Effective Approximate Fault Diagnosis of Systems with Inhomogeneous Test Invalidation

failed components in the system. In large multiprocessors a circuit-level fault diagnosis is impractical, therefore system-level diagnosis considers only processor faults. Processors periodically execute tests on each other. If an error was detected, the diagnostic procedure identifies faulty units by analyzing the collection of the test results (called the syndrome). Once these units were identified, they are logically isolated and the system is reconfigured to continue the error-free operation. System-level fault diagnosis is a methodology to identify the failed components in a multiprocessor system. The traditional approach to system-level diagnosis does not take into consideration many important aspects of modern multiprocessor architectures. This paper examines a special class of multiprocessors, called massively parallel computers. As a practical example, the Parsytec GCel system is presented. The paper describes a new method developed for the Parsytec GCel, called local information diagnosis. The diagnostic algorithm is based on the generalized test invalidation model, therefore it is applicable to a wide range of systems, including inhomogeneous ones. Due to the employed syndrome decoding mechanism, the space and computational complexity of the algorithm is also smaller than in conventional methods. 1.1. Generalized test invalidation Fault-free tester units are assumed to test other units correctly (tests are assumed to be complete). Still, faulty testers may produce and distribute incorrect test results that do not reflect the real fault state of the tested unit. In other words, errors affect the validity of test outcomes. This effect, modelled by test invalidation, makes the decoding of the syndrome more difficult. The two most frequently used test invalidation models are the symmetric or PMC model [2], and the asymmetric or BGM model[3]: