Demystifying Internet Traffic

Опубликовано 6 сентября 2016, 17:12

The Internet has seen a tremendous growth since its inception four decades ago. With its increasing importance, there has been a growing emphasis on improving the reliability of the infrastructure. One approach to delivering such reliability is for design engineers, network administrators and researchers to stress test potential solutions against a wide variety of deployment scenarios. For instance, web hosting services would wish to ensure that they can deliver target levels of performance and availability under a range of conditions. Similarly, Internet Service Providers (ISPs) would benefit from understanding future growth in traffic demands at individual routers in its network as a function of emerging applications and expanding user base. I argue that one of the key ingredients required to carry out such studies is a deep understanding of Internet traffic characteristics. This talk will try to uncover some of the mysteries surrounding Internet traffic, including its rich structure. I will thus describe the principles and key insights that led to the development of the Swing traffic generator. Swing is the first tool to reproduce realistic and responsive Internet-like traffic in a testbed. Starting from observing packets across a given link, Swing automatically extracts parameters for its detailed multi-level model. It then uses this model to generate live traffic that looks qualitatively similar to the original traffic. More interestingly, Swing provides the user with meaningful knobs to project traffic demands into the future. This includes changing assumptions about user popularity of applications, planned upgrades to the network as well as change in the semantics of applications.