[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Tunneling vs routing protocol modification
hi folks,
I think Walid outlined a bunch of the important issues, and I have a
few comments. First, I'd like to qualitatively discuss the issue of
"rmp vs relay approaches" at a high-level, considering the
differences.
o I like the rmp approach because it deals with the problem
directly, i.e., it specifies a link as unidirectional and does
the right thing depending upon the routing protocol in use. In
contrast, the relay approach is sneakier. It solves the problem
by semanticly dissolving it. This means network designers and
maintainers must remain conscientious about what's really going
on, i.e. packets which travel back to the feed across the split
virtual interface of the relay are in fact encapsulated. For
this reason, I would classify the rmp approach as more general
than using relays for the purposes of routing over
unidirectional links.
o However, I also see value in providing bidirectional link
semantics on unidirectional links. I feel some networks may
encounter difficulties deploying the modified protocols, or
there may be other "control" protocols out there we have not
considered where bidirectional semantics prove useful. It may
also in some cases ease deployment of new protocols where
unidirectional links were not considered. While use of a relay
constitutes a "hack" of sorts, it is an elegant hack. I think it
deserves treatment--as evidence of this, consider that at least
three groups have arrived at it independently. And I think the
udlr wg is as good a place as any to hash it out.
This brings me to a question. Can we continue to work on both
approaches at the risk of ultimately providing two viable solutions
to the problem? With some more work, we may discover one or both of
them intractable, but I don't think we know enough yet to make that
determination.
Now if you can bear with me a bit longer, I have a few remaining
comments which more directly address some of the issues brought out
in Walid's message.
[On Fri, 4 Apr, Walid Dabbous writes:]
#
# Tunneling vs routing protocol modification
# Elements for discussion
[...]
#
#
# c) easy configuration (without scalability issues)
#
# the RPM solution is auto-configurable. WIDE and Aerospace
# tunneling solutions propose a dynamic configuration to
# address pb1 above. Aerospace solution is based on a solution
# inspired from ICMP router discovery. When a receiver detectes
# a feed it sets a tunnel (entry in table that will trigger
# encapsulation). However, we need a mechanism to "inform"
# the routing protocol about the asymetric metric to use
# for this tunnel (as cited before).
#
This mechanism will depend upon the routing implementation on the
reciever. As was discussed on the list, I think the best we can do
is call more attention to it in the draft.
[...]
# e) Scalability
#
# The problem is the same for both approaches.
# Tunneling solution may benefit from Mutlicast tunneling
# (e.g. a receiver establishing a multicast tunnel with all
# feeds, encapsulating routing information in multicast datagrams).
# Can someone provide more information on that?
The big scalability problems arise in the face of large numbers
geographically distributed receivers. This constitutes a network
architecture not commonly seen in internets, leading to implosion of
state on the feed side. One solution to this problem involves
increasing the number of feeds to share the responsibility of dealing
with all the receivers. I think there are a number of ways to
accomplish this in a relay-based solution.
As you've described, multiple feeds can receive the same encapsulated
returning packets if a common multicast address is used as the relay
endpoint. This should work as long as the encapsulation protocols
is "stateless" and supports class D addresses. If the encapsulation
protocol is not "stateless", you will have a much bigger scalability
problem on your hands anyway due to a state implosion at the relay
endpoints.
Another option is to send multiple packets with duplicated payload
to different relay endpoints. You could argue this is less
efficient the above, especially for large numbers of geographically
dispersed feeds, but this way you don't need multicast all the way
back to the endpoints.
Finally, you could divide up the physical forward medium into a
number of subnets, each with a different corresponding relay
endpoint. Receivers would only encapsulate data bound for a
particular subnet to its corresponding endpoint, but they could
maintain multiple split virtual interfaces on different subnets, and
therefore multiple relays.
Or you might use any combination of the above. The vipre draft
attempts to incorporate the potential existence of all of the above
scenarios, i.e., class D relay endpoints and stateless encapsulation
as well as multiple split virtual interfaces and multiple relays for
each interface.
thanks for listening and to those at the 38th, enjoy the home of
Graceland,
james
--
James Stepanek <stepanek@aero.org> 310/336-7911