Causal-Temporal Analysis-Based Feature Selection for Predicting Application Performance Degradation in Edge Clouds

Abstract

Next-generation networks will enable applications that are expected to be highly reliable, always available, with guaranteed Quality-of-Service (QoS). Distributed, heterogeneous edge clouds are key enablers for these applications. However, applications deployed in such networks may suffer from performance degradation caused by various infrastructure-related faults. Preventing performance degradations by predicting them using Machine Learning (ML)-based analytics and handling them proactively is thus critical for maintaining the application QoS. Predicting performance degradations can be challenging due to the diversity of the underlying causes. In this paper, we propose an automated feature selection system that uses causal-temporal analysis to find the infrastructure metrics that have causal relationships with application metrics. The selected features are further used to train ML models for predicting application performance degradation. We have validated a proof-of-concept of our system on a Kubernetes testbed, where it is demonstrated that the proposed feature selection system is up to 17.9 times faster than Recursive Feature Elimination (RFE). Using the features selected by the proposed system, the ML models could predict performance degradation with up to 12.7% higher F1-score compared to using historical values of application Key Performance Indicator (KPI) to forecast its future values.