Failure Prediction in Cloud Native 5G Core With eBPF-based Observability

Today, the advances in virtualization technologies are dramatically changing network architecture as well as operation style. In particular, cloud-native network functions (CNFs) provide scalability, flexibility, and lightness to network services, which can be leveraged to satisfy vehicle application requirements in the 5G and beyond 5G era. However, their microservice architecture causes a new challenge in the operation of cloud-native networks due to the larger number of compositions and dynamic topology changes. To address this situation, artificial intelligence (AI)/machine learning (ML) technologies are expected to support operational tasks. One of these tasks is failure prediction that is a key enabler for proactive network operation to minimize the failure impact. For reliable failure prediction, not only ML methods but also data used for ML training should be prepared with a fine granularity. In this paper, we utilize the extended berkeley packet filter (eBPF) to collect fine-grained information from the virtual infrastructure where CNFs are deployed. By training a long short-term memory (LSTM) model on the collected data, we develop the failure prediction model that can provide the future transition of key performance indicators (KPIs) in a cloud-native 5G core network (5GC). Our experiment on the lab environment shows that the prediction model trained on eBPF data outperforms other models trained without them.