high performance computing on graphics processing units: hgpu.org

We present a scalable spatial decomposition coloring approach to implement molecular dynamics simulations with embedded atom method (EAM) on multi-core architectures. It effectively solves parallelization of reduction operations on irregular arrays in molecular dynamics simulations. In OpenMP program model, our methodology avoids that the same memory location is simultaneously modified by more than one thread when the short-range forces is calculated, meanwhile our method reduces memory requirements. The methodology comes from the idea of red-black coloring, popular in linear algebra. We developed the spatial decomposition coloring algorithm, and our work applied this algorithm to implement the embedded atom method formalism for molecular dynamic. In this paper we also describe other optimizing methods applied in our serial and parallel implementations. Results show that our method is scalable and can achieve nearly linear speedup. Additionally we also compared it with other methods which can parallelize reduction operations on irregular array, and we discussed them in detail.