Slow down or sleep, that is the question

Abstract

Energy consumption has become a major concern for all computing systems, from servers in data-centres to mobile phones. Processor manufacturers have reacted to this by implementing power-management mechanisms in the hardware and researchers have investigated how operating systems can make use of those mechanisms to minimise energy consumption. Much of this research has focused on a single class of systems and computeintensive workloads. Missing is an examination of how much energy can actually be saved when running realistic workloads on different classes of systems. This paper compares the effects of using dynamic voltage and frequency scaling (DVFS) and sleep states on platforms using server, desktop and embedded processors. It also analyses workloads that represent real-world uses of those systems. In these circumstances, we find that usage of power-management mechanisms is not clear-cut, and that it is critical to analyse the system as a whole, including the workload, to determine whether using mechanisms such as DVFS will be effective at reducing energy consumption. 1