Predicting web server crashes: a case study in comparing prediction algorithms

Traditionally, performance has been the most important metrics when evaluating a system. However, in the last decades industry and academia have been paying increasing attention to another metric to evaluate servers: availability. A web server may serve many users when running, but if it is out of service too much time, it becomes useless and expensive. The industry has adopted several techniques to improve system availability, yet crashes still happen. In this paper, we propose a new framework to predict time-to-failure when the system is suffering transient failures that consume resources randomly. We study which machine learning algorithms build a more accurate model of the behavior of the anomaly system, and focus on Linear Regression and Decision Tree algorithms. Our preliminary results show that M5P (a Decision Tree algorithm) is the best option to model the behavior of the system under the random injection of memory leaks.
Postprint (published version)