Statistical and Machine Learning-Based Performance Evaluation of Hypervisor and Container Virtualization in Linux Systems

Abstract

Virtualization technologies are fundamental to cloud-native and edge computing infrastructures, enabling efficient resource utilization and workload isolation. Although hypervisor-based and container-based virtualization mechanisms have been extensively compared, limited research integrates rigorous statistical validation with predictive modeling to characterize performance overhead and nonlinear scalability behavior. This work develops a reproducible experimental framework for evaluating hypervisor-based virtualization using Kernel-based Virtual Machine and container-based virtualization through Docker and Linux Containers under controlled Linux environments.

CPU-bound, memory-bound, disk I/O, network-intensive, and web-based workloads were executed with 30 independent repetitions per scenario. Performance indicators included throughput, mean and p95 latency, startup time, resource consumption, and relative overhead compared to bare-metal execution. Statistical analysis was conducted using Shapiro–Wilk normality tests, one-way ANOVA, and Tukey post-hoc comparisons. Supervised machine learning regression models were developed to predict performance degradation under concurrent deployment conditions.

Results indicate statistically significant differences in disk and network workloads (p < 0.05), while CPU-bound scenarios show minor variation between container-based approaches. Predictive models achieved coefficients of determination above 0.90, revealing nonlinear degradation patterns as concurrency increased. The proposed methodology contributes a statistically validated benchmarking framework and demonstrates how machine learning enhances interpretability in virtualization performance analysis.