Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
539625 | Integration, the VLSI Journal | 2015 | 11 Pages |
•A dynamic power management framework based on reinforcement learning is proposed.•The framework is model-free and independent of pre-designed policies.•We perform learning and power management in a continuous-time.•The proposed framework can simultaneously consider power and performance.•An effective application-level scheduling is integrated for further energy saving.
This paper presents a hierarchical dynamic power management (DPM) framework based on reinforcement learning (RL) technique, which aims at power savings in a computer system with multiple I/O devices running a number of heterogeneous applications. The proposed framework interacts with the CPU scheduler to perform effective application-level scheduling, thereby enabling further power savings. Moreover, it considers non-stationary workloads and differentiates between the service request generation rates of various software application. The online adaptive DPM technique consists of two layers: component-level local power manager and system-level global power manager. The component-level PM policy is pre-specified and fixed whereas the system-level PM employs temporal difference learning on semi-Markov decision process as the model-free RL technique, and it is specifically optimized for a heterogeneous application pool. Experiments show that the proposed approach considerably enhances power savings while maintaining good performance levels. In comparison with other reference systems, the proposed RL-based DPM approach, further enhances power savings, performs well under various workloads, can simultaneously consider power and performance, and achieves wide and deep power-performance tradeoff curves. Experiments conducted with multiple service providers confirm that up to 63% maximum energy saving per service provider can be achieved.