For our purposes, an ATM network consists of ATM switches connected by communication links. An ATM switch itself consists of input ports, the switching fabric, and output ports. A cell that arrives at an input port of a switch is transported by the switching fabric to an output port, where it is transmitted along the physical link associated with the output port. Messages are segmented into fixed-size cells. This simplifies the traffic analysis because the cell transmission time is constant, and time can be normalized appropriately.
For the purpose of traffic analysis, the network is traditionally decomposed into a collection of servers . Each server provides an abstraction for a network component in the system. For example, the input ports, the switching fabric, the output ports, and the physical links can each be modeled as a server.
We distinguish two types of servers: constant servers and variable servers [4, 9] Constant servers, such as physical links, input ports, and most common switching fabrics, impose a constant delay to each cell and do not modify the traffic flow characteristics of a connection. Variable servers, on the other hand, add a non-constant delay to each cell, and so modify the traffic characteristics of connections. An output port, for example, acts as a multiplexor and may simultaneously receive cells belonging to different connections competing for transmission on the link associated with the port. Thus, cell blocking may occur, and cells may be forwarded in an order that is determined by the scheduling policy adopted by the switch. An output port, which is a multiplexor, must therefore be considered as a variable server.
Constant servers do not affect the traffic flows. Therefore, they need not be further considered to derive the characterization of traffic flows inside a network. Variable servers, however, modify the traffic that flows through them, and their effect on traffic has to be understood, so that the accuracy of a traffic descriptor does not excessively suffer as the traffic traverses one or more variable servers. In the following section, the network will therefore be modeled as a network of variable servers only.