We are in the process of investigating whether or not to build our own MPLS network. We have a fibre network running SDH for our telecomms but obviously the ability to use this fibre for data as well is worth looking at.
We recently had a vendor present a paper to us on the use of this network and they seemed to be suggesting the following
1) Around the edge of the core network would be layer 3 devices which could do the MPLS VPN's etc.
2) Within the core there would be no equivalent as far as i can tell of the traditional P routers.
3) All Layer 3 devices would be meshed together with virtual links.
In effect for an MPLS VPN there would only ever be 2 hops from CE to CE, the first hop is the edge layer 3 device that the originating CE is connected to and the second hop is the egress layer 3 device.
The vendors argument was that the traditional model of an MPLS network where a lot of the layer 3 devices were used within the MPLS core merely as transit routers was an expensive and undesirable option. They also suggested that the industry as a whole was moving away from this model due to the expense and inefficiences of layer 3 devices.
What i wanted to know was what we would lose by removing layer 3 devices from the core of the network and in effect having may p2p links between edge layer 3 devices. They seemed to be suggesting nothing.
Hopefully i have explained this clearly enough. Any comments, insights, explanations would be much appreciated.
Those papers are always concluding the suggested solution is best ;-)
Well, let me put it this way: Why do you design a distribution and core layer in a campus design? Couldnt we avoid all those expensive Switches and fully mesh our access switches?
Yes we can ... only question is, if interfaces are remaining for user access ...
So I would see several issues with a full mesh PE design (with only a couple of minutes to think about):
1) Number of interfaces per PE (logical or physical)
2) Reconfiguration of every PE, when adding another PE
3) Number of IGP (OSPF, ISIS) adjacencies
4) Restricted routing design (summarization?)
5) bandwidth sizing for PE-PE traffic
1) can become an issue, 2) is an operational nightmare if your network grows. The underlying issue is the n(n-1)/2 problem of any full mesh
3) can also become an issue, depending on your number of PEs. Basically your memory and CPU consumption increases and convergence times increase
4) A hierarchical network design MIGHT help you in a very large network to scale your IGP.
This is practically excluded in the suggested design or at least administratively much more difficult.
5) typically bandwidth requirements will change frequently. How easy or difficult is it to scale the core and the PE-PE links to accommodate those changes? MPLS TE unequal cost load sharing would be f.e. a very suitable tool in a L3 MPLS core. In a TDM environment you will loose the advantage of statistically multiplexing traffic, which you have in an IP/MPLS environment.
Other features: MPLS TE, FRR, error detection through traceroute.
So a P router is only evil, if can not offer one ;-)
On the other hand, I have seen small networks (6 PE) starting out as full mesh. No growth was anticipated. So finally it might just depend on your network size - do you deal with a doberman or with Godzilla (Size does matter!)
My 2 cents. Hope they help!