[Nsi-wg] A slightly different take on topology

[Henrik Thostrup Jensen]

Hi (sorry for being a bit close to the meeting with this mail)

I am probably not the only one who would like to have an NSI meeting that
didn't revolve around Jeroen and Jerry debating topology description, in a not
very productive fashion. Guy did a good effort trying to control this last
week, but it sure is hard keeping those two under wraps :-).

So far I've tried to keep (mostly) out of it, but I'll try to give my views on
the topology issue. The first thing we need to figure out is what we want to
do. Are we trying to solve just the VLAN problem or coming up with a solution
for multilayer path finding / creation and what scope of problems are we
working with.

Solving the VLAN problem is quite easy IMHO. Multilayer is significantly more
tricky.

I've read some of the papers Jeroen has linked to. In particular the "Path
selection in multi-layer networks" illustrates what it is Jeroen is aiming for
(I think). The paper is quite clever and displays a couple of nice tricks that
can be done to solve some seemingly impossible tasks if one has (very) detailed
information (I especially liked the graph conversions).

However it is very much a research paper, and I do not believe the 
approach will work - for several reasons. First is complexity, the 
proposed pathfinding algorithms is rather complex, but it could probably 
be implemented. Second reason is that I do not see it feasible to collect 
so detailed topology information, let alone load/availability information. 
Third reason is that the approach completely ignores any policies that may 
be in place in a network. E.g., NORDUnet has a policy for how to do VLAN 
rewriting (always on ingress), and we have no plans of letting other 
organizations decide how it should be done. Furthermore NORDUnet and SUNET 
(Swedish Reseach Network, which is operated by NORDUnet) shares some 
resources, and wether or not one is allowed to use these resource is 
highly dependant on who you are.

One of the really nice things about the current NSI model is that is promotes
encapsulation. This allows us to define the network we want or need without
regard to the underlying physical infrastructure. Going with this direction
means will be cases where there is a potential path, but we do not have the
means to create it for some reason or other (policy or technical). I am okay
with this, as long as it doesn't become the norm.

On the other hand Jerry haven't really managed to give a coherent view of what
it is he wants. I do think that Jerry has some good ideas, mostly because he
realises the need for abstraction and encapsulation in the system, providing a
means to actually comprehend and understand what is going on in the system (and
hence build and debug it). Some of the ideas with recursive domains and
hiarchial STPs are quite good, but they haven't really been fully formed yet
IMHO.

What is clear, is that we will probably not find our final topology model 
at todays meeting, especially not for multilayer. However we are very much 
standing still and need to move forward. Could we, as a _temporary_ 
solution use Tomohiros STP enumeration in the reserve request in order to 
specify a set of potential STPs to connect with in the reservation 
request. This would solve our most pressing problem (VLANs) and give us 
some more time to work out the multilayer pathfinding/creation (which we 
will need). This would not require any protocol or topology changes, but 
an implementation change for parsing the STPs, and it is backwards 
compatible.

Hopefully some progress will be made in Baton Rouge (I'm not betting on it 
though), but we simply need to get topology out of the NSI meetings as 
there are also other issues we need to focus on (security comes to mind).


     Best regards, Henrik

  Henrik Thostrup Jensen <htj at nordu.net>
  NORDUnet