Abstract

Cloud computing is one of the major paradigms in the information technology industry, offering diverse scalable on-demand services over the Internet. Nevertheless, managing and predicting workloads in cloud data centers is a challenging task due to the dynamic nature of cloud services. In order to reduce costs and improve performance while managing cloud resources efficiently, it is essential to obtain highly accurate projections and estimations. Therefore, this thesis proposes a methodological framework that integrates multiple machine learning models to improve estimation accuracy and enable better decision-making within cloud data centers. In terms of clustering, we develop segmentation pipelines that incorporate various clustering techniques with different data preprocessing methods to improve the cloud workload segmentation process. This process aims to reveal hidden patterns within workloads to obtain segmentation based on various data-driven perspectives. In predictive modeling, we delve into the enhancement of prediction precision, focusing on single-output and multi-output forecasting models. For single-output-based prediction, we propose a multilevel learning-based model for resource utilization prediction that leverages anomaly, clustering, and ensemble methods to improve prediction outcomes. Also, we present a proactive regression-based cost estimation approach, navigating the complexities of prediction-based cloud service pricing and the effect of various target transformation methods on prediction accuracy. In addition, we propose a host load prediction, leveraging both imbalance and ensemble learning methods to improve prediction and handle the challenge of the imbalance states within cloud computing systems. For multi-output-based prediction, advanced predictive models are proposed to forecast function invocation patterns at the user, application, and function levels within serverless computing environments. In this thesis, we conducted research that utilizes advanced data analysis techniques, including windowing, dimensionality reduction, and ensemble learning, to enhance the robustness and precision of workload segmentation and predictive models within cloud environments. We evaluated the proposed models based on their efficiency in processing real cloud workloads using various performance metrics. The findings of this thesis hold the potential to revolutionize cloud resource management, leading to more intelligent, adaptable, and cost-effective cloud operations.