A Scalable Architecture for High Performance, Safely Extensible Routers

Data networks have traditionally been based upon a simple routing and forwarding service provided by the network, with all complexity, for services like reliability and congestion control, placed in end-to-end algorithms run by hosts attached to the network. A fundamentally more powerful network architecture is one in which internal network nodes provide a richer service abstraction. The increased power of such an architecture results from the ability to expose and exploit global network context, and from the ability to dynamically place functionality at strategic points within the network. Making a transition to a functionally richer network has important implications for the system architecture and software architecture of internal network nodes, or routers. Extensible routers must simultaneously provide high forwarding performance and efficient general-purpose computation, making them I/O as well as compute intensive systems. Additionally, while the router should support a dynamically composable computation framework, strong safety and performance isolation guarantees must be provided to the system being extended. Our work attempts to address the challenges posed by this transition, by developing a scalable system architecture and a safely extensible computation framework for routers. Towards this end, we first develop efficient algorithms for critical datapath operations of packet classification and link scheduling. Using these primitives, a decoupled architecture is developed that effectively isolates data forwarding from high-level computation. We present and evaluate Suez , a scalable prototype of the proposed architecture. A software architecture for the system is then presented, that isolates minimal functionality in a protected core, and exposes an interface for safe composition of computation. Preliminary investigations are presented to put the design decisions of the computation framework in perspective, and to show that the framework is efficient in terms of control flow.