Research On Load Balance In Data Center And Traffic Prediction Based On Deep Learning

Data center is the center of data transmission,computing and storage.It centralizes all kinds of sofeware/hardware resources and business systems,such as Web services,search engine,on-line shopping,network game and MapReduce.Especially in recent years,data center networks evolve very fast due to the extensive use of virtualization technology and cloud computing.Many revolutions have happened in its composition,structure,function and scale.Data center network has a great difference from traditional wide area network in the deployment of application de-mands as well as the diversity of network traffic.These characteristics require new demand for data center's availability and reliability.To improve the performance of data center network,this paper do some research on network load balance stragtegy and accurate prediction for network traffic.On the one hand,for the traffic in data center network which is unpredictable,load bal-ance strategy based on packet-routing can reduce the network delay and improve the throughput of network.On the other hand,for the traffic which can be predicted,accurate prediction helps to decision-making of global optimal schedule with the network traffic matrix.In the first research project,based on the widely use of fat-tree topology in data center,we focus on the load balance algorithms which suit for fat-tree network and propose a routing algorithm called Global Round Robin(GRR)and its improved version Improved Global Round Robin(IGRR).By providing periodic connection configuration between top swtich and servers,this algorithm can distribute the packets uniformly into all the available paths in the upstream.In the downstream,self-routing strategy is adopted.Through theoretical analysis and simula-tion verification,we find that this algorithm can achieve good load balance performance under different network traffics.Meanwhile,the configuration of GRR and IGRR also reduces the complexity of structure and the hardware requirement for switches.In the second research project,we pay attention to predict traffic more accurate.With an ac-curate traffic prediction,we can generate routing strategy for the divinable traffic in advance and improve the throughout of the network as a result.In this paper,we abstract traffic prediction as a time series prediction problem.Compared with traditional prediction methods using mathemat-ical statistics,we adopt a network traffic prediction method which based on deep learning.We design a recurrent neural network(RNN)which can fit the characteristic of traffic well,and use this model to predict traffic in the next time unit based on the historical traffic data.According to the characteristic of the traffic dataset,we design a RNN model with 5 layers and 100 basic cells in each layer,and train the model using dataset with one-dimension feature and dataset with three-dimension features respectively.The result shows that the model has lower mean square error(MSE)than mathematical statistics methods,and the RNN model trained by data with three-dimension features shows better performance than data with one-dimension feature.