Subj : Re: Lock Free -- where to start
To   : comp.programming.threads
From : Sergey P. Derevyago
Date : Wed Oct 05 2005 08:00 pm

Joe Seigh wrote:
> > Unfortunately, some "lock free" algorithms make long loops (e.g. make some
> > progress, determine that some critical data was concurrently touched and start
> > the "progress" once again) and it's not so obvious that there can be no real
> > progress as well.
> 
> If a lock-free algorithm loops that usually means some other thread is
> making progress.
>
	Usually isn't equal to always, is it?

> Looping would be a problem for wait-free algorithms
> if there wasn't a bound on the number of loops but it's not a problem
> for lock-free algorithms.
> 
> Non lock-free algorithms usually loop waiting for a specific action by some
> other thread.  If lock-free algorihtms loop it's because their waiting for
> some interval of no actions by other threads.
>
	IMHO even correctly implemented lock free algorithms don't solve fundamental
MT problems: if you really have to concurrently modify some data then you
really have to implement some synchronization scheme and this scheme doesn't
come for free.

	Frankly, I consider the lock free approach as "make frequent event faster but
get rare event be slower". Sometimes the whole program runs faster, sometimes
it don't...
--
         With all respect, Sergey.               http://ders.angen.net/
         mailto : ders at skeptik.net

.