Consistent, fair benchmarks with Windows Prefetcher

Question

I am trying to benchmark an application on Windows XP SP3, but I am having trouble with prefetcher's secret sauce.

The best articles I found:
1. http://en.wikipedia.org/wiki/Prefetcher
2. http://www.microsoft.com/whdc/archive/benchmark.mspx
3. http://blogs.msdn.com/ryanmy/archive/2005/05/25/421882.aspx

My understanding is this
1. Application prefetching works for basically all .exe executables
2. Windows engages application prefetching at the time the application executable is started (in contrast to quickstarters like  "Fast Start", OSA, Adobe Quick Launch, etc.)

My questions are
1. Is my understanding (above) correct?
2. Is there a diagnostic/debug log I can viewer to see what prefetcher is doing?
3. Is prefetching analysis cumulative? To my surprise (below), I notice progressive improvements between reboots.

Starting with a clean XP install, my benchmark does this
1. Run the application twice (as suggested by article #2 above).  These times are not recorded.
2. Runs advapi32.dll ProcessIdleTasks (as suggested by article #2 above)
3. Reboot
4. Run the benchmark 10 times.  The first of these 10 is a cold start.  The other 9 are warm.
5. Reboot
6. Repeat steps 4-5 another 15 times

Here are my results starting with step #4.  The yellow parts concern me. 
http://spreadsheets.google.com/pub?key=pgHmoZGs3VSt1VAozuqjPGA

In conclusion, (1) what is going on? and (2)  how to do I get consistent results and a fair benchmark?  For consistency, instead of disregarding the first 2 iterations (as I do now), it seems I should disregard the first ~30?  However, ~30 is no longer a fair benchmark.


Andrew