Adaptive Heterogeneity-Aware CPU Scheduling using Deep Reinforcement Learning for Energy-Efficient Real-Time VR/AR on Mobile Platforms

Abstract

This article presents an innovative approach to CPU scheduling for mobile Virtual Reality (VR) and Augmented Reality (AR) applications utilizing Heterogeneous Multi-Processors (HMPs). The proposed system leverages Machine Learning (ML) and Reinforcement Learning (RL) techniques to develop an adaptive, heterogeneity-aware scheduler that dynamically optimizes task placement and frequency scaling. Traditional schedulers face significant challenges when managing the highly variable workloads characteristic of immersive applications, particularly in balancing real-time performance requirements with energy efficiency and thermal constraints on mobile platforms. The article introduces a comprehensive monitoring framework that collects detailed system telemetry, including CPU utilization, thermal data, battery state, and application-specific metrics, to inform sophisticated scheduling decisions. The system incorporates advanced predictive analytics to anticipate thermal events, workload transitions, and performance degradation with 87.3% accuracy at 100ms lead times, enabling proactive optimization rather than reactive responses. Multiple learning approaches are examined, including offline training, online adaptation, and hybrid methodologies. The evaluation methodology employs diverse workloads and metrics spanning energy efficiency, real-time performance, thermal management, resource utilization, and scheduler overhead. Preliminary results indicate substantial improvements in energy consumption, thermal stability, frame time consistency, and resource allocation compared to conventional approaches, demonstrating the potential of ML/RL techniques to revolutionize scheduling for immersive mobile applications.