[boost] Re: [threads] read_write_mutex performance

From: Michael Glassford (glassfordm_at_[hidden])
Date: 2005-04-25 08:32:43

Peter Dimov wrote:
> What are the design goals of read_write_mutex? (Lock proliferation aside 
> for a moment.)
> 
> My understanding has always been that read/write locks are an 
> optimization, that is, a read/write lock-based algorithm should deliver 
> better performance than the same algorithm with "ordinary" locks. Right?
> 
> Wrong.
> 
> The program at the end of this message demonstrates that in this 
> specific (contrived?) case ordinary mutexes outperform a read/write 
> mutex by a factor of 2.5 or more in almost every scenario, even when no 
> writers are active!
> 
> In some scenarios (16 readers, 4 writers, 10,000,000 iterations) the 
> ordinary mutex case completes in under 8 seconds on my machine, but the 
> read/write case exceeded my patience threshold.
> 
> Am I missing something?

It doesn't help that the Boost read/write mutex, as I've mentioned 
before, is pretty inefficient; I'm not very happy with it. There are at 
least three reasons for the inefficiency: 1) Supporting four scheduling 
policies in one code base makes the code too complex; 2) Trying to 
adhere to the scheduling policies too rigidly is also adding code 
complexity and inefficiency; 3) The complexity makes it harder to 
optimize, and not much time has been spent optimizing it in the first place.

I mentioned elsewhere that I have some thoughts about ways to improve 
Boost.Threads; some of these ideas deal with addressing these three 
points and getting a much better read/write mutex in place.

Mike