A demand aware strategy for a machine learning approach to VNF-PC problem

Network Functions Virtualization is a network architecture concept that creates flexibility in provisioning new services and increases provider profits by replacing physical servers with a virtualized solution. In this environment, one of the challenges is to place-and-chain new network functions instances on virtual machines deployed on physical servers. However, this problem belongs to the NP-hard class, besides being solved in an online environment, making its solution very challenging. This work presents a heuristic approach consisting of the combination of an exact approach, proposed in the literature, and used as a baseline; with a dimensionality reduction of the processed components performed with machine learning and the application of re-train concepts. The retraining step is determined using the Kullback-Leibler divergence. Based on a literature review, a novel set of instances, consisting of network service requests divided into classes is also proposed and used to evaluate the results. Simulations based on real network topologies show that the heuristic reduces the runtime by up to 66% compared to the baseline, while keeping the acceptance rate and profit stable, even with changes in the requests. The experimental results suggest that the application of re-train is essential to maintain a high acceptance ratio.