TransSched: Evaluating Lightweight Learned CPU Burst Prediction for Scheduling Under Controlled Contention

Machine-learning-assisted CPU scheduling is often evaluated primarily through prediction accuracy, although a predictor is useful to an operating system only if its errors translate into acceptable system behavior. This study evaluates a lightweight next-CPU-burst prediction pipeline and its downstream scheduling utility using a 100,000-event synthetic scheduling dataset containing 24,876 processes. Classical regression baselines, a GRU, and a Tiny Transformer were evaluated on a process-isolated held-out test set, after which frozen predictions were supplied to a deterministic four-core Predicted-SJF scheduler. Four contention regimes were fixed before learned-scheduler evaluation. Under the near-saturation stress regime, Tiny-Transformer Predicted-SJF reduced mean waiting time to 614.35 ms compared with 1,167.47 ms for MLFQ and 1,232.58 ms for 5-ms Round Robin, and reduced mean turnaround time to 766.89 ms compared with 1,320.00 ms and 1,385.12 ms, respectively. These gains did not constitute an unconditional scheduler improvement: mean response time increased to 550.62 ms versus 2.90 ms for MLFQ, and tail and fairness behavior remained problematic. Further analysis showed that nearly identical prediction error across five training seeds could produce stress mean waiting times ranging from 612.02 to 812.93 ms. Instrumented READY-set analysis found that approximately 93% of stress dispatches were contested and that the five predictors disagreed on 25–30% of contested choices. Ablation results found no material benefit from sequence history or global system-state features on this dataset, while exploratory quantization showed that coarse prediction signals sharply reduced cross-seed scheduling variance. The results support a narrow conclusion: learned burst information can improve selected scheduling objectives under severe contention, but predictive accuracy alone is insufficient evidence for deployable learning-assisted scheduling. Responsiveness, fairness, stability, deterministic safeguards, and inference overhead must be evaluated jointly.