VPower: Metering power consumption of VM

Abstract

While virtualization is giving great advantages in cloud computing, power metering for virtual machines (VMs) becomes one of the challenges raised. VM power can hardly be measured purely in hardware, so linear power models have been built to infer power consumption at runtime from resource usage. Linear models are simple and of low overhead in implementation. However, power consumption is not exactly linear with certain resource usage and linearity does not necessarily hold across multiple workloads. We propose a new method, VPower, which is effective for more workloads without increasing complexity. In this method, a two dimensional lookup table (LUT) is constructed for each VM, which gives an empirical power value at certain CPU utilization and last-level-cache (LLC) miss rate. Experiments are performed on VMs running SPEC CPU2006 benchmarks. It is shown that the VM power estimated reflects the variability of the two factors effectively, and the per-VM error is about 2.61W.