TOWARDS EFFICIENT VM PLACEMENT: ATWO-STAGE ACO–PSO APPROACH FORGREEN CLOUD INFRASTRUCTURE

Ali Baydoun 1 and Ahmed Zekri 2
1 Department of Mathematics & Computer Science, Beirut Arab University, Lebanon
2 Department of Mathematics & Computer Science, Alexandria University, Egypt

ABSTRACT
Datacenters consume a growing share of energy, prompting the need for sustainable resource management. This paper presents a Hybrid ACO–PSO (HAPSO) algorithm for energy-aware virtual machine (VM) placement and migration in green cloud datacenters. In the first stage, Ant Colony Optimization (ACO) performs energy-efficient initial placement across physical hosts, ensuring global feasibility. In the second stage, a discrete Particle Swarm Optimization (PSO) refines allocations by migrating VMs from overloaded or underutilized hosts. HAPSO introduces several innovations: sequential hybridization of metaheuristics, system-informed particle initialization using ACO output, heuristic-guided discretization for constraint handling, and a multi-objective fitness function that minimizes active servers and resource wastage. Implemented in CloudSimPlus, extensive simulations demonstrate that HAPSO consistently outperforms classical heuristics (BFD, FFD), Unified Ant Colony System (UACS), and ACO-only. Notably, HAPSO achieves up to 25% lower energy consumption and 18% fewer SLA violations compared to UACS at largescale workloads, while sustaining stable cost and carbon emissions. These results highlight the effectiveness of two-stage bio-inspired hybridization in addressing the dynamic and multi-objective nature of cloud resource management.

KEYWORDS
Cloud computing, Green datacenters, Virtual machine placement, AntColony Optimization (ACO), Particle
Swarm Optimization (PSO)

1.INTRODUCTION

Cloud datacenters are the backbone of modern computing, powering a wide range of services across infrastructure, platform, and software layers. While this model provides scalability and flexibility, it also comes with a significant environmental cost. Global datacenters are estimated to consume over 416 terawatt-hours annually—equivalent to more than 3% of the world’s total electricity production—and contribute substantially to carbon emissions due to their continuous power demands. A large portion of this energy is consumed not only for computation but also for cooling and maintaining network infrastructure. For example, cooling systems alone can account for nearly 40% of a datacenter’s power budget, while network switches and routers can draw nonnegligible power even under light traffic conditions [1].

Cloud computing relies heavily on virtualization, which allows multiple virtual machines (VMs) to run concurrently on shared physical infrastructure. This brings efficiency benefits but also introduces new challenges. Poorly optimized VM placement can lead to underutilized servers, excessive energy consumption, and performance degradation—particularly during times of workload fluctuation. To address this, cloud providers implement VM placement and consolidation

strategies to intelligently map VMs to physical machines (PMs), enabling idle servers to be powered down and reducing the overall energy footprint. Designing an optimal VM-to-PM mapping is a complex combinatorial problem. The mapping must respect physical resource constraints (CPU, memory, bandwidth [BW]), satisfy quality of service guarantees (typically defined via Service Level Agreements [SLAs]), and adapt dynamically to workload changes. As a result, a wide variety of optimization techniques have been proposed in the literature, particularly heuristic and metaheuristic approaches such as Genetic Algorithms (GA), Ant Colony Optimization (ACO), Particle Swarm Optimization (PSO), and Simulated Annealing (SA). These methods have been used to address energy efficiency [2], workload balancing [3], and even network-aware VM scheduling [4].

One growing trend is to combine multiple metaheuristics to leverage their complementary strengths. For example, in fog-cloud environments, [5] employed a hybrid FAHP–FTOPSIS approach to achieve efficient load balancing and energy savings, highlighting the broader applicability of hybrid designs. This demonstrates that hybridization is not limited to edge and fog scenarios but can be effectively extended to VM placement, where the joint optimization of static allocation and dynamic migration becomes crucial. Yet, existing approaches often rely on a singlestage process—focusing solely on either initial placement or migration—without fully coordinating both phases of the VM lifecycle. This leaves an important gap in the management of real-world cloud workloads, which are dynamic by nature and demand flexible resource allocation mechanisms.

To address this gap, we propose a Hybrid Ant Colony–Particle Swarm Optimization (HAPSO) framework designed for adaptive VM placement in green cloud datacenters. Like recent sustainability-driven bio-inspired strategies [6], our approach embeds energy- and carbonawareness directly into the placement process. It integrates ACO for global static VM placement and PSO for runtime migration refinement. HAPSO clearly separates placement from migration and embeds sustainability constraints, making it well suited for dynamic, green-aware cloud environments.

