Energy Consumption Evolutionary Optimization At GCC Compile Time Based On Bayesian Network And Random Forest

Energy consumption is a kind of important quality attribute in embedded systems.Up to 80% of energy consumption in embedded systems is reported to be directly related to software execution activities.Therefore,in the case of outdoor embedded system where the power is insufficient and the battery replacement is inconvenient,reducing the energy consumption of embedded software has a more important role.In recent years,the selection of GCC(GNU Compiler Collection)compilation options has become a hot research topic in energy consumption optimization.It has been proposed that lower energy consumption executable code can be obtained by compiling embedded software source code under a given execution platform by choosing a set of optimal compilation options.However,several challenges remain.The GCC compiler provides a large number of compilation options,forming a huge and discrete selection space.Meanwhile,there are potentially complex effects among compilation options and between compilation options and energy consumption.Then the heuristic information of frequent option mode is incomplete and the timeliness is not good.Less attention is paid to the ordering information of the energy consumption target values.These issues potentially pose considerable challenges in terms of improving search efficiency,optimizing quality,and time-consuming of target value evaluation.To solve the above problems,this paper presents an evolutionary optimization method BN-EDA(Bayesian network Estimation of Distribution Algorithms)for GCC compilation based on Bayesian network and random forest.The contents of this study are as follows:(1)A pattern mining method with frequent energy labels is presented.This method reduces transaction table size by replacing reference points and transaction tables one by one.On this basis,a complex compiled option mining algorithm is proposed to obtain more heuristic information.The strategy of generation by generation mining is helpful to maintain the timeliness of frequent option modes,which efficiently improves the convergence speed.(2)A method for energy consumption optimization based on Bayesian networks and random forests is presented.The method is designed to effectively capture potential complex effects among compilation options and between optimization goals and compilation options.Effective modeling algorithm based on Yes Net is designed to support the selection of potentially optimal compilation options under optimization objectives.The concept of order is introduced into stochastic forests to build energy consumption prediction models to help improve solution quality and reduce energy consumption evaluation time,taking into account size ordering information between energy consumption objectives.(3)An evolutionary optimization algorithm for GCC compile-time energy consumption is designed.The heuristic information reflected in energy consumption labeling for a single frequent option is mutated in a single point,which does not fully utilize the heuristic information such as the number of times and energy labeling for multiple frequent options and it is inefficient to mutate.GCC compile time energy evolution optimization algorithm is presented.Further,the multipoint mutation method with maximum frequent pattern matching help is used to improve the optimization quality and speed up convergence.(4)The framework of GCC compile time energy evolution optimization algorithm is constructed.The frequent option pattern mining methods,evaluation methods for Bayesian network and random forest,and compile-time energy consumption evolution methods proposed in this paper are combined to effectively be applied to the integrated development environment so as to facilitate the development of related research.(5)The experimental comparison validation is designed.The BN-EDA method and GA-FP(Genetic Algorithm Frequent Pattern)algorithm are compared in eight cases of different sizes.The results show that the BN-EDA method in this paper can obtain better solution quality and convergence speed.In the evolution process,the random forest prediction model incorporating sequence not only maintains high prediction accuracy but also effectively reduces the evaluation of energy consumption.