Calibrated network measurement
Paul Barford, Joel Sommers · 2007
Empirical measurements of Internet paths and systems are essential for understanding the behavior and performance of the Internet, for managing operational networks, and for designing future Internet systems. In this dissertation we propose a framework and set of calibration techniques and tools that can greatly increase our confidence in the validity and accuracy of network measurement methodologies and the measurements they produce. Our approach is based on performing experiments in a controlled laboratory environment. This setting offers control and repeatability of experiments. It also offers the ability to measure desired aspects of the system or measurement methodology under test, which is critical for evaluating the accuracy of measurements and for gaining insight into system behavior. Our laboratory-based approach also offers realism in terms of end hosts, routers, operating systems, and network protocol implementations. There have previously existed two significant obstacles associated with laboratory-based experiments in the area of realism: the ability to scalably produce a range of realistic traffic conditions, and the ability to recreate Internet-like link and path characteristics. We propose and evaluate two systems for overcoming these problems. We describe three case studies that leverage the control, transparency, and realism offered in our laboratory calibration environment. These studies are in the areas of end-to-end available bandwidth measurement, measurement of packet loss characteristics, and measurement of the impact of router architecture, buffer size, and queuing policy on transport protocol performance. Through these investigations we illustrate the power and capabilities of our laboratory calibration approach. We develop significantly more accurate available bandwidth and packet loss measurement methodologies. We also illustrate the problems associated with recently proposed buffer sizing methodologies and how consideration of quality of service guarantees made by network providers illuminates important issues in this context.