A Virtual Network Resource Allocation Model Based on Dynamic Resource Pricing

Xian-cui Xiao, Xiang-Wei Zheng, Yi Wei, Xinchun CuiIEEE Access 2020 · DOI

Abstract

Traditional Virtual Network Mapping (VNM) often relies on static resource allocation based on peak demand, leading to significant inefficiencies during off-peak periods. This paper introduces GSO-RBFDM, a novel framework that combines Group Search Optimization (GSO), Genetic Algorithms (GA), and Radial Basis Function (RBF) neural networks to enable dynamic, market-driven resource allocation. By predicting user demand and adjusting prices in real-time, the model enhances network efficiency, reduces operational costs, and maximizes provider revenue.

Introduction

In the paradigm of network virtualization, Virtual Network Mapping (VNM) is the critical process of embedding virtual nodes and links onto a physical substrate infrastructure. This task is categorized as an NP-hard problem due to the complex constraints of computing power and bandwidth. Historically, VNM has been approached through static allocation, where resources are reserved based on a user's anticipated peak usage.

Representation of a substrate network with resource allocation ratios.

As illustrated in Figure 3, substrate networks consist of nodes and links with finite capacities. Static mapping fails to account for the inherent volatility of internet traffic, resulting in "resource strandedness" during low-traffic hours. To address this, a more responsive, dynamic allocation model is required.

The GSO-RBFDM Framework

The proposed GSO-RBFDM model utilizes a multi-stage approach to optimize the mapping process:

Node Mapping via Group Search Optimization (GSO)

The framework employs GSO, a metaheuristic algorithm inspired by the foraging behavior of animals, to identify the most cost-effective physical nodes for hosting virtual ones. This approach allows for a more global search of the solution space compared to simple greedy heuristics.

Path Mapping via Genetic Algorithm (GA)

While the Dijkstra algorithm is commonly used for shortest-path routing, it often falls short when link costs are dynamic and nonlinear. The authors implement a GA to evolve and select paths that minimize total cost, accounting for both link bandwidth and the "buffer cost" associated with node congestion.

Demand Prediction via RBF Neural Networks

To transition from static to dynamic allocation, the model integrates a Radial Basis Function (RBF) neural network. This supervised learning component is trained on historical data to predict real-time resource requirements, allowing the system to dynamically adjust allocations.

Dynamic Nonlinear Pricing

The model introduces a "D-asymmetry" pricing mechanism. In this system, the unit price of a resource increases as the total occupancy of the physical infrastructure rises. This creates a market-driven incentive for users to optimize consumption and helps the provider manage congestion by pushing new requests toward underutilized nodes.

Demand Modeling and Simulation

The effectiveness of the GSO-RBFDM model was tested using synthetic data generated via the GT-ITM tool. A substrate network of 100 nodes and 950 links was simulated. Crucially, the authors modeled user demand based on real-world data from the "45th Statistical Report on China’s Internet Development."

Time-distribution curves for various mobile internet services used for demand modeling.

As shown in Figure 5, different internet services exhibit distinct temporal peaks—such as online delivery at midday and short video consumption in the evening. By training the RBF network on these patterns, the GSO-RBFDM model can anticipate shifts in demand and reallocate resources accordingly.

Performance Evaluation

The GSO-RBFDM model was benchmarked against several baseline algorithms, including static (STAC) and random configurations. The results demonstrated several key advantages:

  • Improved Acceptance Rate: By managing resource "inventory" more fluidly, the model accepted a higher volume of virtual network requests.
  • Cost Efficiency: Real-time adjustment to actual demand significantly lowered the average cost of maintaining mappings compared to peak-based static models.
  • Revenue Maximization: The combination of dynamic pricing and efficient resource reclamation led to the highest average revenue for infrastructure providers.

Conclusion

The GSO-RBFDM model represents a significant step toward more efficient and profitable virtual network management. By treating resource allocation as a dynamic, market-driven process rather than a static assignment, providers can better handle the "peaks and valleys" of modern internet traffic. Future research will focus on improving the interpretability of the RBF predictions and refining the pricing models to better reflect complex real-world Service Level Agreements (SLAs).

Summary updated

BibTeX
@article{xiao2020virtual,
  author = {Xian-cui Xiao and Xiang-Wei Zheng and Yi Wei and Xinchun Cui},
  title = {{A Virtual Network Resource Allocation Model Based on Dynamic Resource Pricing}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.3020944},
}
Figure