中国科学院机构知识库网格
Chinese Academy of Sciences Institutional Repositories Grid
首页 机构 成果 学者
  
  1. CAS IR Grid
  2. 计算技术研究所
  3. 中国科学院计算技术研究所
  4. 中国科学院计算技术研究所期刊论文
  5. 英文
ReflexPilot: Startup-Aware Dependent Task Scheduling Based on Deep Reinforcement Learning for Edge-Cloud Collaborative Computing

文献类型：期刊论文

作者Zou, Wenhao2,3; Zhang, Zongshuai1,2,3; Wang, Nina1,2,3; Tian, Yu2,3; Tian, Lin1,2,3
刊名IEEE TRANSACTIONS ON CLOUD COMPUTING
出版日期2025-04-01
卷号13期号:2页码:641-654
关键词Servers Optimal scheduling Collaboration Cloud computing Scheduling algorithms Containers Computer architecture Computational modeling Resource management Training Edge-cloud collaborative computing deep reinforcement learning dependent task scheduling startup latency
ISSN号2168-7161
DOI10.1109/TCC.2025.3555231
英文摘要With the increasing number of devices, the demand for data computation is growing rapidly. In edge-cloud collaborative computing, tasks can be scheduled to servers as interdependent subtasks, enhancing performance through parallel computing. A task is executed in an executor, which must first initialize the runtime environment in a process called task startup. However, most existing research neglects the reuse of executors, leading to considerable delays during task startup. To address this issue, we model the edge-cloud collaborative task scheduling scenario considering executor reuse, task startup, and dependency relationships. We then formulate the dependent task scheduling problem with task startup. To meet real-time demands in edge-cloud collaborative computing, we propose ReflexPilot, an online task scheduling architecture featuring executor management. Building on this architecture, we introduce OTSA-PPO, a task scheduling algorithm based on Proximal Policy Optimization (PPO), and EMA, an advanced executor allocation algorithm. Under constraints of computational and communication resources, ReflexPilot leverages OTSA-PPO for online scheduling of dependent tasks based on current states, while EMA pre-creates and reuses executors to reduce the average task completion time. Extensive simulations demonstrate that ReflexPilot significantly reduces the average task completion time by 31% to 71% compared with existing baselines.
资助项目National Natural Science Foundation of China[62120106007] ; Pilot for Major Scientific Research Facility of Jiangsu Province of China[BM2021800]
WOS研究方向Computer Science
语种英语
WOS记录号WOS:001504051800007
出版者IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
源URL[http://119.78.100.204/handle/2XEOYT63/42349]  
专题中国科学院计算技术研究所期刊论文_英文
通讯作者Zhang, Zongshuai
作者单位1.Nanjing Inst InforSuperBahn, Nanjing 211100, Peoples R China
2.Univ Chinese Acad Sci, Beijing 100049, Peoples R China
3.Chinese Acad Sci, Inst Comp Technol, Beijing 100190, Peoples R China
推荐引用方式
GB/T 7714		 Zou, Wenhao,Zhang, Zongshuai,Wang, Nina,et al. ReflexPilot: Startup-Aware Dependent Task Scheduling Based on Deep Reinforcement Learning for Edge-Cloud Collaborative Computing[J]. IEEE TRANSACTIONS ON CLOUD COMPUTING,2025,13(2):641-654.
APA Zou, Wenhao,Zhang, Zongshuai,Wang, Nina,Tian, Yu,&Tian, Lin.(2025).ReflexPilot: Startup-Aware Dependent Task Scheduling Based on Deep Reinforcement Learning for Edge-Cloud Collaborative Computing.IEEE TRANSACTIONS ON CLOUD COMPUTING,13(2),641-654.
MLA Zou, Wenhao,et al."ReflexPilot: Startup-Aware Dependent Task Scheduling Based on Deep Reinforcement Learning for Edge-Cloud Collaborative Computing".IEEE TRANSACTIONS ON CLOUD COMPUTING 13.2(2025):641-654.

入库方式： OAI收割

来源：计算技术研究所

浏览0
下载0
收藏0
其他版本

除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。