Preferring peers over customers [was: Do Not Complicate Routing

Mon Sep 5 01:54:27 UTC 2011

Forgive my potential lack of understanding; perhaps BGP behavior has
changed or the way people use it has but my understanding is -

Since BGP is used in almost all circumstances in a mode where only
the best path to a prefix can be re-advertised, only one of the
peer or customer path can be used by a 3rd network, and if the peer 
path is used for a prefix for a customer, then a transit provider can't 
easily provide transit for that prefix back to the customer without 
serious routing shennanigans.

So isn't it in practice the case that if a provider prefers a peer to 
connect to a customer instead of the direct customer link, that:

1) The provider will lose the ability to bill for traffic delivered
   to that customer, and

2) The customer will lose redundancy of inbound path, and

3) The customer will almost certainly notice and have the chance to complain

I would expect that most cases of a provider (for a given prefix, which
is almost always a caveat here) preferring a peer to get to a customer 
would be something the customer had some input into via communities,
or by calling and bitching if the provider doesn't have a rich 
communities set or the ability to set them.

One thing one hears every so often (in cycles) is the pressure for
emerging Tier1/2 aspirants to not peer with customers of larger potential
peers who are also providers, to preserve revenue models of said larger 
peers, but that's a different situation.

And -

If applied to customers of customers, I'd think it'd revert to the cases
above.  Network X has customer Y and buys from provider C.  If C prefers
a peer to get to Y (this is all for a given prefix) and it wasn't due 
to policy expressed by X or Y via communities or request of provider C 
by X, then eventually someone's going to figure out that the backup path 
that presumably X and Y think is being paid for, isn't.  Then the people 
that pay money will bitch and action shoudl be taken.

Consistent announcements by a global network to its peers for the prefixes
of a given customer is another level of wonkiness that customers can
definitely influence by doing strange per-prefix communities settings,
but that again is probably another topic as it'd be presumably driven by 
the customer's actions, not the provider's traffic-engineering goals.

Or am I confused here on one, more, or all points?  Certainly possible.

One thing I think everyone can agree on - academic models of the ways 
that people combine routers, money, fiber, contracts, and policy almost
never catch up to the creativity, poltiics, policy, bugs, and stupidities 
that combine to make the Internet as wonderful as it is.

Avi

> On Sep 5, 2011, at 4:03, Randy Bush <randy at psg.com> wrote:
> 
> >> Because routing to peers as a policy instead of customer as a matter
> >> of policy, outside of corner cases make logical sence.
> >
> > welcome to the internet, it does not always make logical sense at first
> > glance.
> >
> > the myth in academia that customers are always preferred over peers
> > comes from about '96 when vaf complained to asp and me (and we moved it
> > to nanog for general discussion) that we were not announcing an
> > identical prefix list to him at east and west.  the reason turned out to
> > be that, on one of the routers, a peer path was shorter in some cases,
> > so we had chosen it.  we were perfectly happy with that but vaf was not,
> > and he ran the larger network so won the discussion.
> 
> The "myth" comes from engineers at large networks saying it is so.
> 
> We could also have a small miscommunication here.  For example, if a custome=
> r were multi-homed to a peer, and the customer and peer were on the same rou=
> ter, and the customer had prepended a single time (making the AS path equal)=
> , by your original statement you would have sent traffic to the peer.  Most p=
> eople would find that silly.  (And please do not point out customers and pee=
> rs do not connect to the same router, this is a simple example for illustrat=
> ive purposes.)
> 
> However, the statement you make above says that you preferred the peer becau=
> se "the path was shorter".  You do not specify if that is IGP distance, AS p=
> ath length, or some other metric, but it implies if the path were equal, you=
>  would prefer the customer - especially since the customer was preferred on t=
> he other coast.  So there may be assumptions on one side or the other that a=
> re not clear which are causing confusion.
> 
> Either way, this seems operationally relevant.
> 
> I would like the large networks of the world to state whether they prefer th=
> eir customer routes over peer routes, and how.  For instance, does $NETWORK p=
> refer customers only when the AS path is the same, or all the time no matter=
>  what?
> 
> Let's leave out corner cases - e.g. If a customer asks you, via communities o=
> r otherwise, to do something different.  This is a poll of default, vanilla c=
> onfigurations.
> 
> Please send them to me, or the list, with this subject line.  I shall compil=
> e the results and post them somewhere public.  If you cannot speak for your c=
> ompany, I will keep your name private.
> 
> Thanx.
> 
> --
> TTFN
> patrick