On component-based communication systems for clusters of workstations

Most of the communication systems used to support high performance computing in clusters of workstations have been designed focusing on “the best” solution for a certain network architecture. However, a definitive best solution, independently of how well tuned to the underlying hardware it is, cannot exist, for parallel applications communicate in quite different ways. In this paper we describe a novel design method that supports the construction of run time systems as an assemblage of components that can be configured to closely match the demands of any given application. We also describe how this method has been deployed in the development of a communication system in the realm of EPOS, a project that aims at delivering automatically, generated application-oriented run-rime support systems. The communication system in question has been implemented for a cluster of PCs interconnected with Myrinet, and corroborates the effectiveness of the proposed design method