Introduction
In the evolving landscape of 6G, the transition toward disaggregated Open Radio Access Network (O-RAN) architectures introduces a critical challenge: control delay. While the near-real-time RAN Intelligent Controller (RIC) enables sophisticated policy tuning, the tens of milliseconds of delay between the RIC and the Distributed Unit (O-DU) can destabilize traditional scheduling logic. Standard approaches like Proportional Fair (PF) or Backpressure (BP) do not treat this latency as a primary design variable, often resulting in fragile delay behavior and poor performance under observation noise.
This paper introduces the Network-Optimised Spiking Neural Network (NOS) scheduler, a novel framework that explicitly accounts for control delays by modeling network scheduling through the dynamics of a compact spiking neural network.
Methodology
The NOS scheduler models each user (bearer) using a two-state spiking kernel. The first state, excitability, serves as a proxy for finite queue buffers, while the second, recovery, mimics a neuron's refractory period to ensure fairness by suppressing repeated grants. Interference constraints are integrated as delayed "spikes" from neighboring users, weighted by the network's interference graph.
The core theoretical contribution is the definition of the spectral margin ($\delta$). This single, explainable parameter compresses the delay envelope, controller gain, and interference topology into a unified metric: $\delta = k_\star(\Delta) - gH\rho(W)$. The authors prove that as long as $\delta > 0$, the system is mathematically guaranteed to be stable (geometrically ergodic), providing a level of auditability and safety absent in black-box Deep Reinforcement Learning (DRL) schedulers.
Key Results
The NOS scheduler was evaluated against PF and delayed BP baselines using deterministic numerical simulations with synthetic compound-Poisson traffic across three symmetric topologies: pair2, line4, and ring8.
- Tighter Delay Tails: As shown in the performance sweep, NOS consistently achieved the lowest 99.9th-percentile (P99.9) delay. While BP performance degraded significantly as control delay increased, NOS remained robust.
- Higher Utilization: NOS achieved superior utilization headroom (measured via AUC) across all tested delays and topologies.
- Mathematical Alignment: Empirical results perfectly matched the analytic sub-Gaussian tail shapes predicted by the spectral margin theory, confirming that increasing $\delta$ effectively thins the delay tail.
Conclusion and Future Work
By unifying network topology and control latency into the spectral margin, the NOS scheduler offers a high-performance, explainable alternative to traditional O-RAN scheduling. While current results rely on numerical simulations with conservative gain calibrations, future work will focus on dynamic gain adaptation and validation on physical O-RAN hardware to confirm real-world computational efficiency and performance.
Summary updated