Exchange Points Are The Key
The Internet traffic is non-local (as was discussed above), meaning that there is no correlation between locations (both geographical and topological) of communicating parties. In the Internet most traffic is carried by several backbones of comparable size, so most packets will traverse one of exchange points.
For a number of reasons described in previous chapters, the number of exchange points is likely to remain relatively small; i.e. the switching or routing equipment at exchange points must by necessity have the highest capacity, and the exchange points are and will remain the most congested places. Unlike congestion related to an under-built infrastructure, which is easy to remove by adequate line capacity provisioning, alleviation of congestion at exchange points requires pushing the envelope of router performance.
Given exponential Internet traffic growth, the growth of traffic crossing exchange points is also exponential, which is exactly the phenomenon observed in practice (a few years ago the most loaded exchange point, MAE-East, was a shared 10Mbps Ethernet). Like the total Internet traffic, the traffic over every exchange point is doubling every 6-12 months. Increasing the number of public or private exchange points incurs unproductive expenses and increases the route flap amplification effect and the difficulties of traffic engineering; so ideally the number of exchange points should be about a dozen worldwide. This means that a radically new solution must be found to accommodate the future growth.
The current IXP architecture is very much like the architecture of backbone POPs:
Each participating internet service provider has its own router co-located at the IXP site and connected to the shared LAN. The reason for co-location is to allow easy addition of private bilateral links to offload traffic from the shared LAN. The routers perform routing policy computations on routing information learned from peers at the exchange point, announce ISP's interior routes to the peers (filtered accordingly to the routing policy), and forward packets to the appropriate members of IXP.
Obviously, the performance of such an exchange point is limited by the performance of the IP routers and bandwidth of LAN connections.
The most frequently proposed "new Internet" architecture assumes that backbone ISPs will have ATM backbones interconnected likewise at several exchange points (the more complicated routing algorithms required for connection-oriented networking make the routing flap problem even worse for global ATM networking), so the resource reservation on switched virtual circuits could be performed across several backbones. However, it is easy to see that given the high non-locality of Internet traffic, most virtual circuits will have to traverse the exchange points; so the computing power of the control units of the ATM switches at exchange points will have to grow exponentially to handle the requests for the creation and removal of virtual circuits. Again, the rate of such growth exceeds the rate of growth of capacity of semiconductor devices; i.e. the collapse of such a network is imminent.
In fact, a global switched-circuit data network cannot be built even at the present day Internet size -- some measurements show that the busiest Internet routers presently handling connections between the US and Europe top at about 0.7 millions of simultaneous TCP sessions (which would have to be encapsulated into virtual circuits). No known technology is able to sustain corresponding rates of creation and removal of the virtual circuits (about 60k events per second).
An ATM-based interconnection of flattened networks looks more feasible, and the corresponding interconnect architecture is shown in the following picture:
Note that there's only one virtual circuit connecting the networks. Introduction of a greater number of such circuits would have the same effect at IP level as multiplication of the number of exchanges (i.e. the network will get amplified route flap). A closer examination shows that all inter-network traffic still has to go though a single router on each side (or, at most a few routers), so the scalability of that approach is no better than that of the conventional IXPs.
It is interesting to note that conventional telephony does not have the exchange point problem, because long-distance companies (the "backbones") do not forward traffic to each other. The current telephony network configuration is shown in the following picture:
This is possible only as long as Regional Bell Operating Companies (RBOCs) have the monopoly on end-user service in their local areas (LATAs). Deregulation will make the telephone carriers face the same problem of capacity of exchange points. (In fact, in those rare cases when long-distance companies do forward traffic to each other, as in the case of international calls, massive machinery is required to serve the amount of traffic concentrated at those points - compare that with several IP routers that can fit in few racks that currently provide comparable gateway functions for transatlantic Internet traffic.)
The conclusion is very simple: delivering data at exchange points requires native IP routers. Those IP routers should be at least as fast as backbone switches; i.e. usage of ATM in backbones does not yield any performance advantage if inter-backbone traffic is considered.