Predicting Cloud Applications Failures from Infrastructure Level Data

Cloud computing assumes a crucial role in the internet applications and services landscape, serving as the default infrastructure for deploying and delivering web applications. To harness the cloud’s technical and financial benefits, one must address the challenges of the increased consequences of software faults. Online failure prediction has emerged as a solution allowing for actions to be taken before the failures. However, conventionally, each target application requires a specific online failure predictor to be trained from application-specific data, which is impractical. In this paper we explore and assess the possibility to achieve online failure prediction for cloud applications from infrastructure-level data, i.e., data from the hypervisor. For this purpose, we present a method to accomplish this type of failure prediction, through fault injection on the hypervisor and machine learning techniques to create failure prediction models. From our experiments we find that it is possible to predict application failures from low-level infrastructural data, i.e., the infrastructure data presented enough application failure predictive power (with an accuracy of 96%, precision of 95% and a recall of 64%), showing the great potential that infrastructural data has to contribute to this task. Additionally, our study also opens interesting future research directions, such as multilevel failure prediction or generic failure prediction models.