ST-RDP: A deep spatio-temporal network model for VNF resource demand prediction of Service Function Chains

Abstract

Virtual Network Functions (VNFs) offer comprehensive network services within Service Function Chains (SFCs), aiming to satisfy the diverse performance requirements of various application scenarios. However, the dynamic and unpredictable nature of the network environment poses substantial challenges for resource allocation across VNF instances, potentially leading to resource under-provisioning or over-provisioning. Consequently, accurate prediction of VNF resource demand is critical for enabling dynamic resource adaptation. To address this challenge, we propose a novel deep spatio-temporal network model, referred to as ST-RDP, for resource demand forecasting. Initially, spatial dependencies among VNFs within the same SFC are captured using an modified Adaptive Graph Convolutional Attention (AGCA) module, which effectively models interdependencies between VNFs. Furthermore, the improved Mamba module is employed to extract time-series features, thereby facilitating accurate spatio-temporal forecasting of resource demand. Experimental evaluations on real-world datasets demonstrate that the proposed approach significantly outperforms existing methods in terms of prediction accuracy and effectiveness.