Concurrent I/O system for the hypercube multiprocessor

Abstract

This paper describes the Concurrent I/O (CIO) system for the Hypercube Multiprocessor. Our system has three components: Architecture, User Interface, and Programming Paradigm. The architecture of the system is based on clusters. Each cluster has a hypercube topology and contains either computing nodes or I/O nodes. This architecture scales well, provides a convenient naming convention and offers independent connectivity between the I/O nodes. The routing algorithm confines the I/O traffic within the I/O clusters. Clusters can be connected in a variety of ways suitable for such diverse applications as graphics and database systems. The system provides a programming paradigm where several nodes of a multiprocessor can cooperate on creating a file.