The MPLS-OPS Archive

["Adrian Farrel" <adrian@olddog.co.uk>]

>  --- Adrian Farrel <adrian@olddog.co.uk> wrote:
> > Sylvia,
> >
> > One "detail" you hide is switching resources.
> > Labels may be sparse resources and switching a fat
> > pipe only takes one label (or one slot in your
> > matrix).
>
> I agree, but that too amounts to stitching, I "switch
> equal capacity on port A to port B"
>
> Now how is that different from "simply hanging a wire
> between end points connected to A and B"

 

If you will only ever have one LSP across your core, I agree that hierarchy doesn't save you much (although the technology boundaries may still make it desirable so that (e.g.) you odn't have to look at packets on a lambda router.

 

The saving comes when you have bulk parallel LSPs across the core. That ain't stitching.

 

(NB. Some people like to see stitching as a special case of 1:1 hierarchy without label stacking. I wouldn't like to comment :-)

 

> > This get more important (but this is not the only
> > reason for the function) when you change granularity
> > of switching commodity. For example, if I want to
> > send some small packet LSPs down a WDM fiber I
> > could:
> > - map each LSP to a lambda, but I would quickly run
> > out of lambdas
> > - wrap all of the packet LSPs in a lambda and demux
> > the packets at
> >    each router, but this is inefficient at the
> > router (which might not even
> >    be able to handle that number of packets). Note
> > that in this case
> >    you are sort of building a single hop hierarchy.
>
> So if I draw out what I understand:
>
> N1-----N2======N3--N4
>         |       |
> N5------        -----N6
>
> N3 and N2 will still allocate individual "lamdas" to
> some entity "which has TE properties" i.e N5 and N6?
>
> In other words, do I:
> a. "switch" lambdas across? As you yourself mentioned,
> that does not make sense.
> b. switch components/fractions out of those lambdas,
> then I anyways cannot do that as N2 and N3 just have
> switching but not TE properties.

 

Your figure is perhpas a rather simplistic network.

Let us insert a whole network cloud between N2 and N3 so that the hierarchical LSP runs N2,n1,n2,n3,n4,N3.

N2 will allocate a single lambda to reach N3.

This lambda will carry all of the traffic from N1 to N4 and from N5 to N6.

The traffic at N3 is easily discriminated so that it can be sent to N4 or N6. But within the core network (at the nx nodes) only one lambda needs to be switched.

 

In TE terms, the LSP N2,n1,n2,n3,n4,N3 is presented to N1, N2, N4 and N6 as if there was a link from N2 to N3. Thus TE path computation can make use of this link and from the perspective of N1, it is establishing an LSP that goes N1,N2,N3,N4. The fact that the nodes n1, n2 etc. are used is transparent.

 

> In other words, there is no problem here between
> "manually provisioning" the lambdas permanently is
> there?
> What could be the purpose of "signalling" needed in
> this case?

 

There is no difference in effect between a manually provisioned LSP and an LSP provisioned by the control plane. The purpose of a control plane is well discussed elsewhere, but includes dynamic utilisation of resources (so that the core can have resources that are used for different purposes at different times), dynamic provisioning (to simplify the management plane interactions), and responsiveness to network utilisation and faults (to allow repair and recovery as well as optimization).

 

Now it may be that you asking (subtly) how and when the FA LSP is provisioned. Is it triggered when the first LSP tries to cross the core? Is it pre-provisioned through management planning? Are new FA LSPs added automatically when load increases?

 

These would be good quesitons. The answers are (of course) implementation and deployment specific. It is quite likely that many FA LSPs are established through amangement action and sit around waiting for traffic (this is like, but not identical to, the provisioning of physical links - it is clearly somewhat quicker!). However, the concept of the Virtual Network Topology (VNT) is now gaining some traction - this involves a more dynamic approach to the provisioning of FA LSPs depending on traffic demand.

 

> I understand that this could be classified as a node
> with switching but no TE properties, but in a
> simplistic view, it is as good as having a set of
> permanent toggle switches which once provisioned, will
> be left as they are.

 

You must see the TE graph as a hierarchical construct. So the nodes you describe as having switching but no TE properties do not appear in the higher TE graph - in our example the "higher" TE graph is

 

N1-----N2======N3--N4
        |       |
N5------        -----N6

 

That is, the nodes n1...n4 are not visible and the FA LSP N2-N3 appears as a TE link.

 

The lower TE graph looks like

 

N2--n1--n2--n3--n4--N3

 

and is unaware of the outer world.

> Is my understanding correct?

 

Partially :-)

> > Now, in your figure, the topology is linear and here
> > your main saving is in the detailed switching and
> > processing at transit routers. But even then,
> > several parallel, tiny LSPs (maybe only 1 kbit) are
> > hard to dynamically fit into an OC-48 switching
> > network without some form of hierarchy or huge
> > bandwidth wastage.
>
> Is that mainly a limitation due to the fact that the
> Oc-48 switch is seperate from the "access" switch?

 

It is a function of the bandwidth granularity that the switching device switches. If we take a more dramatic example: a 10G lambda can only be switched as a 10G lambda; if you stitch a 10Mbit LSP to the lambda LSP, you waste nearly all the bandwidth. If you nest the 10Mbit LSP down the lambda LSP you can fit many others in as well.

> In other words, there is node further down which has
> the capability to switch at kbps level, but can send
> out a max of say 1/4th Oc-48? these sub OCs can be
> bundled on node -B in an Oc-48. But as far as I can
> tell, the quantum is still sub-OCs which are of
> interest, if the single node in the middle cannot
> switch "subs" of oc-48, it really is nothing but an
> extended cross connect?

 

Yes!

You can see the core network as a giant (virtual) cross connect. This is one of the models used to provide layer 1 VPN services.

> Just like the analogy above. It does not have an TE
> properties.

 

As above, it does have TE properties, but only at its specific level in the network.

> > This may become clearer looking at a topology such
> > as:
> >
> > I1          E1
> >   \        /
> > I2-T1-T2-T3-E2
> >   /        \
> > I3          E3
> >
> > and you wish to operate three LSPs I1-E1, I2-E2,
> > I3-E3 then you can save switching resources (labels)
> > on transit routers by building a tunnel from T1 to
> > T3 and placing the three LSPs inside.
> >
> > Another detail that can be saved by the use of
> > hierarchical LSPs is complexity in the traffic
> > engineering database. Consider a packet network
> > where the routers are connected over a TDM network.
> > The packet network wishes to believe that its
> > routers are adjacent even though they are connected
> > by many hops through the TDM network. To do this,
> > you operate the TDM network on its own and establish
> > LSPs between its edges - these LSPs are presented as
> > TE links (FAs) to the packet network (think of
> > virtual connections).
> >
>
> Yes, but we are making the assumption here that the
> TDM network at the base does not participate in
> signalling, hence has no switching/TE capabilities.

 

No, no, no (no, no).

 

It does have signaling and TE capabilities (or rather it may or may not have). The thing is that it is not part of the TE and signaling at the higher level.

 

Hint: hierarchy is not an accidental choice of word.

 

> So from the view point of the routers, 2 links over
> different TDM paths, as long as they terminate on the
> same node, still form a bundled FA between the 2
> nodes.

 

Correct.

> Can I understand the case where a "node may have
> switching but not TE properties", but would *ever need
> to pariticipate in signalling* ?

 

I'm not sure I can answer that question as phrased :-)

 

Again, my point is that these nodes either:

- appear on a (lower) TE graph with full switching and TE properties

or

- are completely invisible on the (higher) TE graph

 

Adrian