PaperlayerView exact revision on arXiv

Source-linked AI summary

A QoE-Driven Framework for O-RAN Network Slicing Based on Empirical Multi-User VR Traffic Modelling

Ahmed Chyad, David Naseh, Maryam Hafeez, Syed Ali Raza Zaidi

arXiv:2610.14771v1eess.SP

TL;DR

VR slicing must meet stringent QoE requirements, while existing QoE mapping, traffic provisioning, and physical-testbed validation leave empirical gaps. The paper uses fitted multi-user VR traffic in an O-RAN slicing framework; its over-the-air testbed reclaims PRBs from a neighbouring slice, and its analysis estimates admission limits across user counts and channel qualities.

  • Problem

    Existing QoE models often use traffic-agnostic heuristic penalties, slice provisioning relies on synthetic or single-user traffic, and much XR slicing validation has used simulation or emulated links.

    Source: §I · p. 1

  • Method

    The framework fits multi-user VR traffic distributions, uses the data-rate CDF for percentile-based QoE, and applies slice-level PRB quotas through a near-real-time RIC xApp.

    Source: §I · p. 1, D. Quality Mapping via the Fitted… · p. 8, §IX · p. 12

  • Results

    Up to 41% of committed PRBs were reclaimed from a greedy neighbour without reducing the protected VR-eMBB slice, while a fixed-rate synthetic source overstated its PRB footprint by 68%.

    Source: §IX · p. 13

  • Takeaways & Limitations

    The analytical capacity study locates an admission boundary across user counts and channel qualities, with graceful QoE degradation beyond it.

    Source: §IX · p. 13

    The analytical capacity envelope still awaits a concurrent-UE hardware campaign, and the SDR-limited delay floor is to be replaced by direct latency measurement.

    Source: §IX · p. 13

Abstract

from arXiv · show

Virtual Reality (VR) requires strict Quality-of-Experience (QoE) guarantees, making radio resource management harder. Static slice allocations are wasteful, while dynamic approaches often rely on heuristic penalty functions lacking an empirical basis. We introduce a QoE-driven network-slicing framework for Open RAN in which resource allocation is guided by empirically derived multi-user VR traffic statistics. Closed-form n-fold convolutions model the aggregate data rate and inter-arrival time of n concurrent users, validated empirically for up to four. The fitted data-rate cumulative distribution function is checked against the measured throughput to give a percentile-based satisfaction indicator; the inter-arrival law sets the slice targets. A near-real-time RIC xApp closes a loop over the E2SM-KPM and E2SM-RC interfaces to apply slice-level Physical Resource Block (PRB) quotas. On an over-the-air testbed it reclaims up to 41% of the cell's committed PRBs from a greedy neighbouring slice while never reducing the protected VR slice. An analytical study then estimates the PRB demand and per-user QoE across user counts and channel qualities, giving the admission boundary beyond which the target QoE is unattainable. A controlled comparison further shows that a fixed-rate synthetic load overstates the VR slice's PRB footprint by 68%, showing the necessity of data-driven slice dimensioning.

I. INTRODUCTION

The paper addresses gaps in empirically grounded QoE mapping, multi-user VR slice dimensioning, and over-the-air validation by deriving its quality model and control parameters from fitted VR traffic statistics. Its xApp reclaims resources from a neighbouring slice while protecting the VR slice.

I. INTRODUCTION

  • The framework derives a percentile-based QoE model and resource-control parameters from empirically fitted VR traffic distributions.It models aggregate inter-arrival time and data rate for concurrent users, with empirical validation across one to four users.Source: §I · p. 1
  • Up to 41% of committed PRBs were reclaimed from a greedy neighbour without reducing the protected VR-eMBB slice, while packet loss remained at 0.037%; fixed-rate load overstated the VR slice's PRB footprint by 68%.The loss was within the IEEE 3079-2020 budget.Source: §I · p. 2

A. VR Quality of Experience

  • The fitted VR data-rate CDF evaluated at measured throughput provides an interpretable, traffic-grounded QoE measure, unlike learned predictors that do not expose the traffic-statistical basis of their score.The controller also derives service-rate and scheduling-gap targets from the fitted model.Source: A. VR Quality of Experience · p. 2

