Y. Peng, C. Knight, P. Blood, L. Crosby, G. A. Voth, "Extending Parallel Scalability of LAMMPS and Multiscale Reactive Molecular Simulations," Preprint ANL/MCS-P2085-0512, May 2012. [pdf]
Conducting molecular dynamics (MD) simulations involving chemical reactions in large-scale condensed phase systems (liquids, proteins, fuel cells, etc) is a computationally prohibitive task even though many new ab initio based methodologies (i.e., AIMD, QM/MM) have been developed. Chemical processes occur over a range of length scales and are coupled to slow (long time scale) system motions, which make adequate sampling a challenge. Multistate methodologies, such as the multistate empirical valence bond (MS-EVB) method, which are based on effective force fields, are more computationally efficient and enable the simulation of chemical reactions over the necessary time and length scales to properly converge statistical properties.
The typical parallel scaling bottleneck in both reactive and nonreactive all-atom MD simulations is the accurate treatment of long-range electrostatic interactions. Currently, Ewald-type algorithms rely on three-dimensional Fast Fourier Transform (3DFFT) calculations. The parallel scaling of these 3D-FFT calculations can be severely degraded at higher processor counts due to necessary MPI all-to-all communication. This poses an even bigger problem in MS-EVB calculations, since the electrostatics, and hence the 3D-FFT, must be evaluated many times during a single time step.
Due to the limited scaling of the 3D-FFT in MD simulations, the traditional single-program-multiple-data (SPMD) parallelism model is only able to utilize several hundred CPU cores, even for very large systems. However, with a proper implementation of a multi-program (MP) model, large systems can scale to thousands of CPU cores. This paper will discuss recent efforts in collaboration with XSEDE advanced support to implement the MS-EVB model in the scalable LAMMPS MD code, and to further improve parallel scaling by implementing MP parallelization algorithms in LAMMPS. These algorithms improve parallel scaling in both the standard LAMMPS code and LAMMPS with MS-EVB, thus facilitating the efficient simulation of large-scale condensed phase systems, which include the ability to model chemical reactions.