Computer Science

Directed Research Project Details

Performance Database Framework

Abstract

Empirical performance evaluation of parallel and distributed systems or applications often generates significant amounts of performance data and analysis results from multiple experiments as performance is being investigated and problems diagnosed. The analysis and understanding of these performance data is a difficult and time-consuming task. In contrast to most current performance analysis research efforts that focus on studying performance results generated from one single program execution, I present a Performance DataBase Framework (PerfDBF) that provides a common foundation for storing, querying, and analyzing performance data from multiple experiments, application versions, or platforms. In this framework, raw performance data and associated meta data are formalized by a Performance Data Meta Language (PerfDML) in an extensible and portable manner. I design a Performance Database (PerfDB) structured in a hierarchy of application/experiment/trial to store performance information. Two transformations, raw-to-PerfDML and PerfDML-to-PerfDB, accomplish the conversion from raw performance data to performance database internal storage. To facilitate the construction of performance analysis tools and to allow performance analysis tools to be easily interfaced with the PerfDBF, I design a Performance Database Toolkit (PerfDBT) that comprises a set of interface and utility modules. A scalability analysis tool is developed to demonstrate the usefulness of the PerfDBT.