AutoThread Predicts Optimal Thread Counts for RL Simulation Loops, Cutting Execution Time up to 83.8%

arXiv 2608.06025·low signalXBlueskyLinkedInCopy link

In simulation-in-the-loop RL, the simulator — not the policy — is often the bottleneck, and static thread configurations cause contention under dynamic workloads. AutoThread (arXiv 2608.06025, Aug 6) identifies the ratio of task execution time to scheduling time as the factor determining optimal thread count, then uses a Physics-Informed Neural Operator as predictor constrained by a finite-source M/M/1 queueing model, with load-aware online fine-tuning to correct prediction error. It reports 18.4% average speedup over static strategies, 1.7x and 1.8x the throughput of XGBoost and Reinforcer respectively, and up to 83.8% execution-time reduction; code and dataset at github.com/suchenjm/AutoThread.