Abstract
For more than a decade, the power consumption of data centers has been addressed from different perspectives. Many solutions have been proposed to reduce (or optimize) this power consumption, such as controlling the operation of the servers in data centers. However, these approaches have not yet reached their optimum goals. Existing power control solutions using CPU frequency with an ad hoc or frequency modulator approach are not sufficient. In this paper, we review the power consumption effects of different configuration settings applied to the server's CPU. We propose our local power controller using frequency scheduling (LPC), which is a server-level power controller that depends on an extended gain scheduling technique. Our proposed LPC considers the impact of different CPU configuration settings that are typically not considered simultaneously, such as the allocated CPU credits and CPU frequency level. Through a real experimental test bed, our LPC exhibits effective power management of different types of machines and outperforms other existing approaches, such as ad hoc and frequency modulation, when the power budget is low. Moreover, our proposed LPC has a very lightweight control actuation overhead compared with other approaches: approximately of the ad hoc approach's overhead and of the frequency modulator approach's overhead. This lightweight control actuation overhead reduces the power consumption overhead caused by the controller, and it could be used by other controllers, such as performance or thermal controllers running on the same server.