Power and Performance Estimation for Fine-Grained Server Power Capping via Controlling Heterogeneous Applications

Abstract

Power capping is a method to save power consumption of servers by limiting performance of the servers. Although users frequently run applications on different virtual machines (VMs) for keeping their performance and having them isolated from the other applications, power capping may degrade performance of all the applications running on the server. We present fine-grained power capping by limiting performance of each application individually. For keeping performance defined in Quality of Service (QoS) requirements, it is important to estimate applications’ performance and power consumption after the fine-grained power capping is applied. We propose the estimation method of physical CPU usage when limiting virtual CPU usage of applications on VMs. On servers where multiple VMs run, VM’s usage of physical CPU is interrupted by the other VMs, and a hypervisor uses physical CPU to control VMs. These VMs’ and hypervisor’s behaviors make it difficult to estimate performance and power consumption by straightforward methods, such as linear regression and polynomial regression. The proposed method uses Piecewise Linear Regression to estimate physical CPU usage by assuming that VM’s access to physical CPU is not interrupted by the other VMs. Then we estimate how much physical CPU usage is reduced by the interruption. Because physical CPU usage is not stable soon after limiting CPU usage, the proposed method estimates a convergence value of CPU usage after many interruptions are repeated.