Data Exchange : High Performance

Current communications tools and libraries for high performance computing are designed for platforms and applications that exhibit relatively stable computational and communication characteristics. In contrast, the demands of (1) mixed environments in which high performance applications interact with multiple end users, visualizations, storage engines, and I/O engines { termed`distributed laboratories' in our research { and (2) high performance collaborative computing applications in general, exhibit additional complexities in terms of dynamic behaviors. This paper explores the communication requirements of distributed laboratories , and it describes the DataExchange communication infrastructure supporting high performance interactive and collaborative applications. 1 COMMUNICATION SUBSTRATES Communication tools and libraries for high performance distributed computing have evolved substantially during the last few years, from systems enabling distributed parallel computing such as PVMM3], to industrial standards being actively implemented and improved such as MPII2], to recent proposals for infras-tructures that can facilitate the construction of nationwide , networked supercomputers. Our work contributes to such research by addressing several spe-ciic issues that arise whenever diverse networked machines are used in research and development settings or even in production environments, by single or multiple end users. The target applications we consider are those in which end users interact with powerful computational tools and with each other across heterogeneous networked machines, termed distributed laboratories. The communication demands of such dynamic computational environments share some com-monalities with traditional distributed systems as well as with the highly dynamic environment of today's web browsers and CORBA-based distributed object systems , but unfortunately these similar environments do not adequately address the high performance needs of some applications. Moreover, these environments' dynamic communication and computational behaviors are not addressed by standards for distributed high performance computing such as MPI. Our work assumes that end users should be able to exploit the potentially high performance of vendor-provided implementations of standards like MPI when appropriate for their applications. In addition, when application requirements can be stated statically, users should be able to exploit the automatic support for parallelization and for eecient inter-task communications ooered by compilers, their runtime support, and specialized communication libraries. Given these assumptions , the Distributed Laboratories Projectt5] at the Georgia Institute of Technology is developing communication support that addresses the interactive and dynamic nature of high performance applications, and enables their use in scientiic investigations or in production environments by single and by multiple end users. Speciically, the DataExchange communication library utilized in the implementation …