💻 computer science

An Adaptive Grey Wolf Optimizer for Energy-Aware Virtual Machine Consolidation in Sustainable Green Data Centers

This paper introduces AGWO-VMC, an adaptive Grey Wolf Optimizer that employs a gradient-feedback decay mechanism, probabilistic encoding, and local search to significantly reduce energy consumption and SLA violations in green data centers compared to existing metaheuristic approaches.

Original authors: Maad khalil, Odiljan Turdiyev, Ibtesam Hawi, Tolaniddin Nurmukhamedov

Published 2026-09-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Maad khalil, Odiljan Turdiyev, Ibtesam Hawi, Tolaniddin Nurmukhamedov

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every day, the world's digital infrastructure hums with a massive, invisible appetite. From the video calls that connect families to the financial trades that move markets, the servers running these services consume staggering amounts of electricity. For the last decade, this demand has grown by roughly ten percent annually, and projections suggest that by 2030, data centers will gobble up nearly four percent of the world's total power. A significant portion of this energy is wasted not because the servers are working too hard, but because they are sitting idle. In many operational centers, physical machines run at only a fraction of their capacity, yet they still draw more than half their peak power just to stay on. To fix this, engineers use a strategy called virtual machine consolidation. Imagine a fleet of delivery trucks where some are half-empty while others are packed to the brim. Consolidation is the act of moving the cargo from the half-empty trucks into the full ones, then turning off the empty vehicles to save fuel. In the digital world, this means shifting software tasks from many underused servers onto a few busy ones, allowing the rest to sleep. However, finding the perfect way to pack these tasks is a puzzle of immense complexity, and doing it poorly can cause the services to slow down or fail.

Researchers Maad Khalil, Odiljan Turdiyev, Ibtesam Hawi, and Tolaniddin Nurmukhamedov have developed a new method to solve this packing puzzle more efficiently than ever before. They focused on a specific type of problem-solving tool known as a metaheuristic, which acts like a smart search engine for complex arrangements. One popular tool, called the Grey Wolf Optimizer, mimics the social hierarchy of a wolf pack to find good solutions. In this system, a few "leaders" guide the rest of the group toward better positions. The standard version of this tool uses a fixed schedule to decide when to explore new possibilities and when to focus on refining the current best answer. The researchers found that this fixed schedule is too rigid for the chaotic, changing environment of a data center. When workloads shift suddenly, the standard tool often gets stuck in a local solution, missing better opportunities because it stops looking too early.

To overcome this, the team created an improved version they call AGWO-VMC. Instead of following a rigid clock, their new system watches the progress of the search in real time. It measures how quickly the quality of the solution is improving. If the search hits a flat spot where no changes are making things better, the system knows to keep exploring widely, refusing to settle. But if the search finds a steep path toward a better solution, it immediately tightens its focus to exploit that discovery. This feedback loop allows the algorithm to adapt its behavior moment by moment, reacting to the actual shape of the problem rather than a pre-set timer. They also added two other refinements: a method to translate the continuous movements of the algorithm into concrete decisions about which server to use, and a periodic check that swaps tasks around to see if a small rearrangement can save even more energy.

The researchers tested this new approach using detailed simulations of real-world data centers, drawing on workload data from Google and Microsoft Azure. They ran the simulation across four different scenarios, ranging from steady, predictable traffic to sudden, massive spikes in demand. In every case, the new adaptive system outperformed the standard version and other competing methods. It reduced the total energy consumed by the servers by between 18.5 and 23.7 percent compared to the standard tool, with even larger savings when compared to older algorithms. Crucially, it achieved these energy savings without sacrificing the quality of service; the rate of service violations remained below 1.9 percent in all tests. The study also showed that the system could make these decisions in under 18 seconds, a speed fast enough to be used in real-time operations.

