform them by 1.2 on micro-jetting simulations on Intel Xeon Phi (KNL).