Pith. sign in

REVIEW 4 major objections 8 minor 54 references

The paper claims that a production cellular MAC scheduler can be decomposed into pluggable policy stages, letting the same AI scheduling applications run in a PHY-less real-time emulator and over the air—demonstrated by a proactive schedule

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 00:53 UTC pith:2EHHDVVO

load-bearing objection Genuinely useful framework — pluggable OAI scheduler plus real-time L2 emulator — but the headline gains and the emulator validation lean on fitted channel models; the frequency-selective use case is emulator-internal until an OTA check. the 4 major comments →

arxiv 2607.26012 v2 pith:2EHHDVVO submitted 2026-07-28 cs.NI

MAC-Gyver: Open, Programmable, Scheduling for AI-RAN 6G Systems

Maxime Elkael , Reshma Prasad , Tamerlan Aghayev , Salvatore D'Oro , Michele Polese , Tommaso Melodia This is my paper
classification cs.NI
keywords AI-RANMAC schedulerOpenAirInterfacenetwork emulationfrequency-selective schedulingproactive schedulingURLLCscheduler dApps
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

MAC-Gyver aims to close the gap between simulated and deployed AI schedulers in cellular networks. It refactors the per-slot Medium Access Control (MAC) scheduler in the OpenAirInterface (OAI) stack into a fixed pipeline of replaceable policy stages, called Layer 2 distributed applications (dApps), which share a per-UE candidate structure carrying channel observations and decisions. A companion PHY-less emulator, mac-emu, runs the unmodified OAI Layer 2 stack for up to 90 users on one host at real-time slot pace with a 3GPP-compliant channel model, so the same policy implementation executes in emulation and over the air without a second rewrite. Using this framework, the paper shows a proactive uplink scheduler that predicts packet arrivals and roughly halves median round-trip time over the air, and a frequency-selective uplink scheduler whose gains (up to 7% aggregate throughput) depend on channel freshness and power-limited operation. A central lesson is that offline optimization favored a more flexible allocator, but in the complete scheduling loop the simpler policy performed better—an outcome the paper uses to argue for evaluating scheduling algorithms inside a production stack.

Core claim

MAC-Gyver establishes that a production RAN MAC scheduler can be made programmable without reimplementing the protocol stack: its scheduler is a fixed pipeline of typed decision stages (rank/precoder, beam, time-domain allocation, MCS, PRB allocation, LCID multiplexing), and each stage is a replaceable dApp that reads the same candidate structure as the default policy. The same dApps run over the air and inside mac-emu, which replaces only the PHY with a calibrated channel model while preserving the real-time slot clock, protocol constraints, and FAPI boundary. With this interface, the paper demonstrates two working schedulers: a proactive UL scheduler that uses observer and candidate-reques

What carries the argument

The key mechanism is the decomposition of the scheduler into a fixed pipeline of replaceable policy stages sharing a per-UE candidate structure (nr_dl_candidate_t / nr_ul_candidate_t). This structure carries channel observations (wideband CQI, per-PRB SRS energy), buffer state, HARQ status, and the decision fields written by each stage, so every dApp sees the same enriched state. Three additional dApp interfaces—candidate request, DL transmission observer, and UL observer—plus a persistent stateful context allow policies to act across scheduling opportunities. The other load-bearing component is mac-emu, which bypasses the PHY and swaps FAPI messages over UDP between unmodified gNB and UE La

Load-bearing premise

The frequency-selective claims assume mac-emu's fitted channel abstraction—Equations (1)-(4) with Richards curves calibrated offline and a hardware-measured error floor—preserves the frequency-selective structure and its Doppler-driven decorrelation of a real radio channel, and those claims are never checked over the air; moreover, the emulator validation is weakened because path loss is tuned per operating point to match the same hardware MCS distributions being compared.

What would settle it

Run the frequency-selective scheduler comparison (FREQSEL-FIXED vs FREQSEL-REFINED vs proportional fairness) over the air or through a hardware channel emulator at matched path loss, Doppler, and SRS periodicity, and compare the aggregate UL throughput ordering and magnitudes to mac-emu's predictions; if FREQSEL-REFINED wins or the emulator's predicted 7% gain does not appear on hardware, the fitted channel model does not preserve the scheduling-relevant frequency selectivity.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A scheduling policy written once as a dApp can be evaluated in controlled, repeatable, cell-scale emulation and then deployed over the air with no second implementation, removing a major barrier between simulation-based AI research and production RAN experimentation.
  • Proactive scheduling based on online-observed MAC events can cut the SR-induced latency component roughly in half, approaching always-on scheduling at 7% overhead, which could benefit URLLC-style traffic with unknown, event-driven arrivals that configured grants cannot handle.
  • Frequency-selective uplink allocation delivers up to 7% aggregate throughput gain over proportional fairness, but only when SRS-derived channel estimates remain fresh; the gain shrinks with mobility and disappears when all users move, quantifying the freshness requirement for sub-band scheduling.
  • Offline allocation-quality simulation can rank policies in the wrong order: FREQSEL-REFINED looked better in isolation, but FREQSEL-FIXED won in the complete scheduler even with oracle MCS, implying that scheduling research should validate algorithmic refinements inside a full protocol stack.
  • mac-emu can sustain a 40 MHz cell with 90 concurrent UEs on a single host at real-time slot pace, with average per-slot execution below 500 microseconds and deadline-violation probabilities around 10^-3, making large-scale controlled RAN experiments accessible without radio hardware.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the frequency-selective throughput results (including the up-to-7% gain and the simpler-allocator-wins reversal) were produced entirely inside mac-emu and never validated over the air, a natural next step is to reproduce the FREQSEL-FIXED versus FREQSEL-REFINED comparison on a hardware channel emulator; if the ordering or gain magnitude changes, the emulator's frequency selectivity is not
  • The paper's own validation hints that per-point path-loss calibration against hardware MCS distributions can mask systematic emulator bias; an independent check would fix a single path-loss model from measured link budgets rather than tuning per operating point, and then compare MCS, throughput, and BLER across a diverse set of channel profiles.
  • The PUCCH DTX model—calibrated to hardware and dependent on delay spread—suggests that control-channel impairments, not just data-channel BLER, shape scheduler behavior; future emulator designs should include similar control-channel abstractions when the research question involves HARQ feedback or retransmission dynamics.
  • The architecture's typed, stateful interfaces could support more sophisticated online learners (neural predictors, bandit-style MCS selection) at per-slot timescales, since the paper demonstrates only EWMA and linear-regression predictors; the framework's persistent context is the natural place to test such models, but their overhead and convergence remain unmeasured.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 8 minor

