Research On Data Center Network Modeling And Optimization Algorithm Based On Machine Learning

With the rapid development of internet,the demand for high quality of service(QoS)is increasing all the time.As the main carrier of network flows,there are huge challenges for data center networks in network resource management.In order to optimize the QoS performance indicators of network flows(such as delay,jitter,packet loss ratio,etc.),it is necessary to model the transmission characteristics of the network.That is to say,only by understanding the complex internal relations among various variables such as network topology,traffic intensity and routing configurations,modeling the network and evaluating the candidate routing configurations of flows by the network model obtained after modeling,can we know how to decide the routing paths of network flows and improve their end-to-end QoS performance indicators.Traditional methods of network modeling are mainly based on heuristics,which usually need to apply existing prior knowledge and expert experience to model the network system through mathematical methods.Then the problems will be solved by approximate algorithms or solvers,and the obtained results could provide guidance for routing optimization.However,these methods are difficult to accurately describe the system characteristics in real network scenarios,and the calculation time is too long to meet the requirements of rapid deployment in the actual network environment.In response to this problem,some scholars have proposed a SDN-based network modeling algorithm,Route Net,which draws on the idea of message passing neural network.And this algorithm can capture the potential transmission characteristics of the network through message passing between the two entities of the link and the path.However,the algorithm does not take into account the importance of different links in the path.In some complex network scenarios,the performance of the model will be bad.On the other hand,because of failing to capture the correlation among the various QoS performance indicators,the algorithm needs to train multiple neural network models for the average delay,jitter and packet loss ratio.In summary,this thesis addresses the above shortcomings and further solves the network modeling and optimization problems of data centers.The main contributions of this thesis are summarized as follows:(1)Aiming at the problem of data center network management,this thesis gives a formal definition of the network modeling problem and provides a general framework for the prediction of network QoS performance metrics based on supervised learning simultaneously.In addition,this thesis also provides multiple use cases of network optimization based on the predictions ?of QoS indicators obtained by the network model;(2)Aiming at the problem of network modeling studied in this thesis,when predicting the end-to-end QoS performance indicators of each path,regarding the correlation between the paths and the links,and considering the different influences of different links in the path on the routing performance of the entire path,this thesis proposes a network modeling algorithm that combines message passing neural network with a self-attention mechanism.The algorithm can distinguish the importance of each link to the end-to-end routing performance indicators of the path and characterize the attributes of links with different congestion levels,so as to generate an effective embedding vector for the path state;(3)In order to solve the task of predicting multiple QoS performance indicators,on the basis of the above neural network model,this thesis proposes a network modeling algorithm based on multi-task learning.In the network model of multiple outputs,the output of each network performance indicator is connected through cross-layers,which strengthens the complementarity and fusion of the features.At the same time,this thesis also introduces a mechanism to learn the weight of the multi-ask loss function automatically,thereby solving the problem of uneven training difficulty among multiple tasks and improving the accuracy of model prediction;(4)According to the typical data center network topology of fat tree and VL2,this thesis designs a network simulation environment based on OMNe T++ and simulates the network scenario of the data packets' sending and receiving.Then this thesis uses the samples generated by the simulation to verify the effectiveness of the algorithm through experiments.Experimental results show that the algorithm proposed in this thesis can improve the prediction accuracy of QoS indicators such as average delay,average jitter and packet loss ratio of network flows,and the network model can be directly applied to real network optimization scenarios.