Machine Learning for Energy-Performance-aware Scheduling
Zheyuan Hu, Yifei Shi
TL;DR
This work reframes CPU scheduling on heterogeneous multi-core systems as a Bayesian optimization problem and demonstrates that an anisotropic Gaussian Process with a Matérn 5/2 kernel can accurately capture the inherently non-smooth performance landscape. By integrating sensitivity analysis (fANOVA) and comparing kernels, the study provides interpretable insights into which hardware resources most affect energy and latency, and uses expected hypervolume improvement to approximate the Pareto frontier for multi-objective optimization. Key findings include the emergence of Race-to-Idle strategies, a structural decoupling where latency is driven by big cores while energy is governed by mid-tier and little cores, and a phase-transition-like robustness shift under high load. The results show substantial sample efficiency compared with baselines and offer practical guidance for offline configuration of post-Dennard systems, with clear paths toward online adaptation and handling task dependencies in future work.
Abstract
In the post-Dennard era, optimizing embedded systems requires navigating complex trade-offs between energy efficiency and latency. Traditional heuristic tuning is often inefficient in such high-dimensional, non-smooth landscapes. In this work, we propose a Bayesian Optimization framework using Gaussian Processes to automate the search for optimal scheduling configurations on heterogeneous multi-core architectures. We explicitly address the multi-objective nature of the problem by approximating the Pareto Frontier between energy and time. Furthermore, by incorporating Sensitivity Analysis (fANOVA) and comparing different covariance kernels (e.g., Matérn vs. RBF), we provide physical interpretability to the black-box model, revealing the dominant hardware parameters driving system performance.