Enhancing Performance in Mixed-Criticality Real-Time Systems Through Learner-Based Resource Management

Abstract

In mixed-criticality (MC) systems, tasks with varying criticality levels share resources, leading to challenges in resource management during mode transitions. Existing approaches often result in suboptimal performance due to resource contention and criticality level inheritance. This paper introduces a novel learner-based resource management strategy that predicts optimal mode switching times and prevents low-criticality tasks from acquiring resources during critical periods. By combining vector autoregressive (VAR) and feed-forward neural network (FNN) techniques, our approach effectively anticipates system state changes and optimises resource allocation. Specifically, the method extracts key system features, including processor temperature, soft error rate, cache miss rate, and task slack time. A hybrid forecasting model then predicts the probability of a mode transition within a specified time horizon. Based on these predictions, the system proactively denies resource requests from low-criticality tasks during periods of high probability of mode transition, ensuring the availability of resources for high-criticality tasks. Comprehensive simulations demonstrate significant reductions in blocking time (up to 75%), miss rate (up to 9.35%), and energy consumption (up to 12.15%) compared to state-of-the-art methods. These improvements enhance system reliability and efficiency, making it suitable for safety-critical applications.