Re: [asio-users] ASIO multiprocessor scalability

[Yuri T. <yti...@gm...>]

>
>
> Just out of curiosity: did you try it with the latest ASIO, which is
> included in 1.47beta? Perhaps, it performs slightly better, due to
> reduced copying of handler functors (especially with MSVC10).
> Besides, try examining your code for cache coherency issues.
> Ensure that you use make_shared instead of direct creation, and pass
> shared_ptrs by const reference.
>
No, I didn't try 1.47.

As I mentioned above CPU load isn't an issue (according to task manager,
each core's load < 50%). And excessive copying usually results in high CPU
load (we also had such issues :) ).

Marat's idea about allocators sounds good since malloc may cause too much
locking. Unfortunately I can't use tcmalloc globally, since I'm on windows,
MSVC 2010 too.

I also suspect that there can be issue with spawning threads. In addition to
receiving video we poll 2 times per second for IO status, and for each
request we create a separate io_service object (it's legacy design, for
video we use single connection with async_read_some calls). However, for
async_connect ASIO spawns a thread (per io_service object) which calls
select. This results in ~40 threads spawned and joined each second (may me
twice more timer thread, I'm not sure).

Now I'm evaluating Intel's tools.

BTW, regarding Marat's note:

> I had to use my own binders because of MSVC10 (2010) needs explicit move
> constructors and boost::bind-created functors
> doesn't have them.

Do std::bind or lambdas have similar issues? You can workaround your problem
by pulling bind from proper namespace depending on C++0x status.

Thanks for ideas, I'll try to keep you informed.
Yuri