Time-series Specialized Neural Networks for Detailed Changepoint Predictions in Cloud Datacenters

Abstract

Cloud computing data centers play a crucial role in providing the necessary resources for computationally demanding fields and applications, allowing for dynamic need-based resource scaling. However, these massive systems often exhibit unpredictable changes in performance. While much of this variability is noise, consistent changes in mean performance, referred to as changepoints, pose challenges for both cloud providers and users. Accurate projection and management of capacity, usage, and performance are essential for reliable service and resource provision and optimization. In this paper, we investigate the feasibility of forecasting CPU performance and detecting changepoints in cloud computing using time-series forecasting methods. This thesis aims to contribute to the understanding of performance variability and provide practical solutions for cloud providers and users. We will test whether these models can identify latent features associated with performance changes to improve the forecasting accuracy.