Research And Application Of Machine Learning In Dynamic Frequency And Voltage Scaling

With the increasing complexity of mobile device applications and processor power den-sity,power consumption and thermal issues become a major problem in the design of smart mo-bile processors.Dynamic voltage and frequency scaling technology(DVFS)is one of the most commonly employed and most effective system-level power management methods in modern processors.However,the current DVFS strategy in smart mobile device operating systems con-siders incompletely,which usually only considers the current load of the processor.It is difficult to achieve efficient and comprehensive power management by such strategy.On the other hand,the exclusive power management strategy deployed by desktop-level processors doesn't apply to mobile devices represented by smart phones.Researching on the DVFS technology of the ARM processor commonly used in mobile devices,we present the DVFS technology based on machine learning,which aims to establish an accurate and efficient DVFS technology.Firstly,we established a framework for the complete simulation of processors and DVFS switching.The framework employs the simulation tool chain composed of GEM5,McPAT and Hotspot.Then We focus on three aspects to improve the framework gradually:processor state modeling,genetic algorithm-based maximum power consumption solving algorithm,and reinforcement learning-based DVFS strategy.In order to grasp the state of the processor,we propose the processor state model in simula-tion,and make the actual measurement on the development board.The power prediction model estimates the current power based on the performance counters.Compared with the real power,the average error is 2.16%.The thermal prediction model employs the current temperature and power to predict the future temperature,with an average error of 0.83%.Load prediction model predicts future load based on LightGBM algorithm,and its average error is 3.35%.The proposed reinforcement learning-based DVFS strategy employs the DQN algorithm according to the processor state model to predict the best voltage-frequency pair at the future moment.This strategy is optimized from four aspects of algorithm details,state space,action space and reward function to solve the inherent defects of the DQN algorithm and improve the training speed and make prediction effect.In addition,the look-up tables is employed to replace the forward calculation of the neural network in the simulation,in order to reduce the running time cost.What's more,we evaluate our proposed strategy and compare with other four strategies,which achieves a performance improvement of 5.3%?7.3%.