Although it is now widely recognized that network traffic frequently exhibits so-called Long-Range Dependent (LRD) properties, the impact of such properties on performance is still not well understood. This thesis presents a thorough discussion of a particular class of traffic models, namely the aggregation of ON/OFF traffic sources, called N-Burst. A special family of Matrix-Exponential (or Phase-Type) distributions -- called Truncated Power-Tail distributions -- is used for the ON time distribution in order to mimic LRD properties, while still remaining tractable for queueing analysis via Matrix-Analytic methods. In order to make the model applicable to realistic scenarios, adequate procedures for parameter estimation are developed, which are then applied to a set of actual data from measurements of inter-cell times in an IP-over-ATM network. The steady-state analysis of queueing models witch such LRD arrival processes reveals distinctively different behavior than for models without LRD properties: All performance parameters show several so-called blow-ups at particular, well-defined points in the parameter space of the model, at which performance deteriorates dramatically. The knowledge about the existence and the location of these blow-up points is crucial for any network design task. Applicable engineering rules for the purpose of performance oriented design are derived from the gained insights. In many scenarios, the steady-state analysis of such LRD models does not provide a satisfactory description of the performance behavior. Therefore, methods for transient analysis are developed and applied. The discussion of the transient results leads to new characterizations of the fluctuations in the performance behavior of such LRD models. Finally, a variation of the basic analytic queueing model is discussed in which the arrival process is throttled when congestion occurs in the network. The results of that analysis provide insights into the behavior of elastic traffic, such as assumed by the Transmission Control Protocol (TCP) in the Internet. The contributions of the thesis are twofold: First, contributions to the methodology for performance analysis using Matrix-Analytic models with Long-Range Dependent properties are made. Second, the results from the performance analysis of the rather general N-Burst model class provide new insights for the network design task. Although the thesis is directed at modeling and performance evaluation of telecommunication systems, the methodologies can easily be transfered to related areas, such as storage systems and CPU job-times, both of which are known to be subject to the impact of highly fluctuating loads.     «

Although it is now widely recognized that network traffic frequently exhibits so-called Long-Range Dependent (LRD) properties, the impact of such properties on performance is still not well understood. This thesis presents a thorough discussion of a particular class of traffic models, namely the aggregation of ON/OFF traffic sources, called N-Burst. A special family of Matrix-Exponential (or Phase-Type) distributions -- called Truncated Power-Tail distributions -- is used for the ON time distrib...     »

Verkehr in modernen, paketvermittelten Telekommunikationsnetzen weist Charakteristika auf, die in klassischen Verkehrsmodellen nicht adäquat berücksichtigt werden: In Messungen wurden sogenannte " Langzeitabhängigkeiten" (auch " Selbstähnlichkeit" ) des Paketstromes festgestellt, deren Einfluß auf die Übertragungsqualität aber noch diskutiert wird. Diese Dissertation entwickelt eine realistische Klasse von Verkehrsmodellen, genannt N-Burst, die langzeitabhängigen Verkehr beschreiben kann. Aus der Diskussion einer Vielzahl stationärer und transienter Leistungsparameter in analytischen Warteschlangenmodellen werden detaillierte Schlußfolgerungen über das Leistungsverhalten solcher langzeitabhängiger Verkehre gewonnen, die beispielhaft auf reale Problemstellungen der Netzplanung angewendet werden. Dabei zeigt sich, daß in diesem Fall herkömmliche (stationäre) Leistungsparameter das Verhalten der Warteschlangenmodelle nur ungenügend beschreiben. Neue transiente Methoden werden deswegen eingeführt und angewendet.     «

Verkehr in modernen, paketvermittelten Telekommunikationsnetzen weist Charakteristika auf, die in klassischen Verkehrsmodellen nicht adäquat berücksichtigt werden: In Messungen wurden sogenannte " Langzeitabhängigkeiten" (auch " Selbstähnlichkeit" ) des Paketstromes festgestellt, deren Einfluß auf die Übertragungsqualität aber noch diskutiert wird. Diese Dissertation entwickelt eine realistische Klasse von Verkehrsmodellen, genannt N-Burst, die langzeitabhängigen Verkehr beschreiben kann. Aus d...     »