[OpenSTA-devel] Re: [ opensta-Bugs-1084184 ] Incorrect CPU usage collection when monitoring HT

[Daniel S. <dansut@OpenSTA.org>]

Hi Joe,

It's almost amusing that Shahar posted this bug report now - Just
before you posted I had been doing some investigating which triggered
my memory of you reporting the problem on the users list and
providing the fix here.  I am just about to update Shahar's bug
report to report the problem as you describe it (clipping to 100%)
rather than Shahar's symptom description.

This should reiterate the importance of reporting problems as SF bugs
though - there was a chance that I would have forgotten about Joe's
fix and it wouldn't have made it into the upcoming release...

Shahar Kedar wrote...
> > The problem is that the machine on which the process ran, is using
> > a hyper threding CPU.  This means that I get 100% for that process,
> > although in reality it takes only 80% or less.  My guess is that
> > OpenSTA collected the information from only 1 "virtual processor".
> > 
> > I checked the behavior of the Windows Performace tool, and it gives
> > a more reliable number (between the range of 0% and 160%).
> > 
> > Can anyone provide me with a workaround?  Can this behavior be
> > fixed?

Joe Robertson wrote:
> See my submission to the development list Fri 5/9/2003 11:15 AM.
> 
> The solution requires a change to the OpenSTA source code in
>   ..\src\NTPerf\Executers\NTPerf\ntp.cpp then rebuilding OpenSTA
> (at least the TEexecuter_ntp.exe).  This tells the Windows
> Performance gathering mechanism not to cap the %CPU at 100% 
> (by ORing PDH_FMT_NOCAP100 with PDH_FMT_LONG).
> 
> The original code segment (from ..\src\NTPerf\Executers\NTPerf\ntp.cpp)
> is:
>   pdhStatus = PdhGetFormattedCounterValue( m_hCounters[iIndex],
>               PDH_FMT_LONG, NULL, &value );
> I changed the above line to read:
>   pdhStatus = PdhGetFormattedCounterValue( m_hCounters[iIndex],
>               PDH_FMT_LONG | PDH_FMT_NOCAP100, NULL, &value );

I still have some worries about including this fix into the upcoming
1.4.3 release.  I mentioned my concerns on the users list when you
initially detailed this issue and your fix and no-one seemed to care
so I'll most likely drop it in now - especially now it has a bug
report.

As a last ditch attempt to see if anyone has any objections here
before I commit the change:  Removing the clipping will potentially
change results and so "spoil" the possibility of comparing results
from test runs pre-1.4.3 to those afterwards.  Normally I would say
this isn't acceptable - especially as you can work around this by
logging the percentage values seperately and total yourself...

My preferred solution would be not to change the default behavior but
to add a switch to get the "correct" updated behavior.

Does anyone else care, or have any thoughts on this or similar
situations?

> Hope this helps,

It probably won't help Shahar as I don't think he's on the developer
list, but he should get the message when I update his bug report...

Cheers
/dan
-- 
Daniel Sutcliffe <DanSut@OpenSTA.org>
  OpenSTA part-time caretaker - http://OpenSTA.org/