Scalasca is a software tool that supports the performance optimization of parallel programs by measuring and analyzing their runtime behavior. The analysis identifies potential performance bottlenecks – in particular those concerning communication and synchronization – and offers guidance in exploring their causes.
Scalasca supports profiling of MPI, OpenMP and hybrid MPI+OpenMP applications.
There are currently two versions of Scalasca 2.0 modules installed on Anselm:
- scalasca2/2.0-gcc-openmpi, for usage with GNU Compiler and OpenMPI,
- scalasca2/2.0-icc-impi, for usage with Intel Compiler and Intel MPI.
Profiling a parallel application with Scalasca consists of three steps:
- Instrumentation, compiling the application such way, that the profiling data can be generated.
- Runtime measurement, running the application with the Scalasca profiler to collect performance data.
- Analysis of reports
Instrumentation via " scalasca -instrument" is discouraged. Use Score-P instrumentation.
After the application is instrumented, runtime measurement can be performed with the
scalasca -analyze command. The syntax is:
scalasca -analyze [scalasca options][launcher] [launcher options][program] [program options]
An example :
$ scalasca -analyze mpirun -np 4 ./mympiprogram
Some notable Scalasca options are:
- -t Enable trace data collection. By default, only summary data are collected.
- -e <directory> Specify a directory to save the collected data to. By default, Scalasca saves the data to a directory with prefix scorep_, followed by name of the executable and launch configuration.
Scalasca can generate a huge amount of data, especially if tracing is enabled. Please consider saving the data to a scratch directory.
Analysis of Reports¶
To launch the analysis, run :
scalasca -examine [options] <experiment_directory>
If you do not wish to launch the GUI tool, use the "-s" option :
scalasca -examine -s <experiment_directory>
Alternatively you can open CUBE and load the data directly from here. Keep in mind that in that case the preprocessing is not done and not all metrics will be shown in the viewer.
Refer to CUBE documentation on usage of the GUI viewer.