Subj : Re: Lock Free -- where to start
To   : comp.programming.threads
From : David Schwartz
Date : Thu Sep 29 2005 01:44 pm


"Joe Seigh" <jseigh_01@xemaps.com> wrote in message 
news:QradnVuGiqT2o6HeRVn-og@comcast.com...

> So you're saying lock based solutions will scale better then lock-free 
> ones?

    It depends upon which of the dozens of possible meanings of "scale" you 
intend. In general, if by "scale" you mean show improved real world 
performance as load, in general, increases, then yes, lock-based solutions 
will scale better than lock-free ones.

    Consider a case where you have four threads. Threads A and B want to 
access the same collection and threads C and D want to access the same 
collection. With only a single CPU, both a lock-free and lock-based solution 
will run one thread at a time and it will run at full speed.

    Now, let's scale this to two CPUs. The lock-based solution will never 
run thread A and B at the same time. It will always run one of A/B and one 
of C/D. The scaling is perfect and the performance doubles since contending 
threads never run.

    The lock-free solution will randomly match threads. Whenever it happens 
to schedule A and B or C and D at the same time, each thread will run at 
some small fraction of its maximum as the FSB bounces the shared/modified 
data back and forth.

    Granted this is an extreme example, but the point is that contention is 
bad and lock free algorithms tend to maximize contention.

    DS

.