Deep learning-based multivariate resource utilization prediction for hotspots and coldspots mitigation in green cloud data centers

Dynamic virtual machine (VM) consolidation is a constructive technique to enhance resource usage and is extensively employed to minimize data centers' energy consumption. However, in the current approaches, consolidation techniques are heavily relied on reducing the actively used physical servers (PMs) based on their current resource utilization without considering future resource demands. Also, many of the reported works for cloud workload prediction applied univariate time series-based forecasting models and neglected the dependency of other resource utilization metrics. Thus, resulting in inaccurate predictions, unnecessary migrations, high migration costs, and increased service level agreement violations (SLAVs) may nullify the consolidation benefits. To efficiently address this issue, we propose a multivariate resource usage prediction-based hotspots and coldspots mitigation approach that considers both the current and future usage of resources with O(sk) time complexity, where s and k denote the number of PMs and VMs, respectively. The proposed technique uses a clustering-based stacked bidirectional (Long Short-Term Memory) LSTM deep learning network to predict the future memory and CPU usage of PMs and VMs with high accuracy and O((Q(Q+W)*Theta) computational complexity, where Q, W, and Theta represent the number of hidden layer cells, outputs, and training epochs, respectively. Through extensive simulations based on Google's cluster workload traces, we demonstrate that our proposed method obtains substantial improvements in terms of prediction performance, energy-efficiency, actively used PMs, VM migrations, and SLA violations over the benchmark approaches.

Automated summary

Dynamic virtual machine consolidation techniques, which focus on reducing actively used physical servers based on their current resource utilization, may lead to inaccurate predictions and high migration costs. To address this issue, a new prediction method is proposed that considers both current and future resource usage, resulting in significant improvements in accuracy and computational complexity.