Abstract
The SDN concept of separating and centralizing the control plane from the data plane has provided more flexibility and programmability to the deployment of the networks. On the other hand, the separation of the planes has raised some scalability and performance questions, being that the SDN controller is the bottleneck. In this paper we present an implementation of a GPU SDN controller. The goal of this paper is to mitigate the scalability problem of the SDN controller by offloading all the packet inspection and creation to the GPU. Experimental evaluation shows that the controller is able to process 17 Million flows/s in the worst case scenario using just off-the-shelf GPU’s.