RE: [boost] State of giallo, (sockets and demultiplexing)

From: Jeff Garland (jeff_at_[hidden])
Date: 2004-09-16 10:33:54

On Wed, 15 Sep 2004 21:20:56 -0400, Hurd, Matthew wrote
> > The only exception I can think of is if the data is to be 'multicast'
> down 2
> > TCP connections...
> 
> Another exception for you:
> 
> In place messaging with a "fast message" format would be nice for
> support.  Some network hardware, including TCP via TOE cards and
> infiniband and the like, support very low latency messaging, e.g. < 
> 10 microseconds.  You don't want to copy, but also you don't want to 
> serialize.  In fact, for many applications you are prepared to sacrifice
> architecture portability on the altar of speed and you might send a
> contiguous c++ object, padding from alignment and all, directly as a
> message block.  The CPU never has to see the whole object.

Sure, but I think it is pretty rare in practice. I've seen a few network apps
and none of them has needed this -- at least so far ;-)  And if you are
concerned enough about performance to need this, I'm wondering if you will
even buy into using C++ instead of C. 

Jeff