Performance Modeling of Virtual Machines Hosted on Xen

DOI: 10.5176/978-981-08-5837-7_223

Authors: Shing H. Doong, Chih C. Lai, Shie J. Lee, Chen S. Ouyang, Chih H. Wu

Abstract:

Virtualization is a key technology in cloud computing to render on-demand provisioning of virtual services. Xen, an open source paravirtualized virtual machine monitor (hypervisor), has been adopted by many leading data centers of the world today. A scheduler in Xen handles CPU resources sharing among virtual machines hosted on the same physical system. The default scheduler in current Xen release (3.1) is the Credit scheduler, which is non-preemptive, supports both work-conserving and non work-conserving modes, and does a global load balancing scheduling in multi-core computers. Credit uses two parameters (weight and cap) to fine tune CPU resources sharing. Previous studies have shown that these two parameters can impact various performance measures of virtual machines hosted on Xen. In this study, we present a holistic procedure to establish performance models of virtual machines. The calculation power (‘cal’) and network throughput (‘netperf’) of virtual machines were simulated under various settings of weight and cap in an environment similar to the intended use of virtual machines. We then employed a powerful machine learning tool (multi-kernel support vector regressions) to learn the ‘cal’ and ‘netperf’ performance models from the simulated data. These models were evaluated satisfactorily by using established procedures in machine learning.

We further used non-dominated sorting genetic algorithms (NSGA-II) to optimize the two-objective problem resulted from the ‘cal’ and ‘netperf’ models. Any Pareto solution from NSGA-II offers users a choice to statically load balance virtual machines by setting proper Credit parameters.