Modern AQM Schemes in Wi‑Fi 6

Updated 27 December 2025

Modern AQM schemes are advanced queue control methods that proactively drop or mark packets to regulate buffer occupancy and mitigate latency.
FQ-CoDel and CAKE algorithms employ per-flow and per-host isolation to adapt dynamically to Wi‑Fi 6’s variable MAC service rates, ensuring fairness.
Empirical evaluations reveal that integrating modern AQM with model-based TCP CCAs like BBRv3 significantly improves throughput, reduces retransmissions, and minimizes delay.

Modern Active Queue Management (AQM) schemes implement advanced queue control strategies in WLAN infrastructures and broadband routers to manage buffer occupancy, mitigate latency, and optimize throughput, particularly in dense multi-user environments. With the advent of OFDMA- and MU-MIMO-enabled wireless standards such as IEEE 802.11ax (Wi-Fi 6), and the widespread deployment of model-based TCP CCAs (e.g., BBRv3), the interactions between MAC scheduling, transport protocols, and queue management have become critical determinants of network fairness and application QoE. State-of-the-art research, drawing on recent empirical and analytical evidence, underscores the necessity of per-flow and per-host AQM algorithms—specifically FQ-CoDel and CAKE—over baseline drop-tail (PFIFO) and first-generation AQM, especially in heterogeneous Wi-Fi 6 deployments (Shrestha et al., 20 Dec 2025).

1. AQM Fundamentals and Modern Workflow

Modern AQM schemes operate at the router or AP queue, interposed between traffic sources (end hosts or applications) and the wireless scheduling layer. The essential function is to regulate buffer occupancy by dropping or marking packets proactively, rather than waiting for queues to fill to capacity. This is particularly salient in Wi-Fi 6, where the intersection of OFDMA resource allocation and TCP flow dynamics generates highly variable MAC-layer service rates.

Key AQM types in current deployment:

PFIFO (Drop-tail): A single shared queue with packet admissions up to buffer capacity $Q_{\max}$ , after which all further packets are dropped ( $p_i^{\mathrm{fifo}}(t) = 1$ if $\sum_j Q_j(t) = Q_{\max}$ ). There is no delay or flow control; large standing queues amplify bufferbloat.
FQ-CoDel: Implements per-flow queueing with Deficit Round Robin (DRR) or GPS-like scheduling, bounding sojourn time. Drops occur when per-flow sojourn $\tau_i^{\mathrm{soj}}(t) = Q_i(t)/s_i^{\mathrm{fq}}(t)$ exceeds a target (default $T_{\mathrm{target}}=5\,\mathrm{ms}$ ). Provides per-flow isolation and fairness restoration.
CAKE: Supersedes FQ-CoDel with per-host aggregation $\alpha_i^{\mathrm{cake}} = \omega_i/\sum_j \omega_j$ and non-linear adaptive marking, targeting sojourn times $\approx 5\,\mathrm{ms}$ and directly managing fairness across hosts.

Such AQMs must track both instantaneous sender pacing $p_i(t)$ and MAC-layer service $\mu_i(t)$ , controlling queue growth to minimize transport-layer RTT inflation and loss synchronization (Shrestha et al., 20 Dec 2025).

2. Cross-layer Effects in IEEE 802.11ax Networks

The IEEE 802.11ax (Wi-Fi 6) standard integrates OFDMA and MU-MIMO, subdividing channels into resource units (RUs) and allocating per-user slices across both DL and UL. This multi-user parallelism introduces slot-level service granularity, tightly coupling queue input rates (from CCAs) and the stochastic service process $\mu_i(t)$ enforced by the MAC scheduler (Bellalta, 2015). This increases the likelihood of queue oscillation and transient excess delay if not mitigated by responsive AQM.

Critical innovations affecting AQM design:

OFDMA: Variability in per-STA service rate due to allocation size and grant frequency.
MU-MIMO: Dynamic allocation of spatial streams per user alters effective service $\mu_i(t)$ session by session.
MAC Innovations (TWT, BSS Coloring): Periodic wake schedules (TWT) and spatial reuse mechanisms change the effective buffer drain intervals and further increase cross-layer dependencies.
MAC-Layer Random Access/UL Grant: Uplink scheduling varies data movement per cycle, and non-deterministically batched UL activity induces correlated arrival patterns to bottleneck queues (Bellalta, 2015).

This cross-layer structure requires AQM schemes to distinguish tightly-spaced probe-driven bursts (e.g., from BBRv3's ProbeBW cycles) from sustained congestion, so that packet drops/marks maintain consistent end-to-end RTT control.

