Using a Multilayer Ocean Model on Clusters
versus a Traditional Supercomputer

Craig C. Douglas
Department of Computer Science
University of Kentucky
325 McVey Hall - CCS
Lexington, KY 40506-0045, USA
douglas@ccs.uky.edu

Mohamed Iskandarani
Institute of Marine and Coastal Sciences
Rutgers University
P.O. Box 231
New Brunswick, NJ 08903-0231, USA
mohamed@ahab.rutgers.edu

Abstract

We simulate oceanic overflow problems. The long term goal of this research is to understand both the local dynamics of downslope flows in the ocean and their role in the Earth's global thermohaline circulation. The modeling of these flows and their climatic impact is complicated by the inherent range of spatial scales involved, which extend from the global scale [O(10,000) km] down to the local scale of the overflows themselves [O(1) km], and by the intrinsic three dimensionality of the overflow dynamics.

The Spectral Element Ocean Model (SEOM) offers an elegant solution to these difficulties. It features advanced algorithms, based on h-p type finite element methods, allowing accurate representation of complex coastline and oceanic bathymetry, variable lateral resolution, and high order solution of the three dimensional oceanic equations of motion.

SEOM's geometrical flexibility permits highly inhomogeneous horizontal grids. An added advantage of the technique is its scalability. Most of the computations are carried out at the element level; only interface information needs to be exchanged between elements. The dual characteristic of dense and structured local computations, and sparse and unstructured communication enhances the locality of the computations. The dual characteristic of dense and structured local computations, and sparse and unstructured communication enhances the locality of the computations, and makes SEOM ideally suited for parallel computers.

In this talk we demonstrate what types of cluster machine characteristics are suitable for solving our problems, how much they cost at what point in time, and how they compare to a traditional RISC based supercomputer solution. Our comparisions will note time, money, and aggrevation factors.