Abstract:
Today's networks cannot satisfy the highly diverse and fast changing demands of emerging applications and services. Network protocols and algorithms were not designed to differentiate among services or to adapt to changing services in a timely manner. Combining the paradigms Network Virtualization (NV) and Software-Defined Networking (SDN) can potentially overcome this impasse. Their combination is expected to bring flexible resource sharing with guaranteed performance through NV and programmability through SDN; for the first time, tenants can flexibly program their requested networking resources according to their service demands in a timely manner. However, the combination introduces new challenges in particular for network operators, e.g., low control plane latencies of their virtualization layer. On the other side, tenants of virtual networks are highly demanding in order to fully benefit from the offered flexibility of SDN and NV. Their expectations range from a fast, nearly instantaneous provisioning of their requested virtual networks to a predictable and guaranteed operation of virtual networks.
This talk presents a measurement procedure and a flexible virtualization layer design to identify and mitigate the new challenges introduced by virtualizing software-defined networks – first steps towards predictable and guaranteed virtual SDN network performance. At the same time, it briefly introduces mathematical models for analyzing various design choices of control plane designs of SDN virtualization layers. For a fast and efficient resource management on the data plane, it proposes data-driven optimization systems using machine learning and neural processing of combinatorial optimization problems.
