Predictive Versus Projection-Based Scheduling for O-RAN Fronthaul Deadline Assurance

Abstract

In the Open Radio Access Network (O-RAN) 7.2x split, the Distributed Unit (O-DU) must finish upper physicallayer processing before strict fronthaul deadlines. However, heterogeneous acceleration through CPU fallback, shared acceleration, and dedicated acceleration introduces profile-selection, contention, and queueing delays that static and reactive policies handle poorly under bursty, mixed-numerology load. This paper tests whether learned deadline-risk prediction improves acceleration-aware scheduling. We model upper physical layer (High-PHY) tasks as deadline-constrained jobs served by three concurrent acceleration profiles. Building on this model, we conduct a leakage-hardened ns-3 study comparing a predictive risk-gating scheduler against projection-based earliest-deadlinefirst (EDF), static, and reactive baselines. The predictive scheduler improves on static shared acceleration and prevents queue buildup, but it does not outperform projection-based EDF, which attains lower deadline-miss ratio, tail latency, and active energy than the predictive scheduler in every scenario. Tuned reactive switching is also stronger in most scenarios and uniformly more energy-efficient. This gap is structural in the studied regime. When service times are near-deterministic, candidateprofile completion times can be projected almost exactly, so scheduling directly on those projections dominates approximating the same outcome from coarse congestion features. A robustness sweep with stochastic service times and a learning-augmented EDF variant both confirm the finding. Conversely, degrading the projection with stale telemetry locates the other side of the boundary, where the learned policy narrowly overtakes EDF only in the heaviest-loaded regime. The result is a reproducible comparison mapping the conditions under which learned risk prediction does, and does not, improve deadline-driven O-RAN fronthaul scheduling.