Cooperative Software-Hardware Power Management for DRAM

Energy is becoming a critical resource to not only small battery-powered devices but also large server systems, where high energy consumption translates to excessive heat dissipation, which, in turn, increases cooling costs and causes machine to become more prone to failures. DRAM is one of the most energy-consuming components in many systems. In this paper, we propose and evaluate a novel DRAM power management technique that exploits the cooperation between the system software and the memory controller hardware, in which the system software provides the memory controller with only a small amount of information about the current state of the system. This enables the memory controller to more intelligently react to the changing state in the system, and therefore, is able to make more accurate and more aggressive power management decisions. The proposed technique is evaluated against previously implemented power management techniques running synthetic, SPECjbb2000 [40] and various SPEC CPU2000 benchmarks [41]. Using SPEC benchmarks, we are able to show that the cooperative technique consumes 14.2–17.3% less energy than previously proposed hardware-only techniques, 16.0–25.8% less than software-only techniques, and 71.6–75.8% less than no power management.