PaperlayerView exact revision on arXiv

Source-linked AI summary

Hidden in Rounds: Predicting the Time Cost of 802.11 Contention in Federated Learning

Satwat Bashir, Tasos Dagiuklas

arXiv:2609.12903v1cs.LGcs.DC

TL;DR

Existing models and tools leave open how contention-based 802.11 affects FL communication time and participation bias. The paper combines ns-3 measurements with a separate FedAvg trainer, finding that communication time-to-target grows with density and load even as rounds-to-target change little with offered load. Its estimator is evaluated against constructed communication time, and the conclusions are limited to the tested configurations.

  • Problem

    Prior channel models treat client losses independently or constrain participation, while the interaction of contention and participation bias with non-i.i.d. data remains a study question.

    Source: §I · p. 1, A. Convergence of FL over Wireless… · p. 2

  • Method

    The study uses ns-3 to measure frame-delivery ratios and saturation throughput, then a separate FedAvg trainer uses delivery ratio as update-admission probability and constructs communication duration.

    Source: §I · p. 1, §VI · p. 6

  • Results

    All 720 runs reached their predefined targets; the Bianchi-anchored estimator had a held-out MAPE of 2.3% to 10.2% against constructed communication time, not independently measured completion time.

    Source: §VI · p. 7

  • Takeaways & Limitations

    Communication time-to-target reflects density and offered-load effects that rounds-to-target alone do not capture in the evaluated configurations.

    Source: §VI · p. 7

    The study covers a single-cell setting with fixed physical-layer rate, and its communication times do not predict production deployment time.

    Source: §VI · p. 7

Abstract

from arXiv · show

Federated learning over IEEE~802.11 shares the wireless channel among clients that send model updates. We use ns-3 to measure the frame-delivery ratio and saturation throughput for different client densities and offered loads. A separate FedAvg trainer uses the frame-delivery ratio as a first-order proxy for the update-admission probability and uses an equation to estimate communication time. The method does not simulate the delivery of a complete model update or measure end-to-end training time. Across 720 evaluated runs with two datasets, two data partitions, six client densities, six offered loads, and five seeds, all runs reached their predefined target accuracy within the round budget. Rounds-to-target changed little with offered load, while communication time-to-target increased by about two orders of magnitude across the client-density range. A Bianchi-anchored estimator produced a mean absolute percentage error from $2.3\%$ to $10.2\%$ on held-out configurations. This error is measured against communication time constructed from the same round-duration equation, not against independently measured completion time. We also compare uniform participation with persistent heterogeneous participation. The study does not detect a statistically distinguishable excluded-class accuracy gap over five seeds, but the confidence intervals are wide. The results apply only to the evaluated configurations and do not provide a general convergence or fairness guarantee.

I. Introduction

The paper studies how 802.11 contention affects federated learning, using separate network measurements and a FedAvg trainer to estimate communication time and examine participation bias. Across evaluated runs, communication time-to-target grows with density and offered load, while rounds-to-target change little with offered load.

I. Introduction

  • IEEE 802.11 contention makes packet loss and delay emerge from shared-channel competition, unlike independent per-client loss models.The trainer summarizes measured channel behavior with frame-delivery ratios and does not preserve joint loss events or temporal dependence.Source: §I · p. 1
  • The paper examines whether persistent under-representation of poorly connected clients interacts with non-i.i.d. data as a measurable participation-bias effect.It compares uniform and persistent heterogeneous participation and reports uncertainty in class-accuracy gaps.Source: §I · pp. 1–2
  • The study separates ns-3 frame-delivery measurements from a FedAvg trainer that uses delivery ratio as update-admission probability and estimates round communication duration.Throughput measurements are compared with the analytical Bianchi model before use.Source: §I · pp. 1–2
  • All evaluated configurations reach their predefined target, but communication time-to-target grows with density and offered load while rounds-to-target change little with offered load.The paper also reports that the estimator is evaluated against a constructed communication-time metric.Source: §I · p. 2

A. Convergence of FL over Wireless Channels

Prior wireless-FL analyses commonly abstract packet loss as independent per-client loss or impose orthogonal resource constraints, leaving contention-driven effects of density and load uncharacterized. Over-the-air aggregation studies address a different transmission setting from packetized Wi-Fi.

