Evaluating virtual machine allocation in clouds

Abstract

Cloud computing has become a cornerstone of modern IT infrastructure, offering scalable and cost-efficient resource allocation. However, inefficient Virtual Machine (VM) allocation can lead to significant resource wastage, impacting both operational costs and performance. This project evaluates the efficiency of traditional bin packing algorithms—First-Fit, Best-Fit, Next-Fit, and Worst-Fit—in optimizing VM allocation using a real-world dataset from Microsoft Azure. Additionally, it explores the integration of machine learning (LightGBM and XgBoost) to predict VM lifetimes and enhance allocation strategies through a hybrid approach, Dynamic Prefer Best Fit Rule (DPBFR).

The results demonstrate that DPBFR, leveraging binary classification to distinguish between long-lived and short-lived VMs, outperforms traditional algorithms, reducing the number of physical machines required by 0.03% compared to the best-performing baseline (Best-Fit). While LightGBM achieved marginally better efficiency than XgBoost, its runtime was significantly faster (2.1 hours vs. 7.7 hours). Challenges included model accuracy (F1 scores: 0.64 for LightGBM, 0.73 for XgBoost) and scalability limitations in deallocation processes.

This study highlights the potential of machine learning to refine resource allocation in cloud environments, though further improvements in predictive accuracy and runtime optimization are needed. The findings contribute to ongoing efforts to balance computational efficiency with cost-effective resource utilization in cloud infrastructure.