Summary. The paper presents MAC-Gyver, an open-source framework that exposes the OpenAirInterface (OAI) MAC scheduler as a pipeline of replaceable typed policy stages ('dApps'), together with mac-emu, a PHY-less emulator that runs the unmodified OAI Layer 2 stack for up to 90 users on one host at real-time slot pace. The claimed contributions are: (i) a pluggable per-slot scheduling interface in a production RAN stack, upstreamed to OAI; (ii) a real-time emulator with a 3GPP-based channel model; (iii) a proactive UL scheduler validated over the air that roughly halves median RTT; and (iv) a frequency-selective UL scheduler evaluated in mac-emu that shows a 5–7% throughput gain over proportional fairness and a reversal of the offline policy ordering (FREQSEL-FIXED outperforms the offline-favored FREQSEL-REFINED in the full scheduling loop).

Significance. If the central claims hold, this is a significant engineering contribution: it provides a genuinely reusable, production-stack scheduling interface with a matching emulation environment, and it demonstrates the value of evaluating scheduling policies inside a full protocol stack rather than in isolation. The paper ships released code, cites an upstreamed OAI merge request, includes concrete scaling data (average slot-thread time under 500 µs at 90 UEs, deadline-violation probability ~1e-3), and provides OTA results for the proactive scheduler. The frequency-selective use case is also a good example of how offline analysis can mislead and why integrated evaluation matters. However, the strength of the headline frequency-selective results is limited by the emulator's channel-model fidelity, which is not fully validated on hardware for the per-PRB SRS structure that the policy exploits.

major comments (4)
  1. [§III.C (Eqs. 1–4)] The channel model used for ACK/NACK decisions is fitted entirely to OAI's own link-level simulators (nr_dlsim/nr_ulsim) plus a constant error floor φ measured at high SNR on hardware. This means the BLER curves reproduce OAI receivers, not commercial UE receivers or the hardware channel. The paper claims a '3GPP-compliant channel model,' but the compliance refers to the fading profiles, not to the resulting error behavior. This is a limitation that should be stated more prominently; the emulator-vs-hardware agreement in Figs. 5–6 is then partly a test of OAI-against-OAI with a fitted floor.
  2. [§III.D (path-loss calibration)] The validation of mac-emu against hardware calibrates mac-emu's path loss per operating point to match the hardware MCS distribution (Wasserstein distance). This is fit-on-target: the comparison in Figs. 5–6 is then not an independent test of the channel model's ability to predict hardware behavior. The authors acknowledge this in the text, but the conclusion that 'mac-emu reproduces the hardware setup' is overstated. The validation only covers single-UE DL wideband metrics; it does not validate the per-PRB SRS channel structure (frequency correlation, Doppler decorrelation) that the FREQSEL scheduler exploits.
  3. [§III.C (Eqs. 6–7)] The PUCCH DTX model parameters (W, σ, t_corr) are calibrated by maximum likelihood against the same hardware DTX rates they are then reported to match. The resulting agreement (0.07%/0.28%/5.4% vs 0.06%/0.77%/5.5%) is therefore not an independent validation. Also, the model injects detector misses independently of the instantaneous channel state; the paper notes the discrepancy on TDL-C300 at low MCS but dismisses it as 'no visible effect.' Since low-MCS is exactly the regime where power-limited UEs operate, this should be quantified rather than dismissed.
  4. [§V.C (Figs. 20–24)] All frequency-selective results — including the 5–7% UL throughput gain and the FREQSEL-REFINED-vs-FREQSEL-FIXED reversal — are produced entirely inside mac-emu. The SRS synthesis is only checked against OAI's nr_srssim (RMSE 0.36 dB), not against hardware. The channel's frequency-selectivity and Doppler behavior, which are the very mechanisms the FREQSEL policy exploits, are never validated OTA. The paper's own validation (§III.D) tests only single-UE DL behavior. An OTA run of the FREQSEL policy would be the natural check; without it, the reversal claim is an emulator-internal result, not a general system conclusion.
minor comments (8)
  1. [Abstract] The phrase 'with a 3GPP-compliant channel model' should be qualified to indicate that the fading profiles follow 3GPP TR 38.901 but the error model is fitted to OAI link-level simulations plus a hardware error floor.
  2. [§II.B (Listing 1)] The field 'srs_rb_energy' is described as 'per-RB channel energy from SRS,' but §V.B says the allocator reads 'complex SRS estimates.' The struct comment should be updated to reflect that complex estimates are exposed, or the text should align.
  3. [§III.C] Typo: 'Y AML' should be 'YAML'. Also, 'accross' appears in §II.B and §VI.A; should be 'across'.
  4. [§III.D] The sentence 'median throughput agrees to −4% (all points within 13%...)' lacks a closing parenthesis and is hard to parse. Also, the 4% PRB-mismatch explanation (SSB placement) is plausible but could be verified by presenting a PRB-count comparison.
  5. [§V.B] The offline policy evaluation (Fig. 19) uses a static channel and equal PF weights; this is useful but limited. The authors should state that the offline comparison does not include frequency-domain correlation or Doppler, which are the dimensions that later affect the reversal.
  6. [§V.C (Fig. 20)] The oracle-MCS experiment uses 'highest MCS whose predicted BLER does not exceed 13%,' but the motivation for 13% is not given. Please provide a justification or a reference.
  7. [§IV.D] The paper disables the Confident-to-Learning fallback during evaluation and reports results in the Confident state only. This is a reasonable isolation, but the reported 'overhead' of 7% at (TW=0,N=1) should be contextualized against the always-on upper bound, which the Learning state represents.
  8. [References] Reference [12] is cited as an upstreamed OAI merge request; it should be verified to be publicly accessible at the time of publication, as it is a key reproducibility claim.

