Research On Resource Allocation Of Data Flow Sampling In Software Defined Networks

With the explosive growth of data traffic,the packet-level sampling is not competent to realize the network-wise sampling under the situation of high data volume and high-speed forwarding link.In addition,in order to improve the network security analysis,application identification,service quality assurance,the system should pay more attention to the collection of each flow in the global network,as well as the acquisition of flow-level depth information.However,the Deep Packet Inspection technology or the current measurement method has its own defects in global continuous packet samplin.Therefore,under the limitation of collector bandwidth resources,a new method is needed to achieve global,high-precision flow-level sampling.In this paper,based on Software-Defined Networking(SDN)technology,we proposed a flow-awareness framework.What's more,we proposed an optimization strategy which combines an adaptive candidate sampling node selection algorithm and deep reinforcement learning(DRL)algorithm to select sampling nodes and allocate sampling resources.By the central control of SDN and the OpenFlow group table,the strategy realizes the global continuous packet sampling with complete payloads.The adaptive selection algorithm combines the flow betweenness centrality and greedy mechanism,which is beneficial to the centralized resources and reduce the interference of the same flow to the neural network training.Then,it can select several vital candidate nodes.The DRL algorithm which uses Deep Double Q-learning Network(Double DQN)can solve the problem of high-dimensional srate space in real environment,and then dynamically allocate sampling bandwidth resources to sampling nodes,so as to improve the global sampling accuracy both of the elephant flows and mice flows.Then,under the flow-awareness framework,we design experiments according to several factors that affect the sampling accuracy.Experimental results show that our strategy has imporved sampling accuracy than the traditional algorithm.In summary,based on SDN and DRL,we propose a two-step data flow sampling strategy: 1)adaptive candidate node selection,and 2)dynamic sampling resources allocation by Double DQN.The two-step method can collect the continuous packets with their poyloads,and improve the sampling accuracy,especially for mice flows.