A. Convergence of FL over Wireless Channels

  • Existing convergence analyses model loss independently across clients or constrain users through orthogonal resource blocks, so they do not describe how contention-based MAC performance changes with density and offered load.Their loss probabilities instead depend on physical-layer parameters.Source: A. Convergence of FL over Wireless… · p. 2

B. Over-the-Air Aggregation

  • Analog over-the-air aggregation assumes synchronized, power-controlled signal superposition over fading channels, unlike packetized digital Wi-Fi with carrier sensing and backoff.The cited work therefore targets a different deployment.Source: B. Over-the-Air Aggregation · p. 2

C. Network-Coupled FL Simulation

Network-coupled FL tools enable experiments under network effects, while channel-aware selection research motivates concern about participation bias. This study positions itself as a compact estimator-based analysis rather than an end-to-end simulation.

C. Network-Coupled FL Simulation

  • Network-coupled simulators and emulators support FL experiments under network effects, while prior analytical work models FL over Wi-Fi CSMA/CA using renewal-reward processes and a Markov chain.This study instead constructs a communication-time estimate from measured frame-delivery ratios and saturation throughput.Source: C. Network-Coupled FL Simulation · p. 2

D. Channel-Aware Selection and Participation Bias

  • Prior work links persistent poor connectivity and non-i.i.d. data to participation bias, while this study samples client-specific participation independently each round and omits temporal dependence and simultaneous exclusions.The study uses persistent client-specific frame-delivery ratios measured in ns-3.Source: D. Channel-Aware Selection and… · pp. 2–3

E. Positioning

  • The method combines frame-level ns-3 measurements, an equation-based communication-time estimate, and a separate FedAvg trainer, without replaying packet traces or integrating the trainer with ns-3.It supports controlled comparison of rounds-to-target and communication time-to-target, but does not validate end-to-end completion time.Source: E. Positioning · p. 3

A. Overview and Decoupled Design

The study separates network measurement from learning: ns-3 supplies frame-delivery ratios and saturation throughput, while a separate trainer constructs participation and communication duration. It evaluates fixed learning and channel configurations and validates saturation throughput against Bianchi analysis.

A. Overview and Decoupled Design

  • The single-cell setup uses one access point and a shared 802.11 channel for N client stations, with no hierarchical edge tier.The network and learning components do not share runtime state; ns-3 supplies frame counts, frame-delivery ratios, and saturation throughput, while the trainer constructs update admission and communication time.Source: A. Overview and Decoupled Design · p. 3

B. Federated Learning Setup

  • The evaluation uses Fashion-MNIST and CIFAR-10 with i.i.d. and Dirichlet non-i.i.d. partitions, five seeds, and a LeNet-style model; the per-client partitions remain fixed across components.A more concentrated Dirichlet split is used only for the bias analysis.Source: B. Federated Learning Setup · p. 3

C. Contention Model and Validation

  • At N=50, frame-delivery ratio falls from 1.00 to 0.39 across offered-load levels; measured saturation throughput agrees with Bianchi predictions within 8.6% mean error and at most 12%.The figure plots analytical collision probability and measured saturation throughput and non-delivery rate versus N; throughput and non-delivery measurements use saturated traffic at a fixed MAC and physical-layer configuration.Source: C. Contention Model and Validation · pp. 3–4, Figure 1 · p. 4, Table II · p. 4

D. Coupling and Communication-Time Estimate

The method combines frame-delivery measurements with a separate FedAvg trainer to estimate per-round communication time. This estimate is a proxy, not end-to-end update completion time.

D. Coupling and Communication-Time Estimate

  • Each round is charged at measured saturated throughput, with its estimated duration scaled by the inverse frame-delivery ratio.The trainer uses that ratio as an update-admission probability; the model update is approximately 178 KB for this network.Source: D. Coupling and Communication-Time… · p. 4
  • The network simulator sends constant-bit-rate UDP traffic and measures frame delivery, not the delivery of a complete synchronous model update.It does not test whether all frames arrive before a round deadline.Source: D. Coupling and Communication-Time… · p. 4
  • The trainer uses frame-delivery ratio only as an update-admission proxy; it does not model application-level reassembly, timeout, or all-or-nothing acceptance.Communication time is constructed rather than measured end-to-end; local computation is excluded, and stale updates are not reused.Source: D. Coupling and Communication-Time… · p. 4

E. Experimental Design and Metrics

The evaluation sweeps client density, offered load, data split, and dataset across five seeds, then records rounds-to-target and constructed communication time-to-target. Targets are based on perfect-channel baselines and are attainable within the round budget.

