Research On Traffic Scheduling Technology In Software Defined Networking

Software-defined networking(SDN)is a new type of network architecture that separates the forwarding plane from the control plane.The SDN controller can perform corresponding flow control according to the network topology information and network traffic information of the entire network.The service requirements can be customized on demand through programming,and the underlying network hardware is only responsible for completing the forwarding function.How to implement an efficient traffic scheduling strategy under this centralized control mode and improve the resource utilization of the network has become a hot research topic.Aiming at the huge number of traffic matrices in the actual network,scheduling the traffic matrix at each moment is very resource-consuming.The issue of deciding when and how to switch scheduling strategy during the evolution of the traffic matrix is discussed.In this paper,a new multi-dimensional traffic matrix clustering method is proposed considering the similarity of traffic values and the similarity in the time domain.This method can effectively extract the key traffic matrix,and then use the coarse-grained scheduling strategy and fine-grained scheduling strategy to schedule the critical traffic matrix.The simulation results show that the coarse-grained scheduling strategy significantly reduces the number of policy switching,and the fine-grained scheduling strategy improves the link utilization.Most of the researches on traditional TE strategies focus on constructing and solving mathematical models.To reduce computational complexity,an experience-driven trafic allocation algorithm based on multi-agent reinforcement learning is proposed.It can effectively distribute traffic on pre-calculated paths without solving complex mathematical models so that fully utilize network resources.The algorithm performs centralized training on the SDN controller,and can be executed on the access switch or router in a distributed way after the training is completed.Frequent interactions with the controller are avoided at the same time.The simulation results show that the proposed algorithm is effective in reducing the end-to-end delay and increasing throughput of the network with respect to the shortest path(SP)and the equal-cost multi-path(ECMP).