Probabilistic diagnosis in wafer-scale systems

Abstract

Studies fault diagnosis based on a realistic probabilistic model for wafer-scale multiprocessor systems. In this model, an individual processor fails independently with probability p. The authors use a comparison testing approach. The testing is performed in multiple stages by the processors. They assume that different testing tasks are executed in different stages, and the coverage of each testing task is imperfect and is represented by a parameter c/sub v/. Imperfect coverage can be used to model intermittent faults where individual test may be incapable of detecting a fault. The authors present an efficient distributed self-diagnosis algorithm that probabilistically identifies faulty and fault free units. They show that our algorithm achieves very high accuracy even when the system is sparsely interconnected, and a large number of faulty units are present in the system.<>