Intelligent Classification And Scheduling Of Network Traffic For Differentiated Services

With the rapid development of technologies such as the Internet of Things and the internet,as well as the ever-increasing variety of residents’online activities,the scale of network traffic and the number of network application types have dramatically expanded in recent years.The limited network infrastructure gradually becomes inadequate in serving such a vast amount of network traffic.Moreover,the contradiction between providing equal services to different applications and the diverse service requirements of different applications in traditional IP networks has become increasingly prominent under the current scarce network resources,causing network congestion and increased transmission delays.Considering the differences in expected network services among various network applications,how to achieve efficient utilization of network resources and balance the differentiated service requirements of different network applications under the current limited network resources has become an urgent problem that needs to be solved.To address the above problems,we conduct research from two aspects of network traffic intelligence perception and scheduling,and design a network traffic intelligent classification and scheduling architecture for differentiated services.The goal is to achieve differentiated scheduling of network traffic based on accurate perception of network traffic.In this architecture,considering the emergence of new application types,open-set network traffic classification models are deployed on network edge exit gateway devices to classify known application types during model training while identifying unknown application types.For unknown application types detected by different nodes,a distributed incremental updating method is adopted to quickly share unknown application knowledge among different network traffic classification nodes.To address the differences in network services required by different network applications,corresponding cache queues are allocated for different applications in the backbone network forwarding devices,and bandwidth and cache resources are dynamically allocated to different applications based on the real-time network state,thus achieving efficient utilization of network resources.The main work and innovation can be summarized as follows:1.Architecture Design for Intelligent Classification and Scheduling of Network Traffic:Considering the variations in required network services for different network applications,We present an architecture for intelligent classification and scheduling of network traffic that caters to differentiated services.The proposed architecture integrates a network traffic-aware plane into various edge gateway devices to label outbound traffic with application types.Additionally,an intelligent scheduling plane is incorporated into the backbone network switching devices to provide customized network services for different network application traffic.Taking into account the real-time,high-bandwidth network traffic on forwarding devices in actual networks,high-bandwidth network traffic feature extraction and online parsing are achieved using packet cross-kernel techniques based on the Data Plane Development Kit(DPDK)and multi-core parallel processing techniques based on CPU-GPU collaboration.This architecture provides an effective platform support for subsequent research on real-time classification and scheduling algorithms for network traffic.2.Research on open-set network traffic classification method:In response to the constantly emerging unknown application traffic in actual networks,we propose an open-set network traffic classification method based on a discriminative variational autoencoder,decoupling the network traffic classification problem in the open-set into two related sub-problems,including known application classification and unknown application detection.By extracting potential features with high discrimination on the original traffic,the known application classification module divides the samples into a known application type based on the potential features and provides the corresponding classification confidence.The unknown application detection module verifies the credibility of the classification results of the known application classification model,and based on this,the final classification result is decided by integrating the confidence given by the known application classification model.The proposed method accomplishes the mutual promotion of known application classification and unknown application detection performance,and achieves more accurate open-set network traffic classification.3.Research on distributed incremental update of classification models:In order to address the problem of unknown application traffic detected by various distributed network traffic classification nodes,considering the low efficiency issue caused by data island effect of the current single-point incremental update method,we propose a distributed model incremental update method based on federated incremental learning,which realizes the fast sharing of unknown application knowledge among different network traffic classification nodes.This method mainly includes two stages:local model update and global parameter federated aggregation.During the local model update stage,distributed detection nodes only participate in a small amount of typical sampled historical data and all new data,and use the old model with rich historical application classification experience knowledge to guide the update of the new model,so that the new model can quickly approach the old model and remember historical data.By considering the amount of historical data and the amount of new unknown application type data contained in different detection nodes,the model parameters after local incremental update by different nodes are federated aggregated globally,achieving fast global update of network traffic classification models.4.Research on differentiated network traffic scheduling:In order to address the differences in expected network services for different application types,we start from the perspective of application queue management and assigns corresponding cache queues to different network applications in backbone network switching devices,modeling the traffic scheduling problem that distinguishes applications as an active queue management problem.By measuring and evaluating the personalized service requirements of different network applications,and dynamically managing the bandwidth and cache resources of different queues according to real-time network status,network resources can be efficiently utilized and personalized service requirements of different applications can be met as much as possible.A queue resource management method based on SAC(Soft Actor-Critic)algorithm is proposed for complex and time-varying network environments,modeling the reward function based on the personalized requirements of different applications in terms of throughput,delay,and packet loss,and making bandwidth and cache allocation decisions for the next time slot based on the current queue network status.The experimental results of the research presented demonstrate that the designed intelligent classification and scheduling architecture for differentiated services in network traffic,as well as the proposed methods,effectively utilize network resources and improve the overall quality of application experience.Therefore,in the context of continuous growth in network traffic,the proposed framework and methods have broad application prospects.