Edge Latency under Mobility — Measurement & Analysis

Data-analysis pipeline that quantifies how user mobility, distance and cellular signal quality affect end-to-end latency in Multi-Access Edge Computing (MEC) scenarios, using real 4G/5G measurements collected in Rio de Janeiro.

This repository holds the analysis behind two peer-reviewed publications (see Publications). The raw measurements were collected with WolfLatency, an Android client–server toolkit, and processed here into the figures and statistics reported in the papers — developed as undergraduate research at GTA/COPPE/UFRJ.

Research question

In MEC, services run at the network edge to minimise latency. But users move. How much does mobility actually degrade latency, and what drives it — physical distance to the edge node, or radio signal quality? To answer that, each request is paired with the device's GPS position and serving-cell metadata, so latency can be correlated with both distance (via the Haversine formula) and signal level.

Key findings

The WGRS'24 campaign alone gathered 15,000+ latency samples (routes from on-campus hops up to a ~67 km intercity trajectory); the vehicular VehiClouds'24 campaign added ~5,000 more, with tail latencies reaching ~15 s.

Signal quality is the dominant driver of latency. Grouping latency by reported signal level (log scale, n per bin) shows the spread widening sharply as the signal degrades — weak signal triggers retransmissions and heavy tails:

Distance alone is a weak predictor. Least-squares fits of latency against client–edge distance have very low R² in both campaigns — the slope is mildly positive on the WGRS routes (3.11 ms/km) and even negative on the vehicular VehiClouds route, confirming that link conditions matter far more than raw distance:

Cell-tower–aware mapping. The client's path and the base stations it connected to (geolocated via the Unwired Labs API), used to also study the client↔base-station distance:

Additional figures (latency/signal CDFs, jitter, per-tower regressions, min-delay fits) are in figures/.

Methodology

1. Collection — the moving client issues one HTTP GET per second to a static server over IPv6; each sample logs latency (RTT), GPS, and cellular context (signal level, RSRQ/RSSNR, cell id, radio technology 2G→5G NR).
2. Distance — client↔edge separation computed from GPS with the Haversine/geodesic distance.
3. Minimum latency — to estimate propagation latency without retransmissions, consecutive samples (ordered by distance) are grouped into small windows and the minimum per window is kept (windows of 5 in the WGRS study; the September/VehiClouds analysis here uses 7).
4. Statistics — linear regressions (distance × latency, with R²), latency confidence intervals, CDFs, jitter, and signal-level boxplots.

Repository structure

.
├── figures/        # Result figures — PNG (rendered) + source PDF
├── Dataset/        # WGRS'24 data, segmented by route length (Curto / Médio / Longo)
├── Janeiro/        # Jan 2024 measurement campaign  → WGRS'24 paper
├── Setembro/       # Sep 2024 measurement campaign  → VehiClouds'24 paper
├── main.py         # Signal-level boxplots, network-tech distribution, latency CIs, Haversine delay
├── latency.py      # Distance × latency regression (min latency per burst)
├── distance_delay.py
├── jitter.py       # Jitter analysis
├── cdf_correlation_analysis.py  # Latency/signal CDFs; RSSNR & signal-vs-distance correlations
├── haversineApi.py # Haversine distance between client/server GPS traces
├── timestamp.py    # Timestamp parsing/alignment
├── draw_map.py     # Measurement trajectory maps
├── api*.py         # External lookups (cell-tower geolocation / mapping)
├── requirements.txt
└── README.md

Each measurement CSV follows the WolfLatency schema (Sequence, Latency, Latitude, Longitude, Signal_level, RSRQ, RSSNR, Mobile_Network, ...).

Reproducing the analysis

pip install -r requirements.txt
python latency.py        # regenerates the distance × latency figures
python main.py           # signal-level / technology breakdowns

These are research scripts: the analysis entry points live at the bottom of each file and point to a specific campaign folder. Adjust the input path to the dataset you want to analyse.

Publications

• L. F. C. Cristino, P. Cruz, L. H. M. K. Costa — Vehicle Mobility Impact on Performance of Multi-Access Edge Computing, VehiClouds 2024 (IEEE CloudNet).
• L. F. C. Cristino, P. H. C. Caminha — Caracterização da Latência de Borda sob Efeito de Mobilidade a Partir de Dados Reais, WGRS 2024 (SBC).

📊 Presentation slides from IEEE CloudNet 2024: paper/VehiClouds2024_slides.pdf.

Authors

Undergraduate research at the Grupo de Teleinformática e Automação (GTA/COPPE/UFRJ) — Luiz Felipe Cantanhede Cristino, advised within the GTA edge-computing line, with Pedro Henrique Cruz Caminha.