[c-nsp] MPLS LDP question

[Oliver Boehmer (oboehmer)]

Frederic,

> I have a question regarding LDP and FEC.
> When activating MPLS LDP, by default all FEC(s) present in the routing
> table have a label bindings ...
> 
> To my understanding, only (BGP) Loopbacks are  involved when deploying
> MPLS services such as MPLS VPN / PSEUDOWIRES / TRAFFIC ENG etc...
> Therefore, design wise, it makes sense to restrict LDP to "service
> loopback".
> 
> So the question is, according to your experience :
> 1) Is it worth to spend some time restricting LDP only to service
> Loopback ?
>     => In terms of resources allocation
>     => Faster convergence
>     => Other consideration ?

Yes and yes, even though resource allocation shouldn't be a major
concern unless you carry huge number of pfx in your IGP (and then you
might have more urgent problems to worry about ;-)
The other consideration is that restricting it to these "service
loopbacks" actually enables you to differentiate between LSPs and
regular Ipv4-switched packets, for example pinging loopbacks (using
MPLS) vs. pinging link addresses (not using MPLS). Granted, the wider
availability of MPLS-OAM has made this argument less compelling.

> 2) Is there any trade off in a scenario where only loopbacks have LDP
> label bindings ?

Not aware of any, apart from the potential of error/incorrect ACL which
ends up filtering label assignments for loopbacks who need to be
label-switched when allocated from a different/new range. To reduce the
margin of error, I've seen deployments where customers just restrict
label advertisments to /32s, which definitly covers all loopbacks, and
possibly a few others which don't strictly need a label. Unfortunately
this approach doesn't work everywhere ("mpls ldp advertise-labels "
expects a standard acl :-(), but newer IOS have "mpls ldp label allocate
global host-routes", which would achieve this..

	oli