turpin@cs.utexas.edu (Russell Turpin) (08/26/89)
My company is engaged in a modeling project, and we just discovered that Unix's gprof is deceptive. From its display, it appears to measure time spent in a procedure separately for each call path to the procedure. In fact, it only measures total time spent in the procedure, and then allocates this total to the different call paths proportionately to the number of times each call path is executed. (We discovered this after being astounded that the time spent in procedures was independent of where they were called -- a surprising phenomenon for the system we are modeling. It turns out that the gprof documentation has a small note on this, but we were fooled by the clearly labeled report it generates.) Does anyone know of a Unix profiler which actually measures what gprof just appears to measure? More generally, does anyone know of a better profiler? We are using Sun 3's and 4's, and a profiler that works just on these machines would be useful even if it didn't work on other Unix systems. We would also be interested in a tool which automatically instuments either Fortran or C code. We would be glad to hear about commercial products as well as software in the public domain. Please e-mail me, as I do not read this group as a matter of habit. Thanks for any information. If it is desired, I will post a summary of what I learn. Russell