All ETDs from UAB

Advisory Committee Chair

Purushotham Bangalore

Advisory Committee Members

Brandon Eames

Elliot Lefkowitz

Anthony Skjellum

Alan Sprague

Document Type

Dissertation

Date of Award

2009

Degree Name by School

Doctor of Philosophy (PhD) College of Arts and Sciences

Abstract

The grid computing paradigm enables access to geographically and administratively distributed networked resources, and delivers functionality of those resources to individual users. Stemming from the core composition and aggregation of individual resources, the grid is primarily characterized by the heterogeneity it offers. Although such heterogeneity is often considered a feature, it also presents an obstacle in terms of application execution patterns and expectations (in terms of job runtime, resource utilization, and/or user Quality of Service (QoS)). Typical users have little or no knowledge about the concrete requirements their application imposes on such resources and thus have to stumble through a sea of options and uncertainties when submitting a job, leading to inefficient use of available resources. In order to alleviate the user from having to understand existing dependencies and make low-level decisions, a grid metascheduling framework has been devised that enables automated application- and user-oriented job metascheduling. In order to enable application-oriented metascheduling, a set of core grid services, Application Information Services (AIS), were designed and developed to provide application metaschedulers with relevant information regarding each application's execution requirements and preferences. With such information, a metascheduler is capable of automatically realizing more job-to-resource mappings. In order to enable user-oriented metascheduling, a novel mode of user-scheduler interaction has been devised that builds on top of AIS. The model is realized in terms of two-way communication between a user and the scheduler enabling strict focus on an individual user and their current job. Overall, this dissertation makes contributions regarding efficiency of use and ease of access for grid resources. Results of grid job metaschedulers implementing the devised framework are shown as capable of consuming application-specific data in a manner that leverages existing heterogeneity and, in turn, automatically deliver effective application-to-resource mappings. Results achieved are two-fold: (1) behavior of application jobs across grid resources has been significantly improved in terms of job execution control, capable of increasing resource utilization and achieving significant runtime reduction (up to 50%), and (2) each job submission is being tailored specifically to an individual user and their respective job, delivering significantly higher QoS to the user.

Share

COinS