Abstract
Reducing energy consumption and optimizing resource usage are essential goals for researchers and cloud providers managing large cloud data centers. Recent advancements have demonstrated the effectiveness of virtual machine consolidation and live migrations as viable solutions. However, many existing strategies are based on immediate workload fluctuations to detect host overload or underload and trigger migration processes. This approach can lead to frequent and unnecessary VM migrations, resulting in energy inefficiency, performance degradation, and service-level agreement (SLA) breaches. Moreover, traditional time series and machine learning models often struggle to accurately predict the dynamic nature of cloud workloads. This paper presents a consolidation strategy based on predicting resource utilization to identify overloaded hosts using novel hybrid value trend generative adversarial network (VTGAN) models. These models not only predict future workloads but also forecast workload trends (i.e., the upward or downward direction of the workload). Trend classification can simplify the decision-making process in resource management approaches. We perform simulations using real PlanetLab workloads on Cloudsim to assess the effectiveness of the proposed VTGAN approaches, based on value and trend, compared to the baseline algorithms. The experimental findings demonstrate that the VTGAN (Up current and predicted trends) approach significantly reduces SLA violations and the number of VM migrations by 79% and 56%, respectively, compared to THR-MMT-PBFD. Additionally, incorporating VTGAN into the VM placement algorithm to disregard hosts predicted to become overloaded further improves performance. After excluding these predicted overloaded servers from the placement process, SLA violations and the number of VM migrations are reduced by 84% and 76%, respectively, compared to THR-MMT-PBFD.