in a data center network, accurately classifying flow is the key to optimal schedule flow. However, the existing classification methods cannot meet the demand of real networks in terms of classification performance, detection latency, and control overhead. Thus, by combining the ability of deep learning to describe the multi-dimensional features, and the advantage of software-defined networking (SDN) in centrally controlling the network from a global viewpoint, this paper proposes a fine-grained flow classification method. This paper uses random forest technology to select eight important features in three dimensions—time distribution feature of flow, real-time feature of flow and packet header feature—for the classification model. First, this paper proposes a 2-classification scheme with a two-level architecture of pre-classification and exact-classification to detect elephant/mice flows. The pre-classification model using deep residual learning + A-Softmax with cost-sensitive is deployed on a SDN switch at the network edge to filter out a large number of mice flows. The exact-classification model using deep residual learning + AM-Softmax is deployed on the SDN controller to accurately identify the elephant flows. Second, this paper proposes a 4-classification scheme based on gated recurrent unit (GRU) to detect elephant/cheetah/tortoise/porcupine flows. Finally, the experiment results show that, when the 5th packet of a flow arrives, the 2-classification scheme can achieve a recall of up to 97.31%, an accuracy of up to 93.6%, a control overhead of 0.1kbps, and a detection latency of 7 ms. At the same time, the 4-classification scheme can achieve a recall, an accuracy, a false positive rate (FPR), a Kappa, a control overhead, and a detection latency of 83.58%, 86.53%, 5.02%, 0.797, 0.61kbps and 7.5 ms on an average, respectively. Compared with the existing methods (FlowSeer, NELLY and ESCA), all performance measures are improved to different degrees. At the same time, we also confirm the generalization of selected features and of the designed flow classification model.

在数据中心网络中，准确分类流是最佳计划流的关键。然而，现有的分类方法在分类性能，检测等待时间和控制开销方面不能满足实际网络的需求。因此，通过结合深度学习描述多维特征的能力以及软件定义网络（SDN）从全局角度集中控制网络的优势，本文提出了一种细粒度的流分类方法。本文使用随机森林技术从三个维度中选择八个重要特征：流的时间分布特征，流的实时特征和数据包头特征。第一，本文提出了一种具有预分类和精确分类两级体系结构的2分类方案，以检测大象/小鼠的血流。使用深度残差学习+ A-Softmax并具有成本敏感性的预分类模型部署在网络边缘的SDN交换机上，以过滤掉大量的鼠标流。使用深度残差学习+ AM-Softmax的精确分类模型被部署在SDN控制器上，以准确识别大象流。其次，本文提出了一种基于门控循环单元（GRU）的4分类方案，以检测大象/猎豹/乌龟/豪猪的流量。最终，实验结果表明，当流的第5个数据包到达时，2分类方案可以实现高达97.31％的召回率，高达93.6％的准确率，0.1kbps的控制开销，检测延迟为7毫秒。同时，四分类方案可以实现召回率，准确性，误报率（FPR），Kappa，控制开销以及检测延迟为83.58％，86.53％，5.02％，0.797、0.61平均分别为kbps和7.5 ms。与现有方法（FlowSeer，NELLY和ESCA）相比，所有性能指标都有不同程度的提高。同时，我们还确认了所选功能和设计的流量分类模型的一般化。所有绩效指标都有不同程度的改善。同时，我们还确认了所选功能和设计的流量分类模型的一般化。所有绩效指标都有不同程度的改善。同时，我们还确认了所选功能和设计的流量分类模型的一般化。