HG-MRLS: A Hierarchical Graph Meta-Reinforcement Learning Framework for Dynamic Scheduling in Heterogeneous Computing Networks

Heterogeneous computing networks formed by modern cloud and edge collaboration pose dynamic resource scheduling as a key bottleneck for system performance. Rule-based traditional methods falter in handling environmental shifts, while most current deep reinforcement learning schedulers use single-decision architectures, struggling to balance global optimization with local real-time responses. Addressing this core issue, this study proposes a novel hierarchical graph meta-reinforcement learning scheduler (HG-MRLS). The framework employs a dual-layer decision mechanism, decomposing complex scheduling into macro strategic planning and micro tactical execution. At the upper layer, we designed a graph attention network-based macro scheduler that learns forward-looking resource allocation via encoding global topology and dynamic features. At the lower layer, we built a meta-reinforcement learning-based micro scheduler for rapid self-adaptation in unknown local environments. To validate effectiveness, we set up a simulation platform based on real workloads. Extensive experiments show HG-MRLS outperforms various classic heuristics and state-of-the-art deep RL methods in key metrics like average job completion time, deadline satisfaction rate, and resource utilization. Especially under high loads and sudden dynamic events, the framework exhibits outstanding stability and environmental adaptability. Its core scientific contribution is in demonstrating a new paradigm for resolving the inherent tension between global strategic planning and local real-time responses in complex, dynamic systems, offering a scalable blueprint for multi-scale intelligent control.