Profile-Based Energy Reduction for High-Performance Processors

To reduce the energy consumption of modern processors, designers have proposed many energy-saving techniques. In many cases, these techniques are dynamically activated and deactivated. In systems that employ these techniques, to adapt to changes in application behavior, profiling can help determine how to manage the activation of techniques to improve a certain metric. In this paper, we propose to use a global approach to such profiling. The idea is to profile the application on a sectionby-section basis but to make the decisions in a global manner, competitively comparing the different sections. Under such conditions, the system can be more effective. For example, we can target for reduced energy consumption in an application without slowing it down. The results show that such an approach can reduce the energy consumption of 7 applications by 12% with negligible slowdown.