Sahu Dinesh, Chaturvedi Rajnish, Prakash Shiv, Yang Tiansheng, Rathore Rajkumar Singh, Wang Lu, Tahir Sabeen, Bakhsh Sheikh Tahir
SCSET, Bennett University, Plot Nos 8, 11, TechZone 2, Greater Noida, Uttar Pradesh, 201310, India.
Department of Electronics and Communication, University of Allahabad, Prayag Raj, Uttar Pradesh, India.
Sci Rep. 2025 Jan 29;15(1):3730. doi: 10.1038/s41598-025-88197-9.
In general, edge computing networks are based on a distributed computing environment and hence, present some difficulties to obtain an appropriate load balancing, especially under dynamic workload and limited resources. The conventional approaches of Load balancing like Round-Robin and Threshold-based load balancing fails in scalability and flexibility issues when applied to highly variable edge environments. To solve the problem of how to achieve steady-state load balance and provide dynamic adaption to edge networks, this paper proposes a new framework that using PCA and MDP. Taking advantage of the stochasticity of PCA classification our model describes interactions between neighboring nodes in terms of a local load thus allowing for a distributed, self-organizing approach to load balancing. The MDP framework then determines each node's decision-making with the focus on load offloading policies that are aligned with rewards that promote per node balance and penalties for offloading a larger load than it can handle.These models are then incorporated into our proposed PCA-MDP system to achieve dynamic load balancing with low variability in resource usage among nodes. By conducting a large number of experiments, we prove that the proposed PCA-MDP model yields a higher efficiency in the distribution of the load, higher stabilities of the reward function, and a faster convergence speed compared to the existing approaches. Key performance parameters, such as load variance, convergence time, and scalability, validate the robustness of the proposed model. Besides optimizing resource exploitation, load harmony in edge computing networks helps provide efficient work progression and minimize latency, thereby contributing to the advancement of the field with respect to real-time applications such as self-driving vehicles and the Internet of Things. The presented work offers an excellent foundation for the next-generation edge-computing load-balancing solution that can be easily scaled up.
一般来说，边缘计算网络基于分布式计算环境，因此在获得适当的负载平衡方面存在一些困难，尤其是在动态工作负载和资源有限的情况下。传统的负载平衡方法，如轮询和基于阈值的负载平衡，在应用于高度可变的边缘环境时，在可扩展性和灵活性方面存在不足。为了解决如何实现边缘网络的稳态负载平衡并提供动态适应性的问题，本文提出了一种使用主成分分析（PCA）和马尔可夫决策过程（MDP）的新框架。利用PCA分类的随机性，我们的模型根据局部负载描述相邻节点之间的交互，从而允许采用分布式、自组织的负载平衡方法。然后，MDP框架确定每个节点的决策，重点是负载卸载策略，这些策略与促进每个节点平衡的奖励以及卸载超过其处理能力的更大负载的惩罚相一致。然后将这些模型纳入我们提出的PCA-MDP系统，以实现节点间资源使用低变化性的动态负载平衡。通过进行大量实验，我们证明，与现有方法相比，所提出的PCA-MDP模型在负载分配方面具有更高的效率、奖励函数具有更高的稳定性以及更快的收敛速度。关键性能参数，如负载方差、收敛时间和可扩展性，验证了所提出模型的鲁棒性。除了优化资源利用外，边缘计算网络中的负载协调有助于提供高效的工作进程并最小化延迟，从而有助于推动自动驾驶车辆和物联网等实时应用领域的发展。所展示的工作为可轻松扩展的下一代边缘计算负载平衡解决方案提供了良好的基础。
