Strange performance problem

[GFL]

Hi all,

  I am doing some performance measurements for my hobby OS, and I am obtaining some weird results in my test machine.  I am performing my measurements using the RDTSC instruction who loads the timestamp counter value to eax.
So say, for example, if under bochs I get 17 clocks for a bunch of code (nothing special there, some mov's and jmps, no CPL changing nor cr3 reloading), under my test machine (an Athlon XP) I get 128 clocks (!!!???).  A back-and-forth task switch measuring 260 clocks under bochs gives me more than 2000 under my test machine, even with TLB flush penalty this seems a lot to me ... I checked caching is enabled and all pages are cacheable, so this is not the issue: if I disable cacheable pages the former task-switch example raises to more than 65000 clocks... any ideas of what I am missing here?

[CodeSlasher]

To get an accurate timestamp value using RDTSC,you need to use an instruction that forces the CPU to finish what each pipeline is performing,ie Serializing instructions.
Check out
http://www.cs.uit.no/inf2200/2003h/stuf ... /RDTSC.htm
and
http://www.math.uwaterloo.ca/~jamuir/rdtscpm1.pdf
and
http://www.google.com/search?hl=en&ie=U ... gle+Search