Paid Peering among Internet Service Providers
Gireesh Shrimali
Stanford University
234 Packard
Stanford, CA, 94305
gireesh@stanford.edu
Sunil Kumar
Stanford University
273 Little eld
Stanford, CA, 94305
skumar@stanford.edu
ABSTRACT
We develop models for Internet Service Provider (ISP) peer-
ing when ISPs charge each other for carrying traffic. We
study linear pricing schemes in a simple ISP peering model
using a two stage sequential Nash game in which self in-
terested providers first set linear prices for carrying peers’
traffic and then choose to route their traffic according to
the prices set and costs incurred by carrying traffic on their
links. Under reasonable cost models, we show that ratio-
nal ISPs will participate in this game. Moreover, we show
that the ISP with the lower marginal cost in the absence of
peering has no incentive to send traffic in a hot-potato fash-
ion and effectively acts as a monopolist. The other provider
strategically routes traffic, splitting between hot-potato and
cold-potato routing. We also show that though this out-
come is inefficient, both ISPs are strictly better off when
compared to not peering at all. Finally, we consider appro-
priate cost models that make the notion of capacity explicit.
Under certain conditions we show not only that the monop-
olist has an incentive to upgrade the capacity of its links but
also that this incentive is higher when the monopolist is in
a peering relationship.
Categories and Subject Descriptors
J.4 [Social and Behavioral Sciences]: Economics; C.2
[Computer-Communication Networks]
General Terms
Economics, theory
Keywords
ISP, peering, pricing, monopoly, Nash equilibrium, subgame-
perfect equilibrium, capacity
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for pro t or commercial advantage and that copies
bear this notice and the full citation on the rst page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior speci c
permission and/or a fee.
GameNets 06, October 14, 2006, Pisa, Italy.
Copyright 2006 ACM 1-59593-507-X/06/10 ...$5.00
1. INTRODUCTION
1.1 Background
Today’sInternetiscomposedofmanydistinctnetworks,
operated by independent network providers, also referred to
as Internet Service Providers (ISPs). Each provider is inter-
ested in maximizing its own utility and the objectives of the
providers are not necessarily aligned with any global perfor-
mance objective. Most relationships between providers may
be classified under one of two categories [12]: transit and
peer. In a transit relationship, a traffic-originating provider
pays a third-party transit provider to carry traffic destined
to nodes outside the originator’s local network. On the other
hand, in a peering relationship two providers agree to accept
and carry traffic from each other.
In this paper, we focus primarily on peering relationships.
In a peering arrangement, a pair of providers agree to in-
stall bi-directional links at multiple peering points to accept
traffic from each other. In today’s Internet, peering relation-
ships are mostly “Bill-and-Keep” [2] due to ease of imple-
mentation. In this arrangement, the providers don’t charge
each other for the traffic accepted on the peering links. This
arrangement is also referred to as “Zero-Dollar” peering or
“Sender-Keep-All” (SKA) peering [6]. Additionally, in to-
day’s Internet, routing traffic destined to reach a peer is of
importance. Since ISPs are interested in minimizing their
local costs, they predominantly use the nearest-exit or hot-
potato routing [10], where outgoing traffic exits a provider’s
network as quickly as possible. However, in some cases,
where the receiver is a bigger player and is able to exert
its market power, the routing is farthest-exit or cold-potato
[12].
Due to the combination of “Bill-and-Keep” peering and
myopic routing, it may happen that only one of the ISPs
benefits from the peering relationship, and may free ride
on the other ISP. In this case, the ISP getting hurt by the
relationship may not want to participate in peering at all
[12]. This observation provides the main motivation for this
paper. Specifically, it motivates us to study whether simple
linear pricing schemes, in conjunction with rational routing,
would alleviate this problem and make it beneficial for both
providers to participate in the peering relationship.
1.2 Our Work
We analyze peering decisions using non-cooperative game
theory [3] in a simple two-provider model. We assume that
customers are bound to ISPs, subject to general, non-linear
marginal costs. We then look at how ISPs could charge each

other in response to the externality caused by their traffic.
We study a two stage sequential Nash game in which the
providers first set linear prices and then route flows strate-
gically according to the prices set and costs incurred by car-
rying traffic on their links. Our charging models may be
interpreted in one of two ways. In this paper, we focus on
the case where ISPs charge their peers for hot-potato traffic
going to destinations within their domains. (For simplicity,
the pricing of cold-potato traffic is assumed to be zero). In
this case, the ISP strategically chooses between hot-potato
(i.e., sending the traffic across) and cold-potato (i.e., keep-
ing the traffic) options. However, there is another context
in which our results apply: when peering ISPs also act as
transit providers to destinations not in their domain. In
this case, we could think of the ISP strategically choosing
between using the peer as transit (i.e., sending the traffic
across) and using a transit provider as transit (i.e., keeping
the traffic).
