[c-nsp] MPLS speakers behind unreliable link

[Oliver Boehmer (oboehmer)]

Pshem Kowalczyk <> wrote on Tuesday, January 13, 2009 01:18:

> Hi,
> 
> We're in early stages of a project that is supposed to connect
> multiple small MPLS-speaking devices over DSL links. Currently we're
> investigating various options (lab trials will come later) and I would
> like to get your input. If you find the idea to be completely
> ridiculous - let me know ;-)
> 
> The whole network will consist of 500-1500 PE devices (28xx and
> possibly some other vendor counterpart) using DSL. Those devices will
> connect the the core MPLS network to provide end-to-end connectivity
> for the end users in various sites. Bandwidth requirements are very
> low (up to 20Mb/s per site).
> 
> We're considering two main scenarios:
> (remote PE is a small device at the end of the DSL line, local P/PE is
> a bigger device connected over GigE or POS to the MPLS core).
> 
> 1. The core network provides a single VPLS domain all remote PE belong
> to. That VPLS domain is used to run separate MPLS between the remote
> PEs. The biggest problem we can see so far is the fact that this will
> create a very big broadcast domain, possibly spanning over 1000km,
> with many devices directly attached to it over DSL and as a result -
> quite unstable. We're not sure if ISIS or OSPF can handle such
> situation gracefully.

Ack, 500-1500 peers on the same LAN is doomed to fail. You could,
however, create multiple VPLS LANs, and interconnect them using
dedicated routers. Still not a nice approach..

> 
> 2. Somehow connecting all the remote PEs to local P/PEs (multiple
> remote PEs connected to one local P/PE) and using local PE as sort of
> aggregation point, that would hid the instability of the DSL network.
> We haven't done anything like this before, so I'm not even sure if it
> can work - using ISIS create L1 domains from the remote PEs, make the
> local P/PEs a L1L2 devices and use L2 to connect to the core. Would
> label distribution work in a scenario like that assuming LDP for the
> next-hop and MP-BGP for vpn information? After all a ISIS L1 is a
> completely stub network, so it shouldn't see any routes from L2. Is
> that the case also for LDP (i.e. LDP will not generate a label for a
> FEC (prefix) that is not advertised into a L1 domain?)

This would work, but you would need to leak the BGP next-hops (or L2 PW
router-IDs) from the L2 into the L1 areas to provide an end-to-end LSP.
I would consider this a reasonable approach. Make sure you use a
dedicated loopback address range for all your remote devices so you can
easily create an ACL for route leaking ("redistribute isis ip level-2
into level-1 distribute-list <ACL>").

Obviously, the result of link flaps (i.e. loopbacks coming and going)
would still be propagated throughout the whole domain, but you can use a
less aggressive prc-interval setting on your nodes. So achieving
aggressive, sub-second ISIS convergence could be a challenge if the
network is not stable.

	oli