Please choose your delivery country and your customer group
This dissertation presents a new and efficient communication architecture specifically designed for high-performance workstation clusters wich require a high-speed network. The goals of this architecture are to achieve a low-latency, high-bandwidth, and scalable communication, while maintaining portability, reusability, and system stability and security. Two programming models are supported: state-of-the-message passing, as well as a new model known as remote store, which is discussed in detail. The implementation of efficient message passing, with the purpose of supporting the widely used MPI (message passing interface) standard, is examined and commented on. Several workstation clusters using the new architecture were built in the Swiss-Tx project. The goal of this project was to design, build and install the first tera-flop supercomputer in Switzerland. With this aim in mind, two different types of network hardware have been developed within this project: The bus based EasyNet, for small machines with up to eight nodes, and the router-based T-NET, for large and complex network topologies. EasyNet in particular serves to prove that the new architecture satisfies the goals set out for it. The results obtained from T-NET confirm this finding.