Article Preview
TopIntroduction
With the rapid development of 5G communication technology, big data mining, and cloud computing, network connectivity is evolving toward a closer, faster, and more valuable(Peng et al., 2020). At the same time, the explosive growth of network traffic has brought great challenges to traditional networks. New Internet services are becoming more complex and increasingly demanding to manage. However, the traditional network with a decentralized structure has exposed many drawbacks in infrastructure, network services, and resource scheduling when facing these novel Internet applications. To meet the changing needs of future networks and address the shortcomings of traditional networks, communication networks must be fundamentally innovated to cope with the impact of the proliferating Internet applications on the networks in the changing environment.
Software-Defined Network (SDN)(Nunes et al., 2014) has a revised network architecture, as shown in Figure 1. From bottom to top are the data forwarding layer, the control layer, and the application layer. Compared with traditional networks, SDN decouples the network control and forwarding, making network routing and transmission rules deployed by SDN controllers. Open interfaces allow applications to invoke network resources through programming, and users can flexibly configure networks and definition rules, all of which are unique advantages of SDN in future network development.
Network traffic prediction is of great importance in high-performance and architecture-innovative SDN (Kexin et al., 2021). The prediction can help network maintenance personnel to make traffic predictions in real-time, and determine whether there is a security risk, thus preventing hacker attacks; it can also help network maintenance teams to adapt traffic distribution through accurate traffic prediction and rational deployment, thus avoiding network failures, message loss, and troubleshooting. Therefore, the research of traffic prediction based on the SDN network will be more informative.
At present, there are more mature traditional methods and deep learning methods for network traffic prediction, and graph neural networks have also been introduced into the research of traffic prediction problems. However, some problems and limitations are still found through the research results,and the summary problems are as follows.
- 1.
Graph neural network prediction models based on time-series data are mostly applied to traffic flow data, and the distance between intersections is often used as the representation of the adjacency matrix, which is because there is a large connection between distance and traffic flow and speed. However, for SDN traffic data, the distance between nodes has a negligible effect on traffic speed, so such a method does not apply to SDN network traffic data.
- 2.
A single graph neural network in the prediction problem of traffic data, the temporal features tend to be weakened with the enhancement of spatial feature extraction, thus reducing the accuracy of SDN traffic prediction.
To solve the above problems, a model named SDN traffic prediction based on mutual information is proposed in this article, which can learn the implied Spatio-temporal features by constructing Graph Build Layer (GBL) with Spatio-temporal convolutional units. Then enable the task of SDN traffic prediction deploying on SDN controllers. In this article, the research on SDN traffic prediction mainly has the following contributions and innovative works: