Alert-Based Network Reconfiguration and Data Evacuation

Resilient Control Plane Design for Virtual Software Defined Networks

“…Although resilient controller placement is a well investigated topic, to the best of our knowledge, our work is the first which considers resilient hypervisor placement and in-band control channel design pro-actively before the failure occurs. A re-active approach called Dynamic Hypervisor Placement Problem was proposed in [26], where the embedding of virtual networks might be changed owing to a disaster alert [27], which requires the redesign of the control plane and hypervisor assignment. Two approaches were investigated [27], namely when the virtual network to hypervisor assignment is fixed but the hypervisor can be migrated, and when the hypervisor assignments are fixed but the virtual networks can be reassigned to different hypervisors.…”

“…A re-active approach called Dynamic Hypervisor Placement Problem was proposed in [26], where the embedding of virtual networks might be changed owing to a disaster alert [27], which requires the redesign of the control plane and hypervisor assignment. Two approaches were investigated [27], namely when the virtual network to hypervisor assignment is fixed but the hypervisor can be migrated, and when the hypervisor assignments are fixed but the virtual networks can be reassigned to different hypervisors. However, these methods are not suitable for the instantaneous reaction to link failures on the control path.…”

“…In contrast to [23]- [27] minimizing for (average) control path lengths with complex Integer Linear Programs (ILPs), in our novel resilient self-driving control-plane design we will propose a polynomial-time placement algorithm without using any task-specific objectives, without relying on a predefined number of hypervisor instances or having a priori knowledge about virtual network demands. As a result the resilient control path lengths will not be optimal in our hypervisor placement but still satisfy the SLAs for all (future) virtual networks where possible.…”

“…Although it would be possible to rerun a HPP algorithm to recalculate a whole switch-tohypervisor assignment, creating new partitions of the network controlled by new hypervisor instances at new locations would cause unnecessary disruptions for all tenants with running virtual networks. Instead of this, we either suggest to migrate the hypervisor to a different location within the partition if the disruption can be forecasted [27], or run an HPP algorithm only to the subset of physical switches which were assigned to the failed hypervisor. Note that, owing to the increased shortest path lengths caused by the node failure, there might be multiple new partitions as well.…”

“…2. network resources should be rapidly evacuated [27] upon receiving a disaster alert (e.g., in milliseconds for earthquakes or in seconds or minutes for tornadoes and hurricanes), or for topologies where placements might be recalculated frequently owing to day-night or hourly traffic fluctuations, Suurballe's algorithm is the better choice to obtain nodeDistance() and the pre-allocated control paths in Alg. 1 compared to the ILP formulations with high computational complexity.…”