Lightweight and Robust Online Learning for Energy-Aware Multicore Real-Time Schedulers

Energy-aware Multicore Real-Time Schedulers leverage Dynamic Voltage and Frequency Scaling (DVFS) and task migration to optimize energy consumption at runtime. However, mapping every performance fluctuation of task-sets becomes impractical given the complexity of the modern multicore platform. Machine Learning, particularly online learning, is a widespread alternative to make energy optimization robust to such fluctuations. Lightweight models are preferred to avoid large overheads and not disrupt critical task execution. Nevertheless, these models might fail to match performance estimations on scenarios that deviate far from the ones used to train them, potentially locking DVFS due to overestimated performance reduction and, therefore, locking online learning. This paper extends an approach for Online Learning Energy-Aware Multicore Real-Time Scheduling (OL-EAMRTS) introduced in prior work by proposing a novel unlocking procedure for online learning. The proposed unlocking procedure overcame a locking event in a new architecture and promoted an average energy saving of 17.30% for the evaluated benchmarks.