Re: [boost] C++ Networking Library Release 0.5

Subject: Re: [boost] C++ Networking Library Release 0.5
From: James Mansion (james_at_[hidden])
Date: 2010-02-02 15:09:10

Peter Petrov wrote:
> Agreed threads are alternative, but only for small servers without high
> scalability requirements (e.g. tens of thousands simultaneous HTTP
> keep-alive connections). The thread-per-connection model is outdated and in
> 2010 it's not a good idea to start a framework which relies on it.
>   
Well, I'd actually argue the contrary now that we have 64 bit systems 
and the old issues
of running out of VM addresses for stacks is gone. Having said that I 
too prefer async
IO - but I'm really not sure many applications need more than 10k 
connections, and
while I'd have baulked at that number of threads, I suspect that all the 
major server
platforms can handle that many these days.

> Coroutines could work, though I doubt the complexity (e.g. using yielding
> iterators with Spirit) or performance will be any better than simply using a
> FSM-based parser.
>   
Well, I'd prefer an FSM based parser, but Spirit sucks for that, since 
its a pull parser
and you are only able to get away with it by making assumptions about 
receiving a
'whole message' before starting the parse. Which implies either an 
assumption about
some alignment of read with content delimiters, or that you pre-scan the 
content to find
the delimiter - and process every byte twice.  A well executed push 
parser need not
do that, and ragel or re2c feeding lemon is the best approach I know of. 
I haven't tried
the ragel author's parser.

James