BGP help

[saimike]

in trying to setup a BGP link between 2 cisco routers, my coworker and i has a little disagreement. perhaps a cisco expert out here can help us clear this up:

the issue is: what ip addr should we use in the "neighbor x.x.x.x ..." line under the "router bgp xxxx" statement?

every doc i've seen (written by cisco or otherwise) use the neighbor's interface ip in the "neighbor x.x.x.x" ... my coworker insists that it's better to try to use the lo0 ip of the neighboring router.

it seems that the ip addr used in the neighbor line needs to be route-able without using igp etc. thus the neighbor's inteface's ip addr works perfectly. but to use the lo0 addr, one solution would require adding a static route.

what would be the most elegant way of doing this?

 

[randal]

using the loopback interface for iBGP is a little bit more reliable (it never goes down) and is considered a best practice. For eBGP, though, it gets a little bit unwieldy but is doable. Usually you use a loopback interface, set your `neighbor x.x.x.x update-source lo0` to force it to use the loopback interface as the BGP source, then, as you said, add a single static routes to both ends. A lot of providers won't do that, though, so your mileage may vary.

IGPs=loopback's best friends
eBPG=hit or miss on loopbacks, usually ends up being directly connect interface

 

[spidey07]

Randal is dead on.

Loopback should be used and considered best practice. You'll see this a lot with iBGP

BUT - with eBGP you're normally talking about point to point links and you would use the interface IP as it truly gives you direct neighbor relationship without relying on an routing/IGP. If the link happens to go down the BGP neighbor relationship is immediately down and can help with quicker convergence.

So as with all networking the answer really is "it depends"

 

[saimike]

sorry, i forgot to specify that these are eBGP links. so would lo0 or the interface be a better way of doing things from the perspective of convergence speed, scalability and realiability?

these cisco routers are internal to a company (our client) so they have total control over it. there are also other cisco routers linked to them in a serial fashion, talking bgp:

router a <-eBGP-> router b <-eBGP-> router c <-eBGP-> router d ...

all routers are in pairs for redundancy, and some have crosslinks to their neighbors (but not all).

i'm inclined to think that using loopback would be more elegant. however, it seems that the replies so far indicate that it might the less elegant interface IP method might be preferable.

our client should be easily persuaded to go either way. so we're mainly looking at it from a purely technical view point here.

 

[spidey07]

well only eBGP should be used between different ASs (as by definition) and in that case the interface will suffice.

so I'm a little confused why a single client would have all the eBGP going on. try posting on forum.cisco.com

 

[randal]

Using the loopback interface is great for iBGP + IGP links because if an interface goes down, the IGP will re-route the traffic accordingly, making it look like the session never died and making convergence extremely fast.

For eBGP, where things are generally serial lines, there is generally not another good path to take, and hence the eBGP session goes down. If you actually do have a mesh-style network with BGP, then using the loopback interface as your update-source will decrease your convergence times considerably, but will add some load to your routers.

From your network infrastructure you described, I'm not sure why you're using eBGP -- especially if it's all the same company. I'm not an IGP master yet (on the way!), so I'll defer to spidey when it comes to assessing your internal network architecture.

 

[saimike]

thanks for all the info guys. i think i got my answer: use the loopback.

i just remembered, the routers all have redundant links with their immediate neighbors ... so as was pointed out, using lo0 will help keep the network up.

 

[saimike]

as an aside, does anyone know the convergence times if i use lo0 vs physical interface.

for the lo0 scenario: assume all routers have 2 routes to get to its peer. one is 1 hop (ie. direct), the other is 2 hops.

for the physical interface, the 2nd route still exists physically. and can be used if configured.

 

[imported_netman]

Actually the results were.. don't use the loopback if you are using eBGP.

 

[spidey07]

Originally posted by: saimike
as an aside, does anyone know the convergence times if i use lo0 vs physical interface.

for the lo0 scenario: assume all routers have 2 routes to get to its peer. one is 1 hop (ie. direct), the other is 2 hops.

for the physical interface, the 2nd route still exists physically. and can be used if configured.

this can only be done in the lab as it depends on table size and a whole slew of other configuration parameters.

Without looking at this from an AS/route policy perspective it is very difficult to help. I can't for the life of me think of a good reason to use loopbacks for eBGP. In eBGP if the neighbor is unreachable you want it to be immediately declared down - which contradicts the use of loopbacks as an always up interface. The only other reason is possibly flapping and being penalized but there are other methods you can use to combat flapping/dampenig.

I realize you're trying to prove your position here and possibly to the client but in all reality "it depends" and I cannot make a blanket statement one way or the other without fully understanding what the goals are.