Assessing Measurement and Analysis Performance and Scalability of Scalasca 2.0

摘要: 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 more traditional Linux clusters. Its primary focus is support for C/C++/Fortran using MPI OpenMP, mixed-mode combinations thereof, offering detailed call-path profiles each process thread produced by runtime summarization or augmented with wait-state event traces. A new generation (2.0) uses the community-developed infrastructure comprising Score-P associated components, while continuing previous functionality. By comparing version its predecessor, from NPB3.3-MZ-MPI benchmark suite, we validate core functionality assess overheads scalability. Although adequate general use, various aspects are identified further improvement, particularly larger scales.