An Event-Driven Approach to Multiprocessor Diagnosis

For constructing fault tolerance mechanisms in large massively parallel multiprocessor systems, a scalable fault diagnosis is necessary, which works efficiently even if there are several thousand processors in the system. In this paper we present an event-driven, distributed system-level diagnosis algorithm, based on a general diagnosis model which does not limit the number of simultaneously existing faults. In particular, the relation between error detection and fault localization as well as two different methods for distributing diagnostic information are examined in detail. Furthermore, we give measurements concerning how does our diagnosis algorithm affect application performance