Policy Routing

Mailman List Mirror (Read Only)
1

Hello,

I am a network engineer at a regional southeast USA
NSP. I am looking for some recommendations concerning
a scenario that has been presented to me.

My company is attempting to obtain company X's
Internet transit traffic, which will be BGP-4 peering
over either a T-3 or OC-3. Due to financial reasons,
my upper management has proposed that I route company
X's Internet traffic via a specific NSP that we peer
with, we'll call them NSP-A. Apparently, NSP-A has a
substantially cheaper rate than our other upstrem
providers and it is anticipated that this customer
will be sending a full T3 or OC-3's worth of traffic
to us.

Redirecting inbound traffic to company X via NSP-A can
be accomplished very easily through use of AS path
prepending, however, coming up with a solution for
egress traffic from company X to NSP-A, via our AS,
has proven a bit more challenging :-).

The only feasible solution that I've been able to come
up with is to stick customer X directly on the router
that peers with NSP-A and employ the use of policy
routing, which would enable me to set the next hop for
company X's traffic to the peering address on NSP-A.

Our NSP-A peering router is a Cisco 12016, running IOS
12.0(16)S2 and it has 256MB of DRAM.

Additionally, it is configured with NetFlow and dCEF
switching.

I've never employed policy routing in this type of
environment and I am concerned about the overhead that
it might place on the router or on the traffic
traversing the interface.

I've also thought about MPLS TE, however, our core
backbone does not run MPLS and even if we did, I
believe I would still have to policy route the traffic
to NSP-A once the MPLS label was popped off the last
router in the path in transit to the NSP-A peering
router.

Any ideas or comments would be greatly appreciated.

Thanks in advance,

Jeff

catesjl9394@yahoo.com

2

I would be very upset if I were "Company X" and I found out that you were
policy-routing my traffic to the "cheap" connection vs the best
connection.

Is it just me or do others on the list believe that in the absence of full
disclosure this would be shady at best?

3

John,

I appreciate your opinion, however I would like to
keep the responses to my question on a pure technical
level. I can assure you that there will be full
disclosure if this solution is implemented.

Thanks for your response :slight_smile:

--Jeff

4

I would be very upset if I were "Company X" and I found out that you were
policy-routing my traffic to the "cheap" connection vs the best
connection.

Is it just me or do others on the list believe that in the absence of full
disclosure this would be shady at best?

I had the exact same thing proposed to me at one point or another in a
different life, and refused outright. If the salespeople in question are
like most, they've already hyped the quality of whatever transit circuits
the SP in question has ... even if it hasn't been specifically mentioned,
piping a given customer's traffic out your cheapest transit point is
something I'd consider ethically questionable. I'm sure the customer won't
appreciate it either.

-travis

5

Replying to my own post with a bit more. (Forgive me!)

Rereading your post, one would believe that since "Company X" is a BGP
customer of yours, you're going to be sending them a full view. Unless
there is a knob that I'm not familiar with, that means that you're going
to be sending them the _BEST_ routes that you see in your core and not
just those from "NSP A" to which you are proposing to policy-route all of
"Customer X's" traffic. If this is indeed the case, I would think that
policy-routing the customers traffic destined for "prefix Y" via a
path other than the path listed in the NLRI you're sending "Customer X" on
their BGP feed is outright fraud.

Again, this is in the absence of full disclosure and it is my (non
esquire) opinion.

6

Given that there wasn't any indication that the "cheap" connection
was any less usable/reliable, I'd *expect* that the provider I was
signing up with would (as a sheer business matter) attempt to maximize
the difference between what I was paying and what transit was costing,
and as a result route me through th cheapest provider that met my SLA.

If I was a clever customer, I'd either phrase my SLA if I cared which
transit got used, or try to negotiate that some of the savings be
passed along to me :wink:

/Valdis

7

I've had the same sort of thing proposed. If the customer knows in
advance that you're going to send all their traffic through your cheapest
peer/transit path available, what's the problem? They get what they pay
for...and for this sort of service, they'd obviously be paying noticably
less than best path routed customers.

I haven't actually set one of these up yet, but I don't see why they'd
need to be directly on the router to peer-A...unless you're figuring the
policy routing overhead will be substantially lower if you're applying on
their interface rather than a generic interface (with additional traffic)
on the peer-A connected router.

8

Suppose you apply an output route-map on their BGP session that prepends
the heck out of paths where peer-A isn't your next as-hop? That way,
you're giving them full routes and making it clear via BGP that if they
have another path via another provider for non peer-A routes, the other
will be better.

9

John,

First: I agree with you at your main point 110% so my other
       comment is strictly technical in nature.

Second: Correct me if I am wrong, but I believe that if you send
        to company X full view over EBGP there is no technical
        reason to forward packets over different AS path.
        After all, you are advertising reachability via NEXT_HOP,
        which will be your border router.

Before you flame me, please let me reiterate that I agree with you
on the main point, that making a false/misleading AS_PATH advertisements
is bad. But I am just curious if it would work provided that you are
able to forward packets based on some 'coloring' scheme,
so please consider my comment more as a question then questioning :slight_smile:

Thanks,

Przemek.

10

Jon,

Unless I am misunderstand how BGP (and routing at all)
works something very seriously wrong with your suggestion.

Your border router will advertise to a customer only best
and active routes. You cannot advertise to BGP peer two routes
to the same destination. So modifying AS_PATH will not work,
because you will send to Customer X only best selected route
from your border router.

The only way to advertise suboptimal route would be to have
it selected as 'best' on your border router.