B. Network Slicing for VR

  • A prior near-RT RIC xApp released resources while regulating VR latency, but used synthetic traffic over a wired 4G link and a heuristic control law.The present work instead uses empirically collected traffic on a 5G NR air interface and derives targets from fitted distributions.Source: B. Network Slicing for VR · p. 2

C. VR Traffic Modelling

  • Existing traffic models are single-user; this work derives and empirically validates closed-form n-fold convolutions for aggregate inter-arrival time and data rate across concurrent VR users.Prior measurements and standardized models do not characterize packet-level multi-user demand; simplified baseline models can also underestimate effective VR bitrate by roughly 25%.Source: C. VR Traffic Modelling · p. 2

III. VR TRAFFIC MODELLING AND MULTI-USER AGGREGATION

The section fits separate distributions to single-user packet inter-arrival times and data rates, then assumes independent, identically distributed users so their aggregate traffic can be modeled by convolution.

A. Single-User Distribution Fitting

  • The captured traffic has dense within-burst packet arrivals separated by much larger inter-burst gaps; its mean packet size is 1080.3 bytes and measured mean wire rate is 151.5 Mbps.The bursts carry about 110 packets on average, while the unconditional mean IAT includes idle gaps.Source: A. Single-User Distribution Fitting · p. 2
  • The packet IAT is modeled by a Generalized Extreme Value distribution, which gave the lowest DKL among candidate distributions.The fitted negative shape places it in the bounded Weibull domain, with within-burst support ending at 9.727 µs.Source: A. Single-User Distribution Fitting · pp. 2–3
  • The asymmetric data-rate distribution, with a heavier left tail, is best fitted by the Generalized Logistic distribution.The fitted parameters and derived convolution constants are listed in Table I.Source: A. Single-User Distribution Fitting · p. 3, Table I · p. 3

B. Aggregation Model

  • Concurrent users are assumed i.i.d. because a fully coupled model of user behavior and instantaneous radio state is computationally intractable; aggregate IAT and rate are sums modeled by n-fold convolution.The same convolution framework is specialized to the fitted IAT and data-rate laws.Source: B. Aggregation Model · p. 3

C. Generic Bounded-Support Convolution

The section uses a generic bounded-support n-fold convolution for both traffic quantities, then specializes it to the fitted GEV inter-arrival and GL data-rate laws.

C. Generic Bounded-Support Convolution

  • The generic n-fold convolution integrates over n−1 variables while enforcing each summand's bounded support, giving aggregate support [nℓ, nh] for either traffic quantity.At n = 4, the stated aggregate supports reach [0.004, 0.0389] ms for IAT and [280, 1000] Mbps for data rate.Source: C. Generic Bounded-Support Convolution · p. 3

D. Inter-Arrival Time Aggregation (GEV)

  • For GEV inter-arrival times, the two-user density is evaluated over the overlap of the single-user supports, and the non-factoring exponential term makes the integral numerical for n ≥2.The aggregate support is [ℓXn, hXn].Source: D. Inter-Arrival Time Aggregation (GEV) · p. 3

E. Data-Rate Aggregation (Generalized Logistic)

  • For GL data rates, a closed-form amplitude factor leaves a smooth product of logistic kernels, making its aggregate density cheaper to evaluate than the GEV density.Aggregate means scale linearly with user count, and the support bounds the attainable rate; slice sizing uses the aggregate 99th percentile rather than the endpoint.Source: E. Data-Rate Aggregation (Generalized… · pp. 3–4

IV. MULTI-USER EMPIRICAL TRAFFIC FITTING

The fitting results cover four VR-UE density scenarios and treat downlink data rate and packet inter-arrival time as separate traffic characteristics.

