[Paper] How Reliable is Your Service at the Extreme Edge? Analytical Modeling of Computational Reliability

Source: arXiv - 2602.16362v1

Overview

Extreme‑edge computing (XEC) pushes AI‑driven streaming workloads—think real‑time object detection on a phone or a smart camera—onto consumer devices that sit right next to the user. The paper “How Reliable is Your Service at the Extreme Edge? Analytical Modeling of Computational Reliability” tackles a practical question: what’s the chance that a single device, or a group of devices, can keep up with the required processing rate despite their ever‑changing availability? By turning this reliability problem into a set of closed‑form equations, the authors give developers a lightweight way to predict whether a distributed inference (DI) deployment will meet latency and throughput guarantees.

Key Contributions

• Formal definition of computational reliability for streaming AI workloads at the edge (probability that instantaneous capacity ≥ demand at a QoS threshold).
• Closed‑form reliability expressions for two information regimes:
  1. Minimal Information (only declared capacity bounds).
  2. Historical data (Maximum Likelihood Estimation from past observations).
• Extension to multi‑device scenarios with series, parallel, and partitioned workload configurations, including optimal workload‑allocation rules.
• Analytical bounds for device selection, enabling orchestrators to prune infeasible edge nodes quickly.
• Empirical validation using YOLO‑11m real‑time object detection on emulated XEC environments, showing tight agreement between theory, Monte‑Carlo simulation, and on‑device measurements.

Methodology

1. Modeling device capacity – Each edge device’s processing speed is treated as a random variable with known lower/upper bounds (MI) or a parametric distribution fitted from historical logs (MLE).
2. Reliability as a tail probability – The probability that the device’s instantaneous capacity exceeds the streaming demand is computed analytically using the cumulative distribution function (CDF) of the capacity model.
3. System‑level composition – For a set of devices, the authors derive reliability formulas for:
   • Series: the whole pipeline succeeds only if all stages meet demand.
   • Parallel: any device can satisfy the demand, boosting reliability.
   • Partitioned: the workload is split across devices; reliability depends on the allocation vector.
4. Optimization – By differentiating the reliability expression w.r.t. the allocation vector, they obtain simple rules (e.g., allocate more layers to higher‑capacity devices) that maximize overall reliability under a fixed total demand.
5. Validation – Experiments emulate heterogeneous consumer devices (smartphones, tablets, IoT boards) running a YOLO‑11m inference pipeline. Measured frame‑per‑second (FPS) rates are compared against the analytical predictions.

Results & Findings

• Accuracy: Across all configurations, the analytical model stays within ±2 % of Monte‑Carlo and real measurements.
• Scalability: Adding devices in parallel quickly pushes reliability above 0.9, even when individual devices are highly volatile.
• Optimal allocation: The derived rules reduce the required total capacity by ~15 % compared to naïve equal‑split allocation while preserving the same reliability target.

Practical Implications

• Fast feasibility checks: Orchestrators can plug in a device’s advertised CPU/GPU bounds (or a quick MLE from recent logs) and instantly know whether a streaming service will meet its latency SLA. No need for costly simulations.
• Dynamic workload placement: The allocation formulas enable runtime schedulers to rebalance inference layers on‑the‑fly as devices join/leave or their load changes, keeping reliability high without over‑provisioning.
• Edge‑aware service design: Developers can decide early whether to rely on a pure‑edge deployment, a hybrid edge‑cloud split, or a parallel‑edge redundancy strategy based on quantitative reliability targets.
• Resource budgeting: By providing analytical bounds, the framework helps product managers estimate how many consumer devices (or what class of devices) are needed to guarantee a given QoS for a large‑scale AR/VR or video‑analytics rollout.

Limitations & Future Work

• Assumed independence: The model treats device capacities as independent random variables; correlated load spikes (e.g., many devices running a heavy background app simultaneously) could degrade accuracy.
• Static demand model: The current analysis assumes a fixed streaming demand; extending to bursty or adaptive workloads (e.g., variable frame rates) is left for future research.
• Hardware heterogeneity: While the experiments cover a representative set of devices, the framework has not yet been validated on ultra‑low‑power wearables or specialized AI accelerators.
• Security & privacy considerations: The paper does not address how device‑level privacy constraints might limit the amount of historical data available for MLE, which could affect reliability estimates.

Bottom line: This work gives developers a mathematically grounded, yet easy‑to‑use, toolkit for answering the “can my edge fleet keep up?” question—turning reliability from a vague intuition into a concrete design parameter.

Authors

• MHD Saria Allahham
• Hossam S. Hassanein

Paper Information

• arXiv ID: 2602.16362v1
• Categories: cs.DC, cs.NI, eess.SY
• Published: February 18, 2026
• PDF: Download PDF