The unified goal is to minimize energy consumption, reduce carbon emissions, and eliminate unnecessary resource wastage—all while maintaining SLA compliance. Our architecture supports heterogeneous datacenters, including those powered partially by renewable energy sources (e.g., solar or wind), and continuously monitors power, performance, and environmental data to inform optimization decisions.

While hybrid ACO–PSO models have been explored in prior research, most adopt a single-stage integration or lack sustainability awareness. A more detailed review of these approaches is presented in Section 2.

The contributions are as follows: (1) a two-stage hybrid framework that combines ACO for initial VM placement with PSO for dynamic refinement, improving scalability and responsiveness; (2) a system-aware particle initialization strategy that seeds PSO with the current ACO-based mapping to ensure feasibility and accelerate convergence; and (3) a multi-objective fitness function that jointly minimizes the number of active servers and residual resource wastage across CPU, memory, and bandwidth to enhance overall energy efficiency.

This work builds upon our previous ACO-based VM placement approach [4] by integrating PSO to handle dynamic consolidation more efficiently.

The rest of this article is structured as follows: Section 2 reviews the related work, focusing on metaheuristics and hybrid approaches for VM placement. Section 3 formulates the problem and defines the system model, including architectural assumptions, notation, and the multi-objective optimization function. Section 4 describes the proposed HAPSO algorithm, covering its two-stage design, particle representation, velocity and position update mechanisms, algorithmic procedures, and parameter settings. Section 5 outlines the experimental results, including the simulation setup and performance comparison against ACO-only and other baseline approaches. Finally, Section 6 concludes the paper and discusses potential directions for future work

2.RELATED WORK

VMP plays a critical role in optimizing CDC performance by minimizing energy usage, balancing workload, and improving SLA compliance. A related work [7] examined placement in cloudlets within wireless metropolitan networks. While this addresses a different context, it similarly underlines the role of intelligent placement in improving efficiency.

To handle the complexity and dynamic nature of this NP-hard problem, researchers have widely adopted metaheuristic algorithms, with recent trends shifting toward hybrid metaheuristic models, especially ACO–PSO integrations.

2.1. Metaheuristic Algorithms for VM Placement

Metaheuristics such as PSO, SA, GA, and ACO have been extensively applied to VMP due to their ability to explore large search spaces efficiently.

PSO approaches model VM-to-host mappings as particles navigating the search space. Various works ([8], [9]) have demonstrated the potential of PSO in reducing energy consumption and bandwidth usage. More recently, [10] extended PSO with a quantum-inspired formulation (QPSOMOVMP), achieving Pareto-optimal placement solutions that balance energy, SLA, and load distribution. Authors of [11] developed a discrete PSO for VM placement that optimized energy consumption and reduced migrations. However, PSO’s tendency to converge prematurely— especially in large-scale settings—limits its robustness.

SA, another probabilistic method, occasionally accepts worse solutions to escape local optima. Despite this, its slow convergence makes it unsuitable for real-time decisions [12], [13].

GA methods like those in [14] and [15] evolve placements using crossover and mutation, but they require significant computation and careful tuning.

Contemporary heuristic-driven approaches such as MOVMS and MOMBFD [16] also highlight the importance of multi-objective trade-offs in VM placement, emphasizing energy savings and SLA guarantees

ACO algorithms have shown strong results in multi-objective and constraint-based VMP scenarios. These algorithms build feasible mappings via a constructive, pheromone-guided process. Variants like MoOuACO [17], AP-ACO [18], and ETA-ACO [19] target energy, traffic, and server/network joint optimization

Beyond classical metaheuristics, reinforcement learning has emerged as a strong alternative. For example, [20] introduced CARBON-DQN, which combines deep Q-networks with clustering to achieve carbon-aware and SLA-compliant VM placement.

2.2. Hybrid Metaheuristic Models

To overcome individual limitations, hybrid metaheuristics have emerged that combine global and local search techniques.

ACO–GA models like ACOGA [21] leverage pheromone trails with evolutionary recombination to reduce traffic and active server count. The dynamic hybrid ACOPS [22] predicts VM load patterns using ACO and adapts mappings with PSO in real time. The work [23] integrates a permutation-based genetic algorithm (IGA-POP) with a multidimensional best-fit strategy to reduce the number of active servers and balance resource usage.

Another approach is ACO–SA hybridization, as discussed in [24], which applies ACO for global path construction and SA for probabilistic refinement.

More recently, [25] combined ACO with Grey Wolf Optimization (GWO), demonstrating improvements in communication-aware and energy-efficient VM placement.

