Runtime Support for Dynamic Skeletons Implementation

Abstract

Algorithmic skeletons have proved to be a good solution to the problem of implementing parallel applications with specify communication structures. They define the overall structure of the computation, hiding the complex communication details. Nowadays, the different frameworks available offer a fixed set of skeletons. The programmer can implement efficient programs if the computation and communication patterns match the available skeletons. Because of that, the usage of skeleton frameworks has been limited to an important but relative small set of patterns featuring the most common parallel structures, such as map, pipeline, farm, or wavefront. In this paper, we present a programming model that can be used to implement efficient and portable parallel skeletons. We also discuss its implementation and integration into Hitmap, a tool for hierarchical tiling and mapping. This combined proposal allows to develop tailored static and dynamic skeletons while still hiding implementation and communication details. The performance of the implementation is measured against a well known skeleton framework.