Well, as it turns out memory caching was not the problem. I proved this by running the first and last test scenario each immediately after a reboot. The timing results were nearly identical to what I got when I ran all seven of my tests one after the other with only a restart of the application between them.

I was however able to figure out a way to clear the windows memory cache from a script. There are some freeware tools that can do this, but they are windows apps with no (useful) command line interface. The one I used was CacheSet from Sysinternals. Then I wrote a little wrapper module that calls this using Win32::GuiTest. Kinda kludgy, but it got the job done.

At least I learned a new Perl Module.