Human error in configuring routers undermines attempts to provide reliable, predictable end-to-end performance on IP networks. Manual configuration, while expensive and errorprone, is the dominant mode of operation, especially for large enterprise networks. These networks often lack the basic building blocks - an accurate equipment inventory, a debugged initial configuration, and a specification of local configuration policies - to support the holy grail of automation. We argue that migrating an existing network to automated configuration is a rich and challenging research problem rooted in data analysis and in the modeling of network protocols and operational practices. We propose a novel, bottom-up approach that proceeds in three phases: (i) analysis of configuration data to summarize the existing network state and uncover configuration problems; (ii) data mining to identify the network's local configuration policies and violations of these policies; and ultimately (iii) boot-strapping of a database to drive future configuration changes. The first stage reduces the number of errors, the second normalizes the local policies, and the third prevents new errors and reduces the manpower needed to configure the network. We describe the architecture of our EDGE tool for steps (i) and (ii), and present some examples from our experiences applying the tool to several large enterprise networks.
All Science Journal Classification (ASJC) codes
- Computer Networks and Communications