Instrumentation Database for Performance Analysis of Parallel Scientific Applications

Parallel codes, especially scientific applications, have increased the need for performance analysis tools. In this paper we present an approach that addresses some of the difficulties inherent to the analysis of parallel codes. To this end we examine the three components that comprise performance analysis: instrumentation, visualization, and analysis. The essence of our approach is to use a database to capture instrumentation data. We discuss how this approach influences the design of performance analysis components. We further motivate our approach with an example using a sparse matrix vector product code. We conclude by discussing some of the key implementation issues and outlining future plans.