Research And Implementation Of Flow Correlation System Based On Deep Neural Network In Data Center Network Environment

With the development of cloud computing,data centers,as its important infrastructure,are increasing both in number and scale.The data center network is critical for connecting resources and sharing them with users.Through the network,a variety of applications,such as data storage,file download,can be supported.Therefore,network reliability is the key factor to ensure the normal operation of the data centers.In order to improve network reliability,operators need to diagnose network faults effectively to find and deal with network problems in time.As a result,it is necessary to obtain the full view of the network and flow correlation,a traffic analysis technology widely used in anonymous networks,can obtain the full view of the data center network by correlating flows hop by hop and has become a feasible method.There are two main steps in flow correlation--feature extraction and flow matching.Existing work has conducted certain research on these two steps,but the direct application of flow correlation in the data center network environment still has the following challenges: 1)the difficulty of the feature selection caused the heterogeneity of network function nodes;2)the low matching accuracy rate caused by the heterogeneity of flows passing through data center networks.Therefore,how to apply flow correlation to the data center network environment and effectively improve its accuracy is still a problem that needs to be studied,and this thesis starts with the two steps of flow correlation to improve both the effectiveness of feature selection and the accuracy of matching.The specific content includes the following:Firstly,in the feature extraction step,the current feature selection algorithms do not consider the heterogeneity of network function nodes,which leads to the difficulty of feature selection.Thus,this thesis proposes a feature selection method based on multi-source and multi-criteria fusion to select the optimal feature subset according to the feature selection criteria such as relevance,immutability of ingress and egress flows and consistency from original feature set to better reflect the characteristics of different network function nodes and then improve the accuracy of flow correlation.Secondly,in the feature matching step,the current methods do not take the heterogeneity of data center network traffic into consideration which leads to the low matching accuracy of flow correlation.Thus,this thesis proposes the attention-based deep neural network with Siamese architecture,which solves the flow matching problem through the Siamese network and improves the correlating accuracy by introducing the attention mechanism in the model to reflect the attention of different nodes to different flow characteristics.Finally,on the basis of the above two theoretical studies,a deep neural network-based flow correlation system in the data center network environment is designed and deployed on HPE cluster in Southeast University to verify the theoretical results of this thesis.The experimental results show that the flow correlation system proposed in this thesis can quickly and accurately relate the ingress and egress flows in different network function nodes.What's more,under the premise of ensuring certain accuracy,there is no too high time overhead of the proposed system.In summary,the feature selection method based on multi-source and multi-criteria fusion and the attention-based deep neural network with Siamese architecture proposed in this thesis provide effective solutions for flow correlation in the data center environment,which can be widely used in fault diagnosis,location and recovery of data center network with important theoretical and practical application value.