| With the rapid growth of the number of delay-sensitive services(such as VR),deploying the services in the edge server can effectively reduce the overall computing delay,minimize the workload sent to the cloud,save bandwidth and enhance data privacy.However,some factors make it a challenge to deploy services in an edge environment.On the one hand,edge nodes in a cluster may be heterogeneous in terms of CPU,memory,and resource availability.This requires more lightweight virtualization technology,which makes it possible to deploy services quickly and efficiently in the edge computing environment.On the other hand,the type,size,arrival rate and complexity of the user’s request packet vary,and the user may change the location during the request.This requires more efficient resource allocation methods to adapt to the rapidly changing services needs.When allocating resources for delay-sensitive applications in edge nodes,delay is an important performance index to evaluate the strategy.At present,the service delay model often ignores the processing delay and queuing delay of packets passing through the container service,so it is difficult to obtain the real-time end-to-end delay of service request flow.The research shows that the deep reinforcement learning algorithm has good performance in the face of complex edge environment information.Therefore,to solve the above problems,this paper proposes an edge container resource allocation module for time sensitive services.It mainly includes the following contents:(1)In order to improve the accuracy of packet end-to-end delay evaluation,an edge container queuing model based on M/D/1 is proposed in this paper.The process of each service flow reaching the container is approximated as a Poisson process.It mainly considers the difference of packet arrival rate of different services and the impact of processing delay and transmission delay of the previous container on the next container.The delay mainly includes three aspects:processing delay and queuing delay in the container,and transmission delay on the transmission link.(2)In order to meet the resource allocation requirements of delay sensitive services,this paper proposes an edge container resource allocation method based on deep reinforcement learning.Under this method,the average end-to-end delay of packets and resource consumption are used as the optimization objectives of resource allocation problem.This paper uses the idea of multi global model to improve the original A3C algorithm.Each working model thread obtains more search space under different strategies,which is more conducive to find a resource allocation strategy closer to the optimal strategy.(3)In order to simulate and test the delay sensitive container intelligent resource allocation algorithm more quickly and conveniently.This paper designs a delay sensitive container resource allocation simulation module based on docker technology.The module mainly includes edge container task release module,resource allocation decision module and simulation result display module.Finally,the simulation module is developed through web technology,tested and displayed.The experimental results show that the improved A3C algorithm based on multi global model has better delay and resource consumption performance than other algorithms,and it can obtain a better resource allocation scheme in a larger search space;The delay sensitive container resource allocation simulation module can meet the user’s requirements for algorithm simulation after functional test. |