In our models ISPs incur costs, in addition to prices charged
by the peer, when traffic traverses links. We assume ISPs
incur costs for carrying all the (their own as well as their
peer’s) traffic over their own links but do not incur any cost
when their traffic travels over their peers’ links. Note that
this would happen when ISPs are myopic and they worry
only about their own networks. Common examples of this
case are the maximum load on a link (which translates to
reliability and maintenance considerations), the cost of run-
ning and owning a network etc.
Analyzing Nash equilibria, under mild assumptions on the
cost structures, we obtain the following results.
• We show that Nash equilibria exist for our game. In
particular, because our game includes a so-called par-
ticipation constraint in its formulation, this implies
that rational ISP’s would choose to peer and partici-
pate in the game. This participation cannot be guaran-
teed for the ubiquitous Bill-and-Keep peering schemes.
• Even though both peers choose to participate, we show
that that one of the ISPs has no incentive to send any
traffic to the other in a hot-potato manner. We iden-
tify this ISP aprioriby looking at the marginal costs
for the ISP in the absence of peering. Specifically, we
show that the ISP with a lower marginal cost network
before peering is the ISP that keeps all its traffic. We
then show that this ISP acts as a monopolist that ex-
tracts as much social surplus as possible.
• We then analyze this monopolistic behavior under lin-
ear pricing and show that, while it is not possible to
achieve a socially optimum outcome, both ISPs are
strictly better off when compared to not peering at
all.
• Finally, we look at the incentives of the monopolist to
upgrade its capacity, using cost models that allow us
to model capacity. We identify conditions under which
the monopolist has an incentive to upgrade. Further-
more we show that this incentive to upgrade is higher
when peering than if the ISP was not peering at all.
That is, peering increases the incentive to invest in
capacity.
1.3 Related Work
Various aspects of ISP peering have been analyzed by [8],
[2], [14], [15], [7], [16], [5]. [8] was the first paper to ana-
lyze ISP peering in depth from an economic perspective. It
analyzed the impact of symmetric access charge on strate-
gies of the providers and showed that, in a broad range of
environments, operators set prices for their customers as if
their customers’ traffic were entirely off-net. [14] extended
the models in [8] to include the fact that the ISPs are geo-
graphically separated. It thus analyzed the local ISP inter-
action separately from the local and transit ISP interaction.
It also analyzed the economics of private exchange points
and showed that they could become far more wide spread.
Both [8] and [14] used linear pricing schemes assuming fixed
marginal costs. In addition, they assumed hot-potato rout-
ing. [15] extended the models in [8] to include delay costs,
finding that they have a substantial effect on market struc-
ture. [7] used a different model of ISP peering. Similar to
our work, it assumed that customers are bound to ISPs and
looked at how ISPs could charge each other in response to
the externality caused by their traffic. However, it looked
at only one-way traffic exchange.
Our work differs from the above literature in two signifi-
cant ways. First, much of this literature (e.g., [8], [14] etc.)
assumes the ISPs to be symmetric in their cost structures.
This equal and constant marginal cost assumption results
in fixed and symmetric access charges. We analyze a more
general case where the ISPs are not symmetric and, in ad-
dition, have general cost functions. Second, these papers
do not worry about strategic routing once peering is set.
Although our routing is simplistic, it provides significant in-
sights.
Finally, [5] looked at a slightly different but related sce-
nario. It looked at service providers in tandem and showed
that the bottleneck service provider would have monopoly
power and it would extract all the social surplus. It also
established a negative result that the bottleneck link owner
would not have any incentives to upgrade its capacity. Our
work differs from this work in two ways. First, it looks
at peering service providers, who are effectively in paral-
lel rather than in tandem, and establishes that a service
provider may act as a monopoly, regardless of utilization
of capacity, depending on the criteria we mentioned earlier.
Second, and more importantly, we establish a positive result
that, under suitable conditions, the monopolist always has
an incentive to upgrade its capacity.
1.4 Organization
The paper is organized as follows. In Section 2, we de-
scribe the basic model used in this paper and present the
two stage sequential Nash game. In Section 3, we demon-
strate the existence of Nash equilibria and look at its prop-
erties. Finally, in section 4, we look at incentives to upgrade
capacity.
2. THE MODEL AND THE NASH GAME
2.1 The Model
Most of the previous work (e.g., [8] and [14]) looked at
the scenario where the ISPs charge each other in presence of
competition for customers. Since we want to focus on how
the ISPs would charge each other for externalities caused,
similar to [7], we assume that