Performance measurements in packet-switched networks : application to IP networks
MetadataVis full innførsel
The Internet is becoming an important infrastructure. Measurements of network traffic are vital for many areas; including network management, traffic engineering, service differentiation, accounting and billing, the design of new protocols and applications, input to traffic modeling and validation of service level agreements. As service differentiation is introduced in IP networks, and higher QoS is required by customers, performance measurements will become even more important. Apparently, the ability to collect, analyse and understand measurement data will be vital for the Internet evolution. The overall objective of the research presented in this thesis has been to investigate a few central problems, each dealing with different aspects of performance measurements in IP networks. There is a lack of a common framework and model for precise definitions of measurement metrics and methods. The challenges of precisely defining performance metrics are addressed through introducing a formal framework for reasoning about IP measurements. Such a framework is essential to express topics related to measurement metrics and methodologies for IP networks. The conceptual model proposed in this thesis is original in that it combines standard graph and set theory to address network measurements in a way that has not been seen elsewhere. The model uses graph theory to model network topology and applies simple welldefined operations and functions on sets of events that enable precise and consistent definitions of network traffic measurement issues. The conceptual model is used to precisely define fundamental network layer performance metrics and to discuss measurement methods and infrastructure. The conceptual model is obviously an academic and formal concept. Yet, the model provides basic understanding of events that can be observed in packet-switched networks. Systematic gathering of measurement data requires the establishment and deployment of a measurement platform. Fundamental problems related to establishing a platform for performance measurements in an operational network are considered. The topics addressed include investigation of active and passive measurement methods at different protocol layers with a focus on network layer measurement, requirements for a measurement platform and post-processing of measurement data. To assess the capabilities of different measurement strategies, data from actual measurements are included. The most important contributions are the investigation of pros and cons of building blocks of a measurement platform and recommendations considering the designing of such a platform. There are two principal methods to collect measurement data; either actively by insertion of test packets or passively by observing real packets. Active measurements are extensively used for performance measurements in IP networks but have the disadvantage of disturbing the system being measured. A key challenge is to inject test packets often enough to get a sufficient number of observations while avoiding too much disturbances. It is important to know the accuracy of such measurements. This thesis addresses properties of active measurements of waiting time and packet loss in the case where the end-to-end path has a single bottleneck. The measurement accuracy of active measurements of the waiting time is analysed using an M/G/1/ queue. Based on results known from the literature, the variance of the average waiting time as observed by injected test packets is found. Using the mean squared error of the average waiting time as observed by the test packets as an indicator of measurement accuracy, the optimal measurement intensity for active measurements is analysed. It is seen that the optimal measurement intensity depends on the packet length distribution of regular packets (first four moments), packet length distribution of test packets (first four moments) and the load from regular traffic. Numerical studies show that when the load is high, the optimal measurement intensity is low in order to minimise the impact from the measurement traffic. Further, when the load is low the optimal measurement intensity is also low because the relative disturbance from the test traffic is more significant at low loads. It is also seen that at both high and low loads, the measurement period to achieve a certain confidence in estimates of average waiting time must be rather long. Furthermore, an algorithm for adapting the measurement intensity in an operational network is proposed. Finally, the case where the end-to-end path has a single bottleneck where packet losses occur is analysed. The general ideas are the same as for active measurements of the waiting time in an M/G/1/ queue, but now the accuracy in loss is addressed rather than in waiting time. To study the loss accuracy for a finite buffer node is complicated and we therefore resort to an approximation using an M/G/1/ queueing model. A packet is assumed to be lost if the waiting time exceeds a certain threshold. Under these assumptions, based results from the literature an expression for the variance of the average loss ratio as observed by test packets is found. It is found that using the probability that an infinite queue exceeds a threshold as an approximation of the corresponding finite queue may be very inaccurate. However, the results indicate that the approximation has a better accuracy for service time distributions with larger variance. Furthermore, the precision is better for observations made by test packets only than all packets, especially when the measurement intensity is low. Even if the packet loss approximation may be very inaccurate, the results presented increase the basic understanding of active measurements of packet loss in IP networks. In particular, the results developed for active measurements of waiting time and loss allows the importance of the distribution of the data packets on the optimal measurement intensity to be investigated. It must be noted that the models are limited to Poisson arrival process. Empirical studies claim that self-similar processes better model the packet arrival process in a network. However, the results presented increase the fundamental understanding of intrusive measurements in IP networks. Furthermore, the introduction of service differentiation and separation of different traffic types can influence the packet arrival process.