Dynamic Run Time Support for Skeletons

We propose a new implementation model for structured parallel programming models based on skeletons. The model is aimed at overcoming some of the problems observed in traditional, template based implementations of skeleton languages, while preserving all the positive features of structured parallel programming models. The proposed implementation model does not rely on static process networks like the template based implementation model for skeletons. It rather implements skeleton programs by properly scheduling macro data-flow instructions (corresponding to the sequential portions of code appearing in the skeleton code) to a set of macro data-flow interpreters running on the processing elements of the target architecture. We discuss some preliminary experimental results demonstrating the feasibility and effectiveness of our approach.