A hybrid optimization-Machine Learning approach for the VNF placement and chaining problem

Abstract

Virtual Network Function Placement and Chaining Problem focuses on allocation of customers demand’s on the Substrate Network. Among other factors, an optimal allocation of resources is hampered by nature of the online problem and by its complexity. Since we are dealing with an NP-hard combinatorial problem, while the number of network components grows linearly, the computational processing and runtime increase exponentially. Therefore, an application of Machine Learning techniques to reduce the number of components present in the SN is proposed in this work. In particular, we have developed clustering techniques that aim to find groups of promising components to map customer demands and disregard those that are less promising. Two different clustering models are proposed: (i) based on the Spatial Location of the SN components; and (ii) based on SN’s historical resource consumption data. An Integer Linear Programming model is proposed to evaluate the different simulation scenarios. Both approaches reduced the execution time (≈75%) and the end-to-end delay of virtual requests, in addition to keeping the acceptance rate and profit stable compared to the exact approach.