The System-Level Fault Diagnosis Based On Artificial Immune Method

Because the development of scientific research is going to a larger-scale and highercomplexity direction, the requirement of high performance computing (HPC) becomes strongerand stronger. Parallel computer system is a key technology to achieve HPC. However, as thescale of the system becomes larger, the probability that the components of the system be-come faulty tends to increase. How to guarantee the high reliability becomes an importantissue.The system-level fault diagnosis is a kind of method to enhance the reliability of thesystem with no additional cost. The main idea is to make full use of the computation andcommunication ability of the nodes in the system to let them test each other and then get thediagnosis results based on test results. The core problem of system-level fault diagnosis isto design efficient diagnosis algorithms based on test results.In this paper, we have studied the algorithm of the system-level fault diagnosis based onartificial immune method and proposed a new algorithm under the PMC diagnostic model.Firstly, according to the characteristics of the PMC diagnosis model,we have proposedan optimizing method to generate the initial population to improve the quality of the initialpopulation. Secondly, we have defined a new affinity function that considers the whole syn-drome of the fault set. Finally, we have added a shared structure to memorize the optimalantibody in the algorithm and given the new algorithm ?ow chat and variation method. Wehave also proven the correctness and convergence of the algorithm and provided experimen-tal results, which show that the CPU time and the number of generation are superior to theYang's algorithm.In order to further study the effiectiveness of the algorithm, we have applied our algo-rithm to a kind of interconnection network--Cross-cube which is a variant of the hypercubeand was proposed by Haq. We have proved that the t-diagnosable degree of n-dimensionalCross-cube is (n + 1) (n≥4). Experimental results show that our algorithm is effiectiveand applicable. On the other hand, we have also proven that the t₁/t₁-diagnosable degree ofn-dimensional Cross-cube is (2n-2) (n≥4).