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

Key findings from this study

• The study found that proactive threshold adjustment based on spatiotemporal workload prediction reduces both energy consumption and SLA violations compared to reactive consolidation methods.
• The researchers demonstrate that accounting for resource interdependencies through the DMRCIW policy enhances system stability in heterogeneous environments.
• The authors report that physics-aware deep reinforcement learning with action masking achieves superior resource scheduling while respecting hardware constraints.

Overview

The study presents DTCF, a proactive virtual machine consolidation framework designed for cloud data centers. The framework addresses the dual challenge of minimizing energy consumption while maintaining service level agreements under non-steady-state workloads. DTCF integrates workload prediction, dynamic threshold adjustment, and physics-constrained reinforcement learning to optimize resource scheduling across heterogeneous hardware.

Methods and approach

The framework employs a hybrid predictive model combining Wavelet-TCN-LSTM architectures to capture spatiotemporal correlations in workload patterns. The 3D-PADT mechanism dynamically adjusts overload thresholds based on predicted resource demands. The DMRCIW policy evaluates historical volatility and resource interdependencies among VMs to identify high-risk consolidation candidates. The NAP-DRL placement algorithm optimizes VM scheduling through action masking and physically-aware reward structures, constraining solutions to respect hardware limitations.

Results

DTCF achieved 23.2% energy consumption reduction and 43.5% SLA violation reduction compared to state-of-the-art Fuzzy-GWO baselines in high-load scenarios. The framework demonstrated enhanced system stability through synergistic optimization of energy efficiency and operational reliability. Performance gains materialized when evaluated against the Google Cluster Trace dataset under stringent resource constraints.

Implications

The framework's proactive consolidation approach addresses critical limitations in reactive resource management paradigms that characterize existing methods. By incorporating multi-dimensional resource awareness and predictive mechanisms, the study demonstrates that performance variability can be substantially mitigated in heterogeneous cloud environments. The integration of physics-constrained reinforcement learning with workload prediction establishes a pathway for balancing competing optimization objectives under realistic operational constraints.

The substantial improvements in both energy metrics and SLA compliance suggest that modeling resource interdependencies yields measurable operational benefits. Organizations managing large-scale cloud infrastructure may benefit from adopting multi-dimensional resource coupling approaches rather than single-metric optimization strategies. The framework's capability to handle non-steady-state workloads indicates applicability to diverse datacenter deployment scenarios.

Scope and limitations

This summary is based on the study abstract and available metadata. It does not include a full analysis of the complete paper, supplementary materials, or underlying datasets unless explicitly stated. Findings should be interpreted in the context of the original publication.