Access Type

Open Access Dissertation

Date of Award

January 2017

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Electrical and Computer Engineering

First Advisor

Song Jiang

Second Advisor

Shiyong Lu

Abstract

Data-centric workflows naturally process and analyze a huge volume of datasets. In this new era of Big Data there is a growing need to enable data-centric workflows to perform computations at a scale far exceeding a single workstation's capabilities. Therefore, this type of applications can benefit from distributed high performance computing (HPC) infrastructures like cluster, grid or cloud computing.

Although data-centric workflows have been applied extensively to structure complex scientific data analysis processes, they fail to address the big data challenges as well as leverage the capability of dynamic resource provisioning in the Cloud. The concept of “big data workflows” is proposed by our research group as the next generation of data-centric workflow technologies to address the limitations of exist-ing workflows technologies in addressing big data challenges.

Executing big data workflows in the Cloud is a challenging problem as work-flow tasks and data are required to be partitioned, distributed and assigned to the cloud execution sites (multiple virtual machines). In running such big data work-flows in the cloud distributed across several physical locations, the workflow execution time and the cloud resource utilization efficiency highly depends on the initial placement and distribution of the workflow tasks and datasets across the multiple virtual machines in the Cloud. Several workflow management systems have been developed for scientists to facilitate the use of workflows; however, data and work-flow task placement issue has not been sufficiently addressed yet.

In this dissertation, I propose BDAP strategy (Big Data Placement strategy) for data placement and TPS (Task Placement Strategy) for task placement, which improve workflow performance by minimizing data movement across multiple virtual machines in the Cloud during the workflow execution. In addition, I propose CATS (Cultural Algorithm Task Scheduling) for workflow scheduling, which improve workflow performance by minimizing workflow execution cost. In this dissertation, I 1) formalize data and task placement problems in workflows, 2) propose a data placement algorithm that considers both initial input dataset and intermediate datasets obtained during workflow run, 3) propose a task placement algorithm that considers placement of workflow tasks before workflow run, 4) propose a workflow scheduling strategy to minimize the workflow execution cost once the deadline is provided by user and 5)perform extensive experiments in the distributed environment to validate that our proposed strategies provide an effective data and task placement solution to distribute and place big datasets and tasks into the appropriate virtual machines in the Cloud within reasonable time.

Download

Share

COinS