Performance Portability of Parallel Programs on Networks of Workstations

Combining easy-to-use parallelism, portability and efficiency is a very hard task when traditional programming models are applied to networks of workstations. Even if standard message-passing libraries allow to write code that runs on various machines (code portability), the parallel application risks to be inefficient on any other platform for which a different domain decomposition would be the best. If we want a parallel application to be portable with effciency (performance portability) on heterogeneous and nondedicated distributed platforms, the best domain decomposition and communication pattern cannot be chosen during implementation. In this paper, we describe the new architecture of the Dame system that hides to the programmer many details of the actual computing platform, and makes programs self-adaptable to the platform without additional efforts to the parallelization itself