Sunday, February 9, 2014

IP Interconnection is Changing, Because the Commercial Relationships and Traffic Flows are Changing

IP network interconnection periodically erupts as a business issue between two or more interconnecting IP domains, and the problems will grow as the types of interconnecting domains diversify.

The interconnection issue further is complicated by the types of domains. Interconnections can occur between scores of thousands of “autonomous systems,” also called “routing domains.”

Though most of the autonomous systems are Internet service providers, interconnections also occur between enterprises, governmental and educational institutions, large content providers with mostly outbound traffic such as Google, Yahoo, and YouTube, as well as
overlay content distribution networks such as Akamai and Limelight.

In other words, end users, application, service and “access” and “wide area network” providers now are among the entities interconnecting, complicating any potential frameworks for regulating such diverse entities in ways that promote investment and innovation.

Where separate “common carrier” regulation arguably was easier, in the sense that only licensed “carriers” could interconnect, these days, application providers including Google, Apple, Netflix and others operate their own IP networks, interconnecting with carriers and non-profit entities alike.

The interconnection of IP networks historically has been a matter of bilateral agreements between IP network owners, with a tendency to interconnect without settlement payments so long as traffic flows were roughly balanced (the same amount of sent and received traffic on each interconnecting network).

As you can imagine, highly asymmetrical traffic flows such as streaming video will upset those assumptions. That matters, as a practical business matter, since interconnection costs money if traffic flows are not equivalent, or if domains are of disparate size.

Historically, the major distinction among different ISPs was their size, based on geographic scope, traffic volume across network boundaries or the number of attached customers. But symmetry encouraged the “peering” or “settlement-free interconnection” model.

Those assumptions increasingly are challenged, though. Today, a smaller number of large networks exchange traffic with many smaller networks. And there are cost implications.

In an uncongested state, a packet that originates on a network with smaller geographic scope and ends up on the larger network might be expected to impose higher delivery costs on the larger network (which must typically carry the packet a greater distance).

A larger network would presumably have more customers, and this might be seen as giving the
larger network more value because of the larger positive network externalities associated with being part of their networks.

Perhaps more important, even networks of similar size have different characteristics. Consumer-focused “access” providers (cable and telcos) are “eyeball aggregators.” Other ISPs, such as Netflix, are content stores. That has practical implications, namely highly asymmetrical traffic flows between the “eyeball aggregators” and “content stores.”

Also, there are natural economies of scale for a wide area network-based ISP than for an “access” ISP that has to supply last mile connections. Even when traffic flows actually are symmetrical, network costs are unequal.

The point is that settlement-free peering worked best when networks were homogenous, not heterogeneous as they now have become. Like it or not, the traditional peering and transit arrangements are less well suited to today’s interconnection environment.

For that reason, “partial transit” deals have arisen, where  a network Z sells access
to and from from a subset of the Internet prefixes.

For instance, Z may sell A only the ability to send traffic to part of the Internet, but not receive traffic. The reverse may also occur: a network may be allowed to receive traffic but not send traffic.

That arrangement is intended to reflect asymmetrical traffic flows between content store and eyeball aggregator networks.

Those changes in traffic flows, which bifurcate along content store and eyeball aggregator roles, inevitably will disrupt interconnection economics and business arrangements, leading to demand for imposition of common carrier rules for interconnection of IP networks.

Oddly enough, the logic of common carrier rules might lead to precisely the opposite “benefit” some expect.

Disagreements by parties to a bilateral interconnection agreement can lead to service disruptions, if one network refuses to accept traffic from another network on a “settlement free” basis.

So some might call for mandatory interconnection rules, to end peering disputes. Such rules could make the problem worse.

Interconnection disagreements today are about business models and revenue flows. Content stores benefit from settlement-free peering, since they deliver far more traffic than they receive.

Eyeball aggregators often benefit from transit agreements, since they would be paid for the asymmetric traffic flows.

Unless the assumption is that network economics are to be disregarded, the way common carrier rules would work, if applied to IP networks in a manner consistent with common carrier regulation  is that a network imposed an asymmetric load on a receiving network would have to pay for such access.

Disputes over “peering” between IP domains sometimes leads to service disruptions viewed as “throttling” of traffic in some quarters. It is not “throttling,” but a contract dispute.

The relationships between discrete IP networks take three forms. Large networks with equal traffic flows “peer” without payment of settlement fees.

Networks of unequal size tend to use “transit” agreements, where the smaller network pays to connect with the larger network, but also gets access to all other Internet domains. Also, in many cases one network pays a third party to provide interconnection.

Peering and transit rules are going to change, if only because the business interests of IP domain owners are distinct. The issue is whether such change will change to reflect the actual commercial interests, or take some form that is economically irrational.

No comments:

"Tokens" are the New "FLOPS," "MIPS" or "Gbps"

Modern computing has some virtually-universal reference metrics. For Gemini 1.5 and other large language models, tokens are a basic measure...