Scalasca

Scalasca is an open-source toolset for the performance analyzes of parallel applications helping to identify optimization opportunities. It has been specifically designed for the use on large-scale systems including Cray XT/XE, but is also well-suited for small- and medium-scale HPC platforms. Scalasca supports an incremental performance-analysis procedure that integrates runtime summaries with in-depth studies of concurrent behavior via event tracing, adopting a strategy of successively refined measurement configurations. A distinctive feature is the ability to identify wait states that occur, for example, as a result of unevenly distributed workloads.

Developer: JSC Platforms: Category: Performance Analyzer License: New BSD Website: Scalasca homepage

Usage

To use Scalasca start by loading the respective software modules:

# on HAWK:

module load scalasca scorep cube

# on Vulcan

module load performance/scalasca performance/scorep

Then, instrument your application using Score-P by prefixing all MPI compiler and linker commands with the scorep command:

scorep mpicc myapp.c -o myapp

Next analyze your application using the scalasca -analyze command with your usual mpirun command:

export MPI_SHEPHERD=true # this is mandatory when running with mpt on HAWK
scalasca -analyze mpirun -n ${NP} myapp

After the run there will be a folder prefixed with "scorep_*" containing the messurment results.

Warning: The scalasca -analyze command evaluates the provided MPI command line and will try to determine the used MPI implementation, number of MPI ranks (based on '-n' option), additional MPI options, the application name, and the application's arguments. If you encounter problems at this point, please consult the scalasca documentation for details.

The last step is to explore the generated reports by using the scalasca -examine command, with the previously created "scorep_*" folder as the experiment name:

scalasca -examine <experiment_name>

This will open Cube with the generated report automatically if you have enabled X forwarding. You can also inspect the generated *.cubex report files, manually starting the Cube GUI:

cube <file>.cubex

Usage on Hawk

Warning: It is necessary to set

export MPI_SHEPHERD=true

for the ""scalasca -analyze"" step.

See also

• Software Development Tools, Compilers & Libraries

External links

• Scalasca homepage