Back to search

University of Patras

Στοχαστικά μοντέλα για την ανάλυση δεδομένων στο διαδίκτυο

Abstract

dc:description

The purpose of this dissertation was to study the characteristics of internet traffic and to develop appropriate models and techniques that address traffic load forecasting, dynamic bandwidth allocation based on predictions and generation of synthetic traffic that resembles real motion video. In addition, the connection between networks and their topology with time series has been explored so that the properties of the resulting time series can be used as means to classify the corresponding networks; a process which is particularly important since the networks can be more or less resistant to malicious attacks depending on their type. Internet traffic as we know is generated by the different types of media connected to it and it involves audio, video, images, Ethernet, LANs, WANs, etc. For all these different types of traffic, their modeling is necessary to solve their different arising problems. Τhe topics covered by the dissertation are of interest both in terms of stochastic modeling and in terms of of telecommunication engineering where the modeling of traffic is necessary for other tasks such as bandwidth allocation and generation of synthetic traffic.The first chapter of the dissertation presents a general overview of the Internet that concerns both a historical retrospect of its creation and organization, its modeling approaches, its movement and its topology. In the second chapter we study the basic characteristics of the Internet, both in terms of distributions for the various metrics (static characteristics) and its time dynamics . In terms of the time series models, we present only those with which Internet traffic can be modeled and which also constitute the basis of the methodologies that will be developed later in Chapters 3, 4 and 5. Finally, we present basic concepts and elements of network science that are necessary both for the understanding of the features of internet topology and for the development of the methodology which is proposed in Chapter 6 and it refers to the categorization of networks on scale-free or non scale-free.The first subject of the dissertation is presented in Chapter 3 and is the prediction of internet traffic. The most accurate forecast is necessary to resolve issues such as resource allocation of a system and network design. Based on models such as FARIMA, GARCH and Artificial Neural Networks, predictive schemes are proposed through their combinations, along with a model selection scheme that is based on testing for non-linearity and also hybrid models that combine structurally individual models. Lastly, a framework for the comparison of the different forecasting methodologies is suggested in order to decide about the most appropriate forecasting approach. In Chapter 4, the previously developed predictive approaches are used to dynamically allocate bandwidth to an end-to-end network, a very important problem that is part of the overall network resource management. The availability of bandwidth in general can be realized either statically, which does not take into account fluctuations in demand, either dynamically, so the prediction plays an essential role. For this chapter a process was also designed to select a prediction model that has the ability to be accurate in prediction, but at the same time it can be successfully applied to bandwidth allocation schemes. The candidate models were FARIMA and FARIMA / GARCH, the hybrid models FARIMA - MLP and FARIMA / GARCH - MLP as well as two prediction strategies where one selects a model depending on the result of a test for non - linearity in the dependence structure of data and the other combines predictions from different methodologies. Three different dynamic schemes are proposed for bandwidth allocation and are evaluated using metrics. From their comparison the best forecasting methodology is detected and significant conclusions are drawn on the proposed bandwidth allocation schemes. The third subject of this dissertation is presented in Chapter 5 and refers to the generation of synthetic traffic through time series models. The ability to simulate Internet traffic can contribute to more efficient network management and performance analysis. Because real-network performance analysis can be costly and measurements are difficult, the mathematical models that can be used to generate traffic are very important. In this context, FARIMA and FARIMA / GARCH time series models are considered in order to achieve synthetic traffic with characteristics similar to those of the real. However, due to the fact that their effectiveness as sole generators proved to be limited, in this chapter we propose a process that combines FARIMA models with the Lognormal distribution and produces synthetic traffic that approximates quite accurately the behavior of the real traffic. Finally, in Chapter 6 an attempt is made to solve the problem of classifying a network as scale-free or non scale-free. Nowadays networks are complex structures and graphically are represented by nodes, which are the entities of interest, and edges that form the relationships between them. Internet and every small private subnetwork that connects with it can be represented as such network with nodes the servers and edges their links. On the other hand the resilience of a network to various types of attacks can be judged also by its type. In particular, according to our experimental study scale-free networks appear to be resistant to random attacks but very vulnerable to targeted attacks. So a credible network classification process could be an important tool for their managers. In this dissertation it is proposed to transform each complex network into a time series. For this transformation, 3 specific metrics of the nodes are calculated, which then contribute to the definition of a final value for each node. Taking into account the connections between the nodes, we create a sequence of data which then treat as time series. With the help of parameters such as the Hurst and the maximum Lyapunov exponents for the time series, it is possible to characterize the network from which it originates as scale-free or non scale-free. The proposed transformation is applied to both simulated and real networks for which random and targeted attacks are also simulated in order to evaluate the classification of the networks.To summarize, in this dissertation new methods, which are mostly based on time series models, are proposed for the solution of problems related to internet traffic. Based on the experimental results presented in the dissertation we conclude that some of the proposed methods for solving these specific problems are much more effective and others are simply better than their alternatives.

Degree

thesis:*
Grantor dc:publisher
University of Patras
Year dc:date
2017

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Katris, Christos
  • Κάτρης, Χρήστος

Subjects

dc:subject × 20

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/42020

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Katris, Christos; Κάτρης, Χρήστος. Στοχαστικά μοντέλα για την ανάλυση δεδομένων στο διαδίκτυο. University of Patras, 2017. http://hdl.handle.net/10442/hedi/42020