Molecular dynamics is a key method for the study of microscopic systems. We
have developed two scalable massively parallel algorithms for molecular
dynamics. The first, which is intended for studies of condensed phases,
uses a spatial domain decomposition technique and is the most efficient
algorithm yet found for the study of large solid-state and fluid-state
systems. Its cost scales linearly as the problem size increases; and its
speed also increase linearly as the number of processors increases. It
exhibits good load balance on condensed phase and gaseous systems. Thus it
is close to optimal for those systems. It will allow study of systems with
tens of millions of atoms on machines such as Sandia's Intel Paragon. The
second algorithm addresses the problem of treating large chemical and
biological molecules effectively on parallel machines. Here, although the
cost increases linearly with the problem size, the most straight forward
techniques incur communication costs on parallel machines that grow
linearly with the number of processors. So the speed of the problem
ultimately reaches an asymptote independent of the number of processors.
That is, the method does not scale. The spatial decomposition method does
scale but it occurs high overhead due to the inhomogeneity of typical
molecules. The inhomogeneity contributes to poor load balancing for these
problems. We have developed a novel force-based algorithm, whose
communication requirements scale like the square root of the number of
processors. It thus shows speed up like the square root of the number of
processors. While this is sub-optimal, the algorithm has low overhead and
is well load balanced. Thus for even quite lage numbers of processors over
a large range of molecular sizes it is faster than any other current
method. We have applied this method in collaboration with the Wright
Laboratory to the analysis of complex biological liquid crystals for laser
optic applications. Problems hundreds of times as complex as previously
accessible have thus been studied.
Click here for movie.