A multi-objective approach to virtual machine management in datacenters.

ABSTRACTVirtual machine (VM) consolidation has become increasingly important for improving efficiencies of resource usage and power consumption in datacenters. Previous work has addressed the problem of placing and replacing VMs in servers, in order to optimize resource management for different criteria, including performance, power and temperature. However, most work has focused on optimizing only one or at most two of these criteria, by separately managing either the platform layer (e.g., power and thermal management) or the virtualization layer (e.g., application performance management). In this paper, a cross-layer control system is proposed to manage the dynamic mapping of VMs to physical resources. The controller unifies the information from different layers to determine control actions such as when, which and where VMs need to be moved, optimizing multiple and potentially conflicting goals. A prototype of the proposed controller and two other competing ones (one without stabilization and the other using a single-objective approach) are implemented on an IBM BladeCenter. Experimental evaluations are conducted using a mix of types of workloads to emulate the variety and dynamics of datacenter workloads. The results indicate that the proposed multi-objective optimization with stabilization significantly reduces unnecessary VM migration by up to 80%, avoids unstable host selection, and also improves the application performance by up to 30% and the efficiencies of power usage by up to 20%.