Circularity Check

3 steps flagged

Emulator validation contains fit-on-target steps (DTX-rate calibration, per-point path-loss matching, SRS-noise fit), but the central OTA and in-loop scheduler results are not reduced to their inputs; no load-bearing self-citation chain.

specific steps
  1. fitted input called prediction [Section III.C, Channel Model, Equations (6)-(7) and following paragraph]
    "We calibrate the model against the DTX statistics measured on the OAL-based testbed: W and σ by maximum likelihood on the per-channel DTX rates, and t_corr on the burstiness of the resulting retransmissions. We obtain W=475 ns, σ=140 ns, and t_corr=100 ms, which yields DTX rates of 0.07%/0.28%/5.4% on TDL-A30/-B100/-C300, against the 0.06%/0.77%/5.5% measured on the hardware testbed."

    W and σ are selected by maximum likelihood on the per-channel DTX rates, so the reported 'yields ... against ... measured' is a goodness-of-fit of the calibrated model to its own calibration targets. The agreement in marginal DTX rate is therefore true by construction and does not independently validate the PUCCH DTX submodel.

  2. fitted input called prediction [Section III.D, Validation, first and second paragraphs]
    "we calibrate mac-emu's path loss per operating point, rerunning the experiment to find the closest MCS-distribution match with the hardware setup (via the Wasserstein distance). ... Across the 12 points, the median Wasserstein distance between the pooled MCS distributions is 0.35 MCS steps (worst 0.94)"

    The MCS-distribution agreement is the objective used to pick the per-point path-loss parameter: the emulator is tuned so that its MCS distribution matches hardware, and that same agreement is then reported as evidence of equivalence. Throughput and BLER comparisons inherit this calibration, so the closed-loop validation is partly constructed rather than an independent prediction.

  3. fitted input called prediction [Section III.C, Channel Model, SRS synthesis paragraph]
    "To model this error, we measure it offline with the SRS receiver in isolation (OAI's nr_srssim) and inject noise of matching variance into the synthesized matrix, which is derived from the current channel state; the calibrated synthesis matches the simulator with an RMSE of 0.36 dB."

    The SRS synthesis is calibrated to match nr_srssim and then reported as matching nr_srssim; this is a fit to the same simulator used as ground truth. The frequency-selective scheduler consumes this per-PRB SRS structure, so the check establishes internal consistency with the simulator but not validated fidelity against real hardware.

full rationale

The paper's main derivation chain is not circular in the strongest sense. MAC-Gyver's staged scheduler architecture is implemented and upstreamed, and the proactive-scheduling use case is measured over the air on a real testbed; the FREQSEL-FIXED-vs-REFINED reversal is an emergent outcome of the complete OAI scheduling loop under a fixed CDL-B channel, not a parameter equal to its input. The self-citations [12]-[14] describe contributed code, prior dApp work, and a derivative system; they do not carry the load-bearing argument, and no uniqueness theorem is imported from the authors. However, several validation steps are fit-on-target: the DTX model parameters are maximum-likelihood fitted to the DTX rates that are then reported as an agreement, the Section III.D closed-loop validation calibrates mac-emu's path loss per operating point to minimize the Wasserstein distance to the hardware MCS distribution and then reports that distance as the match, and the SRS-noise calibration matches the same simulator it was fit to. These steps make the emulator-equivalence claim partly constructed. They do not, by themselves, produce the headline frequency-selective throughput gains or the OTA latency reductions, so the central claims retain independent content. The AWGN-fitted Richards curves and hardware error floor are model parameters; using them inside mac-emu is not circular unless the target result is the same fitted value, which is not the case for the scheduler-loop outcomes in Section V. Overall, partial circularity in the emulator validation, but not a collapse of the central derivation.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The paper introduces no physical entities (no new fields/particles/forces); its inventions are software abstractions (dApps, candidate structure, mac-emu). The emulator reads its entire channel behavior from a chain of fitted quantities — 6,148 Richards curves from OAI link sims, a hardware error floor, and a three-parameter DTX model — so the '3GPP-compliant channel model' is a parameterized post-fit abstraction, not an independent first-principles model. The central framework claim does not depend on these fits; the frequency-selective use case does.

