A Machine Learning Based Mapping Approach For Heterogeneous Multi-processors System

With the advancement of technology and the development of times,embedded systems are used in almost every method of our lives,such as home electronics,communications,automotive and avionics.From these applications,we can observe the new trend of embedded system development.Firstly,functional diversification.Secondly,low power consumption and high performance.Thirdly,adaptive.In order to meet requirements of the above embedded systems,traditional single-core processors have been limited by power density,and it has been difficult to meet these demands by increasing the clock frequency of the processors.Heterogeneous multicores Processors(Heterogeneous MultiProcesses,HPMs)have become the mainstream solution for modern embedded systems.For heterogeneous multi-core processors,how to make full use of its parallelism to improve system performance and reduce power consumption,an important issue to be solved is the dynamic mapping(or scheduling)of tasks.The key to the dynamic mapping problem is to be able to evaluate the performance of the system's resource allocation and scheduling scheme.The traditional performance evaluation methods include the establishment of cost model and simulation,but they are subject to accuracy and efficiency.Restrictions cannot be widely used.A machine learning-based approach that automatically learns and analyzes system behavior and produces solutions that are designed to solve the problem of how to create a performance assessment that improves accuracy and efficiency.The burden of personnel.This paper analyzes the development trend and problems of current heterogeneous multi-core systems,focusing on the application of dynamic mapping of heterogeneous multi-core systems,and proposes an online mapping scheduling solution based on machine learning prediction model.First of all,this paper adopts two kinds of offline supervised machine learning methods: one is classification learning using support vector machine and adaptive enhancement technology,and the other is regression learning using artificial neural network,and three kinds of construction can be quickly predicted differently.The performance prediction model of the mapping scheme performance;then,for the dynamic mapping and scheduling problem of the runtime task,the performance prediction model and the genetic algorithm are combined to construct an online task dynamic scheduling method,and through experiments and optimal mapping methods and Common polling scheduling,sampling scheduling methods for analysis and comparison.