基于图强化学习的应用感知算力网络调度机制
Application-aware Scheduling Mechanism for Computing Power Network Based on Graph Reinforcement Learning
李珺玥 1汪硕1
作者信息
- 1. 北京邮电大学信息与通信工程学院,北京 100876
- 折叠
摘要
针对现有算力网络任务调度研究中忽视业务差异化需求及算网资源耦合导致调度效率低下的问题,本研究构建了一种面向应用感知的算力网络调度机制。首先,结合 APN6 技术设计应用感知的算力网络模型,并将计算卸载与路径选择联合建模为马尔可夫决策过程。为应对算网状态空间复杂及环境动态变化的挑战,提出了一种基于图强化学习（GNN+DQN）的调度算法,利用图神经网络提取全局拓扑特征,结合深度 Q 网络实现长期回报最优的智能决策。仿真实验表明,该算法将平均任务处理时延降低了约 35\%,实现了算网资源的长期负载均衡,并在异构网络拓扑中展现出优异的泛化性与鲁棒性,显著优于传统启发式与深度强化学习算法。
Abstract
To address the low scheduling efficiency caused by neglected differentiated service requirements and coupled computing-network resources in existing research on computing power network task scheduling, this study constructs an application-aware scheduling mechanism. Firstly, an application-aware model is designed using APN6 technology, and the joint optimization of computation offloading and routing selection is modeled as a Markov Decision Process (MDP). To tackle the challenges of a complex state space and dynamic environment, a graph reinforcement learning-based scheduling algorithm (GNN+DQN) is proposed, which uses GNN to extract global topological features and DQN to make decisions that achieve optimal long-term returns. Simulation results show that the algorithm reduces the average task processing delay by approximately 35\%, realizes long-term load balancing of computing-network resources, and exhibits excellent generalization and robustness in heterogeneous topologies, significantly outperforming traditional heuristic and deep reinforcement learning algorithms.关键词
计算机体系结构/算力网络/网络调度/图强化学习Key words
Computer System Architecture/Computing Power Network/Network Scheduling/Graph Reinforcement Learning引用本文复制引用
李珺玥,汪硕.基于图强化学习的应用感知算力网络调度机制[EB/OL].(2026-03-02)[2026-03-03].http://www.paper.edu.cn/releasepaper/content/202603-21.学科分类
计算技术、计算机技术
评论