Collecting measurement data from HPC applications can be difficult. Score-P presents a generally uniform approach to collecting profiling and tracing data that can be applied to a broad range of HPC applications and allow measurements to be collected in many common HPC environments. In order to use this measurement system effectively, users must be aware of both the common performance engineering issues that affect all performance measurements and the specifics of how to use the various components of Score-P to collect accurate and relevant performance data.

• Able to instrument applications including one or more parallel paradigms from the following:
  • MPI
  • SHMEM
  • OpenMP
  • Pthreads
• Able to instrument applications including at least one specialized form of measurement from the following:
  • Accelerator usage
  • I/O performance
  • Hardware counters
  • Memory usage
• Able to score measurements and effectively describe whether a measurement includes unacceptable perturbation
• Able to create and use filter files to improve measurement quality and relevance
• Able to manage measurement data through environment variables and scripting
  • Able to automate naming/filing of experiments in appropriate locations
  • Able to read and understand configuration and manifest files generated by Score-P

• Bert Wesarg, William (Bill) Williams, ZIH Tools Team @ TU Dresden