4e0
Subj : Re: SMR hazard pointers w/o memory barrier
To   : comp.programming.threads
From : Chris Thomasson
Date : Thu Apr 14 2005 06:38 am

> I think it could be done.  Whether it should be done
> is another question.

You have to take into consideration that SMR is/can be used to prevent the 
ABA problem in certain lock-free algorithms; memory barriers might be a 
concrete requirement for the "original" SMR algorithm that "currently" 
exists...

;)


This is because ABA detection has be performed at the "hazard pointer 
acquire" stage, "before", the lock-free algorithm even has a chance to use 
the hazard. This is shared to all threads interested in the lock-free 
algorithm, so you would have to use some sort of barriers...




> The hazard ptr load is only about 3x slower than a raw
> ptr load which is a lot faster than with a membar in
> there.

You mean an "uncontended" hazard pointer load...

However, if you take into consideration that SMR is so "fine-grain" that the 
hazard pointer collisions, "retries" should really not be a problem. I am 
finding that SMR based reference counting can be a lot more light-weight 
than DWCAS based atomic_ptr ( depends on SMR "node" allocation scheme ).

.

0