Learning Based DVFS for Simultaneous Temperature, Performance and Energy Management

Hao Shen1,  Jun Lu2,  Qinru Qiu1
1Syracuse University, 2Binghamton University


Dynamic voltage and frequency scaling (DVFS) has been widely used for energy reduction in the modern processors. How to select the optimal frequency that minimizes energy dissipation for the given performance constraint at runtime is a nontrivial problem. The problem becomes more complicated if temperature needs to be constrained (or minimized) simultaneously. The temperature, performance and energy have different nonlinear relationships with frequency/voltage scaling ratio and this relationship is closely related to the characteristics of hardware and applications. In this paper, we design a reinforcement learning algorithm to tackle the problem of simultaneous temperature, performance and energy management. The proposed approach allows continuous tradeoff among these three quality measurement of a computer system. It also enables us to set two of the measurements as constraints and optimize the third one. The proposed approach is validated on an Intel Core 2 processor running Linux system.