AI in telecom research: Identifying bottlenecks in cloudified mobile networks
One of the main goals of monitoring self-driving networks is to quickly identify and fix any faults with minimal operating expenses. Here, we present our advanced monitoring solution to do just that.
A cloudified mobile network is expected to deliver a multitude of services, on different slices, with reduced capital and operating expenses.
When it comes to 5G mobile systems, it is important to ensure that the requirements of a customized end-to-end service are met. It’s crucial to monitor not only the network itself, but also the performance at its radio interfaces and user equipment. With millions of user equipments, a central monitoring architecture is not feasible.
A framework for advanced monitoring
CRNA proposes advanced monitoring solutions to identify any performance degradation in the network, while attempting to balance high detection rates with minimal monitoring overhead costs. A 2-stage distributed telemetry framework to identify and attribute bottlenecks in a cloudified mobile network and its last mile. This system monitors both the mobile network and its user equipment.
Monitoring user equipment assists in triggering the identification of bottleneck events that impact the user’s experience. The inclusion of user equipment not only relieves the mobile network from continuous computation of data analytics, but also helps the monitoring system catch a bottleneck that is beyond the internal scope of the mobile network, such as radio interference. Mild bottlenecks at the mobile network, however, may go unnoticed by a user's equipment when they do not degrade its application performance.
To assess our 2-stage distributed framework, we have built a mobile network testbed based on the cloudified architecture.
High accuracy in detection, location, and cause of bottlenecks
By leveraging measurements in user equipment, our variational autoencoder-based model accurately detects different types of bottlenecks with 85 percent accuracy. To attribute the cause and location of the bottlenecks, our classification model achieves 89 percent accuracy. Overall, the bottleneck identification accuracy of our distributed framework is comparable to that of a centralized approach, making it a better choice due to the non-feasibility of a centralized system.
Working with a combination of monitoring approaches, our study further reveals that in-band network telemetry can be the potential future alternative for active monitoring of mobile network links. In this study, we have provided a proof-of-concept of our distributed telemetry framework using generic application traffic in the user equipment, virtual P4 (Programming Protocol Independent Packet Processors) switches, and a cloudified mobile core based on 4G virtual network functions.
Next step: real applications and online adaptation
In the future, we plan to focus on real-world applications of our framework. We will upgrade our testbed with 5G virtual network functions, realizing a 5G standalone core and hardware P4 switches. This will allow us to evaluate our framework on use cases where the user equipment is using novel applications with very high requirements, such as live broadcasting and networked music over 5G networks. Finally, we will explore the feasibility of applying the framework online.
For more information, see our paper: Fida, M. R., Ahmed, A. H., Dreibholz, T., Ocampo, A. F., Elmokashfi, A., & Michelinakis, F. I. (2023). Bottleneck identification in cloudified mobile networks based on distributed telemetry. IEEE Transactions on Mobile Computing.
(This blog post is written by Postdoctoral Fellow Azza Ahmed, Research Professor Haakon Bryhni and Maria Normann).