Advanced NICs can shared buffers with the application to minimize data copying
overhead.  If we are going to write pluggable protocols that support such
optimizations we will have to make the CDR buffers available from the
connection.
Since the amount of memory available through that mechanism could be limited it
is important to count on GIOP message fragmentation to send large messages
through the ORB.
It could be hard to integrate this optimizations with other TAO features, such
as zero-copy for octet sequences (and other data types in the future), because
in that case the buffer is allocated by the application.