Reinforcement Learning–Driven Task‐to‐Core Mapping and Adaptive Urgency‐Score Scheduling for Real time Multi-core Systems

Abstract

As multi-core processors scale to tens or hundreds of cores, efficiently mapping dynamic application workloads becomes critical to avoid thermal hotspots, energy waste, and deadline violations. Conventional static or heuristic mapping schemes fall short under run-time variability in task sizes, thermal coupling, and load imbalance. In this paper, a Reinforcement Learning–Driven Task-to-Core Mapping framework that jointly optimizes temperature, energy, performance, and load balance in a closed-loop simulation environment using ARM based multicore architecture is presented. The model is mapped as a state space model: each decision state encodes per-core load and temperature along with the next task’s profiled features; actions assign the task to one of N cores; and a multi-objective reward, dynamically weighted by task criticality and sensitivity parameters, guides learning via Q-learning. To ensure timeliness, the mapping is followed with a novel urgency-score scheduler that recalculates each task’s WCET and deadline based on run-time load and temperature and combines deadline urgency with starvation balancing. Evaluated on MiBench-derived workloads, our approach reduces makespan by up to 13%, saves ~6% energy, lowers peak temperature by ~5.95%, achieves sub-1% deadline miss-rate and delivers smallest jitter—outperforming task assignment and scheduling baselines.