SDN Hypervisors: How Much Does Topology Abstraction Matter?

SDN network hypervisors realize the virtualization of software-defined networks. They intercept the control path between tenant controllers and their respective virtual Software-Defined Networks (SDN). Over-utilizing SDN hypervisor resources (i.e., CPU) can degrade the control plane performance of the tenants. Although many hypervisor proposals exists, a detailed performance modeling of SDN hypervisors is missing in literature. A precise modeling of the required SDN hypervisor resources, however, is crucial for predictable and reliable operation of virtual software-defined networks. In this paper, we measure and evaluate how topology abstraction can affect the SDN hypervisor CPU utilization. We consider two topology abstraction cases: the (1) transparent and (2) big-switch abstraction. Our measurements taken from a real testbed indicate that the big-switch abstraction can reduce the SDN hypervisor CPU utilization up to $\sim 4\times $. Further, we evaluate different functions to model the SDN hypervisor CPU utilization based on our measurement results. Our evaluations show that a polynomial function provides the lowest fitting error. Motivated by our measurements, we conduct a first-step investigation of the impacts of topology abstraction on the Virtual Network Embedding (VNE) problem. Our initial simulation-based evaluations indicate that different topology abstraction procedures impact the results of the VNE problem.