[ih] EGP vs. BGP

[Louis Mamakos]

BGP also carries around explicit policy information for each of the
routes being announced.  Previously, an EGP peer might attempt to
apply policy based on the peer from which the route was announced, but
more complex topologies become problematic.  Consider that at the time
with the emergence of the NSFNET and some other agency networks, there
were upcoming acceptable use policies and other related politics to
attempt to control what networks would provide transit for what
others.  EGP hadn't the tools available to pursue this in a direct
way, not having really anticipated the problem.

The other significant difference is that BGP is run within a network,
as well as between peer networks.  EGP was an routing domain edge
protocol, and a network operator loss whatever policy information that
he might have gleaned (e.g., I learned this route from the NSFNET)
when the route was injected into the IGP.  Thus, the heroic efforts
that Dave mentioned with transforming metrics between the EGP and the
IGP to at least avoid generating routing loops.  However, the policy
information was lost.  BGP would allow this policy information to be
propagated from edge-to-edge of a transit routing domain and enable
more interesting network topologies (like multiple "backbones") to be
built and operated.

>From a technological view, the EGP transport was such that a complete
EGP announcement of routes was contained within an IP datagram.  As
the Internet grew in size, the announcement had to be fragmented to
fit over the available link MTU and there was a concern on the maximum
size datagram that could be transmited, fragemented and then
reassembled at the EGP peer.  Of course, loss of a single fragment
would likely cause the loss of the entire multi-fragement update.

There was work on an new version of EGP to address some of the
transport issues, but that was overtaken by BGP and the greater
toolset available to operators therein.  Too bad we hadn't the
foresight to include classes route prefix announcements in BGP; that
would have saved some very exciting transition work some years later.

I wonder if we could do another major transition, like EGP to BGP, on
the Internet system today?  I recall the joke at one time that
whatever came after BGP would still be BGP.  Looking at that protocol,
there is extensive accomodation for incremental new feature
deployment.  Can't just schedule a flag day.

Louis Mamakos

On Tue, Aug 18, 2009 at 07:31:45PM +0000, David Mills wrote:
> Dave,
> 
> The EGP specification rfc904 includes a hop count field, but not how to 
> use it. The Fuzzball implemented a Bellman-Ford routing algorihtm with 
> split horizon and hold down. Late in life it connected about 1500 
> networks and the ARPAnet implementation. The major change with BGP was a 
> serious approach to loop prevention. The Fuzzballs had a weaker approach 
> to loop prevention using crafted metric transformations between the RIP, 
> GGP, EGP and Hello routing protocols.
> 
> Dave
> 
> Dave CROCKER wrote:
> 
> >
> >
> > Vint Cerf wrote:
> >
> >> The design contemplated multiple networks and alternate paths from the
> >> start. The configuration of the 3 net test artificially forced 
> >> traffic from
> >> PRNET to ARPANET toSATNET to ARPANET again. The routing protocol in 
> >> ginny
> >> strazisar's gateways was distance vector and I believe would handle 
> >> multiple
> >> paths and backbones. V
> >
> >
> >
> > OK.
> >
> > So, did EGP reflect this flexibility?
> >
> > (I am remembering the anecdote of Bob Bradent's working from UCL back 
> > to ISI over a satellite link, having it freeze when the line go down, 
> > go off to have lunch or dinner, and return to a resumed connection; 
> > TCP doesn't have timeouts and it was only later that o/s 
> > implementations made them common. I am wondering whether the 
> > constraint to a single backbone for the Internet was an implementation 
> > choice, rather than being mandated by the standard routing protocol.)
> >
> > Hence, what were the incremental benefits provided by BGP?
> >
> > d/
> >
> > ps. This list is
> 
>