Communication and computational cost on parallel algorithm of PDE elliptic type

Abstract

High performance computing is widely use in diverse industries and well known as an efficient solver of grand challenge problems. This high speed processing ability enables complex tasks to be accomplished within microseconds. Electronic-chip industry is one of such that clearly demands this parallel computation advantages. To fulfill this industrial needs, this research is focusing on the prediction of power density versus temperature distribution for multilayer full-chip structure that will be solve using parallel red-black Gauss Seidel and Alternating Group Explicit (AGE) methods. The parallel algorithms of 2-dimensional Partial Differential Equation (PDE) elliptic type for the prediction will be executed using distributed memory of heterogeneous cluster platform on LINUX-based environment. The distributed memory architecture and message passing paradigm among processors naturally dealing with communication cost. Additionally, large sparse of matrix that is resulted in the elliptic discretization will contribute to high computational complexity to the problem under consideration. Therefore, it is important to investigate which method resulting in moderate communication and computational cost and at the same time maintains the accuracy of the prediction. Due to this needs, this paper presents the parallel performance evaluations of algorithms that will be discussed in term of communication and computational cost.