I played the following piece of code on both Windows and Linux

// Windows with Visual Studio 2008
unsigned __int64 GetCycleCount()
{
__asm RDTSC
}
...
t = GetCycleCount();
t2 = GetCycleCount();
wprintf(TEXT("cycle diff: %d\r\n"), t2 - t);
...

// Linux with gcc
unsigned long long int GetCycleCount()
{
asm ("rdtsc");
}
...
t = GetCycleCount();
t2 = GetCycleCount();
printf("cycle diff: %d\n", t2 - t);
...

while I always got "cycle diff: 88" under Linux, I got 308 under Windows XP (SP3) occationally (most time I got 297)..

I am wondering why there may be a very small difference under Windows ? just 11 cycle ? If there was a context switch or something, there would be much much bigger difference that 11.

Have you looked at the code generated by the compiler?

Microsoft recommends using QueryPerformanceCounter instead of RDTSC
http://www.strchr.com/performance_me...nts_with_rdtsc
An alternative solution with rdtsc
http://groups.google.com/group/micro...ce7086d2e7549e

Thanks guys,,

I didn't look at the code generated. I think whatever code was generated, the difference should be a constant unless it was interrupted between the two rdtsc instructions. In a case of interruption, I would expect much bigger difference than just 11 cycles. I didn't check Intel manual though..

The links are very useful..thx..

I guess it could be caused by cache warming up..

I don't know if it is cache warming up. as if that were the issue, you should have seen the same numbers the first time the code was run. After all, code is not there in the caches when it runs the first time. In case it indeed is cache warming up, you could run the code in a loop to see if the numbers start matching.

Also, I am just thinking aloud here, could it be powersave features of the processor kicking in, perhaps, better implemented on Linux? If you are running the test on a laptop, you could set the CPU policy to performance to see if that changes anything.

Also, if possible, you could try running this code from a kernel module with interrupts disabled (I think on windows, cli/sti or something equivalent can be used). Although, context switches should be longer, but just to take them out of the equation altogether.