Evaluation of Dynamic Vector Bin Packing for Virtual Machine Placement

TL;DR

This paper evaluates various online algorithms for dynamic vector bin packing in virtual machine placement, focusing on minimizing total machine usage time using real-world Azure data.

Contribution

It compares state-of-the-art algorithms across different knowledge settings and introduces new or improved algorithms for VM placement.

Findings

Clairvoyant algorithms outperform non-clairvoyant ones in efficiency.

Learning-augmented algorithms show promising results with predicted data.

Insights into algorithm structures that enhance practical performance.

Abstract

Virtual machine placement is a crucial challenge in cloud computing for efficiently utilizing physical machine resources in data centers. Virtual machine placement can be formulated as a MinUsageTime Dynamic Vector Bin Packing (DVBP) problem, aiming to minimize the total usage time of the physical machines. This paper evaluates state-of-the-art MinUsageTime DVBP algorithms in non-clairvoyant, clairvoyant and learning-augmented online settings, where item durations (virtual machine lifetimes) are unknown, known and predicted, respectively. Besides the algorithms taken from the literature, we also develop several new algorithms or enhancements. Empirical experimentation is carried out with real-world datasets of Microsoft Azure. The insights from the experimental results are discussed to explore the structures of algorithms and promising design elements that work well in practice.