【摘要】 数据中心网络中的流可以分为大象流和老鼠流,预测流类型是实现优化调度各种流的基础,而现有方法在开销、准确性、预测时间等方面都有或多或少的缺点.为此,利用深度学习具有的刻画多维度特征的能力以及软件定义网络(SDN)具有的全局集中控制的优势,提出了"边缘预分类+中心精分类"两级大象流预测机制.该机制包括以下步骤:首先,利用随机森林技术筛选出流在3个维度(时间分布特征、流的实时特征、数据包头部特征) 10个用于构建预测模型的特征.然后,部署在SDN交换机上的预分类模型使用残差网络算法+带代价敏感性质的Softmax交叉熵损失函数,过滤掉大部分老鼠流.最后,部署在SDN控制器的精分类模型使用残差网络算法+Additive M argin Softmax交叉熵损失函数,准确地识别出大象流.面向公开数据集的实验表明,当流的第5个包到达时,所提机制的召回率可达91%,准确率可达93%,开销低至0.1kbps,预测时间低至7ms.与Flow Seer、ESCA、NELLY等现有主要方法对比,所提机制的各评价指标均有改善,马修斯相关系数MCC是NELLY的2.52倍,开销降低到ESCA的0.046%,预测时间减少到Flow Seer的0.35%.更多还原

【Abstract】 In the data center network,there are elephant flow and mouse flow. Accurately predicting flow type is the key to achieve optimal flow scheduling,while the existing prediction methods have some shortcomings,such as pot great accuracy,high overhead Jong prediction time and so on. Therefore Jbased on the multidimensional feature characterization ability of deep learning and the ability of software-defined networks(SDN) centralized controlling the network Jthe two-evel elephant flow prediction mechanism of "edge preclassification + center fine classification" is proposed. The mechanism includes the following steps: first,the time distribution characteristics of flow,the real-time characteristics of flow,and the characteristics of packet head are screened out by the random forest algorithm. Then Jthe pre-classification model deployed on the SDN switch at network edge uses the residual network algorithm + Softmax cross-entropy loss function with cost-sensitive properties to filter out most of the mouse flows. Finally,the fined classification model deployed in the SDN controller uses the residual network algorithm + Additive Margin Softmas cross-entropy loss function to accurately identify the elephant flow. Experiments from the public data set show that,when the fifth packet of a flow arrives,the proposed method’s recall could reach to 91% Jthe accuracy could reach 93%,the cost was 0.1 kbps,and the prediction time was 7 ms. Its performance also is better than the existing schemes(such as,FIowSeer,ESCA and NELLY). The Matthews correlation coefficient was 2.52 X to NELLY Jthe prediction time was reduced to 0.35% of FlowSeer,and the overhead was reduced to 0.046% of ESCA.更多还原