In adopting a spatial decomposition approach, we found a significant improvement in performance of our codes despite the additional complications of communicating the random forces3, implementation of the Lees-Edwards boundary condition, and accounting for objects that can extend over many processor domains. Clearly, the main bottleneck of such an approach is the message passing between processors. As such technologies improve, we expect corresponding improvements in the computational performance of our algorithms.
Speedups like this on parallel architecture computers also allow us to systematically explore regions of parameter space (e.g., different solid fractions, broader particle size and shape distributions and other boundary conditions) that would be prohibitive on single processor computers. We also note for the record that this technique has proven effective in a shared memory environment [14] where the speedups were a factor of 29 on 32 processors of an SGI Origin 3000 system and a factor of 50 on 64 processors.
3 The random force in the DPD formalism of particle i on particle j has to be equal and opposite of the force of particle j on particle i.