Weaver: Realizing a Scalable Management Paradigm on Commodity Routers

While there is agreement on the drawbacks of centralized management, many approaches that address those do not scale well to large networks. We believe that effective management of future large-scale networks requires decentralized but coordinated control. In our recent work, we introduced the paradigm of pattern-based management, an approach that formalizes the use of graph traversal algorithms for controlling and coordinating lightweight agents that perform computations and data aggregation inside the network. We have shown analytically and through simulations that such a management system potentially scales to tens of millions of nodes, without significant performance problems regarding execution time and traffic overhead. In this paper, we report on a first implementation designed to realize the paradigm. Our system, Weaver, consists of active nodes constructed from small, low-cost Linux computers that are deployed onto a network of commodity routers. Management programs are written in C++ and can be validated and tested for performance on a simulator before being deployed. From the design of Weaver, we derive a simple performance model that allows us to predict the execution times of management operations on this platform. We evaluate the model through measurements on a laboratory testbed and demonstrate the efficiency of the platform. Finally, we use the model to predict the performance of a management operation running on a Weaver system for a large-scale network and thus show that our system is likely to meet the scaling potential of the paradigm.