Predicting live migration performance of virtual machines using machine learning

Abstract

Virtualization is a widely used technology these days as most of server computing environments are rapidly shifting to cloud computing. Live migration, one of the most compelling features in system virtualization, has been an active area of research. Attempts to predict migration performance were made, but most of those were limited to analytical approaches with relatively unstable prediction errors or not easy to extend to realistic environments as more parameters are identified and considered. In this thesis, a novel data driven approach based on the support vector regression method providing flexibility and extensibility in parameter selection is introduced to predict performance metrics such as total migration time, downtime and the total amount of transferred data, especially on QEMU which is hardware virtualization platform that is open-source and the method of this thesis is easy to adapt to various purposes. It will facilitate automated system administration with live migration more efficiently.