Parallel Algorithms for Molecular-Dynamics Simulations of Coulombic Systems

Abstract

In molecular-dynamics simulations for the long-range Coulomb interaction, a great deal of effort is devoted to reducing the computational complexity of the usual N2operations in the direct calculation. For bulk systems, we have designed a parallel algorithm based on the domain-decomposition strategy for the Ewald summation. The performance of the algorithm is evaluated on the in-house iPSC/860 system. We find that this algorithm reduces the computational complexity to O(N). For a 64,000-particle plasma in three dimension, the execution time on an 8-node system is 27.4 sec per MD time step. The interprocessor communication is a small fraction of the total execution time. We find linear speedups and a parallel efficiency of 0.85. For comparison, parallel algorithms are also designed for the Fast Multipole Method (FMM) - a divide and conquer scheme in which the system is divided into cubic subdomains and interactions between distant charged regions are calculated with a truncated multipole expansion. The performance of the FMM on Touchstone Delta machine is discussed.