Ph.D. Proposal of Enis Afgan

• Speaker: Enis Afgan
• Location: Campbell Hall 430
• Date: Wednesday, May 16
• Time: 11:00 am - 12:00 pm

Seminar Abstract

Grid Computing has emerged as the next-generation computational infrastructure for developing and deploying distributed high performance computing applications. Grid environments are perplexed with heterogeneous and dynamic resource availability inherently leading to resource and application dependencies as well as constraints. Scheduling (i.e., selecting appropriate mappings between resources and application) jobs on available resources is a function of resource heterogeneity, availability and application options. As the user of the grid, one is confined by resource heterogeneity and availability leaving application options as a way to observe desired performance. A way to coordinate such environment is through knowledge about application and possible mappings and dependencies to underlying resources. Because of involved complexities and dependencies, a typical user is ineffective in grid job scheduling, leading to unsatisfactory performance of the grid from both, user and resource provider perspectives. This proposal presents work dealing with grid job scheduling and aims at alleviating some of the requirements imposed on users through guided and automated job parameter selection.

Rather than focusing on applications and resources as whole, single units, by breaking each one into smaller components, we are able to observe dependencies and constraints more closely. As such, the focus of our work deals with global job parameter optimization during scheduling of grid jobs, as dictated by user utility (i.e., the change of user satisfaction over time). Through creation of application profiling system, we are able to obtain necessary knowledge about application and resource dependencies. Following data collection, we are developing optimization techniques to process collected data and obtain more efficient job-to-resource mappings. Based on the idea that an important prerequisite for effective scheduling exists in interaction between scheduler and user, we also propose new levels of support for users in terms of user and scheduler interaction.

The realization of this proposal is a scheduling system composing and coordinating individual components into a structured framework. The overall approach consists of (1) facilitating user utility function in grid job scheduling, (2) focusing on job parameter and parameter value optimization, (3) creation of application performance models based on application categories, and (4) constructing a system to coordinate and guide users during the job submission process. For the validation process, we plan to implement individual component(s) into real world application validating our devised guidelines for application modeling and scheduling. The overall framework will be tested using a representative case of applications from matching application categories through simulation. This work is an attempt to contribute to the long-term research goal of user centric scheduling intertwined with grid economic aspects through automation and optimization. Components of this work build on work already begun, some of which has already been published with additional works awaiting publication.