Return to search

Resource dimensioning in a mixed traffic environment

An important goal of modern data networks is to support multiple applications over a single network infrastructure. The combination of data, voice, video and conference traffic, each requiring a unique Quality of Service (QoS), makes resource dimensioning a very challenging task. To guarantee QoS by mere over-provisioning of bandwidth is not viable in the long run, as network resources are expensive. The aim of proper resource dimensioning is to provide the required QoS while making optimal use of the allocated bandwidth. Dimensioning parameters used by service providers today are based on best practice recommendations, and are not necessarily optimal. This dissertation focuses on resource dimensioning for the DiffServ network architecture. Four predefined traffic classes, i.e. Real Time (RT), Interactive Business (IB), Bulk Business (BB) and General Data (GD), needed to be dimensioned in terms of bandwidth allocation and traffic regulation. To perform this task, a study was made of the DiffServ mechanism and the QoS requirements of each class. Traffic generators were required for each class to perform simulations. Our investigations show that the dominating Transport Layer protocol for the RT class is UDP, while TCP is mostly used by the other classes. This led to a separate analysis and requirement for traffic models for UDP and TCP traffic. Analysis of real-world data shows that modern network traffic is characterized by long-range dependency, self-similarity and a very bursty nature. Our evaluation of various traffic models indicates that the Multi-fractal Wavelet Model (MWM) is best for TCP due to its ability to capture long-range dependency and self-similarity. The Markov Modulated Poisson Process (MMPP) is able to model occasional long OFF-periods and burstiness present in UDP traffic. Hence, these two models were used in simulations. A test bed was implemented to evaluate performance of the four traffic classes defined in DiffServ. Traffic was sent through the test bed, while delay and loss was measured. For single class simulations, dimensioning values were obtained while conforming to the QoS specifications. Multi-class simulations investigated the effects of statistical multiplexing on the obtained values. Simulation results for various numerical provisioning factors (PF) were obtained. These factors are used to determine the link data rate as a function of the required average bandwidth and QoS. The use of class-based differentiation for QoS showed that strict delay and loss bounds can be guaranteed, even in the presence of very high (up to 90%) bandwidth utilization. Simulation results showed small deviations from best practice recommendation PF values: A value of 4 is currently used for both RT and IB classes, while 2 is used for the BB class. This dissertation indicates that 3.89 for RT, 3.81 for IB and 2.48 for BB achieve the prescribed QoS more accurately. It was concluded that either the bandwidth distribution among classes, or quality guarantees for the BB class should be adjusted since the RT and IB classes over-performed while BB under-performed. The results contribute to the process of resource dimensioning by adding value to dimensioning parameters through simulation rather than mere intuition or educated guessing. / Dissertation (MEng (Electronic Engineering))--University of Pretoria, 2007. / Electrical, Electronic and Computer Engineering / unrestricted

Identiferoai:union.ndltd.org:netd.ac.za/oai:union.ndltd.org:up/oai:repository.up.ac.za:2263/24762
Date24 January 2006
CreatorsRoon, Selwin Jakobus Emiel
ContributorsProf W T Penzhorn, selwin@tuks.co.za
Source SetsSouth African National ETD Portal
Detected LanguageEnglish
TypeDissertation
Rights© 2005, University of Pretoria. All rights reserved. The copyright in this work vests in the University of Pretoria. No part of this work may be reproduced or transmitted in any form or by any means, without the prior written permission of the University of Pretoria.

Page generated in 0.0027 seconds