Network Delays Turn Lognormal as Traffic Loads Push Queues Toward Saturation

Every time a video call freezes, a cloud game stutters, or a self-driving car’s sensor message arrives a few milliseconds too late, the culprit is usually the same invisible phenomenon: packets waiting in a queue inside a network device. A new study published in Cluster Computing by Ramazan Enisoglu of Kirikkale University, Burak Alp Inan of Hacettepe University, and Saadettin Yavuz Ugurlu of Akdeniz University takes aim at one of the oldest and most stubborn problems in network engineering, namely predicting exactly how long those waits will be when traffic arrives in the bursty, unpredictable bursts that characterize modern wireless networks. The team built a discrete-time queueing simulator based on the N×On-Off/D/1 model, validated it against established theory, and then subjected its delay statistics to some of the most demanding statistical scrutiny in the field.

The N×On-Off/D/1 model describes a single server with deterministic service times, fed by N independent sources that each alternate between active ON periods, during which they emit packets, and silent OFF periods. This structure matters because real network traffic is nothing like the smooth, memoryless arrivals assumed in textbook models. Since the landmark measurements of Ethernet traffic by Leland and colleagues in the 1990s, engineers have known that packet streams are bursty and correlated, and that burstiness is precisely what inflates queues and delays. The On-Off source model captures this correlation in a tractable way, making it a workhorse for teletraffic engineering in broadband and wireless systems alike.

Before trusting any simulator, the researchers had to prove it reproduces known results. Their first validation stage used the M/D/1 queue, the classical model with Poisson arrivals and deterministic service, as a baseline. For this model, the asymptotic decay rate of the queue-length tail distribution is known in closed form, giving an exact mathematical benchmark. Across utilization levels rho ranging from 0.7 to 0.95, the simulator matched the theoretical decay rate to within one percent. That level of agreement is the kind of result that separates a credible simulation framework from a toy, and it established the foundation for everything that followed.

The second validation stage tackled the harder target: the Burst Scale Decay Rate, or BSDR, a closed-form approximation that predicts how quickly the probability of large queue buildups decays when many On-Off sources feed the buffer. Here the simulator tracked the approximation within 2.2 to 12.5 percent across the tested configurations. Crucially, the team did not treat that residual gap as an embarrassment. By running multiple independent replications and constructing confidence intervals, they showed that the discrepancy is a systematic property of the closed-form approximation itself, not random simulation noise. In other words, the simulator is more accurate than the approximation it was being tested against, a subtle but important inversion that gives engineers a clearer picture of how much to trust the analytical shortcut.

With the framework validated, the researchers turned to the central question: what does the distribution of per-packet queueing delay actually look like under bursty load? They examined three different ON/OFF configurations at utilization levels from 0.89 to 0.98, deep in the congested regime where wireless networks spend much of their operating life. Each operating point was simulated with five independent replications totaling roughly 700 million packets, an enormous dataset by any statistical standard. The scale matters because the question they were asking, whether a delay distribution follows a particular mathematical form, is exactly the kind of question where insufficient data produces misleading answers.

The verdict came from four independent distance measures: the Kolmogorov-Smirnov statistic, the Anderson-Darling statistic, the coefficient of determination between cumulative distribution functions, and the Bhattacharyya overlap coefficient. All four measures told the same story. As the load increases, the queueing delay distribution steadily approaches a lognormal shape. The Kolmogorov-Smirnov distance shrank to 0.051 at the highest loads, and the fixed-sample Anderson-Darling statistic, which is especially sensitive to the tails of a distribution, dropped by factors of three to seven. For network engineers, this is a genuinely useful regularity: the lognormal distribution, whose parameters can be estimated from a mean and a variance, offers a compact closed-form description of delay behavior that grows more accurate precisely where delays matter most, near saturation.

Yet the study is equally emphatic about the limits of that regularity. Bootstrap-corrected goodness-of-fit tests rejected exact lognormality at every single operating point, no matter how close the visual and distance-based agreement appeared. The authors therefore present the lognormal as an approximation with quantified accuracy rather than an exact law, a distinction that reflects a mature statistical philosophy. To test whether some other two-parameter distribution might do better, they systematically compared the lognormal against Weibull, gamma, log-logistic, and Lomax alternatives. The result was nuanced: at moderate loads the lognormal could lag behind the best competitor for a given configuration, but as load increased it became competitive with, and in some cases the best among, all the two-parameter candidates.

Configuration mattered too. The team’s sensitivity analysis showed that OFF-dominated or symmetric traffic patterns, where the silent period is at least as long as the active period, produced better lognormal agreement at moderate loads. Near saturation, however, this dependence weakened, suggesting that extreme congestion washes out the details of the source structure and drives the delay distribution toward a common heavy-tailed form regardless of how the bursts are shaped. This finding has practical implications for traffic engineering, because it tells operators when they can rely on a simple lognormal fit and when the specifics of their traffic mix demand a more careful analysis.

Perhaps the most consequential result concerns the tails, the extreme delays that determine whether a system meets the stringent latency budgets of 5G, the tactile internet, and ultra-reliable low-latency communications. Above a utilization of roughly 0.95, the fitted lognormal consistently overestimated the empirical 95th, 99th, and 99.9th percentile delays in every configuration tested, at P99.9 by factors ranging from 3.6 to 9.6 times the observed values. At moderate loads, by contrast, the sign of the tail error depended on the traffic configuration, sometimes underestimating the true risk. The authors frame the heavily loaded overestimation as a virtue: the lognormal acts as an empirically conservative closed-form tool for delay budgeting, erring on the safe side exactly in the regime where network dimensioning decisions are made.

The broader significance of the work lies in its marriage of large-scale simulation with rigorous uncertainty quantification. Rather than declaring a new universal law of network delay, the study delivers something more useful: a validated simulator, a calibrated understanding of when the BSDR approximation can be trusted, and a quantified recipe for using the lognormal as a delay-budgeting tool with known error bounds. As wireless networks absorb everything from factory robots to remote surgery, the difference between a guessed delay distribution and a statistically verified one translates directly into either wasted capacity or violated service guarantees. This research gives the engineers building that future a sharper instrument, and a candid account of its precision.

Subject of Research: Queueing-theoretic modeling and statistical validation of packet delay distributions in bursty wireless network traffic

Article Title: Validation of burst scale decay rate and analysis of lognormal delay distribution in NxOn-Off/D/1 queueing models for wireless network performance

Article References: Enisoglu, R., Inan, B. A., & Ugurlu, S. Y. (2026). Validation of burst scale decay rate and analysis of lognormal delay distribution in NxOn-Off/D/1 queueing models for wireless network performance. Cluster Computing, 29(13), Article 780. https://doi.org/10.1007/s10586-026-06591-y

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06591-y

Keywords: queueing theory, N×On-Off/D/1 model, burst scale decay rate, lognormal distribution, packet delay, QoS provisioning, wireless networks, tail probability estimation, goodness-of-fit testing, network dimensioning, discrete-event simulation, 5G latency