Design And Implementation Of Network Flow Classification Method In Data Center Based On Neural Network

With the booming development of emerging technologies such as cloud computing and artificial intelligence,data centers are facing the challenges of business diversification,scale expansion and traffic complexity.The network traffic in data centers is exploding,the elephant flows occupy the precious bandwidth of data center networks,and the mice flows can easily be crowded out by the elephant flows,which leads to the increase of transmission delay and even affects the business performance.Therefore,in order to effectively improve the bandwidth utilization,reduce the transmission delay and improve the application performance,it is necessary to accurately identify elephant flows and mice flows in real time,so that the subsequent transmission can be regulated timely.However,the relationship between network flow attributes is complex,and the traffic scenarios in data centers are variable.Faced with the challenges of complexity and the demands of intelligence of classification,traditional classification methods are gradually replaced by machine learning methods,but it still faces the problem of poor real-time performance.For this purpose,this thesis proposed an online classification method of network flows in data centers based on neural network.The main research contents are as follows:Firstly,this thesis designed and implemented a neural network classifier based on FPGA.Because the online classification method requires that the training of the neural network model has a high speed,FPGA can fully utilize the advantages of their parallelism when combined with neural network.Therefore,this thesis designed and implemented an FPGA-based neural network classifier under two different excitation functions of the hidden layer.We discussed the overall structure design and hardware implementation details of each sub-module concretely,verified the usage of hardware resources and compared the model training time and model accuracy between FPGA-based and CPU-based implementation through simulations and experiments.Secondly,this thesis designed and implemented the online classification method of elephant flows and mice flows based on neural network.On the basis of the realization of the FPGA-based neural network classifier,according to the characteristics of network flows in data centers,this thesis designed an online classification method for elephant flows and mice flows which obtains classification features from the attributes of the first few packets of a flow,and implemented the high-speed data transceiver,packet access and frame field analysis,flow attribute acquisition,statistical feature calculation and related functions based on FPGA hardware platform.By designing experimental scenarios,this classification method was applied to the classification of elephant flows and mice flows in three different traffic distribution scenarios in data centers and the adaptive classification of elephant flows and mice flows based on dynamic changes of traffic,to verify its classification accuracy,real-time performance and adaptability.The results show that the classification accuracy can reach up to 99.8%,and the recognition ability of elephant flows is 100%;after the dynamic changes of traffic,the classification accuracy can be improved to more than 90%in a short time.