On Workload-driven Router Designs

Abstract

Over the past 15 years, the number of networks connected to the Internet, also called Autonomous Systems (ASes), grew at least by a factor of ten. Accordingly, estimations of the global IP tra c further emphasize this trend with a growth by four orders of magnitude. All this puts persistent pressure toward the core infrastructure of the Internet which is required to keep up with the increasing tra c demand. Consequently, Internet routers need to maintain dynamic forwarding information for a large number of networks in a very fast memory, such that a destination lookup operation can be performed on each incoming packet at speed to comply with the low-latency requirements in Internet Service Provider (ISP) networks. The hardly foreseeable, indispensable capacity of this type of specialized memory and the di culties involved in handling the churn in its contents are primary obstacles in Internet router scalability, notably their price tag and energy consumption. In this thesis we rely on insights derived from Internet measurements to design a router architecture that relaxes the requirements for fast destination lookup memory. Firstly, a very small fraction of all networks (IP prefixes) capture most of the tra c — an opportunity for tra c o✏oading. Secondly, the Forwarding Information Base (FIB) can be compressed — an opportunity to reduce FIB memory requirements. Based on these observations we propose a novel IP router design that follows the split architecture paradigm; separating control from data plane functionality. The control plane is a complex, software-centric system that runs routing protocol daemons and acts as the data plane controller. The data plane on the other hand is a specialized, hardware-centric forwarding engine that provides the required primitives for IP routing. A communication channel constitutes a bridge connecting the two entities. Advances in open-source software routers as well as programmable switches enable us to build an IP router prototype that follows our design principles. The emergent developments in the Software Defined Networking (SDN) community and the OpenFlow protocol in particular match the needs of the communication channel and let us demonstrate the feasibility of our design. To complete the study of our prototype, we implement a framework that allows to spot critical performance impairments of OpenFlow-enabled switches. In this dissertation we raise the point that e cient and scalable router designs can be found by leveraging knowledge about router workload properties, i.e., Internet tra c. Accordingly, we conduct an extensive study of Internet tra c as seen by a large European Internet Exchange Point (IXP). The business model of an IXP is to provide a switching fabric that allows a large number of ASes to establish peering relationships and to exchange tra c. Member ASes connect their routers to the IXP’s switching fabric and announce IP prefix-based forwarding information to one another. The large number of neighboring routers that a single member router peers