FDIO: A Feedback Driven Controller for Minimizing Energy in I/O-Intensive Applications

The relatively low utilization of servers in data-center environments when running I/O-intensive applications is a key concern for efficiency. Energy optimization, by throttling power consumption, is an essential operational goal. Since processors are the most demanding of the components constituting a server, energy optimization has focused on regulating processor consumption. However, more recently memory and storage are increasingly becoming more demanding, collectively accounting for more than 40% of the overall energy consumption in typical system configurations. We argue that this trend necessitates tracking overall energy consumption rather than focusing on any single component. Although currently only processors expose energy-related controls at a fine granularity, we demonstrate that with a more holistic approach we can obtain significant efficiency benefits. Specifically, our feedback-based controller for Linux detects I/O-intensive phases in workloads, and adjusts processor operating frequencies accordingly, in a more effective manner than the standard CPU governors.