Predicting Critical Problems from Execution Logs of a Large-Scale Software System

The possibility of automatically predicting runtime failures in large-scale distributed systems such as critical slowdown is highly desirable, since this way a significant amount of manual effort can be saved. Based on the analysis of execution logs, a large amount of information can be gained for the purpose of prediction. Existing approaches - which are often based on achievements in Complex Event Processing - rarely employ intelligent analyses such as machine learning for the prediction. Predictive Analytics on the other hand, deals with analyzing past data in order to predict future events. We have developed a framework for our industrial partner to predict critical failures in their large-scale telecommunication software system. The framework is based on some existing solutions but include novel techniques as well. In this work, we overview the methods and present initial experimental evaluation.

Keywords: Predictive Analytics, Complex Event Processing, predicting runtime failures, machine learning, execution log processing.