Data Traffic Modeling—A Signal Processing Perspective - Nonlinear Signal and Image Processing: Theory, Methods and Applications - page 196

dramatically, causing the classical teletraffic theory to be challenged by several
new aspects. The traffic (data traffic) that is defined here as bytes per unit time
corresponds to packetized voice, video, images, and computer data files. The
data transfer sessions no longer fit the Poisson model well. For example, dur-
ing an Internet browsing session, the end user and the server may exchange
small and large files in a bursty fashion. The number of transfer requests may
come in bursts during a sustained time period. That may be followed by a
period of “silence,” and then again by another “busy” period. Thus, the times
between transfer requests are no longer drawn from an exponential distri-
bution. Second, because of the multimedia nature of the data that are be-
ing exchanged, there is usually high variability in bandwidth requirements
during a single session. These characteristics of data traffic give rise to some
very interesting properties, namely, self-similarity and impulsiveness , both
of which can have significant consequences in network design and

management.
Developing analytical models that capture the new characteristics of data
traffic has considerable significance in several aspects of traffic engineering,
such as admission control, flow control, and congestion control. For example,
in data networks, switches allocate a certain bandwidth to a group of admit-
ted traffic flows. In general, the allocated bandwidth is not the sum of the peak
rate of all flows, but is computed according to the statistical features of the
incoming traffic and some predefined QoS. Such practice is referred to as “sta-
tistical multiplexing.” An analytical model for traffic would be indispensable
in allocating bandwidth to meet QoS guarantees. Analytic models would also
be used to evaluate the performance of large-scale networks, whose analytic
characterization is rather intractable, before they are deployed. This is typi-
cally done using traces collected from real networks. However, those traffic
traces are shaped by the characteristics of the network in which they were
collected, and thus could be inappropriate when used in a different network.
Furthermore, real traffic traces may require enormous storage space, given
that today's broadband data traffic flows at a speed of giga bits per second.
Analytical models can allow testing of large-scale networks via on-the-fly
generation of traffic traces. Also, by providing traffic traces with various con-
trollable characteristics, they can offer more insight into the reasons for certain
network behavior.
In this chapter, we concentrate on statistical analysis and modeling of data
traffic. In particular, we will focus on the two important salient features of data
traffic, namely, self-similarity and impulsiveness. The chapter is organized as
follows. Mathematical preliminaries are introduced in Section 6.2. Section 6.3
illustrates the self-similar and impulsive nature of data traffic based on real
traffic measurements. Traffic models are discussed in Section 6.4, and param-
eter estimation techniques are reviewed in Section 6.5. Finally, Section 6.6
provides some concluding remarks.