Abstract

Abstract Network measuremnts these days are conducted in a mostly ad hoc manner. Users quite often perform one-time measurement experiments in an unprincipled approach, in many case leading to inefficient measurements and, even worse, biased measurements. Vendors have developed and standardized a number of measurement tools, e.g., NetFlow from Cisco. However, these tools are typically inefficient in that they often yield statistics with large error ranges.

In this talk we argue for the need for a network measurement science that ccan deal in a principled way with the issues of measurement efficiency and measurement bias. To deal with measurement efficiency, we advocate the use of Fisher information during the design of measurment experiments and measurement tools. Briefly, Fisher information is a measure of the anount of information that a single measurment provides to the computation of a statistic such as loss rate. We illustrate its application to the problem of estimating flow size distribution based on packet sampling as found in NetFlow. In the context of measurement bias, we shift our attention to measurements leading to the characterization of graphs as commonly found in the Internet and on-line social networks. We review several flawed studies where the measurements were biased and then describe how measurements based on random walks through graphs can be used to remedy these deficiencies.

Biography I am a Professor in Computer Science at the University of Massachusetts. My interests lie in the area of modelling and analysis of computer and communication systems. I have been visiting MSR for the past couple of months. You can find more information about me here, http://www-net.cs.umass.edu/personnel/towsley.html.