Abnormal behavior detection using resource level to service level metrics mapping in virtualized systems

摘要

Recent years have seen a pronounced growth in research interest in resource management of virtualized systems. In that context, several techniques have been proposed. However, this task is complicated by the heterogeneity and virtualization of resources and the scalability and high-performance service requirements of Cloud Computing. Moreover, unpredictable situations such as fluctuations in resource demand or abnormal changes in systems behavior due, for instance, to emergency situations or special events, cause critical service performance degradation and costly SLA violations. Existing approaches are not generic, and are based mainly on excessive allocation of resources, a prospect which is not efficient, and involves resource loss. Further, they propose monitoring systems without a validation, using correspondence between service levels and resource levels and pattern detection. To overcome these limitations, in this work, we propose an effective, dynamic and real-time detection of abnormal behavior. Our approach is based on outlier detection techniques and employs a mapping between service level metrics and resource level metrics, as well as periodicity detection and effective system state analysis and notification. Experiments conducted on several types of workloads showed that our algorithm is capable of detecting unusual changes in both service level and resource level metrics. It is also able to map detected changes and provide corresponding notifications with an accuracy ranging from 92% to 100%. (C) 2019 Elsevier B.V. All rights reserved.