Asymmetric virtual machine replacement: a multi-objective optimization approach for energy efficiency in cloud data centers

In cloud data centers, resource overcommitment is a key strategy for improving energy efficiency, yet managing it to prevent Service Level Agreement (SLA) violations remains a persistent challenge. Virtual machine replacement (VMrP)—the periodic redistribution of running VMs across servers—is central to this ongoing management. However, existing VMrP approaches balance workloads based on fixed provisioned allocations rather than actual usage, failing to reflect real resource consumption, and distribute VMs symmetrically across servers, overlooking power optimization opportunities inherent in server heterogeneity. This paper proposes AVMrP, a multi-objective optimization framework that addresses these gaps through an asymmetric VM replacement strategy. Rather than distributing VMs uniformly, AVMrP leverages real-time usage observations to permit deliberate, controlled overcommitment across heterogeneous servers, exploiting their distinct power characteristics to achieve energy-efficient VM replacement. Critically, the maximum overcommit ratio is treated as an explicit optimization objective alongside power consumption and migration cost, enabling the framework to sustain energy-efficient asymmetric placements without exposing any individual server to SLA risk. The VMrP problem is NP-hard; AVMrP solves it using a multi-objective evolutionary algorithm (MOEA) with warm-start initialization from the current placement state to minimize migration overhead in periodic operation. Simulation experiments in heterogeneous, dynamic environments show that AVMrP reduces power consumption and significantly lowers migration cost compared to rule-based strategies, while maintaining system stability. Comparison with a reinforcement learning baseline using weighted-sum reward optimization further confirms that simultaneously satisfying competing objectives—power efficiency, migration cost, and overcommit control—proves inherently difficult under scalarized formulations, whereas the Pareto-based approach consistently achieves a well-balanced solution across all three objectives. These results confirm AVMrP as an effective and operationally practical framework for cloud resource management.