free parameters (6)
  • Richards BLER curve parameters (a,b,c) = one curve per direction, MCS, and PRB count (6,148 curves for the 40 MHz config)
    Eq. (3) fitted offline to AWGN waterfalls from OAI's nr_dlsim/nr_ulsim link-level receivers; they define mac-emu's ACK/NACK outcome and CQI synthesis, so every emulator scheduling result inherits them.
  • Error floor φ = 1.2e-5
    Section III.C: measured average error floor on the OAL hardware chain (9.6M + 7.2M blocks) and applied globally in Eq. (4). Hardware-derived, but it is an emulator input with a single fixed value.
  • PUCCH DTX parameters W, σ, t_corr = W=475 ns, σ=140 ns, t_corr=100 ms
    Section III.C: fit by maximum likelihood to the same hardware DTX statistics (TDL-A30/B100/C300) they are then compared against — a fit-on-target validation.
  • Per-operating-point path-loss calibration = chosen per operating point (16/22/28/34 dB × 3 TDL profiles)
    Section III.D: mac-emu path loss is tuned so its MCS distribution best matches the hardware (Wasserstein distance); the agreement in Figs. 5–6 is thus partly constructed, not predicted.
  • Proactive scheduler state-machine thresholds = 3 consecutive misses, 10% false-positive rate, 100 idle slots
    Section IV.C: operator-chosen thresholds trading overhead against the latency guarantee; not swept or optimized in the paper.
  • Learner parameters (EWMA / linear regression) = not reported
    Section IV.C: the EWMA smoothing factor and regression window are not stated, so the latency/overhead trade-off points cannot be recomputed exactly by a third party.
axioms (5)
  • domain assumption 3GPP TR 38.901 TDL/CDL channel models adequately represent real multipath channels for evaluating schedulers
    Section III.C: mac-emu's frequency selectivity and mobility behavior — and therefore the frequency-selective scheduler results — come from these generative models.
  • domain assumption TDD channel reciprocity: DL and UL share one channel realization per UE
    Section III.C: 'As the system is focused on TDD, we assume that the channel is reciprocal.' Affects UL scheduling conclusions drawn from shared realizations.
  • standard math EESM with per-MCS β_m from [24] is an accurate compression of per-subcarrier SNR into one effective SNR
    Section III.C Eq. (2): the entire link abstraction rests on this prior calibration carried from the literature.
  • domain assumption OAI's L2 implementation (RRC/PDCP/RLC/MAC) is correct enough that policies valid there are valid in production
    The paper's central claim is that dApps run in 'the production stack'; OAI's own correctness is taken as given.
  • domain assumption Per-UE channel realizations are independent with the specified synthetic mobility models
    Sections III.C and V.C: multi-UE cell behavior and the mobility results assume these generative models (quasi-static, fixed-speed, Lévy-walk).
invented entities (1)
  • PUCCH timing-error Gauss–Markov state n_i no independent evidence
    purpose: Reproduces the observed burstiness of missed HARQ feedback (DTX) in mac-emu
    Introduced in Section III.C (Eqs. 6–7) solely to match hardware DTX statistics; no physical measurement validates the state dynamics beyond the two fitted rates and one correlation time.

pith-pipeline@v1.3.0-alltime-deepseek · 74 in / 19139 out tokens · 164701 ms · 2026-08-01T00:53:59.973117+00:00 · methodology

0 comments
read the original abstract

Cellular networks are integrating Artificial Intelli- gence (AI) into radio access network control. The MAC scheduler is a promising target because it allocates a limited resource, spectrum, at every slot, under competing latency, throughput, and reliability requirements. However, most learning-based sched- ulers are evaluated only in simulation. Production schedulers are difficult to modify, and realistic stress tests require more radio hardware than most laboratories can provide. We present MAC-Gyver, an open-source framework for developing and evaluating scheduling applications that execute directly inside the OpenAirInterface scheduler. It exposes scheduler observations and controls through typed interfaces while preserving the underlying protocol and real-time execution paths. The same applications run over the air and in mac-emu, a PHY-less emulator that executes the unmodified OpenAirInterface Layer 2 stack for up to 90 users on one host at real-time slot pace, with a 3GPP-compliant channel model. To showcase the flexibility of MAC-Gyver, we evaluate two use cases. A proactive uplink scheduler predicts packet arrivals and roughly halves median round-trip latency. A frequency-selective uplink scheduler selects contiguous sub-bands from per-PRB sounding observations and is evaluated across mobility and power-limited operating points against an offline scheduling ceiling. Together, they show how the same production stack can be an AI playground that supports implementation, controlled evaluation, and over-the-air validation through complementary scheduling use cases.

Figures

Figures reproduced from arXiv: 2607.26012 by Maxime Elkael, Michele Polese, Reshma Prasad, Salvatore D'Oro, Tamerlan Aghayev, Tommaso Melodia.

