Often one can find in the technical press and literature predictions about how IP routers will eventually replace all circuit switches [186,150,21,99,35,37,142,87,156,128,110]. These articles extend the original arguments for adopting packet switching in the early days of the Internet (namely, efficiency and robustness), by adding the simplicity, cost advantage, and ability to provide QoS of IP. These are some of the sacred cows of IP, and in Chapter 2, I evaluate them one by one to demystify the ones that do not hold up to scrutiny and to identify the ones that really apply.
One key claim of packet switching is left for Chapter 3; namely that the statistical multiplexing of packet switching consistently delivers a lower response time than circuit switching while downloading information. This is indeed the most relevant performance metric for end users, and thus it gets its own chapter.
The conclusion of these two chapters is that packet switching is very attractive in Local Area Networks (LANs) and access networks, because of the poor end-user response time of dynamic circuits. On the contrary, circuit switching is more attractive in the core of the network because of its higher capacity, its perfect QoS, and a response time that is similar to that of packet switching. In the future, one can expect a dominant role of IP in the edge of the network, whereas various forms of circuit switching will dominate the core of the network. This partially validates the hybrid network architecture that we currently have and that is shown in Figure 1.5.
However, in the current Internet these two distinct parts are completely decoupled; the edges switch packets independently of the circuits used in the core. Chapter 4 presents a network architecture (TCP Switching) that allows the integration of circuit switching in the core of a packet-switched Internet in an evolutionary way. The chapter starts with a description of what a typical application-level flow in the Internet is, as observed on access points to the Internet of several universities and research institutions. A key observation is that despite the connectionless nature of IP, our use of the network is very connection oriented, and this fits well with the use of circuits. TCP Switching is based on the idea of IP Switching [129], and it maps each application flow to a lightweight circuit. This proposal encompasses a family of solutions, with several design choices. Also in this chapter, I choose one solution based on what constitutes a typical flow in the Internet.
One potential problem with such fine-grain circuits in the core, as thin as 56 Kbit/s, is that they might not fit well with many circuit switch designs. Most core circuit switches have interconnects that only use channels of at least 51 Mbit/s. In addition, optical switches only forward wavelengths carrying channels of over 2.5 Gbit/s. The signaling of these switches might be heavyweight because of the slow reconfiguration of the switch fabrics or because of a signaling mechanism that requires circuit creation confirmation. In Chapter 5, I present another technique for controlling the coarse-grain, heavyweight circuits in the core by monitoring user flows rather than tracking packets or queue lengths. I show the requirements for different circuit setup times. These results could be used in Generalized Multi-Protocol Label Switching (GMPLS), a technique that uses heavyweight circuits to adapt the network capacity dynamically between edge routers.
In Chapter 6, I describe some of the related work in the area of high speed switching in the core of the network. Some proposals include the use of circuit switching in the core (GMPLS [7], OIF [13], ODSI [53], Zing [181]), while others attempt to extend packet switching to all-optical switches (Optical Packet Switching - OPS [186] -, and Optical Burst Switching - OBS [188]). Some emphasis is placed on the comparison of TCP Switching with OPS and OBS. Two metrics are used for the comparison: the loss and blocking probabilities for a given network load, and the complexity of the overall network.
Chapter 7 concludes the Thesis, restating how we would benefit from more circuit switching in the core of the network, and how we could integrate this circuit switched core with the rest of the network in an evolutionary way.