Traffic Classification And Scheduling Algorithms In DiffServ

With the continuous development of network technology and the continuous prevalence of user terminals,both the number of application types and network traffic is also increasing rapidly.On the one hand,it facilitates and enriches people’s life.On the other hand,it also brings unprecedented pressure to the current network infrastructure.However,improving the carrying capacity through network expansion and equipment update still can not meet the increasing demands for the network resources.How to guarantee the service quality of different applications under the limited resources has become an urgent problem to be solved.To solve the aforementioned problems,a Differentiated Service(DiffServ)network framework based on Software Defined Networking(SDN)is constructed in this dissertation.A traffic classification algorithm based on variational autoencoder is proposed to classify the network traffic.And domain adaptation is leveraged to deal with the problem that the classification accuracy decreases caused by the difference in cross-domain traffic features distributions.Finally,a priority-based queue management method using deep reinforcement learning is adopted to schedule the classified traffic differentially and make full use of the network resources.The main research contents are summarized as follows:1)For the problem that traffic labels are scarce in the practical environment and traditional flow-based classification methods do not apply to the online task,a semi-supervised traffic identification algorithm based on variational autoencoder is proposed.By extracting features from the network packets automatically and classifying these features,the proposed classification algorithm can be leveraged to identify the traffic with only a few labels.Finally,the performance of the proposed algorithm is proven to be effective on open-source datasets.2)For the problem that classification accuracy decreases due to the difference in cross-domain traffic features distributions,a cross-domain traffic classification algorithm using unsupervised domain adaptation is proposed in this dissertation.By constructing a feature transformation module to implement both marginal and conditional distribution adaptation to achieve cross-domain traffic identification.Finally,there are several experiments conducted on cross-domain traffic datasets to prove the effectiveness of the proposed algorithm.3)For the difficulty in allocating the network resources dynamically which aims to achieve differentiated services on multi-queue management,we put forward a priority-based queue management algorithm based on deep reinforcement learning.By evaluating real-time network states at the sender port,a deep reinforcement learning algorithm is leveraged to make decisions on bandwidth and queue cache resources allocation to provide differentiated services for the traffic with different priorities.Finally,the simulation experiment results prove that the scheduling algorithm can schedule the traffic differentially and make full use of the network resources.