A FAULT-TOLERANT HIERARCHICAL DIAGNOSTIC NETWORK FOR MASSIVELY-PARALLEL PROCESSING SYSTEMS

Massively parallel processing systems consist of a large number of pro cessing nodes to provide high performance primarily for data-intensive applications. In a system of such dimensions, high availability canno t be achieved without relying on redundancy and reconfiguration. An im portant aspect of highly available design is rapid diagnosis and grace ful degradation in the event of failures. This paper presents a hierar chical diagnostic network for locating faults in parallel processor sy stems comprised of a large number of identical processing nodes. In th e case of a single fault, the network can locate the fault at the time it is detected. Even in the case of multiple faults, it can significa ntly reduce the test time as compared to the well-known binary search. Unlike the existing self-diagnostic circuits, the diagnostic network requires small hardware overhead and may tolerate a fault in the netwo rk itself. (C) 1998 Elsevier Science Ltd. All rights reserved.