1.3 Circuit and packet switching

If one had to give a very succinct description about how the Internet works, one would say it as being composed of end hosts and routers interconnected by links, as shown in Figure 1.4. In more detail, the Internet is a packet-switched, store-and-forward network that uses hop-by-hop routing and provides a best effort network service. This technology was chosen because it enabled a robust network that made an efficient use of the network resources [11,43,172,14].

Figure 1.4: Simple architecture of the public Internet as described in textbooks.

However, the real Internet is more complex than this, and if we look closely, we will find that there is plenty of circuit switching in the Internet, as shown in Figure 1.5. We have circuit switches both in the access networks (leased lines, DSL and phone modems), and in the core of the network (SONET/SDH and DWDM). This figure also shows the market sizes in the year 2001, and it shows how the market sizes for the segments that use circuit switching are significant.

The current mix of packet and circuit switching in the Internet is due to historical reasons. In the early days of the Internet, when two Internet Service Providers (ISPs) in different and distant locations wanted to interconnect with each other, they leased a connection from the only companies that had a continent-wide network, that is the long-distance telephone companies, and these companies have always based their service on pure circuit switching. Similarly, the circuits in the edges were one of the few options for an ISP to get to its customers, namely, by using the existing infrastructure of the local telephone company.

Figure 1.5: Architecture of the public Internet as found in the real world. The figures in the boxes represent the world market sizes in the year 2001. [161,158,157,60,61]

Given the current situation, one may ask two related questions. First, is this hybrid architecture the right network architecture? If we were to rebuild the Internet from scratch and with unlimited funds, would we choose a solution based on only packet switching, only circuit switching, or a mix of the two? Second, given that it would be too costly to build a brand new network, how can the current legacy Internet evolve in the future? Will the network still follow a hybrid model as today, or will it change? These two questions are the focus of the first part of the Thesis. I will conclude that it makes more sense to use circuit switching in the core and packet switching in the edges of the network.

Currently, the circuits that we find in the Internet are considered by IP as static, layer-2, point-to-point links. In other words, the circuit and packet switched parts of the network are completely decoupled, and changes in IP traffic patterns do not prompt an automatic reconfiguration of the circuits over which IP travels. It is usually the case that circuits are manually provisioned by either the network operator (circuits in the core) or the end user (circuits as access lines). This also means that the time scale in which circuits operate is much larger than that of packets.

We would make a more efficient use of the network resources if we could integrate the world of circuits with that of packets in such a way that circuits follow in real time the fluctuations of the packet switched traffic. In this Thesis, I make two proposals of how to integrate these two technologies in an evolutionary manner, without changing existing end hosts and routers. One approach uses fine-grain, lightweight circuits; the other uses coarse-grain, heavyweight circuits (such as optical wavelengths).

1.3.1 Virtual circuits

There is a third family of networks, which uses virtual circuits, such as ATM or MPLS. This family attempts to get the best of two worlds: on one hand, it takes the statistical multiplexing of packet switching. On the other hand, the traffic management and quality of service of circuit switching. Despite their name, virtual circuits are essentially a connection-oriented version of packet switching; it forwards information as packets (sometimes called cells), but it keeps connection state associated with each flow. In contrast, IP is based on the connectionless switching of packets, where no per-flow state is kept. Switches using virtual circuits are hard to design; they have the scalability issues of both the data path of packet switching and the state management and signaling of circuit switching. Therefore, virtual circuits will not be studied any further in this Thesis.

In the early 90's, there was a race between IP and ATM to dominate data networks. In the end, IP routers prevailed over ATM switches partly because the former were simpler and thus faster to hit the market and easier to configure. In contrast, MPLS works just below IP, rather than competing with it, and it is an attempt to do simple traffic engineering in the Internet. Only recently has MPLS started getting sizable deployment with some backbone carriers [34].