Special purpose high performance computers are expensive and rare, but workstation clusters are cheap and becoming common. Emerging technology offers the opportunity to integrate clusters into a single high performance computer - a computational Grid. The acceptance of computational Grids, however, is seriously hampered by the difficulty of efficiently managing the parallelism in such heterogeneous clusters, with characteristics radically different from a conventional high performance computer. To program this complex and dynamic architecture effectively we propose to use a language with high-level constructs, GpH, and to extend its runtime environment, GUM.
The first contribution of this thesis is to develop Grid-GUM1, an initial port of GUM to computational Grids. Systematic evaluation shows that Grid-GUM1 delivers acceptable speedups on relatively low latency and on homogeneous computational Grids. However for high latency or heterogeneous computational Grids poor load scheduling limits performance.
We next present an adaptive runtime environment Grid-GUM2, which includes monitoring mechanisms that determines static and dynamic properties of the underlying clusters and an adaptive scheduling mechanism that dynamically modifies parallel execution accordingly. To the best of our knowledge, Grid-GUM2 is one of the first fully implemented virtual shared memory runtime environment on the Grid. Evaluating Grid-GUM2's performance demonstrates that virtual shared memory is feasible on computational Grids and that it can deliver good speedups if combined with an aggressive dynamic load distribution mechanism.