Real-Time ML Integration with CFS to Improve Power Efficiency in Non-Hybrid Linux Systems

Abstract

Modern non-hybrid Linux systems rely on the Completely Fair Scheduler (CFS) and Dynamic Voltage and Frequency Scaling (DVFS) to balance performance and energy efficiency. However, this default approach has key limitations: CFS distributes tasks evenly across cores without distinguishing between workload types, while DVFS adjusts frequency based on aggregate load. As a result, CPU-bound and I/O-bound tasks often share the same cores, leading to unnecessary frequency boosts and wasted energy for tasks that don’t require high frequencies. This thesis proposes a machine learning-enhanced scheduling framework that integrates task-type awareness into non-hybrid Linux systems. A decision tree classifier predicts whether tasks are CPU-bound or I/O-bound using lightweight runtime features such as execution time, waiting time, context switches, priority, and niceness values. Classified CPU-bound tasks are consolidated onto selected cores running at higher frequencies, while I/O-bound tasks are grouped on separate cores maintained at lower frequencies. To prevent thermal hotspots and performance degradation, periodic rotation of CPU-bound and I/O-bound core groups is implemented. The scheduler is evaluated against the default CFS using controlled CPU-intensive, I/O-intensive, and mixed workloads generated with Sysbench and FIO. Energy consumption, performance, and Energy-Delay Product (EDP) are measured using perf and turbostat. Experimental results show that the ML-enhanced scheduler reduces energy consumption and improves EDP significantly, whilemaintaining performance levels comparable to the default scheduler. These findings highlight the potential of ML-based task classification to improve DVFS utilization and deliver more energy-efficient scheduling in general-purpose, non-hybrid Linux systems.