The OpenMP Application Programming Interface is the emerging standard for shared memory parallel programming. It has been widely endorsed by industry and the Department of Energy. Because it is easy to use, OpenMP is also popular among application programmers. However, OpenMP has previously been available only in implementations for shared memory multiprocessors. I have extended support for OpenMP to networks of workstations (NOWs), allowing users of OpenMP to take advantage of cost effective personal computers and local area networks. This work aims to provide scalable and flexible support for OpenMP on NOWs.

The key idea in this work is the novel integration of compile-time and run-time approaches, which traditionally have been used in isolation. I demonstrate that this combination achieves better performance than either the compiler or the run-time approach used in isolation. This approach is able to optimize irregular applications, those that the compiler cannot analyze precisely. This work substantially improves the scalability and reduces the communication cost of running OpenMP programs on NOWs. The compiler/run-time integration also enables efficient extensions to the OpenMP system, such as supporting OpenMP on networks of shared memory multiprocessors, and allowing a variable number of computing nodes during the execution.