Of course, technically, you can advertise one thing
(AS_PATH via NSP A) , and forward it to a different
destination (NSP el'cheapo), but this is another story.

Please correct me if I am wrong -- it is 1:40 am
and I had a long day...

Thanks,

Przemek

11

For the record, this "cheap path" is an OC-12 to a
well connected Tier 1 provider that we got a
"sweatheart" deal on, and, it's only 2 percent
utilized.

Again, I want to emphasize that I wholeheartedly agree
with those who have commented on the concept of full
disclosure in a scenario such as this. I'm just
looking for technical opinions on how this can be
accomplished most effectively.

--Jeff

12

full disclosure, no problem. I've never personally seen a sales rep go in
and disclose something like this, and the original message was worded in
such a way ("[my] upper management has proposed that I route company X's
Internet traffic via a specific NSP ..") that did not indicate the customer
had any idea what was being proposed.

-travis

13

Jeff,

Please let me understand it a little bit more:

If this OC12 path is so good, why you will simply
not modify your LOCAL_PREF in such way that ALL your
traffic will go over this link?

In such way you will simply send to your customer
your BGP routing table and everything will be simple:
What you have in your routing table will be what your
customer will receive from you and (more importantly)
it will also represent how you are really forwarding
packets.

Przemek

14

I'll bet a small pizza that he's got an OC12 of capacity,
an OC12 customer, and doesn't want the OTHER 6 OC12's worth
of customers going that way...

Aint traffic engineering wonderful? :wink:

15

Unless I am misunderstand how BGP (and routing at all)
works something very seriously wrong with your suggestion.

Yeah...I should have put a few more seconds thought into that reply :slight_smile:

Your border router will advertise to a customer only best
and active routes. You cannot advertise to BGP peer two routes
to the same destination. So modifying AS_PATH will not work,
because you will send to Customer X only best selected route
from your border router.

This would basically tell the customer "don't send us traffic" since
assuming the cheapest path isn't the best path, the majority of routes
you send the customer would be non-cheapest peer paths with lots of
prepending.

Just doing simple policy routing based on the customer's source address to
a next-hop of the cheap peer could cause all sorts of problems as well.
What if the cheap peer doesn't even have routes to the destination? i.e.
the C&W/PSI peering issue a few months ago. Say C&W is your cheap peer
and you send customer-X traffic to C&W desined for PSI. C&W can't deliver
it and just drops it. Sucks to be customer-X.

It looks like this problem is much easier to solve for incoming traffic
than for outgoing.

16

Przemyslaw,

If I'm reading your post (between the lines) correctly, you intended to
say "if you send to company X full view over EBGP there is no technical
reason *NOT* to forward packets over different AS path." I may be wrong.

In any event, there IS a reason NOT to advertise full routes to a customer
that you're NOT going to honor (via the advertised AS path). Although the
immediate nexthop to the customer is going to be your router, we're
talking about a BGP speaking customer. If this was a static routed
customer, it would be much less of an issue.

With a BGP customer, one would think that either the customer is being
irresponsible to the routing table growth and speaking BGP as a
single-homed stub-AS, or (as I imagine in this situation) they're truely
multihomed.

In the multihomed case, if you're showing them full routes, the forwarding
path (AS path) should match what you're sending them in the NLRI. Why?

Say the customer sees the following:

You send them an NLRI for 10.1.1.0/24 via 65530_65531_65532

Their other provider sends them an NLRI for 10.1.1.0/24 via
65535_65534_65533_65532.

Your route looks better to the customer and if they don't color it in some
way, it's going to win because the AS-PATH is shorter.

Now, we throw the policy-routing into the mix.

Although you're advertising the NLRI via 65530_65531_65532, you're forcing
their traffic to go via "NSP-A" and that AS path is actually
65534_65400_65451_65501_65530_65531_65532.

You advertised a path to them that looked better, they used it, they got a
MUCH worse route than what their other provider showed them in all
actuality because you're not sending the traffic via the route you're
advertising to them.

There is no way for the customer to tune this because they never know
which routes you're going to be lieing about.

17

If the OC-12 connection is via a well connected provider and it's that
much less expensive than your other providers, and it's seeing that little
traffic, I would suggest a MUCH simpler alternative: Tune your route-maps
to pref more traffic towards that provider OVERALL. That way, you'll save
money on ALL of your customers and not just this one special case. :wink:

18

Again, I want to emphasize that I wholeheartedly agree with those who
have commented on the concept of full disclosure in a scenario such as
this. I'm just looking for technical opinions on how this can be
accomplished most effectively.

then you posted to the wrong mailing list. as you have discovered, this
list is mostly used to air engineers' social, legal, and business opinions.
try a technical/ops list, such as <your-router-vendor>-nsp@puck.nether.net,
e.g. juniper-nsp@puck.nether.net or cisco-nsp@puck.nether.net.

btw, may i presume you are trying to steer predominantly outbound traffic?
inbound can be steered by only announcing the customer's routes to the
desired provider, though that provides weaker redundancy. sadly, bgp
offers poor knobs for this, asn prepending is popular, albeit grotty.

randy

19

I don't see any policy routing in this example. ou don't need polic routing if you
want to direct ALL traffic
by one provider or if you want to have preferences affected ALL traffic.

You need polit routing if, for example, you have two providers, ISP-A and ISP-B,
and two customers, C-A and C-B, and
(by any reasons) want C-A to work b NSP-A and C-B to work by NSP-B (for IN and OUT
traffic).

Else, it's quite enougph to configure localpreferences and metrics.