Dynamic tuning of parallel programs

Abstract

Performance of parallel programs is one of the reasons of their development. The process of designing and programming a parallel application is a very hard task that requires the necessary knowledge for the detection of performance bottlenecks, and the corresponding changes in the source code of the application to eliminate those bottlenecks. Current approaches to this analysis require a certain level of expertise from the developers part in locating and understanding the performance details of the application execution. For these reasons, we present an automatic performance analysis tool with the objective of alleviating the developers of this hard task: Kappa Pi. The most important limitation of KappaPi approach is the important amount of gathered information needed for the analysis. For this reason, we present a dynamic tuning system that takes measures of the execution on-line. This new design is focused to improve the performance of parallel programs during runtime.