Optimization of instrumentation in parallel performance evaluation tools

Abstract

Tools to observe the performance of parallel programs typically employ profiling and tracing as the two main forms of event-based measurement models. In both of these approaches, the volume of performance data generated and the corresponding perturbation encountered in the program depend upon the amount of instrumentation in the program. To produce accurate performance data, tools need to control the granularity of instrumentation. In this paper, we describe developments in the TAU performance system aimed at controlling the amount of instrumentation in performance experiments. A range of options are provided to optimize instrumentation based on the structure of the program, event generation rates, and historical performance data gathered from prior executions. © Springer-Verlag Berlin Heidelberg 2007.