MOLECULAR-DYNAMICS SIMULATIONS OF COULOMBIC SYSTEMS ON DISTRIBUTED-MEMORY MIMD MACHINES

Parallel algorithms are designed for molecular-dynamics simulations of bulk systems with Coulomb forces. The conditionally convergent series for the 1/r interaction is treated by the Ewald method of summation. One of the algorithms involves the equipartition of an N-particle system into p subsystems of N/p particles each, with a one-to-one mapping between subsystems and p processors. The other algorithm is based on the spatial decomposition of the volume of a system into p equal parts which are geometrically mapped onto processors. The performance of these algorithms is tested on the in-house 8 node Intel iPSC/860 system. Execution times for these algorithms are comparable, and in both cases the computation time dominates the communication time. For a 64000 particle one-component charged plasma in three dimensions, the execution time for a single molecular-dynamics time step is 27.4 s. Execution times increase linearly with an increase in the size of the system and they are inversely proportional to the number of processors. Parallel efficiencies of these algorithms are close to 0.85.