Research On Fault Diagnosis Approaches Of System Symptom In Multicores Processor

As the continuous improvement of the processor’s miniaturization andintegration, processor system becomes more complex, accompanied with theincreasing requirement of reliability and fault tolerance. Therefore, to establish anaccurate fault diagnosis model becomes essential. Using the method of establishinga simulation environment for processor system, injecting fault, capturing symptomsreaction in fault injection process, analyzing of experiment data, we can create amodel of fault diagnosis, to improve the fault identification and diagnosticscapabilities of CPU.In this paper, we research on the fault diagnosis method of system levelsymptoms. We use the full system simulators and co-simulation platform which areprovided by the SUN’s OpenSPARC open source project. Then we build simulationenvironment, inject fault, and analyze the data, observe the fault propagation andperformance in structure, operating system and application, establishing faultdiagnosis model.First of all, we analyze characteristics of SAM, a full system simulator, inorder to study the hierarchy and multi-core simulation of Ultrasparc T2, runningdriving mode and the configuration of instruction set. We analyze the faultpropagation mode in different hierarchies, design the fault injection platform of fullsystem, and conduct the experiment. In addition, we study the co-simulationplatform. With the use of Verilog PLI, we implement a co-simulation based faultinjection platform, and conduct the experiment. Then, we make the featureextraction and statistical analysis of the experimental data. We search thepropagation of the fault and the symptoms distribution. Finally, we use BP neuralnetwork, RBF neural network and PNN diagnosis model to conduct the faultdiagnosis. Meanwhile, we compare and analyze the performance of the threedifferent models.