Joint Optimization Of Cloud-Network-Edge Resources Based On Federated Learning

A variety of new intelligent applications have emerged with the rapid development of the Internet of Things(IoT)and artificial intelligence(AI),while the traditional cloud computing architecture cannot meet the requirements of computing,latency,and rate for the complex applications.Multi-access Edge Computing(MEC)came into being,which offloads some services to the edge,reducing the delay and the transmission load,and becoming the key technology of fifth generation mobile communication(5G).However,the multi-dimensional key performance indicators of the 5G networks(such as latency,spectral efficiency,energy efficiency,etc.)cannot be satisfied by the edge intelligent devices,because of the limited communication,computing,storage,and control resources.Therefore,it is urgent to explore the architecture and resource allocation algorithm for the integration of cloud-edge distributed intelligent computing and communication networks.In addition,as an emerging distributed intelligent algorithm,Federated Learning(FL)has the advantages of protecting users’ privacy and security,breaking the data silos,reducing the amount of data transmission,and improving the utilization of computing resources.To solve these problems,this thesis investigates the FL-based resource allocation algorithm for cloud,network,and edge.Firstly,a cloud-network-edge intelligent collaborative architecture is proposed to support the integration of cloud-edge computing and communication networks.Cloud-network-edge network functions(Network Function,NF)are orchestrated by Kubernetes to provide customized services.Then,a joint optimization of task offloading and wireless resources algorithm is provided,thereby improving the communication quality and computing efficiency.The details of this thesis are as follows:(1)A micro-service-based network architecture for intelligent collaboration of cloud,network and edge is proposed.Firstly,we provide a distributed network architecture with features of resources deeply decoupling,scalability and reconfigurability,and abstract the wireless network into "three domains,four layers and three planes".Three domains refer to the cloud domain,edge domain and network domain,where the cloud domain and the edge domain provide distributed intelligent computing capabilities for users,the information interaction and coordination between cloud and edge are realized by the network domain.Each domain contains four layers,i.e.,the application layer,function layer,virtualization layer and infrastructure layer.In addition,each domain contains a control plane,an intelligent plane,and a management orchestration plane,with a total of "three planes".Secondly,inspired by the idea of the training and inference process of distributed intelligent algorithms,and the micro-service architecture,the distributed intelligent algorithm is abstracted into several virtual network functions to realize artificial intelligence as a service.The Hypertext Transfer Protocol(HTTP)is used among the network functions to complete the acquisition and provision of the services.Then,a suitable template is designed according to the specific service requirements to complete the instantiation process,and perform realtime monitoring and management of the whole life cycle.Finally,a micro-service-based cloud-network-edge intelligent collaborative experimental platform relying on the open source software and hardware as well as self-controllable domestic servers is built.The experimental results show that the proposed architecture can improve the reliability of the intelligent plane in the system(such as the success rate of federated learning)and the network transmission load and business service delay can be reduced.(2)A federated learning-based Double Deep Q Network(DDQN)algorithm(FL-DDQN)is used in this thesis to solve the joint task offloading and cloud-network-edge resource allocation problem,which is modeled as minimizing the utility function(the weighted sum of the total delay and total energy consumption for all the users tasks to complete)while satisfying the maximum tolerable delay and the total resources constraints.Furthermore,we use FL-DDQN to solve this problem,with the advantages of fast convergence and high model training accuracy.For the task offloading problem,the state space includes the resources required for each task,the maximum tolerated delay,cloud-edge computing resources and radio resources.The action space includes the decision variants to indicate whether the task is processed locally,at the edge or in the cloud.The reward function is defined as the weighted sum of time and energy consumption.For the resource allocation problem,the state space includes the resources required by each task,maximum tolerable delay,task queue length of cloud-edge,transmission power and radio resources.The action space is the number of subcarriers.The reward function is defined as the time required to complete the task and weighted sum of energy consumption.Simulation results show that the proposed FL-DDQN algorithm reduces the total task completion delay and total energy consumption effectively,compared with the distributed DDQN and distributed Deep Q Network(DQN)algorithms.