Improving Enterprise VM Consolidation with High-Dimensional Load Profiles

Abstract

Modern enterprise data centers take advantage of virtual machine consolidation to allocate virtual machines to virtualized servers to increase energy efficiency. One key problem is to minimize the number of virtualized servers required while maintaining service quality. A promising approach is to exploit recurring load patterns exhibited by enterprise VMs for increased allocation efficiency. This paper shows that bin packing heuristics can deliver the same allocation quality as integer linear programs if calculation time is constrained. There were no significant differences between vector bin packing heuristics in simulations based on CPU load profiles obtained from enterprise data centers. We further show that consolidating in clusters of a few hundred virtual machines is sufficient as solution quality does not improve with larger clusters.