Scholarlink

Data centers have emerged as significant energy consumers, considerably contributing to global carbon emissions due to their intensive operation and computational demands. This paper presents a comparative analysis of two prominent machine learningMachine learning architectures—Long Short-Term Memory (LSTM)Long Short-Term Memory (LSTM) and Graph Neural Networks (GNN)Graph Neural Networks (GNN)—for predicting virtual machine (VM) power consumption, directly addressing data center carbon footprintCarbon footprint tracking. Employing a robust experimental setup involving VM operational metrics such as CPU utilization, memory usage, network traffic, disk I/O, and execution time, we evaluated both architectures using standard metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), and the coefficient of determination (R2). Results indicate that the LSTMLong Short-Term Memory (LSTM) architecture achieved stable predictive performance, demonstrating reliability in handling sequential data, while GNNGraph Neural Networks (GNN), despite initially higher error rates, exhibited substantial potential for effectively capturing relational and structural data complexities intrinsic to data center operations. The comparative findings underscore the critical importance of model selection aligned with data characteristics. The findings also highlight future research directions, including hybrid modeling and advanced graph feature engineering to enhance the sustainabilitySustainability and efficiency of data center operations.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative Analysis of LSTM and GNN for Predicting Power Consumption and Carbon Footprints in Data Center Virtual Machines

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Morris Kaburu,
  • Patrick Gikunda,
  • Juliet Moso

摘要

Data centers have emerged as significant energy consumers, considerably contributing to global carbon emissions due to their intensive operation and computational demands. This paper presents a comparative analysis of two prominent machine learningMachine learning architectures—Long Short-Term Memory (LSTM)Long Short-Term Memory (LSTM) and Graph Neural Networks (GNN)Graph Neural Networks (GNN)—for predicting virtual machine (VM) power consumption, directly addressing data center carbon footprintCarbon footprint tracking. Employing a robust experimental setup involving VM operational metrics such as CPU utilization, memory usage, network traffic, disk I/O, and execution time, we evaluated both architectures using standard metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), and the coefficient of determination (R2). Results indicate that the LSTMLong Short-Term Memory (LSTM) architecture achieved stable predictive performance, demonstrating reliability in handling sequential data, while GNNGraph Neural Networks (GNN), despite initially higher error rates, exhibited substantial potential for effectively capturing relational and structural data complexities intrinsic to data center operations. The comparative findings underscore the critical importance of model selection aligned with data characteristics. The findings also highlight future research directions, including hybrid modeling and advanced graph feature engineering to enhance the sustainabilitySustainability and efficiency of data center operations.