Application-Independent DVFS Method Based on OS Metrics Monitoring for Cloud Servers

This paper proposes a Dynamic Voltage and Frequency Scaling (DVFS) method for cloud servers that is application-independent and minimally impacts performance. As cloud computing continues expanding, power saving of its servers is one of the most critical issues. Prior power-saving methods through scheduling and configuration optimization based on application characteristics are promising. However, these methods are not suitable for use on cloud servers. Since server administrators and application operators differ in cloud computing, server administrators do not have knowledge of individual applications. On the other hand, the application operator has no control over much of the configuration of the server machine. For these characteristics of cloud servers, Linux CPUfreq governor, one of the de facto implementations of DVFS, is a promising technology because it uses only generic information: CPU usage. However, our preliminary experiments showed that the Linux CPUfreq governor significantly degrades performance in specific system configurations. This paper proposes a DVFS method that suppresses performance degradation by detecting it by monitoring metrics obtained from the operating system while using existing CPUfreq governors. A significant advantage of our method is that it can be applied to cloud servers without requiring application-specific implementation or configuration. Experiments with a web server running on a container showed that the proposed method balances power savings and application performance. It reduced CPU power consumption by up to 3.8% while degrading throughput by less than 0.2% compared to the case where the frequency is fixed at maximum in a configuration where a pure CPUfreq governor significantly degrades performance.