[ih] Exterior Gateway Protocol

[Jack Haverty]

Hi Noel,

Yes, I can confirm that MIT was one of the culprits motivating the
creation of EGP.   I don't remember the details of the particular
incident, but I do remember that MIT and Dave's Fuzzballs were notorious
sources of disruption.

At the time, there were two conflicting goals, running a stable core
Internet as a service, and providing a "real" environment in which to
try out new ideas for our list of problems-to-be-solved.  EGP was
intended as a way to have your cake and eat it too.

The motivation for a stable service was obvious - to prove that TCP/IP
worked and could be deployed as a DoD standard.   But there was also a
strong desire to continue research aggressively, because of various
scenarios that would be poorly handled by the initial routing mechanism.

I probably don't remember them all, but there were a bunch of such
"routing problems" beyond the basic desire for efficiency and
stability.  At the ICCB meetings, we used to keep a list of "things to
be worked on", many of which were associated with routing.  A few I
remember:

- Type-Of-Service Routing: how to route packets differently for
different needs, in particular low-latency for voice/video, versus
high-throughput for file transfers.  That's why we put the TOS field in
the IP header.  Another example would be handling ICMP packets with
priority over others.

- Expressway Routing: how to route packets through a separate network,
even if both ends were attached to a common network.  The example was
how to route traffic between two ARPANET hosts through gateways to the
WideBandNet, rather than just sending packets directly through the ARPANET.

- Policy Routing: how to route packets depending on who "owned" them.  
The example here involved the two paths between the US and Europe at the
time.  The X.25 path cost money to use, and ARPA wanted only
ARPA-sponsored packets to use the satellite (SATNET) path which ARPA was
paying for.

- Multipath Routing: how to route packets along more than the one "best"
path, if two (or more) independent paths led from source to
destination.   This was coupled with another problem - how to take
advantage of Multi-homed Hosts that had connections to multiple
networks.   The goal was to achieve higher host-host throughputs by
simultaneously using all viable paths.

- Capacity Routing: how to route packets depending on the sustainable
bandwidth of possible paths, rather than hops or delay.  This would
allow things like file transfers to finish sooner, even though the
individual packets might not get there in the fastest way.

These usage scenarios were in addition to the common goals of using
delay-based metrics instead of hops, which were well understood to be a
poor metric. 

At the time, the gateways simply had too little memory and CPU to
implement anything fancy, and IIRC the core gateways didn't have
appropriate clocks either.  So it was tough to see how to measure things
like gateway queueing or network transmission delays.   Dave Mills was
especially interested in dealing with and measuring delays, which I
think motivated him to drive the creation of NTP.  

EGP was intended to keep the "operational" part of the Internet running
smoothly, while enabling the research community to solve those
problems.  I've always been curious about how, and if, those "routing
problems" ever were addressed in any subsequent IGP, or whether they
just disappeared into the fog of history.

/Jack Haverty

On 9/2/20 7:12 PM, Noel Chiappa via Internet-history wrote:
>     > From: Craig Partridge
>
>     > Could Bob have been thinking about splitting off DoD IP networks to
>     > another provider ... and using EGP for that purpose?
>
> I was involved with an early incident that _may_ have been part of the
> impetus for EGP; I was out of the loop that started EGP, so I can't say for
> sure. Well, let me tell the story, FWIW.
>
> MIT got a router (gateway as we called them then) between its collection of
> LAN's (mostly CHAOS speaking at that early stage) and the ARPANET somewhat
> early on; later than would have been optimal, but there were no spare IMP
> ports at MIT, and we had to wait for the installation of IMP 77 (one of the
> first C/30 IMPs, I think) to get connected to the Internet. (Jerry Saltzer
> worked out when that was, by looking in hardcopies of CSR progress reports
> that he had in his garage; I can look it up if it's important. My very vague
> memory is that it was roughly around 1980.)
>
> The first ARPANET gateway at MIT was bodged 'Port Expander' code (I found a
> version of it in that CSR dump, if anyone cares; I'm not sure it's the
> version that was in service, all that stuff is pretty disorganized). To
> exchange packets with the rest of the Internet, we needed to let the BBN
> gateways (pretty much all there were at the time) know that MIT-GW was the
> route to net 18. We decided to do that by hacking up a small MOS process to
> run in the same box, which would send a small, short GGP routing update that
> just said 'net 18. this way' every so often.
>
> So I coded it up, installed it in MIT-GW, and apparently as soon as I started
> it, the BBN gateways started falling over. IIRC, I got a phone call from
> someone at BBN, asking what I was doing that was causing their gateways to
> crash. This was news to me, I had no idea I was doing that. Anyway, after some
> investigation, we figured out what had happened: BBN had changed the format of
> GGP routing update, but hadn't documented the change. So when I dutifully took
> the routing update format from the GGP spec online, it was malformed in a way
> that killed the BBN code.
>
> I have this memory that when this tale was told to Vint at the next Internet
> WG meeting, he took out a little notebook and made some notes, and shortly
> thereafter we found out that the Internet router work at BBN had moved
> between Div 4 and Div 6 - we were under the impression at the time that that
> change was a result of those router issues.
>
> I don't know if the start of EGP was also connected, but those events
> certainly showed the downside of having new routers trying to talk GGP to the
> BBN routers.
>
> 	Noel