Program Overview

Abstract: When measuring and modeling different aspects of the Internet and reporting our findings, we frequently hear critical comments such as “all models are wrong, but some are useful.” Made famous by the British Statistician G.E.P. Box, this aphorism is a constant reminder that when it comes to modeling highly-engineered systems such as the Internet, the goal is not perfection but utility. However, the critiques rarely dwell on what makes a given model “wrong but useful” (and in what sense) or “wrong and not only not useful but harmful” (and why). To prevent Box’s aphorism from becoming merely a discussion-ending cliche, I will discuss in this talk what makes the networking domain unique and challenging but also rewarding from a measuring and modeling perspective that aims at assessing the degree to which models are “wrong but useful”. In particular, I will detail the “data-to-model-to-utility” workflow for a number of different use cases that include the “good old” self-similar models of measured Internet traffic (and their utility for today’s Internet) as well as the latest AI models and their intended use for networking. I will demonstrate that carefully scrutinizing these workflows provides a surprisingly effective means for answering in a principled manner contentious questions like “How wrong does a model have to be to not be useful?” The ability to answer such questions takes on new meaning for networking in the age of AI which is in danger of being viewed as mere “tinkering” and no longer as a scientific discipline.