Prediction of performance anomalies in web-applications based-on software aging scenarios

The topic of this paper is about prediction of performance anomalies caused by software aging. We have developed a framework for detection of performance anomalies that is targeted to web and component-based applications. In this study, we selected some amount of historical data previously collected and we conducted a correlation analysis with this data. The resulting dataset was then submitted to some Machine-Learning (ML) classification algorithms. The best algorithms were selected according to the accuracy and precision. In a second step, we induced some synthetic aging scenarios (memory leaks and CPU contention) in the application and we tried to do estimation of the system parameters by using time-series analysis. With the estimated values we conducted a classification with the three previous ML algorithms. From the initial results we observed that combining the estimation of parameters supported by time-series models with ML classification techniques provides some good results on the prediction of performance anomalies. We also observed that there is no single ML algorithm that can be applied effectively to predict the response time for all the web-transactions.