Power proportional characteristics of an energy manager for web clusters

Energy consumption is a major issue in data centers operating 24 hours a day, 7 days a week. The power dissipated by a web cluster is not proportional to the numbers of incoming requests if only DVFS (Dynamic Voltage Frequency Scaling) is used. This is because of the nonlinear power efficiency of DVFS, the large load fluctuation in web services and the typical CPU utilization rates of a server. This paper presents a system level controller making a cluster of low-power servers power proportional by managing the resources on the platform. Our controller uses sleep states to switch on or off CPUs in order to continuously match the current workload with the system capacity. Methods from control theory are used to drive the CPUs from and into sleep states. The power proportional characteristics of the proposed energy manager are studied for different workload patterns. Results from system simulation show that power proportionality is obtainable but only with appropriate parameters set on the controller.