Beyond Mean: Spatio-Temporal Modeling of Queue Utilizations and Flow Latencies Using T-GNNs

Network planning and control require precise, reliable, and dynamic digital network models to easily obtain performance metrics. One central performance metric in any network is the end-to-end latency of connections which can be inferred from queue utilizations along its path. Models take a variety of forms: simulation, emulation, stochastic and deterministic formal methods, and machine-learning-based or -assisted approaches. Simulation and emulation require either too much computational time or too many hardware resources, while formal methods often have a high computational complexity leading to poor scalability. Machine-learning-based methods scale better to larger problem spaces, however, current approaches mainly concentrate on mean performance metric predictions. We show that such an approach can be extended to predict queue utilization and end-to-end latency behavior over time in dynamic networks. This is achieved by utilizing Temporal Graph Neural Networks (T-GNNs) which can model spatio-temporal dependencies. The approach achieves a mean queue utilization error of 5.5% and a flow-level end-to-end latency MARE of 5%-55% depending on time resolution over 100 random topologies. We show that this approach outperforms a non-temporal, static Graph Neural Network (GNN) on the same task in terms of capturing dynamic network behavior such as queue build-up and draining. The approach performs similar to related work while increasing flow rates by up to three orders of magnitude—this improvement is bought with a trade-off in supported scheduling mechanisms and traffic patterns. Our results show that such a T-GNN approach can be useful for performance modeling of high data rate flows in dynamic networks.