Showing papers by "Steven J. Plimpton published in 2015"

Direct simulation Monte Carlo investigation of the Richtmyer-Meshkov instability

Abstract: The Richtmyer-Meshkov instability (RMI) is investigated using the Direct Simulation Monte Carlo (DSMC) method of molecular gas dynamics. Due to the inherent statistical noise and the significant computational requirements, DSMC is hardly ever applied to hydrodynamic flows. Here, DSMC RMI simulations are performed to quantify the shock-driven growth of a single-mode perturbation on the interface between two atmospheric-pressure monatomic gases prior to re-shocking as a function of the Atwood and Mach numbers. The DSMC results qualitatively reproduce all features of the RMI and are in reasonable quantitative agreement with existing theoretical and empirical models. Consistent with previous work in this field, the DSMC simulations indicate that RMI growth follows a universal behavior.

Maintainability and Performance for LAMMPS.

Abstract: Sandia National Laboratories is a multi-­‐program laboratory managed and operated by Sandia Corporation, a wholly owned subsidiary of Lockheed Martin Corporation, for the U.S. Department of Energy's National Nuclear Security Administration under contract DE-­‐AC04-­‐94AL85000. Over the last few years, molecular dynamics (MD) simulation codes have been at the forefront of supporting new hardware architectures. For example, most of the major older codes, including AMBER, Gromacs, NAMD, and LAMMPS, have provided significant support for GPUs for several years, and newer codes like HOOMD have been developed specifically for GPUs. We believe the older codes all initially approached this task by creating CUDA variants of their most important MD kernels specifically for the GPU. Some of these codes now support running completely on GPUs, with no regular data transfers to the CPU other than for MPI communication. Others offload only specific key computations to the GPU. Many also support hybrid execution where, for example, non-­‐bonded interactions are calculated on the GPU at the same time bonded interactions and long-­‐range Coulombics are calculated on the CPU. In general, performance improvements over many-­‐core CPU-­‐only clusters have been quite good. For LAMMPS, this approach of creating variants of key kernels tuned for each architecture (CPU/MPI-­‐only, OpenMP, GPU, Phi, etc), has become increasingly difficult to sustain over time as the code has grown and new architectures proliferate. As a general materials simulation code with models at the atomic to meso to continuum scales, LAMMPS currently has non-­‐bonded kernels for ~120 different functional forms. There are likewise ~40 different functional forms for bonded interactions (bonds, angles, dihedrals, impropers) and several variants of long-­‐range Coulombic models. The code also has 100+ options for add-­‐on calculations which affect the dynamics such as thermostats, barostats, different time integration schemes, force constraints, interactions with boundaries and other objects, external forces, etc., as well as several dozen optional diagnostics, which involve loops over atoms or more complex calculations. Overall, we estimate LAMMPS thus contains around 500 unique kernels, which if not optimized, could become bottlenecks in a particular simulation on specific hardware. The challenge this creates is reflected in the current GPU capabilities of LAMMPS, which only support a small fraction of all these LAMMPS features. The problem is compounded because LAMMPS input scripts allow users to combine these capabilities in flexible and unpredictable ways. As a consequence, just porting the most important kernels to a given architecture may still result in …