The team confirmed that the key to this success was the adaptive feedback mechanism itself. By removing this feature in their tests, they saw energy savings drop significantly, proving that the ability to listen to the search process was the primary driver of improvement. They also compared their method against a sophisticated learning-based approach that requires months of historical data to train. While that learning-based method performed well on familiar patterns, it struggled when faced with sudden, unexpected changes in workload. The new adaptive system, which requires no training data and reacts instantly to current conditions, proved more robust and reliable across all scenarios. By releasing their code as open-source software, the researchers have provided a transparent benchmark for others to verify and build upon, offering a clear path toward greener, more efficient cloud computing.

Technical Summary: An Adaptive Grey Wolf Optimizer for Energy-Aware Virtual Machine Consolidation in Sustainable Green Data Centers

1. Problem Definition

The paper addresses the Virtual Machine Consolidation (VMC) problem in cloud data centers, a critical challenge for reducing the global energy footprint of computing infrastructure. With data center electricity consumption projected to reach 3–4% of global usage by 2030, the primary inefficiency lies in physical servers running at low utilization (15–20% of rated capacity) while consuming 50–60% of peak power even when idle.

VMC aims to pack workloads onto the minimal set of active physical machines (PMs) to allow idle servers to be powered down. However, this is an NP-Hard multi-dimensional bin-packing problem. The core difficulty involves balancing three conflicting objectives:

Energy Minimization: Reducing the number of active hosts and their power consumption.
Service Level Agreement (SLA) Compliance: Avoiding resource contention that leads to performance degradation.
Migration Overhead: Minimizing the energy and performance costs associated with live VM migrations.

The paper argues that standard metaheuristic solvers, specifically the Grey Wolf Optimizer (GWO), often fail in dynamic, non-stationary workload environments. Standard GWO employs a linear decaying control parameter ( $a$ ) that shifts from exploration to exploitation at a fixed rate. This rigid schedule causes premature exploitation during workload stagnation (trapping the algorithm in local optima) and insufficient exploitation during productive convergence phases.

2. Methodology: AGWO-VMC Framework

The authors propose AGWO-VMC (Adaptive Grey Wolf Optimizer for Virtual Machine Consolidation), a framework designed to dynamically adjust the exploration-exploitation balance based on real-time optimization feedback.

Core Innovations

The framework integrates three complementary innovations:

Gradient-Feedback Adaptive Decay (Primary Innovation):
- Instead of a linear decay based solely on iteration count, the control parameter $a(t)$ is modulated by a smoothed fitness gradient over a trailing window ( $k=5$ iterations).
- Mechanism: If the fitness improvement rate (gradient) is near zero (indicating stagnation), the decay is slowed to encourage further exploration. If the gradient is strongly positive (indicating productive convergence), decay accelerates to intensify exploitation.
- Mathematical Formulation: $a(t) = 2(1 - \frac{t}{t_{max}})^\lambda \times (1 - \gamma \bar{f}_{avg}(t))$ , where $\lambda$ is a decay exponent and $\gamma$ is a sensitivity coefficient.
Probabilistic Discrete Encoding:
- Since GWO operates in continuous space but VMC requires discrete PM assignments, the authors replace simple rounding with a temperature-controlled softmax mapping. This maps continuous wolf positions to integer PM identifiers, reducing quantization artifacts and allowing for controlled stochasticity in assignment.
Periodic 2-Opt Local Search:
- A local search is applied exclusively to the leading solution ( $\alpha$ -wolf) every 5 iterations. It performs pairwise VM swaps to refine the configuration, improving the best-known solution without compromising the diversity of the overall population.

Fitness Function

The algorithm optimizes a multi-objective scalarized fitness function weighted via the Analytic Hierarchy Process (AHP):
$F(\phi) = 0.5 E_{norm} + 0.3 SLA_{viol} + 0.15 E_{mig\_norm} + 0.05 U_{imbalance}$
Where the objectives are normalized server energy, composite SLA violations (SLATAH $\times$ PDM), migration energy, and load imbalance.

