Overcoming the Challenges of Cooperating Control Planes

There are many scenarios where dividing control of a networked system into multiple independent control planes is natural and desirable. For example, inter-domain routing should not require a single centralized control plane, but rather utilize independent control planes, allowing each autonomous system to define policies for their own network. This division of control propelled Internet adoption through inter-domain protocols like BGP. However, control plane division often comes at a cost: as no single entity has full information of all resources and policies, correctness, performance, and fault tolerance may suffer. BGP, for example, has been shown to provide non-optimal routes as well as convergence issues during failures. This is further complicated as multiple independent control planes make multiple independent decisions which may have conflicting resource allocation. Clearly, multiple independent control planes can work in harmony, as we’ve seen many examples in the wild and in research, such as multiple BGP instances, OSPF areas, and OSPF fibbing. However, such examples have ad hoc designs. To our knowledge there has been no comprehensive look into characterizing the types of divisions that define multiple control planes, as well as examples of how independent control planes communicate to achieve their overall goals despite these divisions. We argue that there are three fundamental ways to divide control planes: 1) spatially (are data plane resources shared, overlapping, or disjoint?), 2) temporally (do the control planes operate at different timescales?), and 3) administratively (can the control planes share policy information?). Using these three divisions as the axes of a design space, the amount of variety of across different problems and solutions quickly become apparent. In this work we examine three points within the design space that are not covered by existing split control planes in practice or recent research, to examine how each axis affects cooperation. We look at 1) delivering live video over the wide-area in CDNs, 2) optimizing content delivery across multiple CDNs using a content broker, 3) and rearchitecting the datacenter network stack to take advantage of reconfigurable network topologies. Due to their differing points in the design space, we find a different mechanism aids in cooperation for each. Namely, prioritization and isolation, an ad exchange, and cross-layer optimization.