Addressing the challenges of high performance, safely extensible router design

The fact that the Internet has continued to scale to reach its current size is a testimony to its design, which provides a simple and stateless forwarding service in the network and delegates complexity to end systems. However, network-resident computation can provide performance optimizations and functionality to network applications that cannot be provided using end-to-end computation. The legacy of Internet's simple design dictates that it is not possible to place such computation in the network even if it respects Internet's design principles by not keeping critical state in the network. Legacy routers thus lack the cannot provide high performance to a workload consisting of packets bound to computation.; The goal of this thesis is to address the challenges involved in supporting extensibility in the network with high performance and without compromising the robustness of the network. We first present two novel router extensions which validate the premise that network-resident computation can provide significant performance benefits. We then develop a system architecture for network routers that supports safe extensibility with high performance. The building blocks of this architecture are two novel algorithms for route lookup and link scheduling, which are key operations in the router's forwarding path. We incorporate these algorithms in a scalable system architecture that efficiently decouples the generic packet forwarding path from the path of packets bound to computation. A software framework for robust extensibility is then developed, that provides strong memory safety and performance isolation guarantees between core router tasks and router extensions. Two systems resulting from this work are Suez and Srishti. Suez is a scalable prototype of the proposed router architecture built using simple network processors, general-purpose processors and a fast switched backplane. Srishti provides OS support and; By providing a robustly extensible and high-performance system architecture for network routers, we believe that this research lays the groundwork for the deployment of network-resident functions that optimize the performance of network applications, while preserving the Internet's robustness.