Re: [boost] [context review] Performance questions

Subject: Re: [boost] [context review] Performance questions
From: Oliver Kowalke (oliver.kowalke_at_[hidden])
Date: 2012-01-09 14:22:01

Am 09.01.2012 19:35, schrieb Antony Polukhin:
> 1) What if we reimplement the ASIO example without Boost.Context
> (server class will look like this):
>
> class server : public boost::enable_shared_from_this<  server>
> {
> private:
>      boost::asio::ip::tcp::acceptor      acceptor_;
>      boost::array<  char, 1024>  buffer_;
>
>      void do_()
>      {
>              boost::asio::ip::tcp::socket socket( acceptor_.get_io_service() );
>              acceptor_.async_accept( socket, boost::bind(&
> server::do_1_, this->shared_from_this(), _1) );
>      }
>
>      void do_1_(boost::system::error_code ec)
>      {
>              if (!ec) {
>                  socket.async_read_some(
>                      boost::asio::buffer(buffer_),
>                      boost::bind(&  server::do_2_,
> this->shared_from_this(), _1, _2) );
>               } else {
>                  do_();
>               }
>      }
>
>      void do_2_(boost::system::error_code ec, size_t n)
>              if (!ec) {
>                  boost::asio::async_write(
>                      socket,
>                      boost::asio::buffer( buffer, n),
>                      boost::bind(&  server::do_1_,
> this->shared_from_this(), _1) );
>              } else {
>                  do_();
>              }
>         }
>      }
>
>      server( boost::asio::io_service&  io_service, short port) :
>          acceptor_( io_service, boost::asio::ip::tcp::endpoint(
> boost::asio::ip::tcp::v4(), port) ),
>      {}
>
> public:
>      typedef boost::shared_ptr<  server>      ptr_t;
>
>      static ptr_t create( boost::asio::io_service&  io_service, short port)
>      { return ptr_t( new server( io_service, port) ); }
>
>      void operator()( boost::system::error_code /*ec*/, size_t /*n*/)
>      {
>          do_();
>      }
> };
>
> Looks like this implementation will be faster (it has the same amount
> of io_service posted functions, no dynamic allocations, but has no
> context switches). Am I missing something?

Beside the stack allocation (which could be preallocated and/or reused) 
the boost.context function will not be slower.

With 'context switches' I don't mean context switches in the case of 
threads!
'context switches' in boost.context consist of saving/restoring relevant 
registers like stack pointer, instruction pointer, non-scratch registers 
etc = the code generated by the compiler does the same (which registers 
are saved/restored by a function call is defined by the 'call 
convention' like cdecl etc.).

The ASIO example using boost.context is equivalent to using from 
functions do_<xyz>_() in this sense.

The benefit in using boost.context is, that the code can be written more 
straight forward. Your logic isn't scattered among several functions 
(do_<xyz>_()).
I think Chris might explain this issue better than me in its article 
'Thinking Asynchronously in C++' at 
http://blog.think-async.com/2009/08/secret-sauce-revealed.html. I 
adopted Chris example.

Oliver