Voice latencyStarting
—ms Mean RTT
WLAN target 100 msP99 ref. —
Access point simulation
IEEE 802.11be · Three-network scenario · 60-second session
Network topology
Three BSSsNeighbor priorityWaiting for traffic · starts at t = 2 s
Scroll horizontally to view AP2, AP1 and AP3.
AP1 · Under study
6 stationsNo prioritization · BE
All traffic uses AC_BE
AP1Under study
STA 014K videoWaiting for traffic
BESTA 02Video callWaiting for traffic
BESTA 03GamingWaiting for traffic
BESTA 04VoiceWaiting for traffic
BESTA 051440p videoWaiting for traffic
BESTA 06TCP downloadWaiting for traffic
BENo priority marking
VO · VoiceVI · Video / gamingBE · Best EffortOverlapping wireless coverage
Moving dots represent wireless packets between each AP and its stations. Overlapping areas indicate shared-channel contention, not a propagation model.
01Observe5 s window
→02Classify7 features
→03Mark DSCPOpen loop
→04PrioritizeEDCA queues + TF
60 s
Live values and traces are illustrative. P99 badges use the draft's scenario results. Each study run used a fixed configuration; mode changes here demonstrate transitions.
Video qualityStarting
—STA 01
Buffer 0.0 s0.0 Mbit/s
Gaming latencyStarting
—ms Mean RTT
WLAN target 20 msP99 ref. —
TCP downloadAC_BE
—Mbit/s
Illustrative throughputP99 ref. —
Video playback
Emulated playbackOWIN6G · Logo motionWaiting
Buffer 0.0 s
OWIN6G logo · Motion follows the emulated buffer.TCP flow · MPEG-DASH
AP queues
DownlinkAC_BK receives no traffic in this scenario. Queues and packets are visual representations.
WLAN round-trip time
Illustrative mean RTTWLAN targetms · log scale
Enable QoS-ML to observe the change in priority.0–60 s
EventsPreparing the coexistence scenario.
Reference results
Results reported in the draft for the selected scenario, using five random seeds. RTT and throughput are evaluated from t = 7 s. Change the scenario in the Simulation tab.
| Service / metric | Best Effort | QoS-ML | QoS-App | Evaluation criterion |
|---|
Mean RTT is the mean of the five per-seed means. P99 uses RTT samples pooled across those seeds. TCP throughput is the mean across five seeds, from t = 7 s to the last received packet. The values are rounded labels from Figures 9–11 and 13–15 of the supplied draft.
The video rate criteria are 16.88 Mbit/s for 2160p and 8.44 Mbit/s for 1440p (Table 6). The IP throughput includes protocol headers. Meeting a rate criterion alone does not establish uninterrupted playback: the highest resolution, no stalls and a stable buffer are also required (Figures 12 and 16).
These results are independent of the animated session. QoS-App is a source-marking reference, not a strict upper bound on QoS-ML performance.
Latency
With BE neighbors, QoS-ML meets all three WLAN RTT targets. With QoS-App neighbors, voice meets its target, gaming exceeds 20 ms, and video calls exceed 100 ms with two neighbors.
Video playback
With BE neighbors, both videos retain their maximum resolution. With QoS-App neighbors, QoS-ML restores maximum resolution after the initial observation period. QoS-App retains maximum resolution from the start.
Throughput trade-off
Prioritization redistributes access to the channel. With two BE neighbors, mean TCP download throughput is 49.0 Mbit/s under BE and 41.3 Mbit/s under QoS-ML.
Method and assumptions
The study's classification, marking, and medium-access mechanisms, illustrated in a browser.
Flow observation
In the study, the AP groups TCP and UDP packets carrying data into bidirectional flows. Every second, it analyzes the previous 5 seconds using five packet-size features and two timing features.
Traffic starts: t = 2 s
First complete window: t = 7 s
First complete window: t = 7 s
If a flow has no traffic in the final second of the window, the AP skips reclassification and retains its previous class. This is the abstain rule. Classification also runs in BE, but its predictions are not applied.
Classification and DSCP marking
XGBoost predicts the application. The AP then rewrites DSCP in subsequent downlink packets before placing them in MAC queues.
Voice (VoIP)AC_VO
Video · Video call · GamingAC_VI
TCP downloadAC_BE
QoS-App marks downlink traffic at its source from the first packet. QoS-ML waits for the first complete observation window.
TXOP scheduling
After winning channel access through EDCA and protecting it with RTS/CTS, the AP selects the station with the longest queue in the winning AC.
The Trigger Frame requests uplink transmission from the previous STA in the pairing sequence, provided it has pending traffic. The STA responds after SIFS without contending again.
First DL: AC = winning AC
Later DL: AC ≥ winning AC
Uplink: AC ≥ AC_BE
Later DL: AC ≥ winning AC
Uplink: AC ≥ AC_BE
Round-robin scheduling continues while time remains in the TXOP. All three modes use this scheduler.
Study configuration
- PHY
- 802.11be · 5 GHz · 160 MHz
- Modulation
- Fixed MCS 5 · 1 stream · 800 ns GI
- Theoretical PHY rate
- 576.5 Mbit/s
- Channel
- No channel loss or hidden nodes
- Scenarios
- Home: 6 STA · AP2: 7 · AP3: 5
- Duration
- 60 s · study metrics from t = 7 s
EDCA parameters
| AC | CWmin | CWmax | AIFSN | TXOP |
|---|---|---|---|---|
| VO | 3 | 7 | 2 | 2080 µs |
| VI | 7 | 15 | 2 | 4096 µs |
| BE | 15 | 1023 | 3 | 2528 µs |
| BK | 15 | 1023 | 7 | — |
Scope of this emulation
Based on the draft Enabling EDCA through machine learning at the access point in IEEE 802.11be WLANs, file “ML-activated EDCA_Draft_Final.pdf”, Sections 3–5 and Tables 5–6. The study uses ns-3.40 and an XGBoost classifier coupled through ns3-ai. This browser model illustrates the mechanisms and reported trends; it does not run ns-3 or a trained classifier. Each study run uses a fixed configuration. Switching modes during a session is an illustrative transition added for this demonstration.
Packets, queues, RTT traces, throughput, and video playback are synthetic. The live RTT curve illustrates variation around the reported mean; it is not computed from packet measurements. Reference P99 values come from the draft and are independent of the live traces. RTT covers the WLAN segment, not the full delay of a call. The visualization does not decode real video or measure your Wi-Fi. The original logo from the OWIN6G project website is animated locally to illustrate video playback.
This emulation assumes correct AC assignment, consistent with the evaluated runs. The reported 99.6% accuracy applies to the classifier test set and is not used as an error probability here. In QoS-App, video TCP ACKs use AC_VI; all other uplink traffic remains in AC_BE.