Data center multidimensional management strategy based on descending neighborhood DBSCAN algorithm in unsupervised learning

Abstract

Cloud users rent virtual machines (VMs) with varying parameters tailored to their unique business requirements. These diverse VM parameters add complexity to data center (DC) management strategies. Among the crucial parameters are CPU and memory, which must be optimized to ensure efficient physical resource utilization and decreased DC energy consumption. This article proposes three algorithms to manage and optimize VMs. Firstly, the density-based spatial clustering of applications with noise (DBSCAN) algorithm is enhanced, leading to the introduction of the descending neighborhood DBSCAN (DNDBSCAN) algorithm. This algorithm facilitates the clustering of physical machines (PMs). Secondly, the cluster center nearest classification algorithm (CCN) is proposed, leveraging VM attributes and the remaining capacity of the cluster center to classify the VMs for deployment. Additionally, the avoid hot spot time correlation algorithm (AHTC) is introduced to handle VM mapping, deploying VMs on the most time-relevant PMs while mitigating hot spots. Lastly, these three algorithms are integrated into a DC multidimensional management strategy based on the DNDBSCAN algorithm within the framework of unsupervised learning (DND). When compared to other algorithms, the DND algorithm demonstrates significant improvement in PM balanced utilization and reduction of DC energy consumption. The average balanced utilization of PM of the DND algorithm is 86 %, which is an average improvement of 11 % compared to the comparative algorithm. The average total energy consumption of the DND algorithm is 124 kW•h, which is an average reduction of 41 % compared to the comparative algorithm.