A queueing-based model of connected-vehicle services found that higher modeled arrival-time jitter and traffic variability were associated with lower probabilities of meeting latency deadlines and higher 99th-percentile latency. The modeled differences were most pronounced in cloud execution, while traffic variability generally showed a stronger association with deadline satisfaction than jitter. The results are reported as modeled relationships.
Tracing the whole delay path
The paper asks how stochastic temporal variability affects the ability to provide end-to-end deterministic service levels across the IoT-edge-cloud continuum. It focuses on arrival-time jitter, meaning variation in when packets arrive, and traffic variability, meaning changes in packet arrival rates.
Its central tool is a queueing-based model that combines computing and communication latency into a single end-to-end latency distribution. The distribution includes tail latency, the slower part of the range, so the analysis represents occasional long delays alongside typical performance. The proposed model is openly released.
The evaluation used vehicular V2X service profiles, comparing local processing with edge and cloud execution. Computing demand ranged from 1.5 to 6 MCycles per packet, deadlines from 10 to 100 milliseconds, and packet rate was 1,000 packets per second. The scenarios varied processor workload and cellular link quality.
Arrival-time jitter was modeled with a Gaussian distribution based on in-vehicle radar arrival-time data. Traffic variability was represented by a bounded worst-case uniform distribution of packet arrival rates, keeping instantaneous traffic load within defined physical limits. In practical terms, the model varied both when packets arrived and how much the arrival rate fluctuated.
The slower tail stood apart
In the model-validation setup, radio-access-network latency averaged 4.61 milliseconds at 1 Mbits/s and 7.21 milliseconds at 2 Mbits/s. At 1 Mbits/s, the 90th and 99th percentiles were 10.41 and 16.72 milliseconds; at 2 Mbits/s, they were 14.38 and 18.92 milliseconds. The 2 Mbits/s case therefore had higher modeled latency, and the slower tail stood above the average in both cases.
An end-to-end comparison at 1 Mbits/s showed a large gap between edge and cloud execution. Edge latency averaged 6.33 milliseconds, with 12.24 milliseconds at the 90th percentile and 18.57 milliseconds at the 99th. Cloud latency averaged 40.44 milliseconds, with corresponding values of 58.31 and 75.91 milliseconds. Edge was lower on all three reported measures in that model configuration.
Jitter and traffic did not look the same
Higher modeled jitter was associated with a lower probability of meeting a service deadline. Short deadlines were the most sensitive, and cloud execution showed the greatest sensitivity. Jitter was also associated with higher modeled 99th-percentile latency for every service. The highest values were in cloud execution, with stronger differences for services with greater computing demand.
Traffic variability was associated with a lower deadline-meeting probability than jitter overall. The sensitivity was greater for stringent deadlines, higher computing demands, and services offloaded to the edge or cloud. Traffic variability was also associated with higher modeled 99th-percentile latency, with the largest increases in cloud execution and higher-demand services.
The edge advantage depended on demand
For a service with a 100-millisecond deadline under high processor workload, the model showed a higher deadline-meeting probability for edge than local execution at both 30 dB and 10 dB of cellular link quality.
Demand changed that comparison. At 1.5 Mcycles, edge execution was less affected by jitter, but that advantage was not retained at 6 Mcycles. At the higher demand, cloud execution performed worse than local execution. The comparison is conditional on the modeled demand and does not establish a general offloading rule.
A policy tied to the deadline
The authors conclude that stringent-deadline services favor local processing, while services with more relaxed deadlines are more resilient when executed locally or at the edge. That conclusion is conditional on computing demand and on the communication and computing conditions represented in the model. The comparison shows why an offloading policy cannot be based on execution location alone.
The supplied document identifies itself as an author-created postprint. Funding was reported from the European Union MSCA fellowship program and from MCIN/AEI, UMH and Generalitat Valenciana.
Paper data and sources
Original title: Stochastic End-to-End Latency Modeling of the IoT-Edge-Cloud Continuum: Impact of Jitter and Traffic Variability on Deterministic Service Provisioning
Authors: Keyvan Aghababaiyan, Javier Gozalvez, Baldomero Coll-Perales
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text