Wednesday, January 09, 2013 7:31 AMHello --
I am calling the win/sdk function QueryPerformanceCounter to do timing -- how is
it in some cases that over 2 calls over time that the call in the future has a
lesser value than the one in the past so
start = QueryPerformanceCounter
::: wait some time
finish = QueryPerformanceCounter
then the value of
finish - start
is less than zero -- is this possible?
Wednesday, January 09, 2013 9:48 PM
What kind is the machine? How much is the difference?
Assuming it is not a "normal" time rollover, you have a buggy platform, where QPC is not monotonic. Avoid these platforms or find a fix (BIOS update, etc) if you can.
Wednesday, January 09, 2013 10:19 PM
Hello -- Thanks for the reply. It's win/xp -- there are no more BIOS updates from the vendor MSI -- Its an old MOBO -- from 2006 -- I guess I could set affinity and just run on 1 cpu -- are there any other work-a-rounds?
Wednesday, January 09, 2013 11:12 PM
You are probably hitting a bug on older AMD machines where the RDTSC value is not sync'd between cores.
- Proposed As Answer by Chuck Walbourn - MSFT Friday, January 11, 2013 8:16 PM
Friday, January 11, 2013 8:45 PM
yes -- thank for the info ... If its a hardware / bios issue there is noting I can do about it at this time
I guess its its critical, I could force the process to run on 1 cpu
Saturday, January 12, 2013 1:44 PM
If I remember correctly, this behavior can be seen even on one-CPU machines (I've seen it on a Geode UP system).
Monday, January 14, 2013 8:37 PMIf the RDTSC count isn't monotonically increasing on a single-CPU machine, then that's a pretty nasty hardware bug. Of course, you still have to cope with 32-bit rollover on x86 CPUs.
- Edited by Chuck Walbourn - MSFT Monday, January 14, 2013 8:38 PM