Year: 2025
Type: conference-paper
Abstract: Cloud computing has become a commonly adopted solution to increase business operations efficiency and scalability. However, costs related to cloud infrastructure are increasingly prohibitive for some organizations in long-term deployment scenarios. This expense barrier has led to the emergence of a cloud repatriation trend, where existing software systems are moved back to on-premises environments. Although this initiative enhances cost control management and autonomy, it reintroduces technical challenges. This work investigates, through an empirical study, the applicability of Long Short-Term Memory (LSTM) networks for workload prediction to support auto-scaling decisions in on-premises environments. Initially, we contrasted the prediction accuracy of LSTM with classical methods, utilizing both standard benchmark datasets and access logs collected from an industrial-grade software system. In the second phase, we examined how varying temporal resolutions affect precision and computational expense. Results show the LSTM consistently outperforms classical methods, reducing MAPE by up to 18.02% and RMSE by 8.77% on benchmark datasets. Considering the in-field dataset, it outperforms in long-term predictions, especially regarding MAE and MAPE. Throughout all analyzed temporal resolutions, the 10-minute resolution optimally balanced accuracy with efficiency. Moreover,we noted that LSTM requires less training time than classical methods, emphasizing its applicability for real-world scenarios. Therefore, the results indicate that LSTM emerges as a viable solution to enable proactive auto-scaling within on-premises environments, potentially reducing costs in the context of cloud repatriation.
Authors: Rafael Xavier, Bruno Cafeo, Baldoino Fonseca, Davy Baia, Elder Cirilo
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7085490632