| Despite more than two decades of research effort, the question remains: how can we realize the potential of large-scale parallel machines? It can be done now, but only at great expense (i.e ., development time and effort) and with limited portability, rendering the exploitation of parallelism impractical for most users. Advanced-ZPL (A-ZPL) is a parallel programming language intended to address this problem. It's design was guided by a predictive performance model that clearly defines the role of the programmer and the compiler, called the programmer-compiler separation. representations to a particular machine. This dissertation evaluates the design and implementation of A-ZPL in the light of this design criteria. Specifically, we examine two aspects of the language and the compiler implications: efficient loop generation and pipelining wavefront computations. We find the language is highly effective both relatively and absolutely as a direct consequence of considering the programmer-compiler separation. |
