High performance parallel algorithms for scientific computing with application to a coupled ocean model

Parallel computers are becoming more common and are available on a range of platforms from massively parallel machines with thousands of processors to PCs with two or four processors. The software to exploit the capabilities of these machines is yet to be developed and this dissertation explains some of the problems and solutions to implement scientific applications on parallel machines. Many companies have failed to develop a general purpose parallel computer because the problem of achieving high performance on practical applications was more difficult than previously thought.; We have developed efficient parallel algorithms and tools for scientific applications which can be used on commercial parallel machines. The algorithms we used improved performance by reducing communication overhead for data distribution/collection and grid partitioning. We developed tools to study cache performance on a parallel machine and to visualize communication performance for a distributed application. We tested different problem sizes and processor configurations to use a machine with high parallel efficiency. Communication protocols on two common parallel architectures are evaluated.; We used a coupled ocean model as the experimental testbed to evaluate the algorithms and tools we developed. Ocean Models are computationally intensive applications and are one of the 'Grand Challenge' problems. The coupled ocean model is used to simulate the behaviour of the El Nino phenomenon. Our implementation of the ocean model on a Cray T3D parallel computer obtained a performance rate of 2.4 Gflops which was 6 times faster than the current implementation on a vector supercomputer.