3. Evaluation of Modern AQM Algorithms

A recent cross-layer testbed evaluation of BBRv3 and CUBIC under three AQMs—PFIFO, FQ-CoDel, and CAKE—on a fully wireless Wi-Fi 6 (802.11ax) deployment provides a quantitative foundation for theoretical and practical AQM assessment (Shrestha et al., 20 Dec 2025). The principal measured metrics include:

Throughput (average over 30s steady-state)
Median/99th percentile one-way RTT
TCP retransmissions (loss events)
Flow fairness (Jain's index $F$ )

Key results:

Scenario	AQM	BBRv3 Thpt (Mbps)	CUBIC Thpt (Mbps)	Median RTT (ms)	Jain's $F$
Uplink	PFIFO	3.5 ± 0.4	6.5 ± 0.5	55 ± 5	0.53
	FQ-CoDel	5.1 ± 0.2	4.9 ± 0.3	25 ± 3	0.98
	CAKE	5.0 ± 0.1	5.0 ± 0.1	18 ± 2	0.99
Downlink	PFIFO	5.8 ± 0.5	4.2 ± 0.4	50 ± 4	0.80
	FQ-CoDel	6.1 ± 0.3	3.9 ± 0.2	24 ± 2	0.92
	CAKE	5.0 ± 0.2	5.0 ± 0.2	17 ± 2	0.99
Bidirectional	PFIFO	-- (erratic)	--	80 ± 6	0.60
	FQ-CoDel	9.8 ± 0.4 (sum)	--	30 ± 4	0.97
	CAKE	10.0 ± 0.2	--	15 ± 2	0.99

Interpretation: PFIFO exhibits bufferbloat (RTT $>50$ ms) and severe unfairness, especially under uplink congestion. FQ-CoDel and CAKE reduce medium RTTs to $<$ 25 ms, maintain fairness ( $F\geq 0.98$ ), and tightly control retransmissions. CAKE offers the best tradeoff, keeping delay under $20$ ms and full fairness ( $F=0.99$ ) across all scenarios (Shrestha et al., 20 Dec 2025).

4. Wi-Fi-Specific Interactions and Tuning

CAKE’s aggressive target-sojourn and adaptive marking interact decisively with Wi-Fi 6 MAC layer behavior. During BBRv3 ProbeBW Up phases (e.g., $pacing\_gain=1.25$ ), bursts may cause CAKE to rapidly drain short queues to hit the 5 ms sojourn, resulting in temporally-clustered loss events (retransmission bursts). This stems from fluctuating $\mu(t)$ due to OFDMA TXOP variability ( $\pm$ 20%), which may desynchronize BBRv3’s inflight goal and effective service rate. A recommended mitigation is to reduce BBRv3’s probe gain (e.g., $pacing\_gain\approx1.10$ vs default 1.25), smoothing burst amplitude and aligning delivery with queue control (Shrestha et al., 20 Dec 2025).

A plausible implication is that proprietary MAC scheduling heuristics or STA-client variations in RU allocation could require dynamic re-tuning of AQM parameters (e.g., target, interval) in high-density deployments.

5. Modern Hybrid and Scheduled MAC Protocols: Synergy with AQM

Recent protocol-level research demonstrates substantial gains when integrating modern AQM with hybrid MAC protocols designed for Wi-Fi 6 OFDMA. Proposals such as HTFA, ERA, and PRS leverage fine-grained RU allocation, scheduled access (TF-driven), and proportional fairness (Islam, 2023). ERA and PRS protocols, which architect the MAC for deterministic transmission cycles and minimize random access, yield almost zero MAC-level retransmissions and maximize goodput when combined with aggressive AQM.

Summary table of MAC protocols:

Protocol	Throughput Gain	Goodput	Fairness
HTFA	~49 Mbps vs legacy	High (via eq 6.1–6.2)	$F \approx 0.85–0.95$
ERA	~238 Mbps vs 196–227	Near-zero loss	random-free
PRS	~295.7 Mbps	Matches throughput	$J \approx 0.87$ (Jain)

These protocols orchestrate scheduled and opportunistic uplink with dynamic resource assignment, further stabilizing queue lengths and reducing the prevalence and persistence of deleterious congestion epochs (Islam, 2023).

6. Practical Deployment Guidelines and Future Directions

Deployment recommendations are explicit:

FQ-CoDel or CAKE must be enabled in Wi-Fi 6 home gateways; unmanaged PFIFO/drop-tail introduces severe latency and flow starvation, undermining application-level QoE and fairness (Shrestha et al., 20 Dec 2025).
Key AQM parameters (target, interval, host hash) may require adaptation to the RU allocation profile, channel width, and network density mandated by the AP’s OFDMA scheduler (Bellalta, 2015).
With model-based CCAs such as BBRv3, lower probe and cwnd gains (e.g., $pacing\_gain \approx 1.10$ , $cwnd\_gain \approx 2.00$ ) are advisable to limit loss-induced burst oscillations in low-to-moderate bandwidth deployments.

Emerging avenues include dynamic, Wi-Fi-aware AQM tuning frameworks and cross-layer orchestration between MAC grant scheduling and AQM, potentially via explicit feedback of RU scheduling state or observed channel access opportunities. Synchronizing AQM policy with scheduled MAC protocols (ERA, PRS) may further lower aggregate latency and maximize per-user fairness under dense conditions (Bellalta, 2015, Shrestha et al., 20 Dec 2025, Islam, 2023).