Grid Architecture for High Performance Computing

Abstract

In theory, a grid can harness the power of an arbitrarily large collection of computing resources to meet the needs of compute intensive high performance computing (HPC) applications such as finite element model (FEM) simulations. However, the realization of such potential faces many challenges; including: (1) the degree of coupling between the components of an HPC application; (2) the asynchronous and high latency nature of the grid communication medium; and (3) the presence of distinct administrative domains. High latency poses the most serious limitation to the potential use of grids for HPC applications. In this respect, the paper reports new results on the ongoing development of a service oriented grid architecture to support the execution of HPC applications, and in particular the simulation of large scale dynamical systems. At the heart of the architecture is a computational model that ties the domain decomposition of the simulated system to the topology of the grid and the data exchange process in order to minimize the effect of latency. The paper provides an experiment-based comparative analysis of the proposed grid-HPC computational model against a sequential computing model for a two dimensional heat diffusion process simulation.