IV. MULTI-USER EMPIRICAL TRAFFIC FITTING

  • The statistical fitting results cover four UE-density scenarios, from one to four VR UEs.Source: §IV · p. 4
  • Downlink data rate is modeled as one of the two traffic characteristics.Source: §IV · p. 4
  • Packet inter-arrival time is modeled independently from downlink data rate.Source: §IV · p. 4

A. Data-Rate Distribution Fitting Results

Across one to four VR UEs, Generalized Logistic (GL) best fits downlink data rates, while the Generalized Extreme Value (GEV) distribution provides a flexible model for packet inter-arrival times. The section also establishes a per-PRB capacity model from 5G NR parameters for resource allocation.

A. Data-Rate Distribution Fitting Results

  • GL ranks first for downlink data-rate fitting in all four UE scenarios, with DKL values from 0.0095 (one UE) to 0.0682 (two UEs).Its symmetric, bell-shaped density and heavier tails accommodate variability from codec rate adaptation and scheduling fluctuations.Source: A. Data-Rate Distribution Fitting… · p. 4

B. Packet Inter-Arrival Time Distribution Fitting Results

C. Suitability of the GEV Model for VR IAT

  • GEV models IAT as competing jitter sources perturb quasi-periodic VR traffic, with its shape parameter spanning heavy-tailed, bounded, and light-tailed regimes.The regimes correspond to Fréchet, Weibull, and Gumbel behavior; the parameters also have physical readings for packet gap, jitter spread, and tail regime.Source: C. Suitability of the GEV Model for VR… · pp. 4–5

V. 5G NR PHYSICAL RESOURCE BLOCK CAPACITY

  • The per-PRB capacity model derives gross throughput from 5G NR parameters and subtracts FR1 downlink control overhead to obtain net rate.CQI determines the selected MCS and its modulation and code rate; Table IV evaluates representative MCS indices under the testbed configuration.Source: §V · p. 5, Table IV · p. 6

A. Integration with QoE-Driven Resource Allocation

The framework connects QoE targets and per-UE telemetry to slice-level PRB quotas through a near-real-time xApp. Its testbed uses an Open5GS core, OCUDU gNB, OSC RIC, and two UEs, with measured per-PRB throughput validating the capacity model.

A. Integration with QoE-Driven Resource Allocation

  • The xApp uses per-UE CQI to calculate the instantaneous PRB requirement and combines it with the multi-user traffic model for admission control.A new VR connection is rejected when the sum of its minimum PRB requirements exceeds the cell's PRB count; targets include jitter below 5 ms and packet loss below 1%.Source: A. Integration with QoE-Driven… · p. 6

A. System Architecture

  • The testbed comprises an Open5GS core, OCUDU gNB, OSC near-RT RIC running the xApp, and two UEs served by a USRP X310 radio.The priority VR-eMBB slice carries the VR flow, while a general best-effort eMBB slice carries competing load.Source: A. System Architecture · p. 6, Figure 3 · p. 6
  • Measured per-PRB throughput agrees with the capacity model to within 0.5% at MCS 31.The 100 MHz FR1 testbed uses numerology µ = 1 and 2 × 2 MIMO, selected within the USRP X310's capabilities.Source: A. System Architecture · p. 6

B. E2SM-KPM Monitoring

The xApp combines per-UE KPM measurements and scheduler telemetry to estimate VR service quality and radio efficiency, then sends slice-level PRB quotas through E2SM-RC. Its 1 Hz loop uses offline-fitted traffic parameters and online measurements to guide reallocation.

B. E2SM-KPM Monitoring

  • Four 1 Hz KPM subscriptions provide per-UE downlink throughput, CQI, uplink air-interface delay, and downlink RLC packet-drop rate.Throughput supplies both the reallocation trigger and quality-score input; uplink delay is recorded for compliance but is not a trigger.Source: B. E2SM-KPM Monitoring · p. 6

C. E2SM-RC Control

  • The xApp sends slice-level PRB quotas through E2SM-RC Control Action 6, while the DU's once-per-second KPM indications cap the loop at 1 Hz.The 1 Hz cadence is a platform limit, not a design choice.Source: C. E2SM-RC Control · pp. 6–7

