Measurement dataset and analysis code for "Measurement-Driven Modeling of End-to-End Latency in an Indoor Private Standalone 5G Network"

This record contains the measurement data and analysis code underlying the article "Measurement-Driven Modeling of End-to-End Latency in an Indoor Private Standalone 5G Network" (MDPI Network 2026, 6, 78).

Context and methodology. End-to-end latency and throughput were measured in the private standalone 5G network of the TU Wien IFT TEC-Lab. Measurements were taken at seven fixed indoor locations beneath or near the distributed pico radio units (pRRUs) of the deployment. One-way latency was measured over UDP at six target traffic rates (1, 10, 50, 100, 200 and 500 Mbps) with five repetitions per condition, giving 210 runs; TCP throughput was measured with five repetitions per location, giving 35 runs. All measurements used iPerf 2.2.1 in enhanced-report mode, with 60-second runs and one-second reporting intervals. The latency campaign was carried out between 27 June and 3 July 2025 and the throughput campaign on 27 June and 1 July 2025. Measurements characterise the complete path between a wired campus host and a 5G-attached endpoint, including campus routing and firewall functions, the private 5G core and radio access network, the customer premises equipment, and the wireless local-area link between the CPE and the receiving host.

Technical details. data.zip contains the processed per-trial datasets used in the article, including the achieved throughput parsed from the raw iPerf server reports. raw_latency_logs.zip and raw_throughput_logs.zip contain the iPerf output files; the IP addresses of the sending and receiving hosts are replaced by CLIENT_IP and SERVER_IP, and the files are otherwise unmodified. code.zip contains the MATLAB analysis pipeline in execution order, covering parsing, model fitting, bootstrap confidence-interval estimation, cross-validation, baseline comparison and the trial-level distribution analysis, the scripts for Figures 5–8 of the article, a function that fits the model to other measurements (code/fit_latency_model.m), and an independent Python re-implementation used to verify the published parameter estimates. figure8_measured_vs_predicted_latency.png is Figure 8 of the article, shown in README.md. The MATLAB scripts were developed and run with MATLAB R2025b and require the Optimization Toolbox and the Statistics and Machine Learning Toolbox; code/run_all.m runs the complete pipeline. To run the code, extract data.zip and code.zip into one folder and place raw_latency_logs.zip and raw_throughput_logs.zip, without extracting them, in its subfolder raw/. The Python verification scripts were tested with Python 3.14.6 under Windows 11, using both NumPy 2.4.6 and NumPy 2.5.2; the results agree between the two NumPy versions to at least six significant figures. NumPy is their only third-party dependency, and it is also declared in code/python/requirements.txt. All paths in the scripts are resolved relative to the script location, so the folder can be placed anywhere.

Licensing. This record is dual-licensed. The CC BY 4.0 license applies to all data, that is to data.zip, raw_latency_logs.zip, raw_throughput_logs.zip, figure8_measured_vs_predicted_latency.png and the README. All distributed code, that is the MATLAB and Python scripts in code.zip, is licensed under the MIT License. The full license texts are given in LICENSE.md, and the MIT license is additionally included inside code.zip as code/LICENSE-MIT.txt.

Further details. Three runs (location Q1, 50 Mbps, repetitions 3–5) returned invalid one-way delay values: iPerf reported delays close to 4.29 × 10^6 ms (2^32 µs), which is consistent with negative one-way delays wrapped around in a 32-bit counter, as can occur when the clocks of the two hosts are offset. These runs are flagged by valid_owd = 0 in latency_trials_raw_210.csv and were replaced by the cell mean in latency_210.csv; script step10_imputation_sensitivity.m quantifies the effect. See Section 4.3.1 of the article. Models are fitted using the achieved throughput reported by iPerf, not the configured target rate.

Version 1.0.1. The MATLAB pipeline runs directly from the record files; the IP addresses in the raw logs are masked; the scripts for Figures 5–8 of the article and a function that fits the model to other measurements are included; the description of the measurement locations, the dates of the throughput campaign and the toolbox requirements are corrected. The processed data and all results are unchanged. The changes are listed in README.md.