Int J Performability Eng ›› 2020, Vol. 16 ›› Issue (8): 1262-1270.doi: 10.23940/ijpe.20.08.p13.12621270

Previous Articles     Next Articles

Workflow Scheduling using Graph Segmentation and Reinforcement Learning

Shujun Pei*, Qinggen Zhang, and Xuehui Cheng   

  1. School of Computer Science and Technology, Harbin University of Science and Technology, Harbin, 150080, China
  • Submitted on ; Revised on ; Accepted on
  • Contact: *E-mail address: peisj@hrbust.edu.cn

Abstract: Cloud computing with high availability and good scalability is becoming an important platform to solve scientific workflow problems, and scheduling is the prominent issue for optimal strategy. This paper presents a cloud scheduling algorithm using graph partitioning and reinforcement learning for data-intensive workflow scheduling based on the cloud heterogeneous platform. Aiming at reducing the cost and makespan during the process of task execution, the proposed algorithm firstly evaluates the dual optimization weight of workflow and clusters the tasks that have strong data dependence into blocks by the graph partitioning algorithm. Then, we train the partitioned workflow by iterating the reinforcement learning method such that the match for tasks and resources will meet our expectations. In consideration of performance metrics like total cost and makespan, the partitioning reinforcement learning scheduling algorithm is much better compared to classic algorithms during the simulating experiments conducted on the scientific workflow simulation.

Key words: cloud computing, scientific workflow scheduling, multi-objective optimization, reinforcement learning