TorusNet: a direction-specific graph neural network architecture for congestion-aware adaptive load-balanced routing in torus networks

Traditional routing algorithms in high-performance computing suffer from sub-optimal throughput and latency, mainly as a result of not utilising all available paths in an interconnection network which has been particularly problematic with non-uniform traffic. In this work, we introduce TorusNet, an innovative directionally aware graph neural network that models the unique topological properties of toroidal interconnection networks through a set of unique message passing operators. The explicit use of direction aware feature embedding along with real-time appraised congestion state allows for the creation of highly accurate adaptive routing policies to steer traffic away from areas experiencing congestion. Evaluated on a 4 × 4 toroidal network (16 nodes), TorusNet achieves a 75.1% packet delivery success rate at network saturation compared to 44.6% for deterministic XY routing, while improving contention intensity by 41% and latency variance by 53%. Though limited to a small-scale system, this proof-of-concept demonstrates the potential of topology-aware, congestion-sensitive, learning-based routing techniques for interconnection networks.