[cpp-threads] Alternatives to SC

[Doug Lea]

One broader issue hiding in...

Paul E. McKenney wrote:
>   Raul's version has the advantage of
> permitting the caches to retain read sharing.  

As the number of CPUs increase, it is nearly always better
to stall only one processor waiting out a barrier rather
than to stall an unknown number of them for an unknown
period as they fight for exclusive cache line access to perform
an LL/SC or CAS.

This is one part of the reason that concurrent readability
(i.e., CCCC, but not necessarily SC semantics) is so important
for scalability. Standardized specs will need to be sensible
in the upcoming years when having hundreds of processors is not
uncommon.

-Doug