Skip to content

Score-P, Scalasca, Cube#

Score-P, Scalasca, and Cube are typically used together to instrument, measure, and analyze an MPI application.

Documentation#

Installation#

For installation follow the steps in the documentation of the corresponding component.

Usage#

The usage is divided into three parts. A more detailed example can be found in the documentation of Scalasca.

  1. Instrument code with Score-P by prefixing each compile command with scorep: 

    scorep mpicc ...

    • The command scalasca --instrument instead of using scorep is only provided for backward compatibility.

    See the Score-P documentation for details.

  2. Measure and profile/trace an application run with Scalasca.
    Prefix the parallel run with scalasca --analyse. This generates .cubex files that can later be visualized with Cube. For measuring you can choose between:

    • Profiling (default), just run: 
      scalasca --analyse mpirun ...
    • Tracing enabled via flag -t: 
      scalasca --analyse -t mpirun ...

    See Measure for more options, Filtering to reduce the amount of collected data, or consult Scalasca's documentation for all details.

  3. Analyze/examine results with Scalasca or Cube.

    • without GUI: scalasca --examine -s <experiment dir>
    • with GUI: cube <cubex file>
      • Note there are multiple cubex files inside an experiment directory.

    See the Scalasca documentation or Cube documentation

Measure#

Options for measuring are specified via environment variables. Important once include:

  • SCOREP_TOTAL_MEMORY=<memory>[M|G]: Amount of memory for profile/tracing data to set aside on each process until data must be flushed to disc.
  • SCOREP_METRIC_PERF=cycles,page-faults: Add perf events cycles and page-faults. Other perf events can be specified too.
  • SCOREP_METRIC_RUSAGE=all Use all or dedicated metrics from the rusage (resource usage) system call.

Further information can be found in Score-P's Getting Started guide or Measurement Configuration .

Filtering#

For more details consult Score-P's documentation about Filtering.

Generating a profile or trace can amount to a lot of memory. By filtering specific functions or ranges can be included or excluded. As a starting point check the scorep.score file, located inside the experiment directory. It shows how much memory is required for Scalasca per process and which functions can be filtered.

A simple file to exclude functions by name (Fortran functions, hence the trailing underscore):

SCOREP_REGION_NAMES_BEGIN
        EXCLUDE
                vector_IP_calc_dens
                mod_sort.swap_i8_
                mod_ranges.ranges_l
                mod_sort.swap_i4_
SCOREP_REGION_NAMES_END

With Scalasca use a filter file with the -f <filter-file> flag:

scalasca --analyse -f <filter-file> ...

Score-P issues#

Feature test macros, that are set at the beginning of source files might be ignored, when preprocessed/compiled with Score-P.

  • Score-P prefixes some includes of the Standard C library at the beginning of each file.
  • Definition of feature test macros in your source files are then defined to late.
  • Workaround: Define feature test macros on the command command line when invoking the compiler. For example to define _GNU_SOURCE use -D_GNU_SOURCE.