Before starting our discussion about whether IP can be the basis of all communication networks, I will give some background about the two main switching techniques in use today: circuit switching and packet switching.
Circuit switching was the first switching technique used in communication networks because it is simple enough to carry analog signals. This thesis will just focus on the digital version of circuit switching. Of course, the main example of its use is the phone system [72], but it is also used in the core of the Internet in the form of SONET/SDH and DWDM equipment [81,126]. In circuit switching, the transmission medium is typically divided into channels using Frequency Division Multiplexing (FDM), 2.2 Time Division Multiplexing (TDM) or Code Division Multiplexing (CDM) [172]. A circuit is a string of concatenated channels from the source to the destination that carriers an information flow.2.3
To establish the circuits, a signaling mechanism is used. This signaling only carries control information, and it is considered an overhead. It is also the most complex part in circuit switching, as all decisions are taken by the signaling process. It is commonly assumed that the signaling and per-circuit state management make circuit switches hard to design, configure and operate.
In circuit switching the channel bandwidth is reserved for an information flow. To ensure timely delivery of the data, the capacity of the circuit has to be at least equal to the peak transmission rate of the flow. In this case, the circuit is said to be peak allocated, and then the network offers a connection-oriented service with a perfect quality of service (QoS) in terms of delay jitter and bandwidth guarantees, However, this occurs at the cost of wasting bandwidth when sources idle or simply slow down.
Contention only occurs when allocating channels to circuits during circuit/call establishment. If there are not enough channels for the request, the call establishment may be delayed, blocked or even dropped. In contrast, once the call is accepted, resources are not shared with other flows, eliminating any uncertainty and, thus, removing the need for buffering, processing or scheduling in the data path. When circuits are peak allocated, the only measure of Quality of Service (QoS) in circuit switching is the blocking probability of a call.
To summarize, circuit switching provides traffic isolation and traffic engineering, but at the expense of using bandwidth inefficiently and signaling overhead. It is often said that these two drawbacks make circuit switching highly inflexible, especially in a highly dynamic environment such as the Internet. I will argue in this that these drawbacks are outweighed by the advantages of using more circuit switching in the core of the network.
Packet switching is the basis for the Internet Protocol (IP) [152,172]. In packet switching, information flows are broken into variable-size packets (or fixed-size cells as in the case of ATM). These packets are sent, one by one, to the nearest router, which will look up the destination address, and then forward them to the corresponding next hop. This process is repeated until the packet reaches its destination. The routing of the information is thus done locally, hop-by-hop. Routing decisions are independent of other decisions in the past and in other routers; however, they are based on network state and topology information that is exchanged among routers using BGP, IS-IS or OSPF [148]. The network does not need to keep any state to operate, other than the routing tables.
The forwarding mechanism is called store-and-forward because IP packets are completely received, stored in the router while being processed, and then transmitted. Additionally, packets may need to be buffered locally to resolve contention for resources. 2.4 If the system runs out of buffers, packets are dropped.
With the most scheduling policies, such as FCFS and WFQ, packet switching remains work conserving; it keeps the link busy as long as there are packets waiting to be sent. This allows it to have a statistical multiplexing gain; that is, the capacity of an outgoing link can be much smaller than the sum of its tributaries and still have a packet delay or drop probability within certain statistical bounds. This gain is higher when traffic is more bursty. The buffering needs and the statistical multiplexing are the main characteristics of packet switching, and they will be crucial in its comparison with circuit switching.
In the Internet, the network service is connectionless and best effort; that is, it provides no delivery guarantees. Reliability, flow control and connection-oriented services are provided by end-to-end mechanisms, such as with TCP [153]. Because the underlying service is best effort, there are no guarantees in terms of packet drops, maximum delay, delay jitter or bandwidth.
Much research was done in the early days of computer networking comparing circuit switching, packet switching and message switching (a variant of packet switching, in which the whole information flow is treated as a single switching unit) [96,10,164,97,175,95]. Most of the work was done in the context of packet radio, satellite, and local area networks and shows how in these environments packet switching provided higher throughput for a given bound on the average delay. Packet switching not only made an effective use of the network bandwidth, but it also was robust and resilient to node and link failures.
Later work on different scheduling algorithms and signaling mechanisms, such as Weighted Fair Queueing (WFQ) [62], Generalized Processor Sharing (GPS) [141], Differentiated Services (DiffServ) [16], Integrated Services (IntServ) [20] and Deficit Round Robin (DRR) [113], showed how packet switching can also provide QoS guarantees if the admission of new flows to the network can be controlled.