Performance Characterisation and Predictive Modelling of WireGuard VPN on Kubernetes: Efficient Resource Management Towards Auto-scaling

Abstract

This study investigates WireGuard Virtual Private Network (WG VPN) performance on Kubernetes (K8s), analysing resource use and developing regression models. Metrics (Throughput, CPU, Memory, Jitter) were analysed across vCPU and physical core allocations. We found that over-provisioning vCPUs degraded performance due to increased overhead, with optimal throughput achieved, especially with two vCPUs and two physical cores. Among these optimal matched configurations, one demonstrated superior cost-effectiveness (high throughput/core, low CPU) and minimal jitter, which is vital for real-time applications. Pod-level CPU lacked traffic correlation due to the host kernel reliance on WGVPN, making Node-level Horizontal Scaling more suitable than Pod-level Vertical. Regression models for Node-level CPU (best: Linear Regression, with MAE decreasing by 16.74% compared to Multi-Layer Perceptron Model) and Traffic Demands (best: Instance-Based k-Nearest Neighbours, with MAE decreasing by 9.02% compared to Multi-Layer Perceptron Model) were developed. Findings offer insights for optimising WGVPN on K8s and future auto-scaling.