E. Experimental Design and Metrics

  • The full sweep covers six client densities, six load levels, two data splits, and two datasets across five seeds, totaling 720 runs.Runs train to their round budgets; a run missing its target is recorded as a distinct non-convergence outcome.Source: E. Experimental Design and Metrics · p. 4
  • Rounds-to-target is the first round when accuracy reaches a fixed target and stays there; communication time-to-target accumulates Eq. (1) durations through that point.Targets derive from each dataset-and-split perfect-channel baseline and are set below its attained level.Source: E. Experimental Design and Metrics · p. 4
  • The bias analysis additionally records worst-class accuracy, defined as the global model's lowest per-class test accuracy.For CIFAR-10, targets are 0.48 for i.i.d. and 0.46 for non-i.i.d. splits under the fixed configuration.Source: E. Experimental Design and Metrics · p. 4

A. Evaluated Runs Reach the Target Within the Round Budget

All evaluated runs reached their predefined targets, while estimated communication time-to-target rose substantially with client density. Offered load changed that time measure more than rounds-to-target within the evaluated configurations.

A. Evaluated Runs Reach the Target Within the Round Budget

  • All 720 runs reached their predefined targets; at saturated load, communication time-to-target rose from 6 to 640 seconds as density increased from N=10 to N=100.The observation is limited to the evaluated models, datasets, targets, densities, loads, and five seeds.Source: A. Evaluated Runs Reach the Target… · pp. 4–5
  • At N=50, rounds-to-target stayed at 8.6 as load rose from light to saturated, while communication time-to-target rose from 52 to 133 seconds.Per-round delivery fell from 1.00 to 0.39 over the same load range.Source: A. Evaluated Runs Reach the Target… · pp. 4–5
  • For Fashion-MNIST at N=50, accuracy-versus-round curves were almost unchanged from perfect to saturated channel, while the estimated communication-time axis increased about 1.9-fold.Figure 2 shows accuracy against rounds on the left and estimated cumulative communication time on the right.Source: A. Evaluated Runs Reach the Target… · p. 4, Figure 2 · p. 5
  • On CIFAR-10 (i.i.d.) at N=50, rounds-to-target stayed near 4.7 as load rose from light to saturated, while communication time-to-target rose from 39 to 104 seconds.At saturated load, communication time-to-target rose from 6 seconds at N=10 to 463 seconds at N=100.Source: A. Evaluated Runs Reach the Target… · p. 5
  • Round-based analysis can show little load effect and only a modest density effect without describing the communication-time cost assigned by Eq. (1).The cited comparison is framed against rounds-based convergence analyses,.Source: A. Evaluated Runs Reach the Target… · p. 5

B. Estimating Communication Time-to-Target

The Bianchi-anchored estimator predicts communication time-to-target from a rounds-to-target estimate and communication duration, and is evaluated on held-out configurations. Its reported error is against a time metric constructed using the same timing equation, not independently measured completion time.

B. Estimating Communication Time-to-Target

  • The estimator uses two stages: ordinary least squares estimates rounds-to-target, then a second stage computes communication time.Coefficients are fitted separately for each dataset and data split.Source: B. Estimating Communication… · p. 5
  • A degree-two channel-free reference maps client-density and offered-load indices directly to communication time without using frame-delivery ratio or saturation throughput.It is not an equal-input baseline because it does not receive p_f or R_sat.Source: B. Estimating Communication… · p. 5
  • 2.3%–10.2% MAPE: the Bianchi-anchored estimator’s held-out error across the reported tests, measured against constructed communication time-to-target.The channel-free reference has larger errors on several tests, but the methods use different channel information and are not a controlled comparison of estimator structure.Source: Table III · p. 5, B. Estimating Communication… · p. 5
  • Held-out evaluation includes leave-one-configuration-out testing, density extrapolation to N=100, and offered-load extrapolation to the saturated level.No test configuration is used for fitting.Source: B. Estimating Communication… · p. 5
  • Communication time-to-target is rounds-to-target multiplied by T_comm, so the estimator’s error does not validate Eq. (1) against independently measured synchronous FL completion time.On persistent heterogeneous participation runs, its MAPE is 2.1%, compared with 2.0% for the uniform arm.Source: B. Estimating Communication… · p. 5

A. Persistent Heterogeneous Participation Model

The persistent heterogeneous participation model assigns clients distinct, fixed update-admission probabilities derived from ns-3 frame-delivery ratios. The comparison holds mean admission probability and communication duration fixed, varying the distribution of expected participation.

