Exastamp: A parallel framework for molecular dynamics on heterogeneous clusters

Abstract

Recent evolution of supercomputer architectures toward massively multi-cores nodes equipped with many-core accelerators is leading to make MPI-only applications less effective. To fully tap into the potential of these architectures, hybrid approaches-mixing MPI, threads and CUDA or OpenCL-usually meet performance expectations, but at the price of huge development and optimization efforts. In this paper, we present a programming framework specialized for molecular dynamics simulations. This framework allows end-users to develop their computation kernels in the form of sequential-looking functions and generates multi-level parallelism combining vectorized and SIMD kernels, multi-threading and communications. We report on preliminary performance results obtained on different architectures with widely used force computation kernels.