Joulemeter: Virtual Machine Power Measurement and Management
- Aman Kansal ,
- Feng Zhao ,
- Jie Liu ,
- Nupur Kothari ,
- Arka Bhattacharya
MSR-TR-2009-103 |
MSR Tech Report
The importance of power management has led to most new servers providing power usage measurement in hardware and alternate solutions exist for older servers using circuit and outlet level measurements. However, the power measurement and management capability is severely handicapped when the servers are virtualized because virtual machine (VM) power cannot be measured purely in hardware. We present a solution for VM power metering. We use low-overhead power models to infer power consumption from resource usage at runtime and identify the challenges that arise when applying such models for VM power metering. We show how existing instrumentation in server hardware and hypervisors can be used to build the required power models on real platforms with low error. The entire metering approach is designed to operate with extremely low runtime overhead while providing practically useful accuracy. We illustrate the use of the proposed metering capability for VM power capping leading to significant savings in power provisioning costs that constitute a large fraction of data center power costs. Experiments are performed on server traces from several thousand production servers, hosting real-world applications used by millions of users worldwide. The results show that not only the savings that were earlier achieved using physical server power capping can be reclaimed on virtualized platforms, but further savings in provisioning costs are enabled due to virtualization.