Study On Adaptive Traffic Measurement For Source Routing Over Protocol Oblivious Networks

Routing method takes a huge factor affecting the performance of data plane in SDN networks.Traditional routing methods like LPM and EM not only cause lookup latency,but also may trigger the explosion problem of flow tables,since the large demands for flow entries,resulting in the cost of communication,computing and space resources.Source routing is lightweight and efficient which is different from traditional routing.In source routing,the source switch encapsulates routing information in the packet header,then intermediate switches only match the corresponding outgoing port field to forward packets.There only needs one flow entry for each port of the switch to forward all flows statelessly.Most existing OpenFlow-based source routing methods have issues of flexibility and scalability because of occuping protocol fields of the packet header.Leave aside the advantage of greatly reducing the number of flow entries,there exists difficulty in classifying and counting flows in source routing networks.And if the flow counting task is placed in the source switch,it will undoubtedly impose an excessive burden on the switch.Regarding the issues above,in this thesis,we propose and implements an adaptive flow counting mechanism named FlowWatcher,which is applied in source routing networks based on POF.FlowWatcher detects and marks elephant flows in hosts by monitoring socket buffer.The data plane adopts POF to implement protocol-oblivous processing pipeline.And the control plane count network traffic from two granularities.Firstly,fine-grained elephant flow counting;secondly,coarse-grained mouse flow counting.At the same time,it uses Skip List structure to store the elephant flow statistics and Binary Trie to manage the mouse flow statistics separately.FlowWatcher proposes an adaptive mech-anism to adj ust thresholds of counting elephant flow and aggregate rules of monitoring mouse flow,in order to get statistics of the main traffic in network and respond to traffic changes.FlowWatcher uses a real-time linear prediction model to predict the trend of mouse traffic,and if there detects an abnormal flow,it will adjust from temporal and spatial dimensions.Finally,we implement POF-based source routing in DCNs and val-idate the feasibility and effectiveness of FlowWatcher through a series of functional and performance experiments.