Subj : Re: What is the real costs of LOCK on x86 multiprocesor machine?
To   : comp.programming.threads
From : Peter  Dimov
Date : Sat Jul 30 2005 06:23 am

David Schwartz wrote:
> "Mirek Fidler" <cxl@volny.cz> wrote in message
> news:3kuqnuF10818jU1@individual.net...
> >I am unable to find anwswer to above question. Only tidbit of information
> >was that it is 100 cycles on P4, but perhaps that was just the worst case.
> >
> > Measuring the same on my uniprocesor AMD64 machine, LOCKed instructin seem
> > to be 3 times slower than regular one, but I guess that has only a little
> > relevance on real MP machine.
>
>     On a p4, it's 100-200 clocks.

In practice, the actual amortized cost appears to be much lower than
50-100x. I see a 4x difference on shared_ptr_timing_test (which
basically tests reference count increments and decrements in a somewhat
real-life worst case scenario.)

A P4 can stall for 50 cycles without an apparent reason. ;-)

.