Another recent hybrid integrates Genetic Algorithms with Harris Hawks Optimization, offering energy-aware VM placement through complementary exploration and exploitation [26].

Despite these contributions, many existing hybrids suffer from constraint violations requiring penalty-based correction and neglect of green datacenter factors such as renewable energy, carbon tax, or dynamic cooling models.

2.3. ACO–PSO Hybridization for VM Placement

Hybrid ACO–PSO models aim to integrate ACO’s guided exploration with PSO’s fast convergence

In [27], sequential hybridization uses ACO for initial VM placement and PSO to fine-tune placements for energy and resource optimization.

The framework in [28] applied iterative hybridization in their energy-aware scheduling hybrid algorithm where ACO explored the placement space, and PSO fine-tuned it.

A layered hybrid in [29] combines ACO, PSO, and ABC to assign each algorithm a role in exploration, convergence, and local refinement.

These hybrid strategies validate the synergy between ACO and PSO approaches. However, most still lack domain-specific encoding and are rarely implemented in sustainability-focused cloud environments.

2.4. Positioning of HAPSO

Our proposed HAPSO framework integrates ACO and PSO in a clean two-stage model:

• ACO is used for multi-objective initial placement, incorporating energy, carbon, network bandwidth, and SLA considerations through dynamic PUE and renewable-aware heuristics

• PSO is selectively triggered during runtime for VM migrations, ensuring feasibility and fast convergence using particle-based refinements.

Unlike prior methods, HAPSO 1) embeds domain constraints directly into both phases— in the ACO stage, resource requirements (CPU,RAM,BW) are encoded into the heuristic function and pheromone update rules to guide solution construction; in the PSO stage, the particle initialization, position update, and discretization mechanisms enforce feasibility by incorporating VM resource requirements, ensuring that all explored solutions remain valid throughout the optimization process, 2) operates in a green-aware context, including real-time solar energy profiles, carbon rate modeling, and temperature-aware cooling overheads, and 3) achieves modularity and scalability by explicitly separating placement from migration logic.

Together, these features position HAPSO as a robust and adaptive framework for energy-efficient VM placement in modern cloud infrastructures.

3. PROBLEM FORMULATION

We adopt the same VM-to-PM mapping model and constraints as in our previous work [4], including resource capacity, SLA compliance, and energy cost modeling. In this extension, we also define new fitness function for the dynamic phase, to reduce active server count and minimize resource wastage.

3.1. System Architecture

Figure 1. System architecture of the HAPSO-based cloud management system. The Cloud Broker dispatches VM placement requests to geographically distributed datacenters. Each datacenter’s placement module uses ACO for initial placement and PSO for adaptive VM migration upon detecting host overload or underload. Continuous monitoring provides utilization and environmental metrics to inform the optimization.

The architectural model of the HAPSO-based cloud system is illustrated in Figure 1. The platform is designed to emulate a multi-datacenter cloud environment where user-generate VM requests are submitted to a central Cloud Broker. This broker is responsible for dispatching VM allocation decisions across a network of geographically distributed datacenters. Within each datacenter, a VM Placement Module handles both static and dynamic optimization. As shown in the figure, VM requests are initially handled by an Ant Colony Optimization (ACO) engine that performs energyand constraint-aware placement. During runtime, host utilization is periodically monitored, and

when overload or underload conditions are detected, the Particle Swarm Optimization (PSO) module is invoked to refine VM allocations through migration.

The architecture supports heterogeneous datacenters, some of which are partially powered by renewable sources (such as solar energy) or local generators. This setup enables evaluation of placement strategies under green energy-aware policies. Energy consumption, carbon emissions, and resource utilization are all monitored continuously throughout the simulation.

3.2. Notations and Definitions

Let us assume there are 𝑁 𝑉𝑀𝑠 and 𝑀 𝑃𝑀𝑠. The set of VMs is denoted as 𝑉𝑀 = {𝑉𝑀1,𝑉𝑀2, … , 𝑉𝑀𝑁}, and the set of 𝑃𝑀𝑠 is denoted as 𝑃𝑀 = {𝑃𝑀1,𝑃𝑀2, … , 𝑃𝑀𝑀}. Each VM, denoted as 𝑉𝑀𝑗 ∈ 𝑉𝑀, has 𝐶𝑃𝑈, 𝑅𝐴𝑀, and bandwidth (𝐵𝑊) requirements represented by 𝑉𝑀𝑗 𝐶𝑃𝑈 , 𝑉𝑀𝑗 𝑅𝐴𝑀, and 𝑉𝑀𝑗 𝐵𝑊, respectively. Similarly, each server 𝑃𝑀𝑖∈ 𝑃𝑀 has 𝐶𝑃𝑈, 𝑅𝐴𝑀, and 𝐵𝑊 capacities denoted as 𝑃𝑀𝑖 𝐶𝑃𝑈 , 𝑃𝑀𝑖 𝑅𝐴𝑀, and 𝑃𝑀𝑖 𝐵𝑊, respectively.

