论文部分内容阅读
Cluster systems are common platforms for both high performance computing and cloud computing.As the scales of cluster systems increase,failures become normal.It has been long recognized that (failure) events are correlated instead of independent.The impact of correlated failures on dependability is significant.Though we cannot infer causalities among different events without invasive approaches like request tracing,we indeed can find out correlations among different events through a data mining approach.When it is predicted that an error or an attack would occur,administrators can make appropriate action to avoid ordiminish damage. Logs are often a record of the execution activities in production systems where the data recorded is intentionally chosen by the developer as useful information.The data from the log files can be analyzed to guide product related decisions.Event log files are the most commonsource of information for the characterization of events in large scale systems.However the large size of these files makes the task of manual analyzing log messages to be difficult anderror prone. In this thesis we describe the techniques of parsing log,propose an algorithm which mines association rules by the improved Apriori algorithm,and present an approach for fast predicting events according to rules.As these logging events are ordered in time,direct application of the classical Apriori algorithm cant perform association analysis.In this paper,the Apriori algorithm is improved to adapt these orderly events and a new method named First Event Front Limit is proposed to deal with the problem caused by time-window overlap.The improved Apriori algorithm is able to analyze orderly events.Two kinds of tree structures based on the rule tree and status tree are presented.Through the trees,we can deal with a lot of logic,and make the prediction logic clearer,so the prediction process is more efficient and overcoming faults.Experimental results show that the prediction accuracy of themethod proposed in this paper and the error rate have improved significantly.