Existing Internet And ATM "Clouds"
In the overview of the existing and proposed future Internet topologies we will concentrate on the backbone networks particularly because the current problems mostly affect the high-speed backbones; the low-speed leaf networks are easily upgradable.
The existing Internet is constituted by a number of "parallel" backbone networks of comparable capacity, interconnected at a number of Internet exchange points (IXPs):
An important feature of that architecture is that there is only a limited number of exchange points; and although that number is growing, it is not likely to exceed more than few dozen. There are several reasons for that: the technical reason is that IXPs create alternative paths, thus multiplying the routing information that needs to be processed by each backbone router; the economical reason is that the connections to IXPs are expensive, and do not directly generate any revenue, so it is cheaper to have fewer high-speed IXPs than many lower-speed IXPs; and, finally, the "political" process of selection and negotiating for a place for an IXP facility, and peering arrangements makes introduction of new IXPs slow and far from certain.
Every backbone has a number of points of presence (POPs), each one serving a particular geographical region, so access circuits from customers in that region are concentrated at the point of presence. Each point of presence also has several backbone lines coming in; the customer access lines and the backbone lines are connected to a plurality of IP routers clustered around a high-speed local area network (LAN) switch:
The routers in a POP are usually separated into backbone routers (ones with high-speed backbone lines connected) and customer access routers (servicing quantities of low-speed customer access lines). The reason for such separation is to reduce the number of routers participating in the backbone routing (or to reduce size of iBGP mesh, for more technically inclined readers), so the dynamic routing algorithms would have less information to process and would react to routing changes faster.
The clustered design is fairly redundant. The failure of a backbone router, of a backbone line, or of a customer access router will not cause complete interruption of service to the customers of that region. LAN switches are usually backed up with a secondary LAN.
However, this design has a serious drawback, in that the aggregate switching capacity of POP is limited by the capacity of the LAN switch. Worse yet, the capacity is severely limited by the bandwidth of the LAN connections between the LAN switch and the backbone routers. Internet traffic is highly non-local, i.e. most communications would not be confined in the region serviced by POP. It can be seen easily that an effective switching capacity of a typical cluster composed of 10 Cisco routers (2 backbone and 8 customer access router each with 0.7Gbps backplane capacity) and a DEC GigaSwitch (with 6Gbps of backplane capacity) would be a little more than 400Mbps (the capacity of two full duplex FDDI links!).
A partial solution for the POP capacity problem is to replace FDDI with a higher-speed LAN technology, such as ATM or HIPPI (an example of such a solution is Cisco Fusion architecture). However, the aggregate capacity remains limited by the bandwidth of LAN links, multiplied by factor of 2 or 3. Increasing the number of backbone routers is not a solution, either, because it degrades network stability and increases routing convergence time (We'll explain why this is critical later in the discussion of routing flap).
The currently deployed alternative approaches are the so-called "flattened" networks and all-out desktop-to-desktop ATM networks.
A flattened network has an ATM backbone with IP routers at its edges:
In the flattened network, IP routers are connected with permanent virtual circuits, forming a fully meshed system of interconnections mapped over the physical backbone lines. Obviously, the flattened networks do not have the performance limitations of the conventional clustered-routers IP networks. However, as the reader will see later, the flattened networks are not more powerful if used as components of a global Internet; the performance advantage of such networks is only available for virtual private network-type applications. The flattened networks do not provide any additional functionality, such as ability to carry isochronous traffic; in this respect they are equivalent to native IP networks.
The pure-ATM networks are very much like native IP networks topologically, with POPs consisting of a single large ATM switch (or a large ATM switch surrounded by a collection of smaller fanned-out ATM switches); the virtual connections are dynamic and established when host and desktop computers wish to communicate, and torn down when the communication is complete. The pure-ATM networks allow for end-to-end resource reservation and multiple classes of traffic, a feature that is perceived as the major advantage of pure-ATM networking.