Research And Implementation Of Heterogeneous Computing Method Based On Open Source Wireless Communication

With the increasing demand for communication services and scenarios,the society has entered the era of "Internet of Everything".However,new scenarios and applications have increased the complexity of wireless communication networks,and the closed-source nature of commercial networks has made it difficult to obtain data for Artificial Intelligence(AI)model training,and there is a delay conflict between the real-time requirements of wireless tasks and AI methods.The computing resources are heterogeneous and ubiquitous.The above problems make it difficult to make the network intelligent.Therefore,this thesis addresses the above problems by investigating three aspects:network architecture,wireless AI task scheduling,and real-time guarantee,respectively.First,this thesis delves into the approach of achieving endogenous intelligence from the perspective of network architecture.To address the difficulties of data acquisition and AI deployment,this thesis proposes a network architecture design based on open source wireless communication,and proposes a heterogeneous computing engine mainly for the implementation of intelligent Radio Access Network(RAN),which realizes the functions of wireless AI such as input,output,model management,deployment,wireless AI task scheduling and real-time guarantee through four modules of the engine.An asymmetric task sequence generation algorithm is designed for heterogeneous multitask scheduling of wireless AI,which makes full use of the characteristics of different task types in multitasking to achieve parallel scheduling of different types of tasks in order to improve scheduling efficiency and reduce task time overhead.Based on the task scheduling sequence generated by this algorithm,heterogeneous multitask scheduling is implemented in combination with genetic algorithm,which performs well on the implemented engine.In order to guarantee the real-time performance of wireless AI,this thesis proposes a multi-threaded polling mechanism used in the heterogeneous computing engine,which is able to reconstruct the data correlation broken by time delay through the joint collaboration of multiple threads and an action buffer.The theoretical basis of this mechanism is mathematically defined and derived to enable its use in existing deep reinforcement learning algorithms,and its effectiveness in time-delay scenarios is demonstrated by a wireless AI use case.