An Integrated CPU-GPU Frequency Scaling Governor Based on Deep Recurrent Q-Network for Partially Observable Rendering Applications
Authors:
- 11 June 2023
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-5
Abstract
Dynamic voltage and frequency scaling (DVFS) is an effective approach to balance the performance and power consumption of the CPU-GPU-based video rendering applications. Operating System (OS) based DVFS governors are general in nature and thus lack of regulation effectiveness for workload with different characteristics. Reinforcement learning (RL) enhances power management ability for application-specific DVFS, whereas the regulation robustness is limited due to partial observability of complex SoCs. To mitigate this issue, in this paper, a deep recurrent Q-Network (DRQN) based DVFS governor for rendering applications is proposed where a recurrent neural network (RNN) exploiting historical information is embedded in a Deep Q-Network (DQN). Evaluated on the Nvidia Jetson NX platform with various settings, the proposed DRQN policy achieves an up to ×2.11 improvement in Valid Performance Per Watt (VPPW) compared with Linux default governors and shows superior regulation stability compared with other RL-based state-of-the-art.Keywords
This publication has 13 references indexed in Scilit:
- ML-Gov2017
- An adaptive on-line CPU-GPU governor for games on mobile devices2017
- Application-Specific Performance-Aware Energy Optimization on Android Mobile Devices2017
- Processors and Systems2017
- Redefining QoS and customizing the power management policy to satisfy individual mobile users2016
- Learning Transfer-Based Adaptive Energy Minimization in Embedded SystemsIEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2015
- Human-level control through deep reinforcement learningNature, 2015
- Integrated CPU-GPU Power Management for 3D Mobile Games2014
- Partially Observable Markov Decision Processes2012
- Adaptive power management using reinforcement learning2009