[boost] Re: Performance: shared_ptr

From: Joe Gottman (jgottman_at_[hidden])
Date: 2003-11-14 21:23:19

"Matthew Wilson" <stlsoft_at_[hidden]> wrote in message
news:bovl1d$juh$1_at_sea.gmane.org...
> Hi Boosters
>
> I'm just doing some research on reference-counting, and contrasting
> self-counting types with externally counted types, the latter effected
with
> the rather excellent shared pointer technique.
>
> [As a side question, is it true that Nicolai Josuttis was the inventor of
> the technique?]
>
> Naturally, one of the shared pointer's I've used is boost::shared_ptr. The
> measurements have demonstrated some surprising performance characteristics
> for shared_ptr in particular, and I want to solicit the opinion of the
> expert's here before I make any bold claims, as some results are just too
> surprising for me to trust them.
>
> There are five types tested.
>
> 1. A self-counting type, using an ancient Synesis (my company) template
> ref-counting bolt-in class
> 2. A self-counting type, using a custom-written (for this test) template
> ref-counting bolt-in class
> 3. Externally counting, using boost::shared_ptr
> 4. Externally counting, using a custom-written (for this test) shared
> pointer type
> 5. Externally counting, using a custom-written (for this test) shared
> pointer type, which has a custom allocator for the shared count members
>
>

  results snipped.  Shared_ptr did not do well.

> Naturally, the very big question is: what's happening to
boost::shared_ptr?
>
> In thread times, it performs 2-3 times worse than in single threaded
> scenarios (irrespective of whether they are built for single or multiple
> threads). However, in total time, it performs extremely poorly, up to 100
> times slower than some of the others! Clearly, there is some serious
> contention involved.
>
> There are three possibilities I have thought of:
>
> (i) It performs very poorly in multi-threaded sharing scenarios. If this
is
> the case, can any people more expert in Boost than me proffer an
explanation
> as to precisely the problem, and is any such person motivated to try and
> address the issue?
> (ii) The particular sharing scenario is well away from any real-world
> scenario in which boost::shared_ptr is likely to be used. If this is the
> case, why do four different implementations, representing two very
different
> reference-counting strategies, all maintain a consistent relative
> performance? What are the scenarios for which boost::shared_ptr is
> optimised?
> (iii) There are environment settings for which boost::shared_ptr will
> perform much better. If that's so, can someone help me out with
suggestions
> of what defines to make? Why are these settings not the default.
>
> I'd be very interested in your thoughts.
>
> Many thanks in advance.
>

    Have you tried getting the latest shared_ptr.hpp and
detail/shared_count.hpp from the CVS files?  The version of shared_ptr in
release 1.30.2  has a lot of extra overhead due to weak_ptr support.  Every
time a shared_ptr is copied there is an extra comparison and increment (and
later an extra decrement).  The version in the CVS avoids almost all of this
extra work.

Joe Gottman