[ih] order of source and destination address in IP header

[David P. Reed]

Craig Partridge wrote:

>
>Certainly the original Cerf/Kahn paper clearly states TCP operates
>between HOSTS, with GATEWAYS doing the relaying.
>  
>
I misspoke my point, as I said to Dave Crocker.   I meant that it was an 
inter-network protocol - that is the protocol spoken on the overlay 
network that spanned multiple networks.  So the idea that optimizing the 
"switching" via cutthrough was not even in the design space.

>Also, I note that IP's fragmentation model requires a source address.
>(Or perhaps, better said, would have been extremely difficult to do
>without either a source address or a unique, destination-created, nonce
>that was given to the source and placed in each datagram).
>  
>
Good point.   However, you can do fragmentation/reassembly without 
ambiguity without the source host address... remember that source 
host/dest host still doesn't provide an unambiguous reassembly (there 
may be many simultaneous connections between source and dest at the host 
level.)  It may not be ideal, and is quasi-non-deterministic, but if you 
have a crypto-based authenticator (SHA-1, e.g.) instead of a checksum, 
you can just deliver all fragments to every destination socket, and then 
try combining fragments until the authenticator checks correctly...   
not proposing that for lots of reasons, but just to demonstrate that 
"source host address" is not the critical element needed for 
frag/reassembly.