Making Networks More Robust

A network lies at the heart of many enterprises, connecting the computers that manage the business’s data and support its employees with information and communications. Yet, modern networks are far from a dependable infrastructure and require tremendous human-intensive effort to maintain and operate.

One source of this fragility is the increasingly complex demands placed on networks. Networks of all types and sizes are now being called upon to implement functionality such as sophisticated security policies, network address translation, and traffic engineering. Network designers have responded to these challenges with a wide variety of ad hoc approaches; creating networks that have outpaced our ability to model their function or predict their behavior.

In this talk, I will describe a framework for analyzing operational networks. Results from applying these techniques to enterprise networks that range in size from 10 to 1200 routers show the complexity and diversity present even in supposedly simple networks. Generalizing the lessons from this analysis, I will argue that a completely new way of controlling and managing networks is required, and I will describe the 4D architecture we are designing as an alternative. The architecture refactors the network control system, leaving a wafer-thin set of minimal functionality on each router. All of the decision logic is removed from the routers and collected onto servers where the objectives for the network can be explicitly specified and used to directly control the network.