Problems profiling

[betal]

Hi I’m working with a scientific program in Fortran. It consists of a main file flutter.f90, and several subroutines and modules contained in other files.
I compile the whole thing with the following script

//first I produce the object files
pgf95 -c -Mfixed -Mprof=func *.f90
//delete the unused one
rm flutter.o -f
//then compile the main program using the objects created before
pgf95 -Mfixed -Mprof=func flutter.f90 -o prog *.o

The issues I found are:
1)Time measured by the profiler is different(longer) than the one I measured in the application.
2)Althought the program improves when using vectorization(1.4 times, measured by code inside the program), when I compile it for profiling the results (as reported by the profiler) are different.
It actually says that both programs behave the same and with a time about 30 times greater than the ones that were not compiled for profiling.

Any ideas on what should be the problem will be appreciated.

Thanks in advance

[mleair]

Hi Betal,

You may be noticing some unexpected times due to the fact that -Mprof=func does induce some overhead on the program to perform its profiling. This is especially noticeable on programs that do not run for a long time. If you’re using a linux system, then I recommend trying the -Mprof=time option. There is also a way to profile the hardware counters on linux as well…consult the PGI Tools Guide (http://www.pgroup.com/doc/pgitools.pdf) for more information.

-Mark