3. Experimental Setup

Simulation Environment: CloudSim Plus 7.3.
Infrastructure: 800–1,000 heterogeneous physical machines (HP ProLiant G4/G5) in a fat-tree topology.
Workload Traces: Google Cluster Trace (2019) and Azure Public Dataset (2024).
Scenarios: Four distinct workload patterns were tested: Static, Dynamic Steady (diurnal cycles), Flash Crowd (sudden spikes), and Heterogeneous Mix.
Baselines: The proposed method was compared against Best Fit Decreasing (BFD), Genetic Algorithm (GA), Particle Swarm Optimization (PSO), standard GWO, NSGA-II, and a Deep Reinforcement Learning (DRL) baseline (Double DQN).
Statistical Rigor: 30 independent trials per scenario, evaluated using Wilcoxon signed-rank tests and Cohen's $d$ effect sizes.

4. Key Results

The experimental results demonstrate that AGWO-VMC significantly outperforms existing methods across all metrics:

Energy Efficiency: AGWO-VMC reduced energy consumption by 18.5% to 23.7% compared to standard GWO. Improvements were even more significant against PSO (up to 29.3%) and GA (up to 32.8%). The largest absolute saving was 328.1 kWh over a 7-day window in the Dynamic Steady scenario.
SLA Compliance: The composite SLA violation rate remained below 1.9% in all scenarios, a significant improvement over standard GWO (which ranged from 2.41% to 5.43%).
Migration Reduction: The algorithm reduced the number of migrations by 20.9% to 24.3% compared to standard GWO, directly lowering migration energy costs.
Convergence Speed: AGWO-VMC reached 95% of its final fitness value 13.2% to 20.3% faster than standard GWO and exhibited lower solution variability (standard deviation of 0.084 vs. 0.127).
Comparison with DRL: While the DRL baseline performed competitively on scenarios matching its training data, it degraded significantly (11.7% higher energy) on "Flash Crowd" scenarios due to distributional shift. AGWO-VMC required no training data and maintained superior performance across all novel workloads.
Statistical Significance: All pairwise comparisons yielded $p < 0.001$ with large effect sizes (Cohen's $d > 0.8$ ).

Ablation Study

The study quantified the contribution of each component in the Dynamic Steady scenario by measuring the increase in energy consumption relative to the full AGWO-VMC when a component was removed:

Removing the adaptive decay (reverting to linear) resulted in a 21.4% increase in energy consumption.
Removing the gradient feedback (keeping the non-linear base schedule) resulted in a 12.7% increase in energy consumption.
Removing probabilistic mapping (using hard rounding) resulted in an 11.9% increase in energy consumption.
Removing 2-Opt local search resulted in an 8.4% increase in energy consumption.
Removing AHP weights (using equal weighting) resulted in a 6.5% increase in energy consumption.

5. Significance and Claims

The paper claims that the primary significance of AGWO-VMC lies in its closed-loop adaptation mechanism. Unlike previous adaptive GWO variants that rely on iteration counts or population statistics, AGWO-VMC directly observes the optimization trajectory (fitness gradient). This allows the algorithm to dynamically respond to the non-stationary nature of cloud workloads, specifically by extending exploration during stagnation phases where linear decay would prematurely trap the swarm.

The authors emphasize that the implementation is released as open-source software to serve as a reproducible benchmark for green data center research. They position the work as a practical solution for cloud providers that balances energy efficiency with strict SLA requirements without the heavy training overhead associated with deep reinforcement learning approaches.

Limitations Acknowledged:

Experiments were conducted in simulation; network congestion effects at extreme migration concurrency may differ in physical testbeds.
The fitness function relies on fixed AHP weights, which may need re-tuning for operators with different priority structures (e.g., SLA > Energy).
The stagnation detection assumes a zero gradient indicates a local optimum, which may not always distinguish between a local optimum and an extended basin of attraction.