[c-nsp] mpls route target export question

[Oliver Boehmer (oboehmer)]

David,

> would be interested on some expansion of oli's comments on
convergence,
> and why differing RDs for the same prefix would be a good thing (I'm
> imagining perhaps if you want to do multipath then they are considered
> multiple paths as opposed to simply just selecting one because RD is
the
> same?)
 
Scenario: PE1 and PE2 advertising the same VRF prefix to PE3 via RR. If
the RD is the same on PE1 and PE2, the RR will only pass one of the two
paths to PE3 (as you said). Assume RR will select PE1 (as it has a lower
router-id or the like).
So 
a) PE3 will not be able to load-share his traffic over both PE1 and PE2
(which might not be an issue), but more importantly:

b) If PE1 goes down, PE3 might quickly notice this using Fast IGP, but
has no alternate path to fall over to. So PE1 needs to wait for BGP
control plane convergence (RR notices PE1 going down, runs bestpath and
sends PE2's prefix to PE3), which can take a while (several seconds to
minutes, depending on # of prefixes).
If PE3 had both path, he could converge more quickly. If PE3 has the
BGP-PIC-Edge feature, he can even converge all PE1's prefixes
*instantly* (even if there are thousands) after noticing PE1 going down
via IGP. 

So I strongly advise on using unique RDs, networks grow, number of
prefixes grow, customer expectations on convergences grow. Changing RDs
later (as Phil stated) can be a pain.

Downside: unique RD scenarios take up more memory as each importing PE
now has two copies per path (one with the original RD, and one important
with his own RD), which might or might not be a concern (it's ~150 bytes
or so per pfx, but only on the control-plane (RP memory)).

If memory is a concern: There is a optimal RD assignment: Use the same
RD on all PEs, except on the PEs which advertise a redundant prefix. So
all single-homed customers' pfx share the same RD, only the 2nd
redundant pfx has different one.. But this can be cumbersome to manage
manually..

I see more and more customers using the <ip-address>:<num> (rather than
<asn>:<num>) RD notation, and they simply use the PE loopback as the
address and then a "global" VRF number for the num part of the RD.. This
makes it simple to assign unique RDs per PE.

	oli