On integrating error detection into a fault diagnosis algorithm for massively parallel computers

Scalable fault diagnosis is necessary for constructing fault tolerance mechanisms in large massively parallel multiprocessor systems. The diagnosis algorithm must operate efficiently even if the system consists of several thousand processors. We introduce an event-driven, distributed system-level diagnosis algorithm. It uses a small number of messages and is based on a general diagnosis model without the limitation of the number of simultaneously existing faults (an important requirement for massively parallel computers). The algorithm integrates both error detection techniques like ⟨I´m alive⟩ messages, and built in hardware mechanisms. The structure of the implemented algorithm is presented and the essential program modules are described. The paper also discusses the use of test results generated by error detection mechanisms for fault localization. Measurement results illustrate the effect of the diagnosis algorithm, in particular the error detection mechanism by ⟨I´m alive⟩, messages, on the application performance