A Logically Centralized Approach for Control and Management of Large Computer Networks

摘要

Management of large enterprise and Internet Service Provider networks is a complex, error-prone, and costly challenge. It is widely accepted that the key contributors to this complexity are the bundling of control and data forwarding in traditional routers and the use of fully distributed protocols for network control.udTo address these limitations, the networking research community has been pursuing the vision of simplifying the functional role of a router to its primary task of packet forwarding. This enables centralizing network control at a decision plane where network-wide state can be maintained, and network control can be centrally and consistently enforced. However, scalability and fault-tolerance concerns with physical centralization motivate the need for a more flexible and customizable approach. udThis dissertation is an attempt at bridging the gap between the extremes of distribution and centralization of network control. We present a logically centralized approach for the design of network decision plane that can be realized by using a set of physically distributed controllers in a network. This approach is aimed at giving network designers the ability to customize the level of control and management centralization according to the scalability, fault-tolerance, and responsiveness requirements of their networks.udOur thesis is that logical centralization provides a robust, reliable, and efficient paradigm for management of large networks and we present several contributions to prove this thesis. For network planning, we describe techniques for optimizing the placement of network controllers and provide guidance on the physical design of logically centralized networks. For network operation, algorithms for maintaining dynamic associations between the decision plane and network devices are presented, along with a protocol that allows a set of network controllers to coordinate their decisions, and present a unified interface to the managed network devices. Furthermore, we study the trade-offs in decision plane application design and provide guidance on application state and logic distribution. Finally, we present results of extensive numerical and simulative analysis of the feasibility and performance of our approach. The results show that logical centralization can provide better scalability and fault-tolerance while maintaining performance similarity with traditional distributed approach.ud