University of Houston
Department of Computer Science

In partial fulfillment of the Requirements for the Degree of
Master of Science

Sukhdeep Sodhi

will defend his thesis

AUTOMATICALLY CONSTRUCTING PERFORMANCE SKELETONS

FOR USE IN GRID RESOURCE SELECTION AND

PERFORMANCE ESTIMATION FRAMEWORKS

Abstract

The difficulty of estimating performance of an application on dynamically shared heterogeneous computational resources is a key problem that makes grid computing significantly more challenging than conventional high performance computing. Our approach towards solving this problem is based on the concept of performance skeleton of an application, which can be defined as a synthetic short running program whose execution time always reflects the performance of the application it represents. Thus, to estimate the performance of an application under existing grid conditions, one simply needs to execute the performance skeleton under those conditions.

The central challenge addressed by this thesis is automatic construction of performance skeletons. We present a framework towards this end. Our framework for automatic construction of performance skeletons has been implemented for message passing MPI programs, however the approach used has broad applicability across different programming paradigms. We demonstrate that performance skeletons constructed by our framework can be used to estimate the performance of corresponding message passing applications with reasonable accuracy. We also discuss whether there is a decrease in prediction accuracy with decrease in the execution time of performance skeletons.

Date: Thursday, January 29, 2004
Time: 4:30 PM
Place: 550-PGH


Faculty, students, and the general public are invited.
Thesis Advisor: Dr. Jaspal Subhlok