Figure 1
Figure 1. Figure 1: MAC-Gyver dApp architecture. Numbered decision [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: OAL hardware reference setup. of allocated PRBs; the outcome is one Bernoulli draw per transport block. Parameters (a, b, c) are fitted offline for each link direction (DL or UL), MCS, and number of PRBs allocated to the transmission. We do this using link-level simulations of the OAI UE and gNB receivers. OAI exposes each receive chain in isolation as a link-level simulator, nr_dlsim for the UE’s Physical… view at source ↗
Figure 2
Figure 2. Figure 2: mac-emu architecture. Unmodified OAI L2 stacks run to their FAPI boundary; adapters and the mac-emu runtime bypass the PHY. The software slot clock paces the gNB and the UE processes; the purple arrow carries the Layer 1 measurements synthesized by the channel model. C. Channel Model mac-emu abstracts waveform processing but must repro￾duce its consequences on the scheduler. For each scheduled transport bl… view at source ↗
Figure 4
Figure 4. Figure 4: High-SNR initial-transmission error floor across [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Closed-loop link adaptation in mac-emu and on OAL, one UE, saturated DL traffic. 10−2 10−1 TDL-A30 BLER 16 dB 22 dB 28 dB 34 dB 10−2 10−1 TDL-B100 BLER 0 10 20 10−2 10−1 MCS (initial TX) TDL-C300 BLER 0 10 20 MCS (initial TX) 0 10 20 MCS (initial TX) 0 10 20 MCS (initial TX) OAL (hardware) mac-emu twin [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Initial-transmission BLER versus MCS for the same runs as Figure 5. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Aggregate cell throughput versus concurrent UE count. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Scheduling request procedure. 0 5 10 15 20 25 30 35 40 0 0.2 0.4 0.6 0.8 1 RTT (ms) Default SR-based scheduling Always scheduling in UL [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: RTT with and without always-on UL scheduling. [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: RTT with a DL-aware proactive scheduler at varying [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: The deployment is built on the X5G testbed [17] [PITH_FULL_IMAGE:figures/full_fig_p010_13.png] view at source ↗
Figure 13
Figure 13. Figure 13: Over-the-air testbed for proactive-scheduler evalua [PITH_FULL_IMAGE:figures/full_fig_p011_13.png] view at source ↗
Figure 15
Figure 15. Figure 15: Request-response overhead; bottom labels show [PITH_FULL_IMAGE:figures/full_fig_p012_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: End-to-end latency for the size-dependent edge infer [PITH_FULL_IMAGE:figures/full_fig_p012_16.png] view at source ↗
Figure 19
Figure 19. Figure 19: Simulation-based expected cell spectral efficiency for [PITH_FULL_IMAGE:figures/full_fig_p014_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Aggregate UL throughput in mac-emu. static (0.5 Hz) 1 km/h 5 km/h 31 32 33 34 35 UE mobility UL throughput (Mbit/s) PF FREQSEL-FIXED [PITH_FULL_IMAGE:figures/full_fig_p015_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Aggregate UL throughput versus user speed. [PITH_FULL_IMAGE:figures/full_fig_p015_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Aggregate UL throughput. 0.85 0.9 0.95 1 0 0.2 0.4 0.6 0.8 1 per-UE UL throughput (Mbit/s) CDF PF, static (0.5 Hz) FREQSEL-FIXED, static (0.5 Hz) PF, mobile FREQSEL-FIXED, mobile [PITH_FULL_IMAGE:figures/full_fig_p016_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Per-user UL throughput in a cell with 16 quasi-static [PITH_FULL_IMAGE:figures/full_fig_p016_23.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 2 linked inside Pith

  1. [1]

    Ai-ran: Transforming ran with ai-driven computing infrastructure,

    L. Kundu, X. Lin, R. Gadiyar, J.-F. Lacasseet al., “Ai-ran: Transforming ran with ai-driven computing infrastructure,”IEEE Communications Magazine, 2025

    2025

  2. [2]

    NR; Medium Access Control (MAC) protocol specification,

    3GPP, “NR; Medium Access Control (MAC) protocol specification,” 3rd Generation Partnership Project (3GPP), Technical Specification (TS) 38.321, 2024, release 18, V18.x

    2024

  3. [3]

    Joint Scheduling of URLLC and eMBB Traffic in 5G Wireless Networks,

    A. Anand, G. de Veciana, and S. Shakkottai, “Joint Scheduling of URLLC and eMBB Traffic in 5G Wireless Networks,”Trans. on Net., vol. 28, no. 2, pp. 477–490, 2020

    2020

  4. [4]

    Proportional fair frequency-domain packet scheduling for 3GPP LTE uplink,

    S. B. Lee, I. Pefkianakis, A. Meyerson, S. Xuet al., “Proportional fair frequency-domain packet scheduling for 3GPP LTE uplink,” inProc. IEEE INFOCOM, 2009, pp. 2611–2615

    2009

  5. [5]

    Learn to Schedule (LEASCH): A Deep Reinforcement Learning Approach for Radio Resource Schedul- ing in the 5G MAC Layer,

    F. Al-Tam, N. Correia, and J. Rodriguez, “Learn to Schedule (LEASCH): A Deep Reinforcement Learning Approach for Radio Resource Schedul- ing in the 5G MAC Layer,”IEEE Access, 2020

    2020

  6. [6]

    From simulation to practice: Gen- eralizable deep reinforcement learning for cellular schedulers,

    P. Kela, B. Liu, and A. Valcarce, “From simulation to practice: Gen- eralizable deep reinforcement learning for cellular schedulers,”arXiv preprint arXiv:2411.08529, 2024

    arXiv 2024

  7. [7]

