Introduction
Industry 4.0 is driving the proliferation of the Internet of Drones (IoD), which supports critical applications in logistics, surveillance, and emergency response. However, drones are inherently resource-constrained, facing limitations in battery capacity, processing power, and communication range. To address these constraints while maintaining scalability, Network Virtualization (NV) is employed to decouple logical services from physical infrastructure.
The central challenge in this environment is the Virtual Network Embedding (VNE) problem: the efficient mapping of virtual nodes and links onto a multi-domain physical substrate. This problem is NP-hard and particularly complex when virtual requests must span multiple administrative or geographical domains. Existing algorithms often struggle with high computational latency or get stuck in local optima, failing to find the most efficient global solution.
Multi-Domain Architecture
The proposed solution, Multi-domain Particle Swarm Optimization VNE (MP-VNE), operates within a hierarchical framework. A global controller coordinates with multiple local domain controllers to manage resources across the substrate.
As shown in Figure 1, the physical network is partitioned into distinct domains connected by inter-domain links. This structure allows for localized resource management while supporting global connectivity. To maintain domain privacy and reduce control overhead, local controllers generate "pseudo-topologies"—abstracted representations of their internal resources—which are then shared with the global controller for cross-domain decision-making.
The MP-VNE Algorithm
The MP-VNE algorithm integrates several advanced techniques to optimize resource allocation. The core of the approach is a modified Particle Swarm Optimization (PSO) algorithm. While standard PSO is effective at searching large solution spaces, it often converges prematurely to local optima. To mitigate this, the authors introduce a genetic mutation factor, allowing particles to occasionally relocate to random positions, thereby ensuring a more exhaustive search for the global optimum.
The execution flow, detailed in Figure 2, begins with the division of virtual network requests into subgraphs. These subgraphs are distributed to physical domains where local controllers identify candidate nodes. A critical step is the calculation of an Estimated Mapping Cost (PreCost), which evaluates CPU requirements and potential link costs to boundary nodes. This pre-selection phase prunes inefficient candidates early, significantly reducing the computational burden on the PSO engine. Furthermore, the algorithm employs a weighted multi-objective optimization function to balance mapping costs and delays, transforming a complex multi-objective problem into a more manageable single-objective convergence task.
Experimental Evaluation
The performance of MP-VNE was evaluated in a simulated environment consisting of four domains, each containing 30 nodes with resources distributed uniformly. Virtual requests were generated via a Poisson process, with an average of six nodes per request. The simulation results demonstrate that MP-VNE significantly outperforms benchmark algorithms such as MC-VNM, LID-VNE, and standard VNE-PSO:
- Acceptance Rate: MP-VNE maintained a stable request acceptance rate of approximately 60%, whereas competing methods saw rates drop toward 30% under sustained load.
- Mapping Cost: The average mapping cost for MP-VNE was approximately 750 units, compared to over 1100 units for standard VNE-PSO.
- Latency: MP-VNE achieved the lowest average mapping delay at 460ms, nearly half the delay observed in the MC-VNM algorithm (800ms).
Conclusion and Future Work
The MP-VNE algorithm provides a robust framework for resource allocation in multi-domain IoD environments. By combining hierarchical control, pseudo-topology abstraction, and a mutation-enhanced PSO, it achieves superior efficiency in terms of cost, delay, and acceptance rates.
Future research will focus on moving beyond weighted summation toward Pareto-optimal frontiers to better handle nuanced trade-offs between conflicting objectives. Additionally, integrating machine learning models for predictive mapping and testing the algorithm on larger, more heterogeneous networks will be essential for real-world Industry 4.0 deployments.
Summary updated