AN APPROACH TO DATA-PARALLEL MOLECULAR-DYNAMICS FOR LIQUIDS

An efficient approach to large scale data parallel short-range molecular dynamics for liquids is presented. The method is based on the coarse-grained cell method in which the simulation cell is decomposed into equally sized subcells, with the shortest side being larger than the cut-off radius. To avoid a large fraction of the nonproductive calculations we develop a geometric sorting procedure based on particle distances to subcell boundaries. Due to particle migration, the contents of the subcells need to be updated. This is done with a method based only on nearest-neighbor communications. Special ''null-particles'' are introduced, which act as buffers during periodic updates and allow for a globally uniform algorithm during the force calculations. The method should be easy to implement on most massively parallel computers of SIMD or MIMD type. We have implemented our code in CM Fortran on an 8K CM200. Communication cost is around 7% of the total cpu time. The overall speed for one million particles is approximately 5.9mus per MD time step and per particle and 5.5mus for five million particles.