Energy Conservation for Servers P. Bohrer, D. Cohn, E. Elnozahy, T. Keller, M. Kistler, C. Lefurgy, R. Rajamony, F. Rawson, E. Van Hensbergen IBM Austin Research Lab mootaz@us.ibm.com Introduction Power management has been thoroughly studied for applications running on battery-powered platforms [3,5]. We take the position that power management is equally important for server environments, where high performance and reliability have traditionally been the most important design and evaluation factors. This position breaks with tradition and argues for considering energy consumption on the same footing. We base our opinion on several observed trends in the technology and the marketplace. Technology trends for server-class processors predict ever increasing performance at the expense of a rising energy budget [4]. Recent advances also have allowed manufacturers to pack and integrate unprecedented numbers of transistors on a chip, with a corresponding increase in power consumption and cooling requirements. These technology trends coincide with a growing trend in the marketplace to “consolidate” computing services into large data centers, which use the economy of scale to amortize ownership and management costs over a large number of machines. A typical data center thus deploys hundreds or thousands of computers densely packed to maximize floor space utilization, providing the customer with a more cost-effective approach than the alternative of operating the same services in-house. Server consolidation, however, pushes the limits of power supply and cooling systems in a data center. Anecdotal evidence from data center operators already points out to the large contribution of power consumption and cooling to operation costs, and the intermittent failures of computing nodes in densely packed systems due to insufficient cooling. Given these trends, the case for reducing the energy consumption in servers becomes clear: Energy-efficient servers can be packed in denser configurations than inefficient ones, leading to better utilization of floor space, lower energy bills and cooling requirements. These factors thus have a direct financial impact on the successful operation of a data center. . Additionally, by reducing the power requirements for data centers, energy-efficient servers reduce the need for new power generation and distribution facilities and allow the more rapid deployment of servers without waiting for utility company upgrades. Server Workloads and Power Management Servers are typically configured with sufficient processing capacity to handle the expected maximum workload. Additional capacity may also be planned in clustered environments for high availability, where it may be necessary to distribute the load of a failed server over the functioning ones. But actual workloads vary widely depending on the time of the day, time of the year, and nature of the application among other factors. For example, a Web server that provides banking services will likely be idle during nighttime, while a file server farm in an academic campus tends to be busy until the early morning hours. Likewise, retail Web sites in the western world experience their maximum load only during the last two months of the year. Previous work corroborates these observations[2].