    ColO-RAN: Developing machine learning-based xApps for open RAN closed-loop control on programmable experimental platforms,

    M. Polese, L. Bonati, S. D’Oro, S. Basagniet al., “ColO-RAN: Developing machine learning-based xApps for open RAN closed-loop control on programmable experimental platforms,”IEEE Transactions on Mobile Computing, vol. 22, no. 10, pp. 5787–5800, 2023

    2023

  8. [8]

    dApps: Distributed applications for real-time inference and control in O-RAN,

    S. D’Oro, M. Polese, L. Bonati, H. Chenget al., “dApps: Distributed applications for real-time inference and control in O-RAN,”IEEE Communications Magazine, vol. 60, no. 11, pp. 52–58, 2022

    2022

  9. [9]

    Real: Reinforcement learning-enabled xapps for experimental closed-loop optimization in o- ran with osc ric and srsran,

    R. Barker, A. E. Dorcheh, T. Seyfi, and F. Afghah, “Real: Reinforcement learning-enabled xapps for experimental closed-loop optimization in o- ran with osc ric and srsran,” inICC Workshops, 2025

    2025

  10. [10]

    dApps: Enabling real-time AI-based open RAN control,

    A. Lacava, L. Bonati, N. Mohamadi, R. Gangulaet al., “dApps: Enabling real-time AI-based open RAN control,”Com. Net., 2025

    2025

  11. [11]

    The ns-3 network simulator,

    G. F. Riley and T. R. Henderson, “The ns-3 network simulator,” in Modeling and tools for network simulation. Springer, 2010, pp. 15–34

    2010

  12. [12]

    Refactor DLSCH/ULSCH scheduler: Extract the proportional-fair policy behind a function-pointer interface,

    M. Elkael, “Refactor DLSCH/ULSCH scheduler: Extract the proportional-fair policy behind a function-pointer interface,” OpenAirInterface5G merge request !3902, 2026, merged 2026-05-

    2026

  13. [13]

    Auguste: Online-learning dapp for predictive urllc scheduling,

    M. Elkael, M. Polese, Y . Lee, K. Furuedaet al., “Auguste: Online-learning dapp for predictive urllc scheduling,”arXiv preprint arXiv:2606.03664, 2026

    Pith/arXiv arXiv 2026

  14. [14]

    Agentran: An agentic ai architecture for autonomous control of open 6g networks,

    M. Elkael, S. D’Oro, L. Bonati, M. Poleseet al., “Agentran: An agentic ai architecture for autonomous control of open 6g networks,”IEEE Communications Magazine, 2026

    2026

  15. [15]

    Characterizing and modeling mobile networks user traffic at millisecond level,

    P. F. P ´erez, C. Fiandrino, and J. Widmer, “Characterizing and modeling mobile networks user traffic at millisecond level,” inProceedings of the 17th ACM Workshop on Wireless Network Testbeds, Experimental evaluation & Characterization, 2023, pp. 64–71

    2023

  16. [16]

    Network simulations with the ns-3 simulator,

    T. R. Henderson, M. Lacage, G. F. Riley, C. Dowellet al., “Network simulations with the ns-3 simulator,”SIGCOMM demonstration, vol. 14, no. 14, p. 527, 2008. 18

    2008

  17. [17]

    X5G: An Open, Programmable, Multi-vendor, End-to-end, Private 5G O-RAN Testbed with NVIDIA ARC and OpenAirInterface,

    D. Villa, I. Khan, F. Kaltenberger, N. Hedberget al., “X5G: An Open, Programmable, Multi-vendor, End-to-end, Private 5G O-RAN Testbed with NVIDIA ARC and OpenAirInterface,”IEEE Transactions on Mobile Computing, 2025

    2025

  18. [18]

    Colosseum: Large- scale wireless experimentation through hardware-in-the-loop network emulation,

    L. Bonati, P. Johari, M. Polese, S. D’Oroet al., “Colosseum: Large- scale wireless experimentation through hardware-in-the-loop network emulation,” inProc. IEEE Int. Symp. on Dynamic Spectrum Access Networks (DySPAN), 2021

    2021

  19. [19]

    Open ai cellular (oaic): An open source 5g o-ran testbed for design and testing of ai-based ran management algorithms,

    P. S. Upadhyaya, N. Tripathi, J. Gaeddert, and J. H. Reed, “Open ai cellular (oaic): An open source 5g o-ran testbed for design and testing of ai-based ran management algorithms,”IEEE Network, 2023

    2023

  20. [20]

    5gperf: profiling open source 5g ran components under different architectural deployments,

    C. Wei, A. Kak, N. Choi, and T. Wood, “5gperf: profiling open source 5g ran components under different architectural deployments,” inProceedings of the ACM SIGCOMM Workshop on 5G and Beyond Network Measurements, Modeling, and Use Cases, 2022, pp. 43–49

    2022

  21. [21]

    Evaluating open-source 5g sa testbeds: Unveiling performance disparities in ran scenarios,

    M. Rouili, N. Saha, M. Golkarifard, M. Zangooeiet al., “Evaluating open-source 5g sa testbeds: Unveiling performance disparities in ran scenarios,” inNOMS 2024-2024 IEEE Network Operations and Man- agement Symposium. IEEE, 2024, pp. 1–6

    2024

  22. [22]

