RIGEL: Real-time Optical Anomaly Diagnosis with Stateful In-Network Inference based on Distributed On-switch GNNs

The rapid growth of data-intensive applications has significantly complicated optical network management, making real-time optical anomaly diagnosis an essential capability. Existing approaches predominantly rely on centralized data analytics, which introduce inherent latency and overhead due to message exchanges between data and control planes. These limitations hinder the ability to promptly detect and localize anomalies, potentially leading to service degradation and network failures. To address these challenges, this work proposes RIGEL, a novel system that realizes stateful distributed in-network inference through collaborative Graph Neural Networks (GNNs) deployed directly on Tofino switches. RIGEL is designed to be fully in-network, eliminating the need for centralized processing and enabling real-time diagnosis. The system leverages a software-hardware co-design to preprocess high-dimensional spectral data, making it suitable for hardware-based inference. By combining an autoencoder with a GraphSAGE-based GNN, RIGEL achieves highly accurate anomaly detection and localization, as demonstrated in a realistic packet-over-optical network testbed. This digest presents the key aspects of RIGEL following the IMRAD structure.

RIGEL's methodology centers on a software-hardware co-design that enables stateful distributed in-network inference on Tofino switches. The system preprocesses high-dimensional spectral data to reduce dimensionality and extract relevant features suitable for hardware implementation. The core model combines an autoencoder for unsupervised feature learning with a GraphSAGE-based GNN for collaborative inference across switches. The autoencoder compresses the input spectral data into a latent representation, which is then processed by the GNN to capture spatial and temporal correlations among network nodes. To adapt the model to Tofino switches, a generalizable method is proposed that maps the GNN operations onto the switch's programmable pipeline, ensuring efficient execution. The distributed inference is stateful, meaning that each switch maintains and updates its local state based on incoming data and messages from neighboring switches, enabling collaborative anomaly diagnosis without centralized coordination. The system operates entirely within the data plane, avoiding control plane overhead. The mathematical formulation of the autoencoder can be expressed as: