A proactive virtual machine consolidation framework based on multi-dimensional workload awareness and deep reinforcement learning

In contemporary cloud data centers, striking an effective balance between minimizing energy consumption and fulfilling service level agreements (SLAs) presents a critical challenge that impacts system sustainability. Although virtual machine (VM) consolidation has been widely implemented to improve resource utilization, most existing methods overlook multi-dimensional resource constraints, such as disk I/O, or depend on reactive, delayed response mechanisms. This often results in performance variability and resource imbalances when managing non-steady-state workloads and heterogeneous environments. To address these issues, this study introduces a framework called Dynamic Threshold Control and Three-Dimensional Resource Coordination Optimization Framework (DTCF) for VM consolidation. The framework employs a hybrid model, Wavelet-TCN-LSTM, through the 3D-PADT (Three-Dimensional Predictive Adaptive Dynamic Threshold) mechanism. This model captures the spatiotemporal correlation features of workloads and dynamically adjusts overload thresholds to proactively prevent overloads. Additionally, the DMRCIW (Dynamic Multi-Resource Coupling Impact Weight) policy takes into account historical volatility and the interdependence of resources among VMs to identify and reallocate high-risk workloads, thus enhancing system stability. Lastly, the NAP-DRL (Noise-Aware Physics-Constrained Deep Reinforcement Learning) placement algorithm optimizes resource scheduling by using action masking and a physically-aware reward structure, which helps to fully exploit heterogeneous hardware capabilities while adhering to strict resource constraints. Experimental results on the Google Cluster Trace, demonstrate that DTCF outperforms the state-of-the-art Fuzzy-GWO, reducing energy consumption by 23.2% and SLA violations by 43.5% in high-load scenarios. These results confirm the framework's capability to achieve synergistic optimization of system energy efficiency and operational stability under stringent constraints.