The MPLS WG Archive

[Curtis Villamizar <curtis@avici.com>]

Noel,

Your lengthy message was trimmed before replying.

In message <200003311927.OAA16024@ginger.lcs.mit.edu>, "J. Noel Chiappa" writes
:
> 
> Then you could do what Curtis alluded to - something less than full N^2.
> This has the limitation that paths through the fabric between routers X and
> Y may actually exist - but such a path is not represented in our
> "virtualization" of the ATM network, and the only way to get from X to Y is
> to find some router Q which *does* show a pair of virtual links, one to X,
> and one to Y, and take an "extra hop" through it.

In optical networks there is a good reason to do this.  You may have
very little traffic from A to C and not that much from A to B and B to
C either.  If you only set up a path from A to B and B to C, you can
still get from A to C without congestion, and you use one less lambda.

In IP over ATM if each POP i has two routers for redundancy, called
R_{i1} and R_{i2}, then you need to connect R_{in} and R_{jn} for all
i!=j and for n={1,2} and you will also want to connect R_{i1} to
R_{i2}.  This provides redundancy for a network with N POPs without
creating 2N(2N-1) adjacencies.

> You can go on for a while, trying different things, and you soon discover a
> simple truth: in basically all real networks, there's *no* virtual
> representation of a part of the topology graph which is both i) simpler
> than the full representation, and ii) accurately and fully describes the
> connectivity through that part of the graph.

This is not a problem.

> So, where does this leave us, here in the real world, with the real tools at
> our disposal - tools which unfortunately don't allow us to create any virtual
> representation, but just a few selected ones?
> 
> Well, the main conclusion is that if you're producing non-optimal results,
> it's not necessarily because your tools are bad - it's an impossible problem
> even with the best possible tools.

Depends on your view of optimal.  If you are optimizing your use of
the available lambdas, then not creating a path from A to C in the
above example is optimal because the extra hop cost didn't cost
anything (you needed to have those router interfaces at A-B and B-C)
and saved a lambda that did cost something (you didn't burn another
lambda and didn't neeed an extra pair of router interfaces at A and
C).  Whether this optimization is best done offline is still an open
issue.

> I will also stick to my earlier conclusion: the problem will be easier to
> tackle in a system which uses a uniform routing system at both the local
> and higher layers. Whatever information flow you *do* decide to do (i.e.
> whatever virtual representation of the local topology you do decide to
> propogate out to the higher layer), you will have more flexibility in doing
> so if the two are using the same routing system.

Another way of thinking of this is that bypassing LSRs with optical
paths and not doing the electrical routing will always be making
slightly less efficient use of the lambdas that it is using than if
the lambdas were terminated and routed.  OTOH if lambdas are available
and cheaper to use than additional routing capacity, then from a
dollar cost standpoint, the network would be better off bypassing some
of the routers with PSC tunnels that are optically switched.

> 	Noel

The point of my earlier message was that we need a signaling mechanism
that does not constrain us down the road and some of the internet
drafts seem to allow that.  For example:

  An offline NMS can configure explicit paths on the edge LSR and the
  edge LSR can set up an LSP and use it as a PSC through the optical
  core signaling the explicit path.  This gives us a static overlay
  that is signaled from the edge.  This can also be accomplished by
  configuring each network element separately (as was done in the old
  days with PVCs) but constrains us to this approach only.

  The edge LSR can signal an explicit path for an LSP with one strict
  hop to the immediate optical device and a loose hop to the far end.
  The LSP can be used as a PSC.  This provides the signaled overlay
  (same functionality as the open model).

  If we get to the point where the peer model is feasible, the loose
  hop becomes a set of strict hops.

There is still plenty of work to be done to make the peer model work
given the optical network's characteristics but the signaling won't be
a restriction if it is designed to accomodate all three methods.
Additional information will have to be flooded later for the peer
method to work but the setup is not likely to need change.

Curtis