C3S: Towards Generalizing Coordinated Congestion Control Across Flows and Topologies

We revisit C3, a centralized reinforcement learning–based approach to coordinated congestion control, and introduce C3S, a modular, permutation-invariant neural architecture that enables scalable guidance for an arbitrary number of flows.
Our results demonstrate effective scaling and fair bandwidth sharing, while highlighting the importance of topology-aware aggregation for generalization to complex network settings.