Pseudorandom number generator for massively parallel molecular-dynamics simulations.

A class of uniform pseudorandom number generators is proposed for modeling and simulations on massively parallel computers. The algorithm is simple, nonrecursive, and is easily transported to serial or vector computers. We have tested the procedure for uniformity, independence, and correlations by several methods. Related, less complex sequences passed some of these tests well enough; however, inadequacies were revealed by tests for correlations and in an interesting application, namely, annealing from an initial lattice that is mechanically unstable. In the latter case, initial velocities chosen by a random number generator that is not sufficiently random lead quickly to unphysical regularity in grain structure. The new class of generators passes this dynamical diagnostic for unwanted correlations.