Using Machine Learning for Accelerating the Detection Time of Unreliable Failure Detectors

Abstract

This study evaluates Machine Learning-based Failure Detectors for detecting faulty nodes in distributed systems, focusing on metrics like Detection Time and Average Mistake Rate. Random Forest and Support Vector Machine Failure Detectors emerge as top performers. However, Machine Learning-based Failure Detectors exhibit higher error rates, especially with longer detection times. To address this, three ensemble methods are proposed, with one integrating top Machine Learningbased Failure Detectors and analytical Failure Detectors for better balance. The Precision Distribution module improves accuracy but struggles with completeness. One analytical Failure Detector is used as a fallback to address completeness concerns, although its effectiveness depends on the relationship between the analytical Failure Detector and Precision Distribution module thresholds. Additionally, the Li-Marin Long Short-Term Memory Failure Detector shows reduced detection time but with increased computational overhead, posing challenges in fine-tuning overestimation levels.