The Time Dimension in Predicting Failures: A Case Study

Abstract

Online Failure Prediction is a cutting-edge technique for improving the dependability of software systems. It makes extensive use of machine learning techniques applied to variables monitored from the system at regular intervals of time (e.g. mutexes/s, paged bytes/s, etc.). The goal of this work is to assess the impact of considering the time dimension in failure prediction, through the use of sliding windows. The state-of-the-art SVM (Support Vector Machine) classifier is used to support the study, predicting failure events occurring in a Windows XP machine. An extensive comparative analysis is carried out, in particular using a software fault injection technique to speed up the failure data generation process.