Assessing measurement and analysis performance and scalability of scalasca 2.0

Abstract

The Scalasca toolset was developed to provide highly scalable performance measurement and analysis of scientific applications on current HPC platforms, including leadership systems such as IBM BlueGene/Q and more traditional Linux clusters. Its primary focus is support for C/C++/Fortran applications using MPI and OpenMP, and mixed-mode combinations thereof, offering detailed call-path profiles for each process and thread produced by runtime summarization or augmented with wait-state analysis of event traces. A new generation of Scalasca (2.0) uses the community-developed infrastructure comprising of Score-P and associated components, while continuing to provide the previous functionality. By comparing the new version of Scalasca with its predecessor, using the applications from the NPB3.3-MZ-MPI benchmark suite, we validate core functionality and assess overheads and scalability. Although adequate for general use, various aspects are identified for further improvement, particularly for larger scales. © 2014 Springer-Verlag Berlin Heidelberg.