Research And Implementation Of Probabilistic Fault Localization Algorithm For Dynamic System

With the rapid development of the Internet and a great deal of deployments of various enterprise applications, it is very important to quickly and accurately locate IT service failures in order to guarantee the normal operation of business applications and provide better service for customers. In large and complex networks, types of application failures in the upper protocol stack caused by a variety of different reasons have great uncertainty, that is, failure of these applications may be caused by other logical failures, but also may be caused by a physical failure. The relationship between these failures becomes more and more uncertain. Coupled with the observation noise in the network environment, the Internet has become a complex dynamic system, which a great challenge for traditional deterministic fault localization technique.As one of the most effective theoretical models of probabilistic inference, Bayesian network represents the uncertain relationship between the entities in dynamic system using the knowledge of graph theory and probability theory. At present, fault diagnosis techniques based on static Bayesian network has made considerable progress in algorithm complexity, accuracy and precision. However, it is difficult to exhibit good performance in dynamic system.After the research of fault diagnosis technique based on static Bayesian network including analysis of the network environment and the performance of every algorithms, this paper establishes a fault diagnosis technique based on dynamic Bayesian network which, given the set of failed end-to-end services, discovers the underlying root causes. In order to diagnosis for dynamic systems, this method improved the algorithms based on static Bayesian Network in many aspects. Such as introducing the time-slice information, the mechanism of analyzing relationship between adjacent time-slice to deal with the dynamics, noise filtering to deal with the spurious alarms, model simplification to adapt to large system, fault transmission and probability updating. We used dynamic Bayesian model to handle system dynamics and a simple tool to deal with noise. Our approximate algorithm has taken several measures to reduce the algorithm complexity in order to diagnosis for larges-scale networks. We implement simulation with JAVA and compare our algorithm with algorithm based on BN in accuracy, efficiency and time. The results show that the algorithm can run well and have good performance. This method makes full use of the historical data and current observations to estimate the current system state and complete the fault diagnosis.