Next: Running HPMCOUNT
Up: Using the Hardware Performance
Previous: The toolkit
HPMCOUNT investigates the
performance of an entire application. However, several of the rates
listed among the derived metrics make use of the wall-clock time. A
typical run used for the performance analysis of an application
contains very sizable overheads. The initialisation of the
application and the exchange of initial and final data with the mass
storage are examples for the sources of these overheads. Such
overheads can distort the results for these rates considerably when
comparing short test jobs to proper production run.
Despite these limitations, we feel HPMCOUNT is a very useful tool,
particularly since it is so simple to use.
Global analysis with HPMCOUNT