BOUNDING THE OUTCOME OF PEERING NEGOTIATIONS
One way to try to understand the context of negotiation between CD and A is to speculate about their relative market power. A has a terminating monopoly with respect to its customers, but CD may be hosting valuable content that the customers of A demand.
So which network has the stronger bargaining position? Earlier we noted our assumption that content producers and content delivery networks are unlikely to possess significant market power, but even if one were to relax those assumptions (and we expect that some will argue for just such a case), we do not expect such speculation to be productive: one can likely find specific circumstances in which one or the other outcome seems to hold.31 Ultimately, we believe that the determination of whether and on which side market power exists will depend on empirical facts that may vary case by case.16.4.1 Finding Limits on Payment P
Rather than make assumptions about market power, we seek to identify constraints or bounds on the outcome of negotiation between CD and A that might allow policymakers to infer whether the result of the interconnection negotiation was about a reasonable allocation of delivery-related costs or about an unreasonable allocation of the surplus associated with end users’ willingness to pay for content, above whatever it costs to efficiently deliver that content to the end users.
The model of traffic and payment flows described above is actually overly simplistic because ISPs and content delivery networks typically have multiple options for routing traffic from source to destination nodes. In reality, CD and A are typically embedded in a rich complex of interconnection agreements, and these agreements will limit the bargaining power of the two networks.
16.4.1.1 Transit as a bound on content delivery payments
Network CD may purchase transit from a third network T that subsequently either peers with network A directly or via its own transit agreements.
Other peering agreements can affect the delivery of traffic from CD to the ‘eyeball’ customers of network A. Network A has multiple peering partners, and may have its own transit provider to enable it to sustain its connectivity to the rest of the Internet and provide diverse routing options to enhance reliability and facilitate load balancing.In this case, if network A seeks to extract a large payment from network CD, then CD can choose to route the traffic to A indirectly via a transit connection to network T. Because the market for transit services appears relatively competitive, with prices consistently declining over time, the availability of this alternate routing choice provides a loose upper bound on what CD might be induced to pay A to deliver content to A’s ‘eyeball’ customers. The reason it is a loose upper bound is because A might be able to offer valuable performance enhancements (e.g., caching services) or other quality assurances that CD desires, thereby making CD willing to pay a premium above the transit payment Pt.
While (as we noted above) it is difficult to get internal cost numbers for CA, since most ISPs consider these proprietary, we can speculate that CA (measured in dollars/GB transferred) is larger than current values of Pt, and in many cases substantially larger. Access networks have large outside plant or access networks, and to the extent that these have a usage-sensitive cost component, these are likely to be much larger than transit costs. We speculated above that CA might be between $0.10 and $0.30/GB. Typical Pt for large volume agreements (e.g., with negotiated discounts) are currently as low as $1/Mbps per month (peak rate). Assuming average link loading of 50 percent, 1 Mbps is about 151 GB/month. Put otherwise, $1/Mbps is the same as $.0066/GB - less than one cent. So as a practical matter, we speculate that even if CD is persuaded by A to pay a premium over Pt, they will by no means be able to recover all of their internal costs, which makes it unlikely that there will be a content-related payment that is part of the negotiated payment from network CD to network A.
There is anecdotal evidence that the relationship we predict here is true in practice. Bill Norton, who tracks peering and transit issues closely, reported in January 2011 that: ‘The metered rate [of Comcast paid peering] is rumored to be in the $2-$4/Mbps price range, in the same ballpark as the market price of transit’.32 Our intention in this chapter is to explain why this relationship might hold.
16.4.1.2 Single-hop access
The existence of ‘single-hop interconnection services’ provides another way by which potential paid-peering payments might be moderated. Under this model, another network O negotiates a peering agreement with network A, and then O solicits network CD to interconnect with network O at the same physical location where O interconnects with A. This ‘single-hop’ interconnection arrangement imposes very few costs on O so it will still find it profitable to offer this interconnection option to CD at a very low markup over O’s cost of handing off traffic to network A.33 Those costs may be zero if O and A exchange traffic under a revenue-neutral peering arrangement.
16.4.1.3 Changing the routing restriction
We described two sorts of traditional interconnection: transit, which gives one party access to the entire Internet via the other, and peering, which implies a routing restriction on each party that the traffic exchanged between them is local to them and their customers. Normally, a network would not agree to route traffic coming in from one peering partner out to another peering partner: it would be forwarding traffic without being paid by either partner. However, once paid peering is an option, more variants open up for different sorts of routing restrictions.
For example, network CD might negotiate a partial transit agreement with network T, its transit provider, which guarantees delivery only to a subset of addresses. That subset might include network A’s end customers. Structurally, this would be like a paid-peering arrangement in that it includes a routing restriction.
Network CD would expect to pay less than Pt for a guarantee of delivery to only a portion of Internet addresses. Thus, the option of payment may facilitate more direct connections, and may also lead to greater diversity in the negotiated routing restrictions.16.4.2 The Real Picture
In the real world a large access network such as A might have several tens of peers, and might purchase transit from several providers. A content delivery network CD might have thousands of servers, each able to serve the same content. So CD might have thousands of choices as to how to source content flowing into A and can use these in a very nuanced fashion to control overall flows. This enhances the bargaining power of network CD because this finer-grained control over traffic gives them some control over network A’s costs.34
We noted above that traffic from networks of type CD into networks of type A now represents at least 40 percent of all traffic coming into A (Sandvine, 2011). Because content delivery networks control the routing of this traffic, control of routing has moved away from the low-level routing protocols, and into a space controlled by higher- level business agreements and subtle controls over the dynamics of how the large content delivery networks choose one source rather than another for content. This makes the management of interconnection agreements even more strategically important.
16.4.3 Norms of Negotiation
As we noted in the beginning of the chapter, as the old regime of revenue-neutral peering breaks down and is replaced by less constrained negotiation, we can expect to see the emergence of new norms and regimes of interconnection over time. The previous discussion hints at two sorts of norms. First are criteria by which one ISP would consider agreeing to revenue-neutral peering. One source35 examines a number of existing peering policies, and identifies 25 criteria affecting these agreements, of which perhaps ten are commonly used.
One such criterion is balance of flows, in which the data rates between the two parties are roughly in balance (perhaps no more than 2 to 1 in the peak direction).36 Balance of flows is a rather rough approximation for balance of value, as we discussed above, but it can be used to impose limits on behavior such as single-hop access. If all parties understand up front the maximum amount of imbalance that A will tolerate, this can avoid the pain of after-the-fact attempts to renegotiate a peering agreement.When revenue-neutral peering is not agreeable to both parties, we have speculated that a new norm might emerge to bound the price that might be charged for paid peering, which is that the rate for paid peering would be related to the price of transit. A proposal for a paid peering fee that greatly exceeds the customary price of bulk transit would be seen as evidence that the network proposing that fee does indeed have market power that allows it to distort the market. But a non-zero peering fee is not in itself a signal of such power.
Other norms might emerge, such as other proxies for cost (e.g., average route miles internal to an ISP), or industry average costs for outside plant. Average route miles could be used to bargain over the relative benefit of hot-potato vs cold-potato routing.
16.5