Subj : Re: What is the real costs of LOCK on x86 multiprocesor machine?
To   : comp.programming.threads
From : Chris Thomasson
Date : Fri Jul 29 2005 03:31 pm

>> Surprisingly, on Intel Pentium processors the overhead of
>> a locked instruction is about the same for a single processor
>> as for one in a multiple processor configuration.
>
> Hm, can you give me some numbers? Is it closer to my measurement (3 times 
> slower than non-atomic) or to 100 cycles I have seen somewhere?

There are too many variables to consider for any definitive answers. For 
example, the negative effects of asserting a cpu's #LOCK signal could be 
tremendous. It also tends to have different timing results on multi-cpu 
systems because the #LOCK signal actually locks the entire system bus so 
subsequent requests from any other cpu's are blocked. You can't really know 
how many operations were blocked during a #LOCK signal assertion. Sometimes 
the #LOCK signal assertion can be avoided if the shared data being accessed 
is already cached in the cpu.

This variable behavior can lead to inconsistent timing results.

-- 
http://appcore.home.comcast.net/
(portable lock-free data-structures)

.