ANALISIS KOMPARATIF EFISIENSI KOMPUTASI MODEL XGBOOST PADA INFRASTRUKTUR DOCKER CONTAINER DAN NATIVE OS UNTUK MENDUKUNG KEANDALAN LAYANAN PEMBELAJARAN DIGITAL BERBASIS KECERDASAN BUATAN

Abstract

The reliability of AI-based digital learning services depends heavily on the appropriate selection of machine learning model deployment infrastructure. Educational institutions commonly face computational resource constraints, choosing between a native OS and Docker Container, a critical factor affecting the responsiveness and availability of digital services that underpin academic activities. This study empirically compares the computational efficiency (CPU and RAM) and accuracy of the Extreme Gradient Boosting (XGBoost) algorithm in these two execution environments, using the CIC-IDS 2017 dataset from the Canadian Institute for Cybersecurity, comprising approximately 2.8 million records across 15 classes as a representative large-scale computational workload. A quantitative experimental method with real-time resource monitoring was applied to identical hardware. Results show that the Native OS environment achieved an average end-to-end (E2E) execution time of 209.72 seconds, approximately 9.4 times faster than Docker Container, which required an average of 1,965.83 seconds due to a 2-core CPU restriction. Native OS averaged 90.60% CPU and 403.01 MB RAM, while Docker averaged 24.41% CPU and 683.76 MB RAM with a +280.75 MB memory overhead. Critically, all model quality metrics Accuracy, Precision, Recall, and F1-Score were identical at 99.89% in both environments, confirming that the execution environment does not affect AI model quality. This study concludes that to support the reliability of AI-based digital learning services requiring real-time responses, using a native OS or providing unrestricted resource allocation in a Docker Container is the recommended deployment architecture for educational institutions