VNF Performance modelling: From stand-alone to chained topologies

One of the main incentives for deploying network functions on a virtualized or cloud-based infrastructure, is the ability for on-demand orchestration and elastic resource scaling following the workload demand. This can also be combined with a multi-party service creation cycle: the service provider sources various network functions from different vendors or developers, and combines them into a modular network service. This way, multiple virtual network functions (VNFs) are connected into more complex topologies called service chains. Deployment speed is important here, and it is therefore beneficial if the service provider can limit extra validation testing of the combined service chain, and rely on the provided profiling results of the supplied single VNFs. Our research shows that it is however not always evident to accurately predict the performance of a total service chain, from the isolated benchmark or profiling tests of its discrete network functions. To mitigate this, we propose a two-step deployment workflow: First, a general trend estimation for the chain performance is derived from the stand-alone VNF profiling results, together with an initial resource allocation. This information then optimizes the second phase, where online monitored data of the service chain is used to quickly adjust the estimated performance model where needed. Our tests show that this can lead to a more efficient VNF chain deployment, needing less scaling iterations to meet the chain performance specification, while avoiding the need for a complete proactive and time-consuming VNF chain validation.