3-D land avoidance and load balancing in regional ocean simulation

We present methods designed to enhance the performance of a parallel ocean circulation model (OCM). The underlying code is based on a primitive equations model of the Mediterranean Sea. We examine the characteristics of this basin and study the tradeoffs between minimising computations by avoiding sweeping over land areas vs. the higher computational rates achieved otherwise for a cluster based multiprocessor with hierarchical memory. Numerical experiments on Cedar, an architecture consisting of clusters of multiprocessors connected to a global shared memory, show the performance improvements that can be obtained when land avoidance is combined with a domain partitioning strategy designed to achieve load balancing