Predicting Latency Quantiles using Network Calculus-assisted GNNs

Network digital twins commonly rely on Graph Neural Networks (GNNs) as functional models. They typically predict network performance metrics, such as latencies. Most approaches have one of the following restrictions: they use simulated data, predict mean values, or don't utilize formal method results as inputs. We introduce an approach that: (I) relies on data obtained from a hardware testbed, increasing realism, (II) predicts quantiles in addition to means, increasing flexibility and applicability, (III) uses the formal method of network calculus to obtain input features, increasing prediction accuracy. We show that latencies in hardware testbeds can be predicted at different quantiles with median relative errors between 8% and 29% using a simple GNN architecture. Furthermore, we show that network calculus bounds are especially useful for predicting higher quantiles and that they mostly correct large prediction errors.