The VM-to-PM assignment is represented by a zero-one assignment matrix X, where the element 𝑥𝑖𝑗 indicates whether 𝑉𝑀𝑗 is assigned to 𝑃𝑀𝑖 . If 𝑉𝑀𝑗 is placed on server 𝑃𝑀𝑖 , then 𝑥𝑖𝑗 = 1; otherwise, 𝑥𝑖𝑗 = 0. The assignment must satisfy the following constraints:

• VM Assignment Constraint
  Each VM must be assigned to exactly one PM:

• Resource Capacity Constraints
  The total demand placed on any PM must not exceed its capacity in any resource dimension r ∈{CPU,RAM,BW}:

These constraints ensure that placement decisions are both exclusive (each VM placed once) and feasible (no host is overloaded). Table 1 summarizes the symbols and notations used throughout the optimization model

3.3. Objective Function

Unlike the ACO-based placement phase, which jointly optimizes energy consumption, network performance, and carbon footprint across the entire datacenter, the PSO-based migration phase operates with a more focused and adaptive goal. It seeks to dynamically refine VM placement by reducing resource fragmentation and consolidating workloads onto fewer active physical machines. While both phases ultimately aim to enhance energy efficiency and overall datacenter utilization, the PSO stage achieves this through a distinct fitness function—one that prioritizes minimizing the number of active servers and the total residual resource wastage across CPU, RAM, and BW.

This change reflects the shift in scope between the static and dynamic phases. Static placement aims for long-term energy efficiency, while dynamic phase must react to workload imbalance and avoid resource fragmentation. ACO targets global, sustainability-aware optimization, while PSO addresses time-sensitive runtime consolidation with minimal disruption.

The PSO optimization problem is expressed in equation (3).

Where:

• α and β are weighting coefficients such that α + β = 1

• 𝑃𝑀𝑖 𝑟 is the capacity of 𝑃𝑀𝑖 for resource r (CPU, RAM, BW)
• 𝑉𝑀𝑗 𝑟 is the demand of 𝑉𝑀𝑗 for resource r
• 𝑥𝑖𝑗 is the binary assignment variable (1 if 𝑉𝑀𝑗 is placed on 𝑃𝑀𝑖 , 0 otherwise)
• The inner sum represents the total unused resources (wastage) per active PM

The first term encourages consolidation by minimizing the number of active servers, while the second penalizes underutilization by measuring residual capacity across resources. Together, they guide the PSO search toward compact and efficient VM placements that preserve feasibility and avoid unnecessary migrations.

Table 1. Symbols and Notations

4 . PROPOSED ALGORITHM (HAPSO – HYBRID ACO–PSO ALGORITHM)

4.1. Hybrid Algorithm Overview

The proposed Hybrid ACO–PSO (HAPSO) algorithm integrates the strengths of two metaheuristic techniques—Ant Colony Optimization (ACO) and Particle Swarm Optimization (PSO)—to address both the initial placement and dynamic migration of VMsin cloud datacenters. This hybridization is designed to enhance resource consolidation, minimize active PMs, and reduce overall resource wastage while maintaining system feasibility and energy efficiency. In the first phase, ACO is applied to perform the initial placement of VMs (Figure 2(a)). We reuse the static placement logic introduced in our previous work [4], where the ACO assigns VMs to PMs based on multi-objective heuristics.

ACO probabilistically constructsaVM-to-PMmappingbyleveragingpheromone trails andheuristic factorssuchasenergyefficiency,carbonemissions,andnetworkimpact. This phase ensures that each VM is assigned to a suitable PM while satisfying resource constraints on CPU, RAM, and BW.

The system then transitions into runtime operation, during which a periodic evaluation is conducted to monitor PM utilization. When a PM exceeds a predefined utilization threshold 𝑇ℎ 𝑜𝑣𝑒𝑟or falls below a lower bound 𝑇ℎ 𝑢𝑛𝑑𝑒𝑟, the PSO-based optimization phase is triggered (Figure 2(b)). This second phase serves as a dynamic refinement mechanism, that selectively migrates VMs hosted by the overutilized or underutilized servers.

