Performance Optimization For Network Function Virtualization Based On Reinforcement Learning

Network Function Virtualization(NFV)decouples the traditional network functions based on special hardware through virtualization technology,to deploy them on the general servers in the form of software(virtual network functions).And then NFV forms the service function chain by linking the required software network functions,deployed on the heterogeneous environment,to complete user requests.However,limited by the processing power of the general servers and the deployment of the service chains,the improvement of NFV performance is an important problem to be optimized.Traditional performance optimization schemes based on various assumptions or preconditions and models have great limitations in the actual network environment,so it is necessary to design optimization methods that don't need prior knowledge and can control adaptively.Aiming at the related process of service function chain deployment in NFV,the performance of NFV is improved by combining virtual network functions processing optimization on nodes with service chain deployment optimization.For the optimization of the processing capacity of one or more virtual network functions deployed on one node,the strategy of introducing GPU to accelerate based on the existing research is used.Moreover,according to the characteristics that GPU is sensitive to the batch size of received packets,an offloading scheme adapted to the dynamic network is designed based on the reinforcement learning method.It can effectively reduce the processing delay of virtual network functions on the nodes.The experimental results show that compared with the related static offloading scheme,this method can significantly reduce the processing delay while ensuring the processing throughput.For the deployment process of the service function chains,it is necessary to make dynamic mapping decisions to maximize benefits in the face of dynamic network,real-time changes of node resources,and other complex environments.According to the characteristics of this problem,a dynamic deployment algorithm based on reinforcement learning is designed to reduce the overall cost of average transmission delay and deployment failure.Experimental results show that the algorithm is better than the other methods in reducing the overall cost of transmission delay and deployment failure,and is more effective for performance optimization.