Online learning and adaptation of network hypervisor performance models

Software Defined Networking (SDN) paved the way for a logically centralized entity, the SDN controller, to excerpt near real-time control over the forwarding state of a network. Network hypervisors are an in-between layer to allow multiple SDN controllers to share this control by slicing the network and giving each controller the power over a part of the network. This makes network hypervisors a critical component in terms of reliability and performance. At the same time, compute virtualization is ubiquitous and may not guarantee statically assigned resources to the network hypervisors. It is therefore important to understand the performance of network hypervisors in environments with varying compute resources. In this paper we propose an online machine learning pipeline to synthesize a performance model of a running hypervisor instance in the face of varying resources. The performance model allows precise estimations of the current capacity in terms of control message throughput without time-intensive offline benchmarks. We evaluate the pipeline in a virtual testbed with a popular network hypervisor implementation. The results show that the proposed pipeline is able to estimate the capacity of a hypervisor instance with a low error and furthermore is able to quickly detect and adapt to a change in available resources. By exploring the parameter space of the learning pipeline, we discuss its characteristics in terms of estimation accuracy and convergence time for different parameter choices and use cases. Although we evaluate the approach with network hypervisors, our work can be generalized to other latency-sensitive applications with similar characteristics and requirements as network hypervisors.