Re: Profiling with GPROF considered buggy?

[Carsten Thorenz <XZ550S@gmx.de>]

[-- Attachment #1: Type: text/plain, Size: 4042 bytes --]

Hi Tim!

> I thought that most linux implementations had the same timer resolution as
> Win2K installations.  I don't know that anyone has considered the
implications of gprof on WIn9x.

Linux seems to use 10ms as a standard, the same as WinNT4 (which I'm using)
but it 
can be changed to higher resolutions. 

Anyhow, it seems as if this is not only a problem with profiling.  I've
attached
the source for a _very_ simple test program. If you compile and profile it 
with -pg and gprof, you will immediately see some of Cygwins problems:

- The real total runtime is bigger than the one recorded in the flat profile
- The number of calls and the time per call is missing in the flat profile
- The call graph is completely useless

For comparison, this is what it looks like on a (rather slow) Linux-machine:

Flat profile:

Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total           
 time   seconds   seconds    calls  ns/call  ns/call  name    
 50.01     37.63    37.63   400000 94075.00 94075.00  crunch2
 49.99     75.24    37.61   100000 376100.00 376100.00  crunch1

...

Call graph:

index % time    self  children    called     name
                                                 <spontaneous>
[1]    100.0    0.00   75.24                 main [1]
               37.63    0.00  400000/400000      crunch2 [2]
               37.61    0.00  100000/100000      crunch1 [3]
-----------------------------------------------
               37.63    0.00  400000/400000      main [1]
[2]     50.0   37.63    0.00  400000         crunch2 [2]
-----------------------------------------------
               37.61    0.00  100000/100000      main [1]
[3]     50.0   37.61    0.00  100000         crunch1 [3]
-----------------------------------------------

Here, the call graph shows the correct structure, furthermore the
timing records for the function itself and the children are correct.

Bye, Carsten









> ----- Original Message -----
> From: "Carsten Thorenz" <XZ550S@gmx.de>
> To: <cygwin@cygwin.com>
> Sent: Monday, August 27, 2001 6:37 AM
> Subject: Re: Profiling with GPROF considered buggy?
> 
> 
> > Hi Tim!
> >
> > Tim Prince wrote:
> > > Unless I am mistaken, cygwin doesn't include any libraries built with
> -pg.
> > > When I wish to profile with g77, I build a copy of
> > > libg2c with -pg as well as building all my code with -pg. When
> > > I profile numerical code built with gcc, I use a mathinline.h as well
> > > as a few of my own math functions to avoid spending much time in the
> > newlib libm.
> >
> > I'm using standard ANSI-C, no libraries are linked in right now.
> > What seems strange to me is that the behaviour of the profiler is _so_
> > different when used under Linux, OS/2 or on completely other platforms
> > (I tried HPUX, too). Ok, on HPUX the results differ, because it is a
> > completely different machine, but at least the tendencies are the
> > same for all of the above and provide very useful information. Not
> > so with Cygwin or Mingw.
> >
> > > I can't tell from your message which language you
> > > are using or whether you expect all the time to be spent in your own
> -pg
> > compiled code.
> >
> > As I wrote previously: The program doesn't communicate much.
> > It doesn't use libraries. It simply crunches numbers, but some
> > of the CPU-intensive functions do not appear in the profile. Again my
> > guess: The time spent in each of these functions for a single call is
> > very short, but they are called _many, many_ times. Maybe the timer
> > resolution
> > is too bad to catch this?
> >
> > Bye,
> >
> > Carsten
> >
> >
> >
> > --
> > GMX - Die Kommunikationsplattform im Internet.
> > http://www.gmx.net
> >
> >
> >
> > --
> > Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple
> > Bug reporting:         http://cygwin.com/bugs.html
> > Documentation:         http://cygwin.com/docs.html
> > FAQ:                   http://cygwin.com/faq/
> >
> 

-- 
GMX - Die Kommunikationsplattform im Internet.
http://www.gmx.net

[-- Attachment #2: test.c --]
[-- Type: text/x-c, Size: 388 bytes --]

long i, j, k;
double a, b, c; 
void crunch1(void);
void crunch2(void);

int main(void) {
  for(i=0;i<100000;i++) crunch1();
  for(i=0;i<400000;i++) crunch2();
}

void crunch1(void) {
  for(j=0;j<400;j++)
    for(k=1;k<10;k++)
      a = (double)j / (double)k;
}

void crunch2(void) {
  for(j=0;j<100;j++)
    for(k=1;k<10;k++)
      a = (double)j / (double)k;
}