CEA-GNN: Performance Prediction of Hybrid Network Traffic with Deterministic and Best Effort

The development of the Internet and the mutual promotion of people's needs have posed deterministic demands on the network. But accurately and quickly evaluating the performance of deterministic network systems has become an important obstacle to the promotion and application of deterministic networks. Traditional queuing theory, deterministic network calculus, and machine learning can achieve this goal, but they lack the adaptability for the characteristics of deterministic network systems. The research presented in this paper mainly aims to propose and implement a deterministic network system performance evaluation method based on Graph Neural Network with Complex-Enhanced Attention(CEA-GNN), seeking to achieve accurate and rapid performance evaluation of deterministic network systems and facilitate their implementation and large-scale deployment. This network model starts with the characteristics of deterministic network systems, conveying essential information to the graph neural network through the attention mechanism, and then extracts information from both the graphical spatial domain and the complex frequency domain according to the actual meaning of the graph for GNN updating. Comparative experiments demonstrate that compared to graph neural networks designed for general networks, our method achieves a 3.6-fold increase in speed, while still controlling the relative error to around 3%, and also obtains similar results across multiple different real-world topologies, exhibiting a certain degree of robustness.