Large-scale scientific irregular computing on clusters and grids

Data sets involved in many scientific applications are often too massive to fit into main memory of even the most powerful computers and therefore they must reside on disk, and thus communication between internal and external memory, and not actual computation time, becomes the bottleneck in the computation. The most challenging are scientific and engineering applications that involve irregular (unstructured) computing phases. This paper discusses an integrated approach, namely ideas, techniques, concepts and software architecture for implementing such data intensive applications on computational clusters and the computational Grid, an emerging computing infrastructure. The experimental performance results achieved on a cluster of PCs are included.