Summary of Data Collection
The more detailed, the more intrusive and the more processing effort.
- Use profile and count information to obtain parameter values needed to
complete performance models.
- Measure execution times for a range of processor counts and problem
sizes, and compare results with predictions.
- If observed and modeled performance differ, use profile and count
information to verify basic assumptions of model.
- Message counts and volumes, load imbalances, replicated computation).
- If aspects are still unexplained, use tracing and study performance on a
few processors. Increase processor number as understanding improves.
Author: Wolfgang Schreiner
Last Modification: December 23, 1997