AV Edge-to-Cloud Workload Model

This repository contains the reproducibility artifact for the paper:

Modeling Edge-to-Cloud Offloading Workloads for Autonomous Vehicles

The artifact provides a compact, public version of the workload generation and evaluation pipeline used in the paper. It reproduces data grounding, equal-volume workload controls, stochastic capacity risk, event timing experiments, parameter sensitivity, scheduling of training clips, and rolling evaluation of AP capacity.

Repository Structure

.
├── data/processed/          # Compact processed inputs used by the experiments
├── docs/                    # Data-source and reproducibility notes
├── outputs/figures/         # Generated paper figures
├── outputs/tables/          # Generated CSV summaries
├── scripts/                 # Reproduction scripts
├── requirements.txt
└── README.md

What Is Included

• Processed 24-hour Munich active-vehicle inputs; workload components are regenerated from declared parameters.
• Hourly profile of 6,940 Munich accidents involving personal injury in 2023, used only to determine the timing of candidate training clips.
• Hourly summaries of vehicles associated with the nearest AP in the Munich service region.
• Circular-shift experiments that preserve daily learning volume while varying event-profile alignment with mobility.
• Capacity-exceedance curves for a Poisson baseline and three Gamma--Poisson (Cox) event-count settings.
• Bounded schedules that preserve the complete volume of training clips while minimizing peak workload.
• Five rolling chronological AP fitting/evaluation splits.
• A Python script that regenerates all reported experiment tables and figures.

Reproducing the Figures

Create a Python environment and install the dependencies:

pip install -r requirements.txt

Run the reproduction script:

python scripts/run_nextgcom_experiments.py

The script writes regenerated figures to outputs/figures/ and regenerated CSV summaries to outputs/tables/. Every plotted curve is accompanied by its underlying CSV output.

Interpretation Notes

This artifact supports an exposure adjusted evaluation based on mobility traces. The Munich mobility profile is derived from a SUMO scenario based on origin and destination demand. If $c_h$ is the historical count and $N_h$ is the hourly mean number of concurrent vehicles, the script uses $q_h=(c_h/N_h)/(\sum_j c_j/\sum_j N_j)$. The mobility-weighted mean of $q_h$ is one, so the independent baseline parameter gives exactly one candidate clip per active vehicle per hour over the scenario. The sensitivity analysis varies the mean rate from 0.25 to 4 times its baseline. The event timing experiment circularly shifts $q_h$ and normalizes each shift to preserve the daily training volume of the fleet.

The fixed rate control has the same 24-hour fleet volume as the application model. Stochastic clip counts use a Poisson baseline and Gamma--Poisson alternatives with latent intensity shapes k = 2, 4, 8. Peaks are evaluated against capacity provisioned from the fixed rate peak with 0--20% headroom. AP capacity uses five rolling chronological splits; each uses an eight-hour reference window, evaluates the next four hours, and advances by two hours. Equal and demand informed allocation receive the same total capacity at every factor and split.

The scheduling experiment keeps monitoring and map traffic at their generated hours and permits only training clips to move forward by at most 0, 1, 2, or 4 hours. Every schedule conserves the complete daily clip volume and sends pending traffic by the end of the 24-hour window.

Citation

If you use this artifact, please cite the associated paper once it is available.