While the previous sections introduced the three basic actions supported by the Scalasca Trace Tools based on an abstract example, this section will now guide through the typical analysis workflow using a moderately complex, MPI-based benchmark code: BT from the NAS Parallel Benchmarks (NPB-MPI 3.3.1) [11]. The BT benchmark implements a simulated computational fluid dynamics (CFD) application using a block-tridiagonal solver for a synthetic system of nonlinear partial differential equations and consists of about 20 Fortran 77 source code files. Although BT does not exhibit significant performance bottlenecks—after all, it is a highly optimized benchmark—it serves as a good example to demonstrate the overall workflow, including typical configuration steps and how to avoid common pitfalls.

The example measurements shown in this section were carried out using the Scalasca Trace Tools v2.5 in conjunction with Score-P v5.0, CubeLib v4.4.3, and CubeGUI v4.4.3 on the JURECA cluster at Jülich Supercomputing Centre. JURECA's compute nodes are equipped with two Intel Xeon E5-2680 v3 (Haswell) 12-core CPUs running at 2.5 GHz, and connected via an EDR InfiniBand fat-tree network. The BT benchmark code was compiled using Intel compilers and linked against ParTec ParaStation MPI (which is based on MPICH). The example commands shown below should therefore be representative for using the Scalasca Trace Tools in a typical HPC cluster environment. For convenience, the resulting post-processed Cube files are also available for download on the Scalasca documentation web page [13].

Note: In the following, it is assumed that all Scalasca commands are available in the shell's search path ($PATH), for example, after loading site-specific environment modules. Also, remember that the Scalasca convenience commands use other executables provided by Score-P, CubeLib, and CubeGUI, which therefore need to be available in the search path as well.

Subsections: