Enabling Real-Time Adaptive Cloud Provisioning Via Hybrid Attention Similarity Graph Neural Networks

Authors

  • Nandhini Ponnusamy Author
  • Narendra B Mustare Author
  • Tarak Hussain Author
  • Arun Prasad D Author
  • Hari Jyothula Author
  • Kishore Kumar K Author
  • Johny Renoald Albert Author

Keywords:

Cloud Computing, Workload Prediction, Graph Neural Networks, Resource Provisioning, Metaheuristic Optimization, Autonomic Systems.

Abstract

In order to efficiently provision cloud resources, it is essential to have accurate workload predictions. However, most existing methods rely on fixed thresholding rules or deep learning (DL) models, which poorly model the complex spatiotemporal interactions between virtual machines, have trouble optimising their hyperparameters, and have high rates of service level violations (SLAs) when faced with bursty loads. The current study proposes a novel framework for proactive management of cloud resources called FOX-CA-SGSTGLF, which is designed to address these limitations. The architecture integrates autonomic decision-making, similarity-based graph modelling, robust data preprocessing, and continuous surveillance into a unified pipeline. Combining the Cell Attention-Based Similarity-Navigated Graph Neural Network (CA-SNGNN) with the FOX optimisation algorithm (FOA) to learn spatial correlations between virtual machines (VMs) and the dynamics of their workload over time, the network converges in 42 iterations and achieves better generalisation rates. Results from large-scale testing using actual PlanetLab traces demonstrate that the prediction accuracy is high and outperforms state-of-the-art algorithms with MAE values of 0.015, 0.031, 0.058, and 0.085 for 5-, 15-, 30-, and 60-minute horizons, respectively. Utilising the framework in an autonomic MAPE loop resulted in a 32.4% decrease in energy and resource costs, a 1.9-fold decrease in SLA violations, and an average CPU utilisation of 63.7%. The results show that FOX-CA-SGSTGLF is able to provide cloud resources accurately and at scale for diverse workloads that are constantly changing.

Downloads

Download data is not yet available.

Downloads

Published

2026-09-22

How to Cite

Enabling Real-Time Adaptive Cloud Provisioning Via Hybrid Attention Similarity Graph Neural Networks. (2026). Minimax Theory and Its Applications, 11(2s), 928-952. https://journalmta.com/index.php/jmta/article/view/463