Characterization and Identification of Cloudified Mobile Network Performance Bottlenecks

Georgios Patounas, Xenofon Foukas, Ahmed Elmokashfi, Mahesh K. Marina

Abstract

This study is a first attempt to experimentally explore the range of performance bottlenecks that 5G mobile networks can experience. To this end, we leverage a wide range of measurements obtained with a prototype testbed that captures the key aspects of a cloudified mobile network. We investigate the relevance of the metrics and a number of approaches to accurately and efficiently identify bottlenecks across the different locations of the network and layers of the system architecture. Our findings validate the complexity of this task in the multi-layered architecture and highlight the need for novel monitoring approaches that intelligently fuse metrics across network layers and functions. In particular, we find that distributed analytics performs reasonably well both in terms of bottleneck identification accuracy and incurred computational and communication overhead.

Introduction

The transition toward 5G involves a fundamental shift from specialized hardware to "cloudified" architectures. This evolution leverages virtualized network functions (VNFs) running on shared cloud infrastructure, offering flexibility through network slicing. However, this shift introduces a significant observability gap. In traditional networks, performance drops often had clear hardware-linked causes. In a cloudified environment, a single symptom—such as a drop in throughput—could stem from radio interference, link congestion, CPU starvation in a VNF, or container misconfiguration. Because these layers are abstracted and interconnected, different root causes often produce identical "signatures" at the service level.

The Multi-Layered Observability Gap

To address this complexity, the study identifies three distinct layers within the system architecture: the Service Layer, the Network Function (NF) Layer, and the Infrastructure Layer.

The three-layered architecture of a cloudified mobile network, illustrating the Service, Network Function, and Infrastructure layers alongside Management and Orchestration (MANO).

A primary finding of this research is "Service Layer Blindness." Metrics such as throughput and Round Trip Time (RTT), while effective at signaling that a problem exists, are insufficient for root-cause analysis. The study demonstrates that cross-layer data fusion is mandatory for accurate identification; neither NF data nor infrastructure data is sufficient on its own.

Experimental Methodology and Testbed

The researchers developed a comprehensive prototype testbed to capture the nuances of a cloudified mobile network. The setup utilizes OpenAirInterface (OAI) for the core network and Orion for RAN slicing, all deployed within an OpenStack cloud environment.

The prototype testbed setup, showing the integration of OpenStack hosts for Mobile Core and RAN VNFs, connected via Gigabit switches.

The study employed a data-driven approach to bottleneck identification:

  1. Bottleneck Induction: Five categories of degradation were injected: Interference, Packet Loss, Link Congestion, Computational Resource Stress (CPU/Memory), and Delay.
  2. Multi-Layer Telemetry: 52 Key Performance Indicators (KPIs) were collected across the three architectural layers.
  3. Unsupervised Learning: Hierarchical Clustering with the Soergel/Tanimoto distance metric was used to group similar bottleneck signatures without prior labeling.
  4. Analytics Comparison: The study compared centralized "global" analysis against a distributed approach where the Radio Access Network (RAN) and Core Network (CN) analyze local data independently.

Key Results and Insights

The experimental results underscore the necessity of cross-layer data fusion. When using only service-layer metrics, the clustering "purity"—a measure of how accurately the machine learning model separates different bottlenecks—was only 0.7. However, by combining NF and Infrastructure layer data, purity increased to 0.94.

Furthermore, the distributed analytics approach outperformed centralized analysis, achieving perfect purity (1.0) in identifying bottleneck types within the RAN and Core domains. This success is attributed to the reduction of "noise" from unrelated network segments during local analysis. Distributed analytics also proved more robust when dealing with "composite bottlenecks," where multiple issues occur simultaneously.

Efficiency was also addressed through automated feature selection. The researchers found that a subset of just 11–19 key metrics (such as MCS, TCP errors, and memory utilization) could maintain high accuracy while reducing data overhead by approximately 65%.

Conclusion

This study demonstrates that while cloudified 5G networks present complex observability challenges, intelligent telemetry fusion and distributed analytics can accurately identify performance bottlenecks. While the current testbed uses 4G LTE components to emulate 5G concepts, the architectural principles hold. Future work will focus on scaling these methods to production-level environments and investigating "cascading" failures where small delays trigger larger systemic issues.

Summary updated

Abstract

This study is a first attempt to experimentally explore the range of performance bottlenecks that 5G mobile networks can experience. To this end, we leverage a wide range of measurements obtained with a prototype testbed that captures the key aspects of a cloudified mobile network. We investigate the relevance of the metrics and a number of approaches to accurately and efficiently identify bottlenecks across the different locations of the network and layers of the system architecture. Our findings validate the complexity of this task in the multi-layered architecture and highlight the need for novel monitoring approaches that intelligently fuse metrics across network layers and functions. In particular, we find that distributed analytics performs reasonably well both in terms of bottleneck identification accuracy and incurred computational and communication overhead.
BibTeX
@article{patounas2020characterization,
  author = {Georgios Patounas and Xenofon Foukas and Ahmed Elmokashfi and Mahesh K. Marina},
  title = {{Characterization and Identification of Cloudified Mobile Network Performance Bottlenecks}},
  journal = {IEEE Trans. Netw. Serv. Manag.},
  year = {2020},
  url = {https://openreview.net/forum?id=wN8idMmE0ld},
}

Loading related content...

Figure