A novel GRU-Attention based deep learning architecture for the prediction of idle workloads

-, Anisha Khalam (2025) A novel GRU-Attention based deep learning architecture for the prediction of idle workloads. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (825kB) | Preview

Abstract

Cloud providers face escalating operational costs as data centres expand, with substantial fractions of virtual machines remaining idle, yet consuming energy and resources. Proactive VM consolidation, predicting future idle states to enable advance migration or shutdown, offers cost savings but requires proactive forecasting rather than reactive monitoring. Existing research exhibits gaps: models train exclusively on single cloud platforms without cross-cloud validation, continuous regression outputs require post-processing rather than actionable predictions, severe class imbalance (80%+ idle) remains not addressed causing minority class failure, large architectures constrain resource-limited deployment, and single-architecture proposals lack systematic multi-cloud comparisons. This thesis proposes a GRU-Attention architecture with a triple class imbalance strategy (Focal Loss, class weighting, threshold optimisation) for one-hour ahead binary VM state forecasting (IDLE/ACTIVE) which consist of two stacked GRU layers (64/32 units) with decomposed single-head attention mechanism achieving selective temporal focus. Evaluation against four alternative architectures on Microsoft Azure Public Dataset demonstrates that GRU-Attention achieves an accuracy of 93.65% , f1 macro of 90.89% using 24,162 parameters. The triple imbalance strategy improves the recall of the minority by 20.14 percentage points (78.45% to 98.59%). Cross-cloud deployment to AWS EC2 achieves 83.1% average confidence with successful state transition detection, representing the first empirical validation in the literature of cloud-agnostic model transfer without retraining. Organisations can deploy proactive consolidation systems that provide 60-minute advance warning, maintain operational safety, which represents a reasonable trade-off for cost reduction. This work establishes GRU-Attention as a production-ready architecture for cloud work-load forecasting, addressing all identified gaps by providing empirical evidence that cloud agnostic temporal patterns transferable across heterogenous infrastructure can be learned by deep learning models.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Estrada, Giovani UNSPECIFIED
Subjects:	T Technology > T Technology (General) > Information Technology > Cloud computing Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Cloud Computing
Depositing User:	Ciara O'Brien
Date Deposited:	31 Aug 2026 11:46
Last Modified:	31 Aug 2026 11:46
URI:	https://norma.ncirl.ie/id/eprint/9686

Actions (login required)

View Item