VII. CROSS-LAYER XAPP METHODOLOGY

  • A near-RT RIC xApp connects offline-fitted VR traffic statistics to online telemetry, quality scoring, reallocation triggers, and E2SM-RC quota dispatch.The control loop links application-layer flow statistics to the gNB's physical-layer scheduling policy.Source: §VII · p. 7

A. Offline-Derived Control Parameters

  • The offline packet-IAT fit gives x0.99 ≈ 8.95 µs; the mean source-rate estimate is 151.5 Mbps and agrees with the fitted data-rate mean within 3.1%, while the slice target is 250 Mbps.The target is taken from the upper support of the fitted data-rate law, rather than its mean.Source: A. Offline-Derived Control Parameters · p. 7

B. Online Data Acquisition

  • Online acquisition pairs KPMs with 1 s scheduler telemetry to estimate per-PRB efficiency and mean inter-grant time from the downlink rate, PRB holding, and successful HARQ transmissions.The scheduler feed supplies PRB occupancy, MAC bitrate, HARQ counts, slot counts, MCS, CQI, and rank information.Source: B. Online Data Acquisition · p. 7

C. Radio-Parameter Engine: Per-PRB Efficiency Hierarchy

The xApp selects a per-PRB efficiency estimate from a gated three-tier hierarchy and maps measured throughput through the fitted data-rate CDF to a reported MOS score. The score is relative when offered load cannot reach the slice target and does not gate control.

C. Radio-Parameter Engine: Per-PRB Efficiency Hierarchy

  • A three-tier cascade selects per-PRB efficiency: self-calibration is load- and range-gated, fresh WebSocket telemetry is preferred, and theoretical CQI/MCS efficiency is the fallback.Tier 0 rejects partly filled grants and stale high estimates; Tier 1 dominates in the reported runs.Source: C. Radio-Parameter Engine: Per-PRB… · pp. 7–8

D. Quality Mapping via the Fitted Data-Rate CDF

E. Trigger Evaluation

The xApp reallocates PRBs when VR throughput, grant regularity, or packet-loss thresholds are breached, and reclaims excess allocation when either slice exceeds its target margin. Cooldowns allow throughput to settle between control decisions.

E. Trigger Evaluation

  • VR reallocation triggers when its 10 s average throughput falls below 250 Mbps, its sustained-traffic grant interval exceeds Gmax, or its RLC drop rate exceeds 1%.The grant-gap gate requires fresh WebSocket data and sustained traffic; over-provision reclaim fires at 110% of target for either slice.Source: E. Trigger Evaluation · p. 9
  • A 15 s VR cooldown and 5 s eMBB cooldown let windowed throughput reflect quota changes before another decision, limiting oscillation and control signalling.The 110% over-provision margin was chosen as the smallest margin that reliably keeps eMBB throughput from falling below target after reallocation.Source: E. Trigger Evaluation · p. 9
  • Air-interface delay is excluded as a trigger because the measured uplink delay has a fixed 7–8 ms processing floor; Gmax is set to 2 ms to test grant regularity between healthy and frame-rate bounds.The 2 ms bound exceeds the reported ≈0.84 ms nominal inter-grant interval and stays below the 8.33 ms frame period.Source: E. Trigger Evaluation · p. 9

F. Priority-Based PRB Allocation and Control Dispatch

The allocator converts rate targets to PRB counts using current efficiency estimates, reserves resources for VR-eMBB before serving eMBB, and dispatches slice quotas through E2SM-RC. In the contended UC3 test, it returned 41.1% of committed PRBs to the free pool without reducing the protected VR slice.

