Dynamic Graph Neural Network for Traffic Forecasting in Wide Area Networks

Tanwi Mallick, Mariam Kiran, Bashir Mohammed, Prasanna BalaprakasharXiv 2020 · arXiv

Abstract

Predicting traffic in Research Wide Area Networks (R-WANs) is a critical task for proactive resource management, yet it is complicated by the sporadic, bursty nature of scientific data transfers. This paper presents the Dynamic Diffusion Convolutional Recurrent Neural Network (DDCRNN), a novel architecture that models network connectivity as a dynamic graph. By calculating real-time traffic correlations to update the network's adjacency matrix, DDCRNN captures transient spatial-temporal dependencies that static models miss. Experimental results on ESnet traffic traces demonstrate that the model achieves a ~20% mean absolute percentage error (MAPE), significantly outperforming classical statistical methods and standard deep learning baselines.

Introduction

Proactive congestion management in Wide Area Networks (WANs) requires accurate traffic forecasting. However, Research WANs (R-WANs)—such as those operated by the Department of Energy—exhibit traffic patterns fundamentally different from the consumer internet. While consumer traffic follows predictable daily or weekly cycles, R-WAN traffic is dominated by massive, non-seasonal bursts originating from large-scale scientific experiments.

These "spiky" patterns render classical tools like ARIMA or standard LSTMs ineffective. Furthermore, the influence between nodes in these networks is not merely a function of physical proximity; it changes dynamically based on active data transfers. Consequently, a static representation of the network topology is insufficient for capturing the underlying traffic dynamics.

Methodology

The proposed DDCRNN architecture addresses these challenges through a multi-faceted approach to graph-based deep learning. The model treats the network as a graph where sites are nodes and fiber links are edges, but it moves beyond static connectivity.

The DDCRNN architecture integrates WAN topology and traffic time-series to generate dynamic correlations for an encoder-decoder graph model.

Key Components:

  1. Dynamic Adjacency Matrix: Instead of a fixed distance-based graph, DDCRNN computes a dynamic adjacency matrix using Pearson correlation coefficients between nodes for every input sequence. This allows the model to learn which sites are currently influencing each other.
  2. Diffusion Convolution and GRUs: The model employs diffusion convolution to capture spatial patterns of traffic movement and Gated Recurrent Units (GRUs) to model temporal dependencies.
  3. Nonautoregressive Forecasting: To prevent error propagation common in step-by-step (autoregressive) predictions, the model uses an encoder-decoder structure to forecast an entire multi-hour window in a single pass.

Experimental Setup

The study utilized real-world traffic traces from ESnet (the U.S. Department of Energy’s Energy Sciences Network) spanning the full year of 2018. The dataset includes 48 physical sites mapped to 96 nodes (accounting for incoming and outgoing traffic). Data was collected at 30-second intervals and aggregated into 1-hour windows for training (70%), validation (10%), and testing (20%).

Results and Discussion

DDCRNN achieved approximately 20% MAPE for multi-hour forecasts, outperforming ARIMA, Random Forest, and standard recurrent networks. A key finding was the importance of the forecasting strategy; the authors discovered that autoregressive models fail in R-WAN environments because a single prediction error during a traffic burst cascades through the entire forecasting horizon.

Comparison of R² scores over a 24-hour horizon shows that the nonautoregressive strategy maintains significantly higher stability than the autoregressive approach.

As shown in the performance analysis, the nonautoregressive approach maintains a much higher $R^2$ over a 24-hour horizon compared to the autoregressive strategy. Additionally, nodes associated with major experimental facilities (e.g., SLAC or Argonne) showed higher predictability, as their bulk transfers possess a consistent "signature" once initiated.

Conclusion

The DDCRNN model demonstrates that dynamic graph modeling is essential for navigating the volatility of research network traffic. While the model faces challenges with low-volume or highly random traffic (such as during facility maintenance), its ability to forecast major spikes offers a path toward automated, congestion-free routing. Future work will focus on integrating these forecasts into Software Defined Network (SDN) controllers for real-time traffic engineering.

Summary updated

Abstract

Wide area networking infrastructures (WANs), particularly science and research WANs, are the backbone for moving large volumes of scientific data between experimental facilities and data centers. With demands growing at exponential rates, these networks are struggling to cope with large data volumes, real-time responses, and overall network performance. Network operators are increasingly looking for innovative ways to manage the limited underlying network resources. Forecasting network traffic is a critical capability for proactive resource management, congestion mitigation, and dedicated transfer provisioning. To this end, we propose a nonautoregressive graph-based neural network for multistep network traffic forecasting. Specifically, we develop a dynamic variant of diffusion convolutional recurrent neural networks to forecast traffic in research WANs. We evaluate the efficacy of our approach on real traffic from ESnet, the U.S. Department of Energy's dedicated science network. Our results show that compared to classical forecasting methods, our approach explicitly learns the dynamic nature of spatiotemporal traffic patterns, showing significant improvements in forecasting accuracy. Our technique can surpass existing statistical and deep learning approaches by achieving approximately 20% mean absolute percentage error for multiple hours of forecasts despite dynamic network traffic settings.
BibTeX
@misc{mallick2020dynamic,
  author = {Tanwi Mallick and Mariam Kiran and Bashir Mohammed and Prasanna Balaprakash},
  title = {{Dynamic Graph Neural Network for Traffic Forecasting in Wide Area Networks}},
  year = {2020},
  eprint = {2008.12767},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2008.12767},
}
Figure