FR-EAHTS: federated reinforcement learning for enhanced task scheduling with hierarchical load balancing and dynamic power adjustment in multi-core systems

Scheduling tasks in multi-core systems is essential for achieving high performance and energy efficiency in data centers and portable devices. Traditional scheduling methods struggle with challenges such as communication overhead, varying core activity levels, and dynamic environments, leading to inefficiencies in processing and energy consumption. To address these issues, this work introduces the Federated Reinforced Energy-Aware Heterogeneous Task Scheduling (FR-EAHTS) method, that integrates transformer-based models with reinforcement learning in a federated environment. The key innovation lies in leveraging TransformX for enhanced state-action representation and Adaptive Proximal Policy Optimization (APPO) for efficient policy optimization, improving both adaptability and scheduling performance in heterogeneous multi-core systems. Real-time load distribution and task granularity adjustments are achieved through the integration of Dynamic Heuristic Load Distribution (DHLD) and Parameter Dynamic Adjustment (PDA), enabling adaptive scheduling decisions. Additionally, Dynamic Voltage and Frequency Scaling (DVFS) is incorporated to optimize power efficiency, ensuring improved energy-aware scheduling in heterogeneous multi-core systems. To further enhance stability and convergence in reinforcement learning, the proposed approach integrates a Dynamic Learning Rate (DLR) mechanism, which adjusts learning rates based on task execution variability. Unlike fixed learning rates, DLR ensures adaptive learning adjustments, preventing unstable updates and improving scheduling efficiency in rapidly changing environments. This approach effectively coordinates policy changes, adapts to core workload heterogeneity, and dynamically adjusts to system characteristics. Experiments demonstrate the proposed work’s efficiency using three key evaluation metrics: makespan, energy consumption, and overhead, achieving a 25% reduction in makespan task completion time, 30% energy savings, and 20% lower scheduling overhead, making it highly effective for multi-core systems, data centers, and mobile environments.