Jesus Labarta, Polytechnical University of Catalonia (UPC), Spain
Current practice in performance analysis tends to be based on aggregated numbers from which understanding the actual behavior of the system is difficult. The cost of obtaining and processing such performance data often refrains developers from performing detailed analyses.
The objective of the talk is to discuss some of the issues (flexibility to compute and display performance indices, precission, scalability, instrumentation overhead, methodology) that contribute to the current situation and how they are tackled within the Paraver tools set. This environment supports extremely detailed analysis of the performance of parallel (MPI, OpenMP, mixed) applications and systems.