Re: [boost] Re: Re: Re: Re: Profiling Library suggestion

From: Beman Dawes (bdawes_at_[hidden])
Date: 2005-02-18 17:00:53

At 07:30 PM 2/13/2005, christopher diggins wrote:

 >> No, there is a very particualar problem relating to
 >> QueryPerformanceCounter() on some (old?) chipsets the can cause it go
 >> backwards in time and wrap after approximately 20 minutes. I reported 
this
 >
 >> to microsoft some years ago and it was confirmed an issue. (at the time 

 >> only happened to me on one particualar celeron / motherboard
 >> configuration). By backwards in time I mean you could call the counter 
in
 >
 >> a loop and have a reported time earlier than one got the previous
 >> iteration.
 >
 >Really? That is different than the issue documentation I found at
 >http://support.microsoft.com/default.aspx?scid=kb;EN-US;q274323 which 
stats
 >that the QueryPerformanceCounter() can jump *ahead* a few seconds. I 
wonder
 >if the status of the issue has changed to this new version or the
 >documentation just sux.

There is another problem with QueryPerformanceCounter(); the latency (time 
spent executing the call itself) varies dramatically depending on (1) the 
version on Windows, and (2) which compiler was used. Why the compiler 
should have a strong effect on the latency of an API call is a mystery.

John Maddock and I ran some tests several years ago, and both of us could 
reproduce the effects. I recently downloaded some timing code written by 
someone else, tried it with several compilers and the effect still appears 
to exist with current compilers. So be careful, and be sure to test with 
different versions of Windows and different compilers.

--Beman