Detecting Network Partitioning in Cloud Native 5G Mobile Network Applications

Sammanfattning

With the transition of the 5G core network to a cloud native service-based architecture—composed of network functions operating through microservices communicating over the network—there is an increased risk of network failures causing service downtime unrelated to the applications themselves. In particular, cases of partial and simplex network partitionings have been observed in production systems to produce silent failures causing severe symptoms. Thus, diagnosing these failures have proven difficult. As such, the need of monitoring the network between microservices is of particular interest. In this thesis, we devise a distributed monitoring scheme to identify and classify network partitionings in a Kubernetes cluster. We implement and evaluate two approaches of this scheme based on both active and passive monitoring. While both approaches are feasible for our purpose, we conclude that our approach to passive monitoring struggles with classifying simplex partitions due to TCP being a two-way protocol. Similarly, operating the passive mode requires privileges not necessarily suitable for a shared cloud environment. While the active monitoring scheme is able to infer all types of partitions, it will—unlike the passive alternative—increase the overall load on the network. We further present how to make our proof-of-concept implementation scalable when deployed in larger clusters.