BBN/GTEI

Fri, 21 Aug 1998, Owen DeLong wrote:

> Sure, but only the assymetry that results from BBN customers ASKING for more
> than they OFFER.

Or is it the asymmetry that results from Exodus customers OFFERING more
than they ASK FOR?

Since flow studies are confidential information, I will not quote from them.
However, I will say that I would think it intuitively obvious to almost anyone
on this list that Web Hosters don't OFFER very many packets that aren't a
response to a REQUEST. Now, if we were discussing a large content provider
that hosted alot of push-oriented sites, that might be a different story.

I don't think one of these views has any claim to precedence over the
other. Just because long distance phone calling introduced the purely
artificial concept that the initiator of the transaction pays for it does
not mean we should analyze IP traffic in the same way. In the past we have
considered the initiator of IP transactions to be irrelevant and had
no-charge peering for networks that basically send a similar number of
bytes to what they receive.

Ah, but with the advent of providers that focus primarily on access and other
providers that focus primarily on hosting, the similar number of octets concept
is disappearing. Not because one side is changing the balance, but because
both sides are changing. More and more of the access providers are becoming
larger and larger sinks at the same time that the hosting providers are
becoming larger and larger sources. The shifting of the traffic flows is
changing at roughly the same rate on BOTH sides, not because of one sides
business model.

Also, long distance phone calling didn't introduce the concept, it was around
prior to telephones. Look at how shipping was handled on catalog orders in
days before telephones. Generally, the purchaser pays shipping in almost all
commerce. This has been true for a very long time.

So what do we do when that is no longer the case?

I _REALLY_ don't see that much of a need for changing the way things are.
If the access providers cannot deliver the packets their customers request
from the point they say they want to receive said traffic to their customers
for the price they are charging said customers, then they need to reevaluate
their pricing model. Similarly, it is the responsibility (IMHO) of the
content provider to deliver the packets to the requested handoff point.

Owen

Since flow studies are confidential information, I will not quote from them.
However, I will say that I would think it intuitively obvious to almost anyone
on this list that Web Hosters don't OFFER very many packets that aren't a
response to a REQUEST.

I know nothing of packets being OFFERED or REQUESTED. I only know of two
networks which are peering and exchanging traffic. Neither network OFFERS
or REQUESTS data. They both perform an identical function of routing
packets through an exchange point. I can measure those streams of bits and
determine that one peer has more bits going out than coming in. This is
clearly an asymmetry. The question is that if such asymmetries become more
common and if the spreads between bits in and bits out increases, can we
continue to consider these two networks to be peers in the same way we
considered two networks to be peers in 1994.

Or do we need to update our definition of what peering means and provide
interconnect choices between pure peering and pure transit? The users and
the web servers are largely irrelevant background noise in this discussion
of peering relationships. Yes, we know that they exist and how they use
the network. Yes, we know that if networks specialize in certain types of
customers it will result in asymmetries at exchange points. That is the
context, but I believe that we can factor out that context from the
peering relationship itself and the reason that I believe we can and
should factor out that context is that BOTH TYPES OF CUSTOMER PAY FOR
SERVICE. Web browsing customers pay for service and web hosting customers
pay for service. Their relationship to their network provider is a done
deal. Once their packets enter the network they are merely packets full of
bits that cross exchange points between peers. Forget the end points of
the packet flows. Look at what is happening between the two network peers.

The shifting of the traffic flows is
changing at roughly the same rate on BOTH sides, not because of one sides
business model.

Yes. And I don't pretend to have all the answers here and I don't intend
to imply that only one type of provider should pay the other type. Let's
remember that any two peers are exchanging traffic at several points in
the network. But the only situation in which a dialup provider and a
webhosting provider will not have problematic asymmetry is when both
providers have POPs in the same set of cities and both providers
have private exchange points in every one of those cities. Then, although
there would be asymmetry of traffic flows, if you go by my model where
only inter-city transit has a value, there would be no value asymmetry in
this situation.

Generally, the purchaser pays shipping in almost all
commerce. This has been true for a very long time.

I hesitate to draw parallels between the movement of atoms and that of
electrons.

The users and the web servers are largely irrelevant background noise
in this discussion of peering relationships.

They are the customers who pay backbone operators for transit. I don't
think that that's irreleavnt.

Suppose that networks P and Q peer with each other at an exchange point
X, and network P has a customer A, while network Q has a customer
B. Traffic flows beteen A and B, in both directions (along the paths
A-P-X-Q-B and B-Q-X-P-A), but the traffic from A to B is much more than
the traffic from B to A (and so the traffic P-X-Q is much more than the
traffic Q-X-B). (The same argument extends trivially to asymmetric
routing through two different exchange points X and Y.)

Now, customer A is paying provider P for all the traffic in both
directions A-P-X and X-P-A, while customer B is paying provider Q for
all the traffic in both directions X-Q-B and B-Q-X.

Clearly, provider Q is paid (by its customer B) for every byte of
traffic that flows along the path P-X-Q, and for every byte that flows
along the path Q-X-P. Similarly, provider P is paid (by its customer A)
for every byte of traffic that flows along the path P-X-Q, and for every
byte that flows along the path Q-X-P.

In this model, it makes no sense for provider P to pay provider Q, or
for Q to pay P, because each has already been paid by their respective
customers. The cost of transit through network P (in both directions)
between customer A and exchange point X are paid by customer A, and the
cost of transit through network Q (in both directions) between customer
B and exchange point X are paid by customer B.

Web browsing customers pay for service and web hosting customers pay
for service. Their relationship to their network provider is a done
deal. Once their packets enter the network they are merely packets
full of bits that cross exchange points between peers. Forget the end
points of the packet flows. Look at what is happening between the two
network peers.

I believe that every packet-mile in the network operated by the web
browser's ISP was paid for by the web browser, and every packet-mile
in the network operated by the web server's ISP is paid for by the web
server. This applies even if the number of packet miles in one ISP
greatly exceeds that in the other ISP.

--apb (Alan Barrett)