Assessing the Impact of Virtualization on the Generation of Failure Prediction Data

Abstract

Fault injection has been successfully used in the past to support the generation of realistic failure data for offline training of failure prediction algorithms. However, runtime computer systems evolution requires the online generation of training data. The problem is that using fault injection in a production environment is unacceptable. Virtualization is a cheap sand boxing solution that may be used to run multiple copies of a system, over which fault injection can be safely applied. Nevertheless, there is no guarantee that the data generated in the virtualized environment can be used for training the algorithms that will run in the original system. In this work we study the similarity of failure data obtained in the two scenarios, considering different virtualized environments. Results show that the data share key characteristics, suggesting virtualization as a viable solution to be further researched.