Architecture-Adaptable Finite Element Modelling A Case Study using an Ocean Circulation Sim.pdf

Architecture-Adaptable Finite ElementModelling: A Case Study using an OceanCirculation SimulationSanthosh Kumarany, Robert N. Millerz, and Michael J. QuinnyyDepartment of Computer ScienceOregon State University, Corvallis, OR 97331.zCollege of Oceanic and Atmospheric SciencesOregon State University, Corvallis, OR 97331.AbstractWe describe an architecture-adaptable methodology for the parallel implemen-tation of nite element numerical models of physical systems. We use a model oftime-dependent ocean currents as our working example. The heart of the com-putation is the solution of a banded linear system, and we describe an algorithmbased on the domain decomposition method to solve the banded system. The algo-rithm is represented in a divide-and-conquer framework. The well-dened algebraicstructure of this framework makes it relatively easy to predict the performance ofthe algorithm with high accuracy. Performance prediction makes possible architec-ture adaptability. In the case of the domain decomposition method described inthis paper, performance prediction enables us to make decisions about granularityand the number of processors to use for individual steps of the algorithm. Thedivide-and-conquer framework facilitates easy implementation of various algorith-mic options. The process is straightforward and amenable to automation. Wedemonstrate the validity of this approach using two radically dierent target ma-chines: a workstation network and a supercomputer. Our results show very goodspeedup on both platforms.1 IntroductionOne of the challenges in parallel computing research is to develop ecientparallel algorithms to solve problems which are known to have fast sequen-tial algorithms. An example of such a problem is the numerical solutionof partial dierential equations (PDEs) using the nite element method.The crucial part of this method is the solution of a banded linear system.Developing cost-optimal parallel algorithms to solve such systems is pivotalto ecien