Graph Neural Networks for D2D Link Scheduling: Centralized and Distributed Schedulers Evaluated with Sionna Ray-Traced Channels

Device-to-device (D2D) communication requires a link scheduler that manages interference among simultaneously active pairs, an NP-hard problem even under binary on/off transmission. Learning-based schedulers are mostly trained and evaluated under stochastic channel models that miss site-specific propagation. We study learned D2D scheduling under Sionna ray-traced channels of downtown Munich. Two schedulers are trained with policy-gradient reinforcement learning to maximize the sum-rate: a centralized graph-attention scheduler that observes the full channel matrix, and a fully distributed per-pair policy that uses only locally measurable channel statistics, with no inter-pair message exchange at decision time. On 24,000 snapshots with N=10 pairs, the centralized scheduler reaches 94.6% of the brute-force-optimal sum-rate at 1.1% of its computation time on a GPU (25% on one CPU thread), rising to 97.5% with a local-search refinement, and, without retraining, stays within 94.4–95.6% of optimal across N=2–10, i.e., up to its training size. The distributed scheduler attains 70.7% of optimal (mean of five runs), quantifying the accuracy cost of decentralized execution, which grows with the number of pairs. Under independent channel estimation errors, both retain more than 97% of their accuracy, whereas missing estimates of silent peers affect the centralized scheduler more strongly.