The robust middleware approach for transparent and systematic fault tolerance in parallel and distributed systems

摘要

We propose the robust middleware approach to transparent fault tolerance in parallel and distributed systems. The proposed approach inserts a robust middleware between algorithms/programs and system architecture/hardware. With the robust middleware, hardware faults are transparent to algorithms/programs so that ordinary algorithms/programs developed for fault-free networks can run on faulty parallel/distributed systems without modifications. Moreover, the robust middleware automatically adds fault tolerance capability to ordinary algorithms/programs so that no hardware redundancy or reconfiguration capability is required and no assumption is made about the availability of a complete subnetwork (at a lower dimension or smaller size). We also propose nomadic agent multithreaded programming as a novel fault-aware programming paradigm that is independent of network topologies and fault patterns. Nomadic agent multithreaded programming is adaptive to fault/traffic/workload patterns, and can take advantages of various components of the robust middleware, including the fault tolerance features and multiple embeddings, without relying on specialized robust algorithms.