Modelling uncertainty in cloud workload forecasting: a Hybrid Bayesian Neural Network approach
Loading...
Files
Full Text E-thesis
Date
2024
Authors
Rossi, Andrea
Journal Title
Journal ISSN
Volume Title
Publisher
University College Cork
Published Version
Abstract
The exponential growth of cloud services utilisation has rendered robust and accurate workload forecasting paramount. Predicting future computational demands is essential for efficient resource management, cost optimisation, and guaranteeing service level agreement (SLA). At the same time, the increase in cloud resource demand poses new challenges related to data centres' carbon emissions and their environmental impact. However, these predictions are characterised by inherent uncertainty stemming from both fundamental limitations in data and the dynamic nature of cloud usage.
This dissertation delves into workload forecasting in cloud computing. Our primary focus is quantifying predictions' inherent uncertainty, an important but often overlooked aspect. Traditionally, forecasting models provide point estimates, leaving cloud providers blind to the potential range of future demands. This lack of uncertainty awareness can lead to suboptimal resource allocation, unnecessary costs, and SLA violations.
In particular, our work focuses on quantifying the uncertainty of different natures. The epistemic uncertainty arises from limitations in our knowledge or model, while the aleatoric uncertainty stems from the variability in the data itself, reflecting the inherent randomness of cloud workloads. By capturing both these forms of uncertainty, we gain a deeper understanding of the prediction's confidence interval, empowering informed decision-making.
This work introduces a novel Hybrid Bayesian Neural Network (HBNN) model designed to capture both types of uncertainty in workload forecasting by incorporating a Bayesian layer at the network's end. The HBNN surpasses conventional approaches by estimating the future probability distribution to compute the associated confidence intervals. This quantifiable uncertainty empowers the model for more accurate prediction.
The thesis expands upon the HBNN model by exploring its application in bivariate forecasting, simultaneously predicting processing units and memory demands. This approach proves good accuracy at the cost of more training data required. Furthermore, this work investigates the impact of various factors on forecast accuracy, including training data size, unseen data generalisation, and the potential of transfer learning across different cloud environments.
Furthermore, recognising the potential computational hurdles associated with Bayesian neural networks, this work introduces a clustering-based preprocessing technique. This technique intelligently selects training data points, significantly reducing computational cost while maintaining forecast accuracy. This development enables the HBNN model to be deployed in cloud environments at scale.
The thesis comprehensively evaluates the HBNN model through extensive experiments on real-world cloud workload datasets. The results demonstrate the model's superior performance in quantifying uncertainty, achieving higher accuracy than baseline models and showcasing improved generalisation on unseen data. By quantifying inherent uncertainty, HBNN empowers cloud providers to manage cloud resources more efficiently, reliably and cost-effectively.
Description
Keywords
Time series forecasting , Cloud computing , Uncertainty quantification , Deep learning , Time series snalysis , Deep learning , Workload prediction , Workload forecasting
Citation
Rossi, A. 2024. Modelling uncertainty in cloud workload forecasting: a Hybrid Bayesian Neural Network approach. PhD Thesis, University College Cork.