GRP-0165 Reinforcement Learning for Latency-Aware Priority Boosting in Linux Completely Fair Scheduler

GRP-0165 Reinforcement Learning for Latency-Aware Priority Boosting in Linux Completely Fair Scheduler

https://www.kennesaw.edu/ccse/events/computing-showcase/fa25-cday-program.php

Tail latency remains a persistent challenge in Linux’s Completely Fair Scheduler (CFS), particularly when short, latency-sensitive jobs compete with longer ones. Traditional boosting heuristics improve tail latency but require manual tuning and generalize poorly across workloads. This project evaluates whether a reinforcement-learning (RL) controller can dynamically apply priority boosts more effectively than fixed heuristics. Using a discrete-event Python simulator modeled after CFS, this project compares baseline CFS, heuristic boosting, and PPO-based learned boosting under mixed workloads. Results show that while heuristics achieve the lowest absolute latency, RL achieves competitive tail-latency reduction with significantly better fairness and adaptability. A state-space study further analyzes how observation richness affects RL performance.