Online Resource Mapping for SDN Network Hypervisors using Machine Learning

The virtualization of Software-Defined Networks (SDN) allows multiple tenants to share the same physical infrastructure and to use their own SDN controllers. SDN virtualization is achieved through an SDN network hypervisor that operates between the tenants' controllers and the SDN infrastructure. In order to provide performance guarantees, resource mapping is required for both data plane as well as control plane for each virtual SDN network. In the context of SDN virtualization, the control plane resources include the network hypervisor, which needs to be assigned to guarantee the performance for each tenant. In previous work, the hypervisor resource mapping is based on offline benchmarks that measure the hypervisor resource consumption against the control plane work load, e.g., control plane message rate. These offline benchmarks vary across different hypervisor implementations, e.g., single or multi-threaded, and depend on the capabilities of the deployed hardware platform, e.g., the used CPU. We propose an online approach based on machine learning techniques to determine the mapping of hypervisor resources to the control workload at runtime. This concept is already successfully applied in the context of self-configuring networks. We propose three models to estimate hypervisor resources and compare them for two SDN hypervisor implementations, namely FlowVisor and OpenVirteX. We show through measurements on a real virtualized SDN infrastructure that resource mappings can be learned on runtime with insignificant error margins.