INDIANA UNIVERSITY PURDUE UNIVERSITY INDIANAPOLIS
	Computational Fluid Dynamics Laboratory

Proceedings of Parallel CFD '96, Edited by P. Schiano, et al., Elsevier Science, 1996.

 

Communication Cost Function for Parallel CFD in a Heterogeneous Environment Using Ethernet

Y.P. Chien, A. Ecer, H.U. Akay and S. Secer

Computational Fluid Dynamics Laboratory
Purdue School of Engineering and Technology, IUPUI
Indianapolis, Indiana 46202, USA


Abstract

In order to increase the efficiency of parallel CFD algorithms, a special domain decomposition approach is adopted which divides the problem domain into a number of blocks that are more than the number of computers. Each block is associated with a set of interfaces. Each block and interface is assigned with a block solver and an interface solver, respectively. A software library was previously developed to support this approach. To improved the efficiency of parallel CFDusing this approach, a load balancing algorithm was also developed. The load balancing algorithm requires the availability of a computation cost function and a communication cost function to describe the speed of the computer and network environment using Ethernet with TCP/IP protocol is presented. A pratical real time procedure for obtaining the communication cost function during the execution of parallel CFD is described. This procedure supports dynamic computer load balancing of parallel codes. The experimental results shows that the predicted elapsed times derived from our computation and communication cost functions are very close to the measured elapsed times.