The PSO swarm is initialized using the current live VM-to-PM assignment as the first particle. Additional particles are generated by introducing controlled perturbations to this mapping, ensuring diversity without violating feasibility. To ensure feasibility and heuristic quality of initial solutions, each particle’s continuous position was discretized using a constraint-based mapping step. This step integrates domain constraints (CPU, RAM, and BW) early in the search process, enhancing convergence stability and reducing the need for penalty or repair mechanisms. Each particle is evaluated using a multi-objective fitness function that jointly minimizes the number of active servers and the total residual resource wastage across CPU, RAM, and BW.

Once the PSO converges or reaches its iteration limit, the best-performing particle defines a new VM reallocation plan. Only those VMs with changed host assignments are migrated, thereby reducing overhead and maintaining system stability.

The potential of HAPSO to balance exploration and exploitation—via ACO and PSO respectively—will be validated in the results section, where it demonstrates adaptive performance under dynamic cloud workloads

4.2. Particle Representation

Building upon the optimization model defined earlier, each particle in the PSO phase represents a potential reallocation plan for a subset of VMs. Specifically, particles encode VM-to-PM assignments for VMs currently hosted on over- or underloaded PMs identified during runtime monitoring.

The position of each particle is encoded as a binary matrix 𝑋 𝑡 ∈ {0,1} 𝑀×𝑁, where the element 𝑥𝑖𝑗 𝑡 =1 indicates that 𝑉𝑀𝑗 is assigned to 𝑃𝑀𝑖 at iteration 𝑡, and 𝑥𝑖𝑗 𝑡 = 0 otherwise. Each particle must comply with the constraints introduced in Section 3.2, i.e., each VM must be assigned to exactly one host and no PM should exceed its capacity.

Figure 2. Workflow of the proposed hybrid VM placement and migration algorithm. (a) The complete hybrid approach integrating ACO for initial VM placement and PSO for dynamic migration. (b) The PSObased optimization phase, which refines VM allocation by minimizing energy consumption and resource wastage.

For example, 3 PMs (rows) and 4 VMs (columns) might have the following matrix:

4.3. Velocity and Position Updates

Once the PSO swarm is initialized, particles evolve iteratively through updates to their velocity and position vectors. In the context of VM placement, this evolution corresponds to proposing andrefining VM-to-PM migration plans.

Velocity Update

For each particle i, the velocity vector is updated using the standard PSO formulation in Equation (5).

Where:

• 𝑤 is the inertia weight.
• 𝑐1, 𝑐2 are acceleration coefficients guiding the cognitive and social components.
• 𝑟1, 𝑟2 are random numbers in [0, 1] introducing stochasticity.
• 𝑝𝑏𝑒𝑠𝑡𝑖
  is the best-known position (VM-PM mapping) of particle i.
• 𝑔𝑏𝑒𝑠𝑡 is the best-known position among all particles.

This update rule encourages each particle to move toward both its personal best and the global best positions while retaining some influence from its current trajectory.

To balance exploration and exploitation, we use a linearly decreasing inertia weight as shown in Equation (6):

Where:

• 𝜔𝑡 : inertia weight at iteration t.
• 𝜔𝑚𝑎𝑥: initial inertia (0.9).
• 𝜔𝑚𝑖𝑛: final inertia (0.4).
• 𝑇𝑚𝑎𝑥 : total number of iterations.
• 𝑡: current iteration index.

Early iterations favor wide exploration with high 𝜔, while later ones promote convergence by reducing 𝜔. This approach has been shown to improve solution stability in swarm-based optimization [9].

Position Update and Discretization

Following the velocity update, the new position vector is computed as in Equation (7)

Since VM-to-PM assignments must remain binary, and are constrained by physical resource capacities. In the PSO migration phase, every particle starts from the current ACO-derived VM placement and is lightly perturbed to maintain swarm diversity without violating resource limits. After each PSO iteration, we perform a feasibility check to detect and correct any capacity violations. This ensures that every particle’s position always maps to a valid VM-to-PM assignment. This feasibility check is performed also after each velocity update

The swarm follows standard PSO equations with adaptive inertia, and a multi-objective fitness function that jointly minimizes the number of active servers and residual capacity waste steers convergence. Compared with hybrids that rely on randomized initialization [23] or mutation-based particle diversification [29], our controlled perturbation approach preserves feasibility from the outset and accelerates convergence.

4.4. System Parameters

PSO parameters were adopted from prior works ([8], [30], [31]) with minor empirical adjustments for convergence [4]. A full summary of the HAPSO parameters appears in Table 2.

Table 2/strong>. System Parameters