    Tiny- twin: A cpu-native full-stack digital twin for nextg cellular networks,

    A. Mamaghani, U. Ghosh, S. Shakkottai, D. Bharadiaet al., “Tiny- twin: A cpu-native full-stack digital twin for nextg cellular networks,” in 2026 IEEE International Symposium on Spectrum Innovation (DySPAN). IEEE, 2026, pp. 1–8

    2026

  23. [23]

    Available: https://gitlab.eurecom.fr/oai/openairinterface5g/ -/merge requests/3902

    [Online]. Available: https://gitlab.eurecom.fr/oai/openairinterface5g/ -/merge requests/3902

  24. [24]

    Study on channel model for frequencies from 0.5 to 100 GHz,

    3GPP, “Study on channel model for frequencies from 0.5 to 100 GHz,” 3rd Generation Partnership Project (3GPP), Technical Report (TR) 38.901, 2026, release 19, V19.4.0

    2026

  25. [25]

    New Radio Physical Layer Abstraction for System-Level Simulations of 5G Networks,

    S. Lagen, K. Wanuga, H. Elkotby, S. Goyalet al., “New Radio Physical Layer Abstraction for System-Level Simulations of 5G Networks,” in ICC 2020 - 2020 IEEE International Conference on Communications (ICC), 2020, pp. 1–7

    2020

  26. [26]

    Openairlink: Reproducible wireless channel emulation using software defined radios,

    Y . Deshpande, X. Wang, and W. Kellerer, “Openairlink: Reproducible wireless channel emulation using software defined radios,” in2024 IFIP Networking Conference (IFIP Networking). IEEE, 2024, pp. 1–6

    2024

  27. [27]

    NR; Physical layer procedures for data,

    3GPP, “NR; Physical layer procedures for data,” 3rd Generation Partner- ship Project (3GPP), Technical Specification (TS) 38.214, 2026, release 19, V19.4.0

    2026

  28. [28]

    Multi-generator (MGEN) network test tool,

    U.S. Naval Research Lab., “Multi-generator (MGEN) network test tool,” https://github.com/USNavalResearchLaboratory/mgen, accessed 2026

    2026

  29. [29]

    Service requirements for the 5G system,

    3GPP, “Service requirements for the 5G system,” 3rd Generation Partnership Project (3GPP), Technical Specification (TS) 22.261, 2026, release 19, V19.14.0

    2026

  30. [30]

    Wireless Access for Ultra-Reliable Low-Latency Communications: Principles and Building Blocks,

    P. Popovski, J. J. Nielsen, C. Stefanovic, E. de Carvalhoet al., “Wireless Access for Ultra-Reliable Low-Latency Communications: Principles and Building Blocks,”IEEE Network, vol. 32, no. 2, pp. 16–23, 2018

    2018

  31. [31]

    Enhancing URLLC Uplink Configured-Grant Transmissions,

    T.-K. Le, U. Salim, and F. Kaltenberger, “Enhancing URLLC Uplink Configured-Grant Transmissions,” inIEEE 93rd Vehicular Technology Conference (VTC2021-Spring), 2021

    2021

  32. [32]

    AutoRAN: Automated and Zero-Touch Open RAN Systems,

    S. Maxenti, R. Shirkhani, M. Elkael, L. Bonatiet al., “AutoRAN: Automated and Zero-Touch Open RAN Systems,”arXiv preprint arXiv:2504.11233, 2025

    arXiv 2025

  33. [33]

    LTE frequency selective scheduling performance and improvements assessed by measurements,

    P. ¨Okvist, A. Simonsson, and H. Asplund, “LTE frequency selective scheduling performance and improvements assessed by measurements,” inProc. IEEE 22nd Int. Symp. Personal, Indoor and Mobile Radio Communications (PIMRC), 2011, pp. 1919–1923

    2011

  34. [34]

    NR; Physical layer procedures for control,

    3GPP, “NR; Physical layer procedures for control,” 3rd Generation Partnership Project (3GPP), Technical Specification (TS) 38.213, 2025, release 18, V18.5.0

    2025

  35. [35]

    Multiuser OFDM with adaptive subcarrier, bit, and power allocation,

    C. Y . Wong, R. S. Cheng, K. B. Letaief, and R. D. Murch, “Multiuser OFDM with adaptive subcarrier, bit, and power allocation,”IEEE Journal on Selected Areas in Communications, vol. 17, no. 10, pp. 1747– 1758, 1999

    1999

  36. [36]

    Increase in capacity of multiuser OFDM system using dynamic subchannel allocation,

    W. Rhee and J. M. Cioffi, “Increase in capacity of multiuser OFDM system using dynamic subchannel allocation,” inProc. IEEE 51st Vehicular Technology Conference (VTC 2000-Spring), vol. 2, 2000, pp. 1085–1089

    2000

  37. [37]

    SARA: Sub-band aware resource assignment for 5G-LENA,

    G. Ferreira, B. Bojovi ´c, A. A. Gargari, and S. L ´agen, “SARA: Sub-band aware resource assignment for 5G-LENA,” inICNS3, 2025

    2025

  38. [38]

    Fre- quency semi-selective scheduling for a DFT-S-OFDM based uplink,

    A. Krishnamoorthy, P. J. Sartori, K. L. Baum, V . Nangiaet al., “Fre- quency semi-selective scheduling for a DFT-S-OFDM based uplink,” inProc. IEEE 68th Vehicular Technology Conference (VTC 2008-Fall), 2008, pp. 1–5

