Communication Issues in Parallel Computing across ATM Networks

Cluster-based computing, which exploits the aggregate power of networked collections of computers, has drawn increasing attention in the parallel processing community. The success of cluster-based computing depends largely on the performance of the underlying network. Although Asynchronous Transfer Mode (ATM) was designed specifically to transport integrated multimedia data, the expected ubiquity of such networks will lead to their use for a wide variety of applications, including parallel processing. While ATM networks offer a high-performance communications infrastructure, they must be accompanied by communications software designed to meet the specific needs of parallel applications. This paper is intended to put forward the issues involved in parallel computing across ATM networks and to describe some of the solutions to the accompanying communication problems. Particular emphasis is placed on collective communication operations, such as multicast and global reduction, which typica..