Harnessing the killer micros: Applications from LLNL's massively parallel computing initiative

Abstract

Recent developments in microprocessor technology have led to performance on scalar applications exceeding traditional supercomputers. This suggests that coupling hundreds or even thousands of these “killer-micros” (all working on a single physical problem) may lead to performance on vector applications in excess of vector supercomputers. Also, future generation killer-micros are expected to have vector floating point units as well. The purpose of this paper is to present an overview of the parallel computing environment at Lawrence Livermore National Laboratory. However, the perspective is necessarily quite narrow and most of the examples are taken from the author's implementation of a large-scale molecular dynamics code on the BBN-TC2000 at LLNL. Parallelism is achieved through a geometric domain decomposition — each processor is assigned a distinct region of space and all atoms contained therein. As the atomic positions evolve, the processors must exchange ownership of specific atoms. This geometric domain decomposition proves to be quite general and we highlight its application to image processing and hydrodynamics simulations as well.