Investigation into the energy cost of live migration of virtual machines

One of the mechanisms to achieve energy efficiency in virtualized environments is to consolidate the workload (virtual machines) of underutilized servers and to switch-off these servers all together. Similarly, the workloads of overloaded servers can be distributed onto other servers for a load balancing reason. Central to this approach is the migration of virtual machines at runtime, which may introduce its own overhead in terms of energy consumption and service execution latency. This paper experimentally investigates the magnitude of this overhead. We use the Kernel-based Virtual Machine (KVM) hypervisor and a custom-made benchmark for our experiments. We will demonstrate that the workload of a virtual machine does not have any bearing on the power consumption of the destination server during migration but it has on the source server. Moreover, the available network bandwidth and the size of the virtual machine do indeed. introduce a non-negligible energy overhead and migration latency on both the source and the destination server.