Performance characterisation of IP networks
The initial rapid expansion of the Internet, in terms of complexity and number of hosts, was followed by an increased interest in its overall parameters and the quality the network offers. This growth has led, in the first instance, to extensive research in the area of network monitoring, in order to better understand the characteristics of the current Internet. In parallel, studies were made in the area of protocol performance modelling, aiming to estimate the performance of various Internet applications. A key goal of this research project was the analysis of current Internet traffic performance from a dual perspective: monitoring and prediction. In order to achieve this, the study has three main phases. It starts by describing the relationship between data transfer performance and network conditions, a relationship that proves to be critical when studying application performance. The next phase proposes a novel architecture of inferring network conditions and transfer parameters using captured traffic analysis. The final phase describes a novel alternative to current TCP (Transmission Control Protocol) models, which provides the relationship between network, data transfer, and client characteristics on one side, and the resulting TCP performance on the other, while accounting for the features of current Internet transfers. The proposed inference analysis method for network and transfer parameters uses online nonintrusive monitoring of captured traffic from a single point. This technique overcomes limitations of prior approaches that are typically geared towards intrusive and/or dual-point offline analysis. The method includes several novel aspects, such as TCP timestamp analysis, which allows bottleneck bandwidth inference and more accurate receiver-based parameter measurement, which are not possible using traditional acknowledgment-based inference. The the results of the traffic analysis determine the location of the eventual degradations in network conditions relative to the position of the monitoring point. The proposed monitoring framework infers the performance parameters of network paths conditions transited by the analysed traffic, subject to the position of the monitoring point, and it can be used as a starting point in pro-active network management. The TCP performance prediction model is based on the observation that current, potentially unknown, TCP implementations, as well as connection characteristics, are too complex for a mathematical model. The model proposed in this thesis uses an artificial intelligence-based analysis method to establish the relationship between the parameters that influence the evolution of the TCP transfers and the resulting performance of those transfers. Based on preliminary tests of classification and function approximation algorithms, a neural network analysis approach was preferred due to its prediction accuracy. Both the monitoring method and the prediction model are validated using a combination of traffic traces, ranging from synthetic transfers/environments, produced using a network simulator/emulator, to traces produced using a script-based, controlled client and uncontrolled traces, both using real Internet traffic. The validation tests indicate that the proposed approaches provide better accuracy in terms of inferring network conditions and predicting transfer performance in comparison with previous methods. The non-intrusive analysis of the real network traces provides comprehensive information on the current Internet characteristics, indicating low-loss, low-delay, and high-bottleneck bandwidth conditions for the majority of the studied paths. Overall, this study provides a method for inferring the characteristics of Internet paths based on traffic analysis, an efficient methodology for predicting TCP transfer performance, and a firm basis for future research in the areas of traffic analysis and performance modelling.