Fault Tolerance via Replication in Coarse Grain Data-Flow

Abstract : Recent advances in network technology promise to make gigabit-per-second bandwidth between remote hosts a reality in the near future. This increase in bandwidth paves the way for increased exploitation of distributed computing resources. Coupled with advances in distributed memory parallel compiler technology, there is strong reason to believe that wide-area distributed parallel processing will be an increasingly popular and important programming paradigm. Parallelizing and distributing program sub-tasks has the potential to increase performance for many applications while also improving the overall utilization of system resources. Unfortunately, there is a downside. When a program is partitioned into sub-tasks, each sub-task is distributed to potentially a different processor. As the number of processors employed by an application increases so does the chance that the application will fail due to a host/ processor failure. At the University of Virginia, we have experienced first hand the problems caused by host failures in distributed systems while developing and using a prototype for the Legion project [13][14]. The objective of Legion is to construct the software environment to enable a nation-wide or world-wide virtual computer capable of supporting distributed and parallel applications. Our current prototype, which we call the Campus-Wide Virtual Computer (CWVC), contains a mix of over 90 workstations and an IBM SP-2 multicomputer. Even in this relatively small environment, we are frequently experiencing host failures. On the scale of the envisioned nation-wide system, host failures will simply be a fact of life and must be dealt with accordingly. User applications, especially those that are critical or are composed of many distributed components, must be resilient to host failures. Fortunately developing fault tolerant parallel applications does not need to be difficult.