Power Measurement And Modeling For ARM Android Applications Based On DVFS Technology

As the electronic information technology progresses at a rapid pace, the architecture complexity and hardware performance of the embedded processor are improved constantly. But the power dissipation problem follows as well. For mobile devices represented by the smart phone, higher power dissipation will shorten the work time inevitably, resulting in the worse user experience. On the other hand, Android has already been the main operating system owing to its powerful features and open source. Therefore it is necessary to study the power dissipation of the embedded processor in application scenarios of Android system.A model of the processor's power dissipation has been established on the basis of artificial neural network in this thesis, aiming at predicting the dynamic power dissipation of the processor without the hardware measurement equipment. The frequencies of the relevant performance events are the input of the model, while the output is the predicted power dissipation of the processor. The neural network is trained by the measured data, with the quantitative relation of the two are obtained through the fitting procedure. Firstly, the performance events are filtered by the Spearman correlation coefficient, among which the most related performance events are chosen to be the input of the proposed model. Then the structure of the neural network is designed. The number of network layers, the number of nodes and train method are confirmed as well. Finally, the train to the proposed model is accomplished based on the sample data of the measured power dissipation. Piecewise function is utilized to analyze the performance of the processor under different voltages and frequencies. Every kind of voltage and frequency has the corresponding neural network.Eight typical Android applications, including WeChat, Baidu Map,360 mobile phone defender and so on, are used to verify the accuracy of the proposed model. From the reference to the experiment results, it is obviously that the average relative error of the proposed power dissipation model owns a max value of 13.1%, while the minimum value reaches 10%. These results show that the proposed model can realize the power dissipation prediction very well.