Computer and Information Science at the University of Oregon

Oral Comprehensive Exam Details

Automated Statistical Methods for Parallel Performance Analysis

Abstract

This paper explores the use of process automation to guide a parallel performance analyst through the knowledge discovery process, while providing the ability to customize the analysis process. For example, the input data can be evaluated to determine the distribution of the data, the standard deviation and/or the prevalence of outliers. Different analytical methods assume or require particular distributions or other data characteristics, and process automation would help prevent the misapplication of inappropriate analytical methods. There are four research areas necessary to examine this problem in depth. The first is an examination of the current parallel performance analysis tools and the methods they use, to identify useful data to collect and the types of analysis to perform. The second is an examination of analytical methods used in knowledge discovery. The third is an examination of software engineering practices available for automation and distributed application design. The fourth is an examination of data management methods used in performance tools and elsewhere. Finally, we conclude with a review of the work to date and an exploration of the potential research directions in these combined fields of research.