An integrated approach to programming and performance modeling of multicomputers

The relative ease with which it is possible to build inexpensive, high-performance multicomputers using regular microprocessors has made them very popular in the last decade. The major problem with multicomputers is the difficulty in effectively programming them. Programmers are often faced with the choice of using high level programming tools that are easy to use but provide poor performance or low level tools that take advantage of specific hardware characteristics to obtain better performance but are difficult to use. In general, existing parallel programming environments do not provide any guarantee of performance and they provide little support for performance evaluation and tuning. This dissertation explores an approach in which users are provided with programming support based on parallel programming paradigms. We have studied two commonly used parallel programming paradigms: Processor Farm and Divide-and-Conquer. Two runtime systems, Pfarm and TrEK, were designed and developed for applications that fit these paradigms. These systems hide the underlying complexities of multicomputers from the users, and are easy-to-use and topology independent. Performance models are derived for these systems, taking into account the computation and communication characteristics of the applications in addition to the characteristics of the hardware and software system. The models were experimentally validated on a large transputer-based system. The models are accurate and proved useful for performance prediction and tuning. Pfarm and TrEK were integrated into Parsec, a programming environment that supports program development and execution tools such as a graphical interface, mapper, loader and debugger. They have also been used to develop several image processing and numerical analysis applications.Science, Faculty ofComputer Science, Department ofGraduat