Integrated CPU-GPU Power Management Research For Android Mobile Games

Smartphones grow rapidly in number and improve a lot in performance,which satisfies users' demand but intensifies the contradiction between energy-consumption and battery-capacity.Dynamic Voltage&Frequency Scaling(DVFS)is an effective mechanism to alleviate the contradiction.But traditional strategies depend on hardware feedback to evaluate processors' workloads and work on totally separate CPU/GPU,failing to utilize the unique characteristics of applications to improve energy efficiency.This thesis is guided by 'application-characteristics driven power management of processors' and focuses on android games,trying to find out the characteristics of android games which can be adopted to direct the synergistic CPU-GPU power management while gaming.Contributions of this thesis are as follows:1.Build the power management framework for android games.The framework,taking account of android features and game features,connects the application layer and lower DVFS logic in vertical,helping the DVFS logic feel the workload change of games in time,and connects the CPU and GPU DVFS modules in horizontal,providing support for synergistic CPU-GPU DVFS.The framework is the basis of the following work.2.Propose the User-Interaction-Aware CPU workload estimation and Frame-Complexity-Based GPU workload estimation methods.New methods utilize the message-number caused by user-interaction and frame data(vertex,texture and command)to estimate CPU and GPU workloads separately,helping to improve the promptness and accuracy of workload estimation while gaming.3.Design a synergistic CPU-GPU DVFS logic for android games.The logic targets for good Quality of Service(QoS)and high energy-efficiency.In this logic,the change of QoS and CPU/GPU workloads are described with hierarc-hical state machine to specify the time and direction for DVFS,the 'Producer-Consumer,relationship between CPU and GPU is fully considered.All work above is implemented on ODROID-XU3 platform.In evaluation,the ratio of average processing time(performance)to energy-consumption(energy)of a frame,Q,is used as the criterion to evaluate the energy efficiency.We achieve average 12%,top 34%improvement of Q with good QoS.