An Approach For Fine-grained Traffic Measurement In Programmable Networking

Traffic measurement refers to the statistics of traffic information in networks,and then provides input information for network management applications such as route optimization,intrusion detection,and network anomaly detection.Traditional networking architecture has some shortcomings,such as lack of flexible control capabilities and basic network measurement functions,resulting in high implement complexity,coarsegranularity measurement capability,and low measurement accuracy.In programmable networking paradigm,with the help of programmable switches,the customized network traffic measurement function can be flexibly deployed on switches,which paves the way for fine-grained traffic measurement.This thesis mainly studies the fine-grained flow measurement problem in the programmable network.In programmable networking paradigm,the computing and high-speed storage resources on switches are very limited,and the accuracy of traffic measurement on a single switch will decrease as the number of measurement flows increases.If a cooperative flow measurement mechanism is implemented among switches,so that each switch measures different flows,then the number of flows that need to be measured on a single switch can be effectively reduced,the utilization efficiency of measurement resources and the accuracy of flow measurement.Therefore,this thesis first studies the problem of cooperative measurement problem,and proposes an efficient measurement node allocation algorithm.Experimental results show that the algorithm can effectively enhance the cooperation between switches;while reducing the measurement load of each node,it improves the accuracy of traffic measurement and the accuracy of large flow detection.Existing work shows shown that the characteristics of network traffic generally follow the Pareto's distribution,that is,20% of the flow generates 80% of the network traffic,and this the 20% of the flows are so called the elephant flows,while the remaining 80% flows are called the ant flows.The impact of elephant flows on network link load and delay is much greater than that of ant flows,so identifying elephant flows is one of the main tasks of network traffic measurement.Existing large flow detection methods mainly rely on the results of flow measurement to identify elephant flows,which have poor real-time performance.In order to solve this problem,this thesis proposes a quick detection method for elephant flows based on window statistics and machine learning.This method caches the relevant information of the data packets in a small window,and uses the statistics of the data packets in the window to detect elephant flows using machine learning algorithms.Experimental results show that the method of elephant flow detection can accurately learn the characteristics of large flows in the network,and accordingly improve the efficiency of the use of idle measurement resources.Existing load-balancing algorithms requires fine-grained estimation of traffic information,which cannot be obtained until the very time when single measurement period has passed.This kind of latency reduces a lot the agility of load-balancing.Aimed at the point,this thesis proposes a load-balancing algorithm based on quick detection method for elephant flows.It firstly utilizes the quick detection method to distinguish elephant flows from mice flows on a single node.Given the identification results,we could determine forward strategy for elephant flows directly and thus ensure real-time response toward dynamic traffic in the network.Experimental results show that the algorithm has the advantages of quick response,ensures the real-time performance,incurs less split of flow and performs well load-balancing at the same time.