The current generation of high-performance computers are hierachical multiprocessor machines. These computers can be an effective technology for solving challenging problems. However, their complexity leads to an array of tradeoffs in writing programs that hinders timely implementation. In particular, the built-in assumptions of a flat hierarchy in most existing parallel programming models makes it difficult to express efficient hierarchical algorithms needed on these machines. This project will compare two approaches to programming hierarchical parallel machines: an explicitly hierarchical model that is aware of the machine's structure, and a model that abstracts the communication to hide the hierarchy. Both models will be investigated by extending the KeLP2.0 framework, which provides support for masking communication latency (either hierarchical or not) by pipelining. This supports the design of application-specific overlap techniques. The investigators will carry out detailed performance studies of several applications, including Direct Numerical Simulation (DNS) of turbulent flows. The results of these will be incorporated in alternative implementation policies within the KeLP runtime system. The KeLP system will be made available to interested researchers as a possible platform for further investigation in hierarchical parallel programmig.