A. Persistent Heterogeneous Participation Model

  • Each client receives a persistent update-admission probability q_i derived from its ns-3 frame-delivery ratio, then is sampled independently across clients and rounds.The model does not represent simultaneous exclusions, temporal dependence, packet bursts, or queue dynamics.Source: A. Persistent Heterogeneous… · p. 5
  • Distance and frame error make some clients persistently less likely to participate; distance-to-frame-delivery correlation ranges from −0.60 to −0.82 across cells.Stations are placed uniformly at random in a 20 m-radius disc around the access point.Source: A. Persistent Heterogeneous… · p. 6
  • The uniform and heterogeneous arms share the same mean update-admission probability and T_comm, so the paired comparison changes the distribution of expected client participation.The per-cell matched-mean difference from the target is 0.000 across 18 cells, and the trainer does not replay ns-3 packet traces.Source: A. Persistent Heterogeneous… · p. 6

B. Class-Accuracy Gaps and Uncertainty

Across five seeds, the study does not detect statistically distinguishable class-accuracy gaps between persistent heterogeneous and uniform participation, but wide confidence intervals do not establish that the gaps are zero.

B. Class-Accuracy Gaps and Uncertainty

  • Worst-class accuracy is the minimum of ten per-class test accuracies, while excluded-class accuracy measures the class most common at the lowest-q_i client.The excluded-class metric avoids selecting the minimum after training.Source: B. Class-Accuracy Gaps and Uncertainty · p. 6
  • Every reported 95% confidence interval for worst-class and excluded-class accuracy gaps includes zero, so the five-seed results do not detect a statistically distinguishable gap.For concentrated heterogeneity (α=0.1), worst-class gaps range from −0.9 to −10.2 percentage points and excluded-class gaps from −1.0 to −2.6 percentage points; several intervals are wide.Source: B. Class-Accuracy Gaps and Uncertainty · p. 6
  • Under the evaluated Dirichlet(α=0.1) split, each class remains distributed across many clients, and the lowest-q_i client holds only nine samples of its dominant class at N=100.More concentrated class ownership was not evaluated; the results are limited to the evaluated partitions, client densities, and five seeds.Source: B. Class-Accuracy Gaps and Uncertainty · p. 6

VI. Conclusion and Future Work

All evaluated FedAvg runs reached their targets, while contention-related communication time varied with client density and offered load. The conclusion bounds its timing and participation findings to the modeled setup and calls for packet-level validation.

VI. Conclusion and Future Work

  • ns-3 measured frame-delivery ratio and saturation throughput; the trainer used the ratio as an update-admission probability and constructed communication time with Eq. (1).Source: §VI · pp. 6–7
  • All 720 evaluated runs reached their predefined targets; rounds-to-target changed little with offered load, while communication time-to-target grew with density and offered load.Source: §VI · p. 7
  • The Bianchi-anchored estimator's held-out MAPE was 2.3%-10.2% against the constructed communication-time metric, not independently measured synchronous FL completion time.Source: §VI · p. 7
  • Over five seeds, persistent heterogeneous participation showed no statistically distinguishable worst-class or excluded-class accuracy gap, but wide intervals do not establish a zero gap.Source: §VI · p. 7
  • The study uses one single-cell collision domain, fixed 24 Mbit/s physical-layer rate, and a compact LeNet-style model, limiting topology and deployment-time generalization.Source: §VI · p. 7
  • The trainer omits joint exclusions, temporal dependence, queues, complete-update delivery, and slowest-client completion; packetizing updates in ns-3 is proposed for independent Eq. (1) validation.Source: §VI · p. 7
Loading 2609.12903v1…