Exploiting data-flow for fault-tolerance in a wide-area parallel system

Wide-area parallel processing systems will soon be available to researchers to solve a range of problems. In these systems, it is certain that host failures and other faults will be a common occurrence. Unfortunately, most parallel processing systems have not been designed with fault-tolerance in mind. Mentat is a high-performance object-oriented parallel processing system that is based on an extension of the data-flow model. The functional nature of data-flow enables both parallelism and fault-tolerance. In this paper, we exploit the data-flow underpinning of Mentat to provide easy-to-use and transparent fault-tolerance. We present results on both a small-scale network and a wide-area heterogeneous environment that consists of three sites: the National Center for Supercomputing Applications, the University of Virginia and the NASA Langley Research Center.