A Pipelined Parallelism Approach to Parallel Short-Range Molecular Dynamics Simulations on Multi-Core Platforms
We present a pipelined parallelism approach to implement the short-range force computations in molecular dynamic simulations on a multi-core machine. Our methodology is based on the OpenMP programming model. It uses multiple producer threads and a single consumer thread to adapt the application for pipeline parallelism, and utilizes the high inter-core communication bandwidth. Multiple producer threads compute the short-range force, and the consumer thread modifies the global force-array. Compared with some other methods applied in data parallelism that can parallelize reduction operations on a force-array, our method achieves high performance especially when the simulation system is characterized by irregular geometry or by inhomogeneous atom densities.