Aeria Edge AI Eval

aeria-edge-ai-eval

Dynamic Pricing for On-Demand DNN Inference in the Edge-AI Market — Li et al. (2025) (arXiv:2503.04521, 2025)

What this evaluates

Evaluates an auction-based dynamic pricing and resource allocation mechanism for on-demand DNN inference at the edge. It probes the system's ability to jointly optimize model partitioning, pricing, and resource distribution under varying user requirements and real-world trace-driven conditions.

Datasets

• Multi30K — total 30016; splits: test (-1)
• ImageNet-1K — total ?; splits: test (-1)
• CIFAR-100 — total ?; splits: test (-1)
• CIFAR-10 — total ?; splits: test (-1)
• Shanghai Telecom — total 800; splits: test (-1)

Metrics

• revenue (primary) — range: other
  • Total earnings of the edge AI service provider from allocating inference resources to bidding users, optimized via the auction mechanism.
• inference performance — range: other
  • Composite metric evaluating inference latency and accuracy across multi-exit DNN branches under varying user requirements.

Input / output format

Input: User bids R_i = {beta_i, t_i, sigma_i, w_i} (price, latency requirement, accuracy requirement, DNN model), device computation capacity, geographical coordinates, real-time wireless data rate, edge server capacity, and electricity price.

Output: Resource allocation decisions, dynamic pricing per time slot, and DNN model partitioning/exit selection for each user.

Scoring recipe

# Trace-driven simulation setup
users = load_shanghai_telecom_trace(n=800)
power_prices = load_ontario_power_trace(days=15)
bandwidth = load_ghent_4g_trace(slots=360)

revenues = []
for run in range(120):
    # Run mechanism and baselines (IAO, Edgent, AMR^2, Fixed Profit Rate)
    allocation = run_aeria(users, power_prices, bandwidth)
    rev = sum(user_bid_price for user in allocation if user.wins)
    revenues.append(rev)

avg_revenue = sum(revenues) / len(revenues)
# Compare avg_revenue against baseline averages

Common pitfalls

• Relies on synthetic early-exit branches that equally apportion computation overhead, which may not reflect real ME-DNN training dynamics.
• Uses historical trace datasets (Shanghai Telecom, Ontario power, Ghent bandwidth) for simulation rather than live edge deployments, limiting real-world generalizability.
• Averages results over 120 runs but does not report statistical variance or confidence intervals in the provided section.

Evidence (verbatim from paper)

Differing from AERIA for revenue maximization, it takes a rigid earning target for the AI service providers with the fixed profit-rate objective of 1.0. All the trace-driven experimental results about the AERIA, IAO, Edgent, and Fixed Profit Rate = 1.0 are averaged over 120 runs.

Citation

@misc{li2025aeria,
  title={Dynamic Pricing for On-Demand DNN Inference in the Edge-AI Market},
  author={Li et al. (2025)},
  year={2025},
  note={arXiv:2503.04521}
}

• arXiv: 2503.04521