Few points for making it efficient

1. The call reset () on the CDR should be removed and its a waste since we only 
use a fresh CDR
2. The GIOP class should cache a CDR stream with the GIOP header written the 
first time. This is perfectly valid 

   - the client holds a lock on the transport and so you can use the cached   
version with the same header again and again

   - the server needs to asses the GIOP minor revision before sending the reply 
out and its a single write operation after indexing to the right position. 

for #2 to be possible the ACE/TAO_OutputCDR streams need to have a way to chain 
CDR streams or the internal message blocks.  

This should save CPU cycles in the outgoing path.