Performance Management of Virtual Machines via Passive Measurement and Machine Learning

Virtualization is commonly used to efficiently operate servers in data centers. The autonomic management of virtual machines enhances the advantages of virtualization. For the development of such management, it is important to establish a method to accurately detect performance degradation in virtual machines. This paper proposes a method that detects degradation via the passive measurement of traffic exchanged by virtual machines. Using passive traffic measurement is advantageous because it is robust against heavy loads, nonintrusive to the managed machines, and independent of hardware/software platforms. From the measured traffic metrics, performance state is determined by a machine learning technique that algorithmically determines the complex relationship between traffic metrics and performance degradation from training data. Moreover, the feasibility and effectiveness of the proposed method are confirmed experimentally.