F. Priority-Based PRB Allocation and Control Dispatch

  • The allocator reserves required VR-eMBB PRBs first and serves eMBB from the remainder; guaranteed quotas round up and over-provision ceilings floor conservatively at the PRB-count stage.Under genuine capacity shortage, the lower-priority eMBB slice yields rather than degrading VR-eMBB below target.Source: F. Priority-Based PRB Allocation and… · pp. 9–10

VIII. EXPERIMENTAL RESULTS AND DISCUSSION

  • The four hardware-testbed cases compare isolated VR, a disciplined CBR neighbour, a greedy 500 Mbps UDP neighbour, and fixed-rate 250 Mbps VR traffic with the greedy neighbour.The fixed-rate case explores the oversubscribed regime.Source: §VIII · p. 10

A. PRB Optimisation Framework

  • In UC3, the xApp returned 99 PRBs (41.1% of committed allocation) to the free pool, cutting the greedy slice from 184 to 53 PRBs while raising VR-eMBB from 57 to 89.The protected VR-eMBB slice was never reduced in the contended cases.Source: Figure 6 · p. 10, A. PRB Optimisation Framework · p. 10
  • UC2's disciplined neighbour required no reclaim (98→101 total PRBs), while UC4 freed 84 PRBs (35.6%) under oversubscription.UC4 reduced total allocation from 236 to 152 PRBs; the greedy neighbour fell from 140 to 41 PRBs while VR-eMBB rose from 96 to 111.Source: A. PRB Optimisation Framework · p. 10

B. Realistic Replay versus Synthetic Load

Empirical VR replay uses fewer PRBs than a fixed-rate synthetic load, while testbed results show quota control preserving VR delivery as resources are reclaimed. An analytical capacity envelope estimates the PRBs needed across user counts and channel qualities, with multi-user hardware validation still pending.

B. Realistic Replay versus Synthetic Load

  • 68.4% larger PRB footprint under the 250 Mbps synthetic load than under empirical VR replay, despite essentially similar per-PRB efficiency.The replay occupied 57 PRBs and the synthetic load 96 PRBs before xApp control.Source: B. Realistic Replay versus Synthetic… · pp. 10–11

C. Testbed Validation and Stability

  • Under xApp control, VR delivery ratio was ∼0.98 with ∼0.3 ms jitter, including the greedy-neighbour case UC3.The larger MCS spread in contended UC3 reflected quota-driven changes in PRB grants rather than channel noise.Source: C. Testbed Validation and Stability · p. 11

D. Multi-User Capacity under Varying Channel Quality

E. Analytical and Perceptual QoE Estimation under Varying User Count and Channel Quality

The analysis maps user count and channel quality to PRB demand and per-user QoE, locating where the VR slice begins to degrade. The conclusion reports over-the-air resource recovery and shows that fixed-rate synthetic traffic misstates the slice footprint.

E. Analytical and Perceptual QoE Estimation under Varying User Count and Channel Quality

  • The model assigns each user Rreq while aggregate PRB demand is feasible, then assumes equal sharing after saturation as graceful degradation.The equal-share alternative contrasts with admission control that would refuse excess users while keeping admitted users near MOS ≈5.Source: E. Analytical and Perceptual QoE… · p. 12
  • At CQI 10, MOS was 5.0, 4.9, 2.2, and 1.3 for n = 1 to 4, as per-user rates fell to 250, 215, 144, and 108 Mbps.The QoE knee moves to higher CQI as user count grows; before saturation, MOS remains near 5.Source: E. Analytical and Perceptual QoE… · p. 12
  • The steep knee is a model boundary, not a transient forecast: FR(Rreq) ≈0.997 and stepwise feasibility compress the transition.Real link adaptation, retransmissions, and scheduler cooldown spread the transition across control cycles.Source: Figure 8 · p. 12
  • The analytical surface is empirically grounded at one measured point: absolute MOS is meaningful only at UC4, while UC1–UC3 use a relative signal.For UC1–UC3, VR quality is assessed through delivery ratio, loss, and latency measures.Source: E. Analytical and Perceptual QoE… · p. 12

IX. CONCLUSION

  • 41% of committed PRBs were reclaimed from a greedy neighbour while packet loss stayed at 0.037%; fixed-rate synthetic traffic overstated the VR slice’s PRB footprint by 68%.The analytical study also located the admission boundary across user counts and channel qualities.Source: §IX · p. 13
  • The capacity envelope still needs concurrent-UE hardware validation; multi-cell, handover, and broader VR cases remain future work.The paper also identifies replacing the SDR-limited delay floor with direct latency measurement as future work.Source: §IX · p. 13
Loading 2610.14771v1…