This is the mail archive of the mailing list for the Cygwin project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: gettimeofday() does not returns usec resolution

Ralf Habacker wrote:

 > Hi,
 > for kde2 we are building a profiler lib for profiling complex c++
 > applications (currently found in the cvs areas of
 > (

 > using the high resolution timer of native windows (about usec
 > resolution).
 > This lib could be use for easy profiling of any c++ application and
 > libs like cygwin.dll and so on.
 > While adding unix support (and cygwin) for this lib, I noticed, that
 >  the gettimeofday() function returns only a resolution of 10ms (the
 >  time slice resolution) but mostly other unix os returns a
 > resolution in the usec region.  I have appended a testcase for this.
 > Has anyone address this problem already. I have looked int the
 > cygwin and list and found the only topic

 > In there is a
 > detailed instruction how to use the hugh resolution counter. .
 > $ cat timeofday.c #include <sys/time.h>
 > int main() { struct timeval tp; long a,b;
 > gettimeofday(&tp,0); a =
 > ((unsigned)tp.tv_sec)*1000000+((unsigned)tp.tv_usec);
 > printf("timestamp (us): %d\n",a); usleep(1000); gettimeofday(&tp,0); b
 >  = ((unsigned)tp.tv_sec)*1000000+((unsigned)tp.tv_usec); 
 >  (us): %d (diff) %d\n",b,b-a); }
 > Ralf Habacker
This is a continuing source of consternation, which may be considered OT 
for cygwin. I suspect that linux for ia32 tends to use one of the low 
level cpu tick registers to obtain the microsecond field; I have not 
examined current source. I don't know that it is possible to guarantee 
how well the zero of the microsecond field coincides with the second 
ticks.  On many ia chips, it is possible to use the rdtsc instruction 
directly, for timing intervals at sub-microsecond resolution.  A 
calibration run is required, to measure the tick frequency against the 
lower resolution time of day clock.  linux and Windows, of course, do 
something along this line, when booting up.  Any working Windows will 
report usable results via the QueryPerformance API's, usually with 
better than 10 microsecond resolution, and it seems reasonable for 
cygwin to base its functions directly on Windows API's.  On many chips 
the direct use of rdtsc can produce better than 1 microsecond 
resolution, but then the application takes on the burden of dealing with 
various odd hardware combinations, rather than expecting the hardware 
vendor to make Windows work.

Tim Prince

Unsubscribe info:
Bug reporting:

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]