Evaluation on eBPF-Based Network Failure Prediction Using AutoGluon

Abstract

This study evaluates an extended Berkeley Packet Filter (eBPF)-based network failure prediction method using Autogluon-Tabular to process the fine-grained network information extracted by eBPF. The extracted information is considered as input features of the proposed model, which aims to predict the subsequent packet loss and determine a network failure event before it causes a huge impact. Supervised learning and semi-supervised learning are both adopted in Autogluon. The accuracy and detection time are evaluated as the main criteria. Simulation results show that F1 scores exceed 0.9 for our proposed method, and the proposed method can achieve prediction for potential failure events within 30 and 40 seconds when symptoms such as packet loss occur.