Scholarlink

In expanding cloud clusters with increasing users, resource diversity and job variety multiply, making efficient task scheduling crucial for optimal resource utilization. Traditional methods, relying on heuristic rules or static strategies, struggle to adapt to dynamic settings, achieve global optimization, or handle complex task relationships. Therefore, this paper proposes a scheduling algorithm based on two-stage deep reinforcement learning and residual networks (TD2QN). In the task selection stage, the agent prioritizes tasks based on their attributes and urgency. In the resource allocation stage, tasks are allocated considering dependencies and resource availability for virtual execution. Introducing a residual network into the neural network structure enhances model performance and stability, allowing better adaptation to complex workflows, improved resource utilization, and reduced resource usage costs. Simulation results from real-world scientific workflow datasets demonstrate that compared to advanced algorithms like LB-HEFT, GA-GWO, and DQN, this method excels in cost savings, enhancing resource utilization, and other scheduling aspects across various scales and forms of workflows. When dealing with large-scale workflows, the algorithm reduces execution time by 25% to 28%, saves approximately 18% to 28% in costs, and improves resource utilization by 0.6% to 10.5%. Thus, it offers a viable solution to the challenge of workflow scheduling in heterogeneous cloud environments.

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

Two-Stage Workflow Scheduling Based on Deep Reinforcement Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hongyu Xu,
  • Hu Zhang,
  • Jibin Wang,
  • Yinglong Wang,
  • Ying Guo,
  • Lijuan Xu

摘要

In expanding cloud clusters with increasing users, resource diversity and job variety multiply, making efficient task scheduling crucial for optimal resource utilization. Traditional methods, relying on heuristic rules or static strategies, struggle to adapt to dynamic settings, achieve global optimization, or handle complex task relationships. Therefore, this paper proposes a scheduling algorithm based on two-stage deep reinforcement learning and residual networks (TD2QN). In the task selection stage, the agent prioritizes tasks based on their attributes and urgency. In the resource allocation stage, tasks are allocated considering dependencies and resource availability for virtual execution. Introducing a residual network into the neural network structure enhances model performance and stability, allowing better adaptation to complex workflows, improved resource utilization, and reduced resource usage costs. Simulation results from real-world scientific workflow datasets demonstrate that compared to advanced algorithms like LB-HEFT, GA-GWO, and DQN, this method excels in cost savings, enhancing resource utilization, and other scheduling aspects across various scales and forms of workflows. When dealing with large-scale workflows, the algorithm reduces execution time by 25% to 28%, saves approximately 18% to 28% in costs, and improves resource utilization by 0.6% to 10.5%. Thus, it offers a viable solution to the challenge of workflow scheduling in heterogeneous cloud environments.