Evaluating the Performance of Machine Learning Models for Dynamic Resource Allocation in NFV

Network Function Virtualization (NFV) is an emerging architecture that leverages virtualization technology to decouple software from hardware. The management of virtual network functions (VNFs) poses a significant challenge in NFV, given the dynamic nature of the system. Resource allocation to VNFs must adjust to accommodate varying traffic loads, but this process can lead to substantial delays. To address this issue, resource estimation models can be applied to predict the resources required for each VNF and optimize dynamic resource allocation. Our study utilized machine learning (ML) techniques to train models using real data from VNFs, encompassing performance metrics and resource needs. The models, once trained, were successful in accurately forecasting the necessary resources for handling particular traffic loads concerning each VNF. The evaluation utilizing real-world data demonstrates the importance of selecting a good model for dynamic resource allocation by contrasting seventeen different ML algorithms with a common fixed allocation model. Using FRA in particular could lead to significant resource over-allocation, a high number of VNF instances, and needless high delay, all of which could lead to high prices and poor service quality. Although linear approximation performs admirably in the assessment, more potent techniques like KNN, SVR, or Boosting outperform it because they can more precisely forecast the actual VNF resource requirements. Overall, using ML to allocate resources dynamically can help cut down on resource waste while enhancing service quality.