Best Practices for Loopback addressing (Core routers & VPN CPE)

Mailman List Mirror (Read Only)
1

Hello,
I was wondering what are the choices made by Service Providers on the
loopback addressing.
The context is an IP/MPLS Backbone providing both Internet and BGP-VPN
services.

I have 2 different cases to address :

1) Loopbacks on the backbone routers :
I have the feeling that general practice is to use public IP adresses for
Core routers.

However, considering that these loopbacks are only used for routing
protocols (OSPF,BGP, LDP)
and for network management (SNMP, telnet, ...) and that these addresses
don't need to visible from public Internet
(not seen in traceroute, not seen on Internet BGP announces ...) I am
considering to
use private RFC1918 for a new Backbone deployment.

N.B. : Assumption is that e-BGP sessions with Internet peers are done on
public interface IP, not on loopback IP.

Is there some specific case I am missing where public loopback IP is
required, and therefore
private adressing would break something (maybe some Carrier-to-Carrier
scenario ?) .

I also plan to use RFC1918 addresses for Internet CPE routers loopbacks.

2) Loopback on CPE routers of the MPLS VPN customers.
For this case, the issue is to assign the adresses in a global range for
all the CPE of
all the VPN customers.
In fact, all these loopback will need to be part of the Network Management
VPN for supervision needs.
Using RFC 1918 addresses might create trouble as there is a very high
chance that the VPN customers
are already using 1918 addresses, this might generate addresses conflicts.
Addresses unicity among all the customers is required due to the Network
Management VPN common
to all the customers.
Using public address guarantee unicity, but will create issues with public
registries, considering that
these addresses are used for internal needs.
I am considering to use the 198.18.0.0/15 defined in RFC 2544 and listed in
RFC 3330 as reserved for
lab testing.
I suppose that no VPN customer uses this prefix for its internal IP
addressing, and as these addresses don't
need to be announced on Internet.
Do you suggest to use an other prefix than 198.18.0.0/15 for this purpose ?

If you consider your adressing policy as touchy topic in terms of
security, don't hesitate to reply in private ...
Regards,

2

However, considering that these loopbacks are only used for routing
protocols (OSPF,BGP, LDP)
and for network management (SNMP, telnet, ...) and that these addresses
don't need to visible from public Internet
(not seen in traceroute, not seen on Internet BGP announces ...) I am
considering to
use private RFC1918 for a new Backbone deployment.

Or, you could use a seperate class C or whatever fits yoru backbone for loopbacks and router interfaces.. Just don't advertise that block. That way you use non-rfc1918 on the backbone, and yet outside people cannot get to it since you dont advertise it to the world... It's just me but i am against using rfc1918 on any part of a backbone.

-hc

3

If the BGP Identifier, which is used for connection collision resolution and
path selection (among others seemingly random things) conflicts you'll have
issues.

In my experience even mediocre IP addressing frameworks typically don't have
issues with RIRs when space is appropriately justified (i.e., there
typically aren't issues with loopback and inter-router address space
justification).

What I'd be more concerned with is loopback IP allocation and it's effect on
aggregation, stability, and other network policies (e.g., source-interface
type stuff). For example, using a single contiguous block for all loopback
IPs significantly simplifies filtering policies. OTOH, you may opt to
provide more optimal aggregation and allocate loopback IPs from the same
block as p-t-p IPs (per router, POP, region or other) such that less
information needs to be carried internally in your IGP (or BGP).

RFC 2519 provides some guidelines for inter-domain route aggregation. A
slew of other documents and books provide IP address allocation guidelines
as well.

-danny

4

Consider the situation where you have a peer or customer who needs to do
ebgp multihop peering from loopback to loopback. This happens
infrequently, but it does happen. You need public IP address space to
(reasonably) make this work. I know you are assuming this won't happen,
but the day you need to provision two OC-12s to the same provider or peer,
and want to load balance them effectively...

Thanks,
Dan

5

Well, as you mentioned, it's fairly infrequent that these situations
arise, but it's not unheard of. However, in this case, you can always take
an IP address out of the interface pool and create another loopback
address specifically for the purpose of load balancing to the
peer/customer.

I'm actually not advocating using 1918 address space in the backbone, just
trying to point out a solution to a problem. :slight_smile:

Guy

6

In situations like this, I find it helpful to provision an additional loopback interface for each peer that has more than one connection to the same router. This lets us remain in control of our own destiny rather than relying on outside parties to do reconfiguration when circuits need to get moved between routers.

I wouldn't agree with the original poster's statement that using address space obtained from one of the public registries for internal purposes will be an issue. ARIN, et al shouldn't have a problem with using address space obtained from them for internal uses (i.e. not announced to the public internet) If you have concerns about how they might react to this, I'd suggest engaging them directly for advice on how to structure your request. In my past experience, I've found it easier to discuss my request with ARIN prior to submitting it, just to be sure they understand what I'm intending to do, and that I'm giving them all of the necessary data. The last thing you want to do is have that dialogue once they're already reviewing your request, it's likely to make the process take far longer than necessary.

FWIW, I am loathe to use address space that is not publically routable for anything I can't rip apart and renumber inside of a few hours. There are a many ways to give yourself a headache that are far more enjoyable. :slight_smile:

Chris

Daniel Golding wrote: