Probabilistic cluster fault diagnosis for multiprocessor systems

As high performance computing systems consisting of multiple processors play an important role in big data analytics, we are motivated to focus on the research of reliability, design-for-test, fault diagnosis and detection of large-scale multiprocessor interconnected systems. System-level diagnosis theory, which originates from the testing of VLSI and Wafer, aims to identify faulty processors in these systems by means of analyzing the test results among the processors, while diagnosability as well as diagnosis accuracy are two important indices. The probabilistic fault diagnostic strategy seeks to correctly diagnose processors with high probability under the assumption that each processor has a certain failing probability. In this work, based on the probabilistic diagnosis algorithm with consideration of fault clustering, we specialize in the local diagnostic capability to establish the probability that any processor in a discrete status is diagnosed correctly. Subsequently, we investigate the global performance evaluation of multiprocessor systems under various significant fault distributions including Poisson distribution, Exponential distribution and Binomial distribution. In addition, we directly apply our results to the data center network $HSDC$ and $(n,k)$-star network. Numerical simulations are performed to verify the established results, which reveal the relationship between the accuracy of correct diagnosis and regulatory parameters.