|
Abstract : |
A new multiple-instruction-multiple-data restructuring technique has been developed which can be used to parallelize large three-dimensional simulations of plasma phenomena. This technique is designed to effectively exploit the data parallelism present within these simulations and produces parallel representations which execute over a massively parallel systems which possess distributed shared memory. The technique combines static allocation methodologies, where the problem is partitioned in a manner which is based upon the physical geometry of the simulation space. The technique exploits the fact that the desired simulation volume is often elongated around a single dimension and that the simulation space is assumed to be periodic. The natural topology for such simulations is therefore ring-like in structure, leading to a generic representation which can be effectively mapped onto multiple parallel processing configurations by employing embedding algorithms based upon base-2 reflected Gray encoding. The restructuring technique has been analyzed empirically and analytically to determine how well such simulations can be expected to perform as the geometric domain of the simulation increases in size as the number of processors is increased. The restructuring technique has been used successfully to map this problem onto IBM SP3 using MPI and then compared with previous work. The technique was applied to the 3-D simulation of the cavity formation in the Aurora region near the Earth's ionosphere. This application requires large-scale parallelism to achieve the desired accuracy in a minimum turnaround time and requires the use of state-of-the-art computing platforms which have the aggregate computing speeds approaching 1Teraflop and beyond., |
