[ih] internet-history Digest, Vol 37, Issue 6

[John Day]

That is interesting. I would have thought it was a little bit sooner, 
but then there isn't much room for it to be much sooner.  ;-)

It is in some sense ironic.  We were so concerned about overhead 
(both bandwidth and processing) that it was thought that the 
transport protocol had to operate directly over whatever was at the 
network layer.  The Europeans knew they had to deal with X.25 and the 
pseudo header pretty much prevented TCP from use with X.25 directly. 
And it was thought to be too much overhead to run transport over an 
IP like protocol over X.25.  So WG6.1 chose the older TS protocol. 
which was then carried into OSI as a liaison contribution to become 
TP4.

However, when OSI finally got around to working out what the Network 
Layer was all about (IONL) they found that it was 3 "sublayers." 
Ironically, X.25 became a subnet access protocol at the bottom 
(referred to as 3a) of the network layer (since that is what the 
title of the document said it was!, i.e. a DTE-to-DCE protocol)  ;-) 
Not the network protocol at the top of the network layer. The top of 
the network layer was subnet-independent (referred to as 3c) or what 
has been called the internet layer, where CLNP was connectionless 
subnet independent protocol.  (For some reason the connection 
advocates in OSI never did the connection-oriented subnet independent 
protocol. They were invited to.)  For completeness, 3b was suppose to 
be whatever was required to convert between subnet-dependent and 
subnet-independent.  Although, I don't know of any such protocols 
ever being proposed.

So OSI ended doing what WG6.1 had avoided. Transport (TP4/TS) over an 
internet protocol (CLNP) that names the node over a network protocol 
(X.25 or IP) that names the interface.  In the space of just a few 
years the overhead issue had become a non-issue.

Had we gone this route, we wouldn't even have noticed the issues the 
current silliness emanating from the RRG are struggling with. Where 
the false distinction of loc/id split (you can't identify something 
without locating it and vice versa) has backed into loc/loc split. 
(Like that old joke that Communism was the longest most tortuous road 
from capitalism to capitalism.)  But it just keeps getting messier 
and messier as they try to save IPv6 from its own incompetence.  To 
think this was all well-understood before 1990, actually to some 
extent by 1972.

Of course, once one understands that the view of Transport and 
Network as separate layers is the telecom (beads-on-a-string) view 
and not the systems view and understands the natural structure of 
that class of protocols, you realize that IP is unnecessary as a 
distinct protocol.  In essence, TCP was split in the wrong direction. 
It would have been better to split vertically to separate "control 
from data" which yields a much more powerful, efficient and flexible 
design.

We were all too heavily influenced by the politics of the debate and 
what we thought was the case and not enough by what the problem was 
telling us, i.e. the science.  In some sense, it is too bad that 
software is so forgiving and Moore's Law allowed us to avoid problems 
that we could get by and didn't have to dig deeper.  Now we have 
quite a mess on our hands.



At 6:55 -0500 2009/11/09, Vint Cerf wrote:
>this comports with my recollection. Danny Cohen, David Reed and Jon 
>Postel lobbied strongly for a non-sequential, fast delivery 
>mechanism and TCP 3.0 made reference to a new header that performed 
>this function (routing). We decided later to split off into a 
>distinct protocol document, developed UDP for user access to this 
>mode of operation and put TCP on top. This was all codified in the 
>first TCP and IP v4 documents in 1978.
>
>v
>
>On Nov 9, 2009, at 6:27 AM, Craig Partridge wrote:
>
>>
>>>Also, IEN 21 is dated January 1978.  The split must have occurred
>>>well before that because if *my memory serves* ;-) one of the
>>>discussions in arriving at INWG 96 by Dec 1977 was whether or not the
>>>protocol chosen could be used over something other than a datagram
>>>service and for TCP, IP already existed.
>>
>>A few years ago I went digging to try to find out when and where IP
>>was invented.
>>
>>The simple answer appears to be it was created in a hallway conversation
>>at an TCP meeting at ISI in 1977.  (Full answer is more complex and
>>involves tracing various research efforts and ideas to that hallway...).
>>
>>Thanks!
>>
>>Craig