Scholarlink

<p>A Computing Power Network (CPN) is a cloud–edge–end integrated infrastructure that interconnects heterogeneous computing resources to support latency-sensitive and volatile workloads. A core challenge in CPN task scheduling is to jointly determine task prioritization and computational node selection under dynamic arrivals, resource heterogeneity, and multi-objective conflicts. In this paper, we study dynamic task scheduling in a CPN and aim to (i) minimize average task completion time, (ii) maximize scheduling success rate under deadline constraints, and (iii) improve global resource utilization. We propose CPN-HRL, a hierarchical deep reinforcement learning framework that decomposes scheduling into two coordinated subproblems: a high-level agent (LSTM-PPO) generates global priority weights for queued tasks, and a low-level agent (GAT-PPO) performs topology-aware node selection for each task in the ranked sequence. To further guarantee timeliness for latency-critical tasks, we introduce a two-queue mechanism with an Urgent Queue and a Main Queue, together with a dynamic reordering trigger. Simulation results on multiple real-world topologies show that, under our experimental settings, CPN-HRL reduces average completion time and maintains high scheduling success rate under high loads, while improving overall resource utilization.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CPN-HRL :a hierarchical deep reinforcement learning approach for priority–aware task scheduling in CPN enabled by cloud–edge–end environments

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月31日

查看全文
  • Qiqiang Yue,
  • Le Tian,
  • Xu Feng,
  • Zheng Yuan,
  • Shuai Wei,
  • Jiqiang Xia,
  • Bo Chen,
  • Xuanyan Song,
  • Ying Yao,
  • Yi Liu

摘要

A Computing Power Network (CPN) is a cloud–edge–end integrated infrastructure that interconnects heterogeneous computing resources to support latency-sensitive and volatile workloads. A core challenge in CPN task scheduling is to jointly determine task prioritization and computational node selection under dynamic arrivals, resource heterogeneity, and multi-objective conflicts. In this paper, we study dynamic task scheduling in a CPN and aim to (i) minimize average task completion time, (ii) maximize scheduling success rate under deadline constraints, and (iii) improve global resource utilization. We propose CPN-HRL, a hierarchical deep reinforcement learning framework that decomposes scheduling into two coordinated subproblems: a high-level agent (LSTM-PPO) generates global priority weights for queued tasks, and a low-level agent (GAT-PPO) performs topology-aware node selection for each task in the ranked sequence. To further guarantee timeliness for latency-critical tasks, we introduce a two-queue mechanism with an Urgent Queue and a Main Queue, together with a dynamic reordering trigger. Simulation results on multiple real-world topologies show that, under our experimental settings, CPN-HRL reduces average completion time and maintains high scheduling success rate under high loads, while improving overall resource utilization.