    2008

  39. [39]

    Search-tree based uplink channel aware packet scheduling for UTRAN LTE,

    F. D. Calabrese, P. E. Mogensen, C. Rosa, P.-H. Michaelsenet al., “Search-tree based uplink channel aware packet scheduling for UTRAN LTE,” inProc. IEEE 67th Vehicular Technology Conference (VTC 2008- Spring), 2008, pp. 1949–1953

    2008

  40. [40]

    A joint modulation-coding scheme and resource allocation in LTE uplink,

    A. Gutierrez, V . Rangel, J. Gomez, R. M. Edwardset al., “A joint modulation-coding scheme and resource allocation in LTE uplink,” Elektronika ir Elektrotechnika, vol. 26, no. 5, pp. 50–58, 2020

    2020

  41. [41]

    Es- timating the doppler spectrum of a short-range fixed wireless channel,

    A. Domazetovic, L. J. Greenstein, N. B. Mandayam, and I. Seskar, “Es- timating the doppler spectrum of a short-range fixed wireless channel,” IEEE Communications Letters, vol. 7, no. 5, pp. 227–229, 2003

    2003

  42. [42]

    On the l ´evy-walk nature of human mobility,

    I. Rhee, M. Shin, S. Hong, K. Leeet al., “On the l ´evy-walk nature of human mobility,”IEEE/ACM Transactions on Networking, vol. 19, no. 3, pp. 630–643, 2011

    2011

  43. [43]

    OpenAirInter- face: Democratizing innovation in the 5G cellular ecosystem,

    F. Kaltenberger, A. P. Silva, A. Gosain, L. Wanget al., “OpenAirInter- face: Democratizing innovation in the 5G cellular ecosystem,”Computer Networks, vol. 176, p. 107284, 2020

    2020

  44. [44]

    srsLTE: An open-source platform for LTE evolution and experimenta- tion,

    I. Gomez-Miguelez, A. Garcia-Saavedra, P. D. Sutton, P. Serranoet al., “srsLTE: An open-source platform for LTE evolution and experimenta- tion,” inWiNTECH, 2016, pp. 25–32

    2016

  45. [45]

    Understanding O-RAN: Architecture, interfaces, algorithms, security, and research challenges,

    M. Polese, L. Bonati, S. D’Oro, S. Basagniet al., “Understanding O-RAN: Architecture, interfaces, algorithms, security, and research challenges,”IEEE Communications Surveys & Tutorials, vol. 25, no. 2, pp. 1376–1411, 2023

    2023

  46. [46]

    FlexRIC: an SDK for next- generation SD-RANs,

    R. Schmidt, M. Irazabal, and N. Nikaein, “FlexRIC: an SDK for next- generation SD-RANs,” inProc. 17th Int. Conf. on Emerging Networking EXperiments and Technologies (CoNEXT), 2021, pp. 411–425

    2021

  47. [47]

    POWDER: Platform for open wireless data-driven experimental research,

    J. Breen, A. Buffmire, J. Duerig, K. Duttet al., “POWDER: Platform for open wireless data-driven experimental research,” inProc. 14th Int. Workshop on Wireless Network Testbeds, Experimental Evaluation & Characterization (WiNTECH), 2020

    2020

  48. [48]

    Network simulations with the ns-3 simulator,

    T. R. Henderson, M. Lacage, G. F. Riley, C. Dowellet al., “Network simulations with the ns-3 simulator,” inACM SIGCOMM Conference Demonstration Session, 2008

    2008

  49. [49]

    An E2E simulator for 5G NR networks,

    N. Patriciello, S. L ´agen, B. Bojovic, and L. Giupponi, “An E2E simulator for 5G NR networks,”Simulation Modelling Practice and Theory, vol. 96, p. 101933, 2019

    2019

  50. [50]

    Sionna: An open-source library for next-generation physical layer research,

    J. Hoydis, S. Cammerer, F. Ait Aoudia, A. Vemet al., “Sionna: An open-source library for next-generation physical layer research,”arXiv preprint arXiv:2203.11854, 2022

    Pith/arXiv arXiv 2022

  51. [51]

    NVIDIA Aerial CUDA-accelerated RAN,

    NVIDIA, “NVIDIA Aerial CUDA-accelerated RAN,” https://developer. nvidia.com/aerial, 2024, software and product documentation

    2024

  52. [52]

    5g-emane: Scalable open-source real-time 5g new radio network emulator with emane,

    B. Ryu, R. Knopp, M. Elkadi, D. Kimet al., “5g-emane: Scalable open-source real-time 5g new radio network emulator with emane,” inMILCOM 2022-2022 IEEE Military Communications Conference (MILCOM). IEEE, 2022, pp. 553–558

    2022

  53. [53]

    A First Large-Scale Study of Operational 5G Standalone Networks,

    M. Ghoshal, O. Basit, S. Wang, P. Dinhet al., “A First Large-Scale Study of Operational 5G Standalone Networks,”CoNEXT, 2025

    2025

  54. [54]

    DRL-Based Joint Resource Scheduling of eMBB and URLLC in O-RAN,

    R. M. Sohaib, S. T. Shah, O. Onireti, Y . Samboet al., “DRL-Based Joint Resource Scheduling of eMBB and URLLC in O-RAN,” inICC Workshops, 2024. ACKNOWLEDGMENTS We thank Robert Schmidt from OpenAirInterface for his precious comments during the review of our open-source contributions

    2024