Cloud Computing. 6th International Conference, CloudComp 2015, Daejeon, South Korea, October 28-29, 2015, Proceedings

Research Article

Hybrid Workflow Management in Cloud Broker System

Download
244 downloads
  • @INPROCEEDINGS{10.1007/978-3-319-38904-2_15,
        author={Dongsik Yoon and Seong-Hwan Kim and Dong-Ki Kang and Chan-Hyun Youn},
        title={Hybrid Workflow Management in Cloud Broker System},
        proceedings={Cloud Computing. 6th International Conference, CloudComp 2015, Daejeon, South Korea, October 28-29, 2015, Proceedings},
        proceedings_a={CLOUDCOMP},
        year={2016},
        month={5},
        keywords={Workflow scheduling Virtual machine allocation Cloud resource provisioning},
        doi={10.1007/978-3-319-38904-2_15}
    }
    
  • Dongsik Yoon
    Seong-Hwan Kim
    Dong-Ki Kang
    Chan-Hyun Youn
    Year: 2016
    Hybrid Workflow Management in Cloud Broker System
    CLOUDCOMP
    Springer
    DOI: 10.1007/978-3-319-38904-2_15
Dongsik Yoon1,*, Seong-Hwan Kim1,*, Dong-Ki Kang1,*, Chan-Hyun Youn1,*
  • 1: KAIST
*Contact email: dongsik.yoon@kaist.ac.kr, s.h_kim@kaist.ac.kr, dkkang@kaist.ac.kr, chyoun@kaist.ac.kr

Abstract

In Cloud broker system, workflow application requests from different users are managed through workflow scheduling and resource provisioning. In workflow scheduling phase, most existing algorithms allocate each task on certain VM in serial. In general, single task does not fully utilize allocated resource such as CPU, memory, and so on. When multiple tasks are processed with same resource in parallel, the resource utilization is improved that leads to saving the cost. In order to solve this problem, the Parallel Task Merging scheme in the same VM is proposed, which saves the cost of execution while satisfying SLA deadline. After workflow scheduling, VM resource provisioning is required. Auto-scaling VM resources approach is proposed, which adjusts the number of VMs while the number of requests varies. In this paper, we do experiment the parallel task merging and auto-scaling approaches on different environments to observe on which conditions these two approaches are working well or not.