HK ↔ WUS
Clean and symmetric enough for a baseline path.
- 2 packets lost inbound
- 143 packets lost outbound
Presentation deck
Mapping Azure's Backbone with iperf3
A repeatable 24-hour UDP test for long-haul bandwidth stability, jitter, and packet loss.
Yingting Huang · Nov 2025
Short checks miss long-haul drift
A tiny, steady UDP stream can reveal packet-loss and jitter patterns that quick spot checks never expose.
Each direction ran for 86,400 seconds at 10 Mbps.
Traffic stayed on Microsoft's backbone through global VNet peering.
The output is comparable data, not anecdotal network feel.
One WUS hub, three long-haul spokes
flowchart LR WUS[(West US hub<br/>VM + VNet)] HK[East Asia / HK<br/>spoke VM] KC[Korea Central<br/>spoke VM] UAE[UAE North<br/>spoke VM] HK <-->|global VNet peering<br/>UDP 5201| WUS KC <-->|global VNet peering<br/>UDP 5201| WUS UAE <-->|global VNet peering<br/>UDP 5201| WUS
- Standard_D2s_v5 VMs provided matching endpoints.
- UDP/5201 was opened so iperf3 could run directly.
- Every region pair was tested in both directions.
The filename is the experiment contract
Each run is named by direction, bandwidth, duration, and timestamp so the pipeline can parse intent without extra metadata.
- HK-WUS means sender in East Asia, receiver in West US.
- WUS-HK flips the direction while keeping the same region pair.
- KC and UAE follow the same source-destination pattern.
- Completed JSON files land in 24H/ with the direction prefix intact.
The test is one durable iperf3 stream
#!/bin/bash
SERVER="${1:-172.16.0.4}"
OUTPUT_DIR="./test_results"
TIMESTAMP=$(date +"%Y%m%d_%H%M%S")
mkdir -p "$OUTPUT_DIR"
iperf3 -c "$SERVER" -u -b 10M -i 1 -t 86400 -J \
> "$OUTPUT_DIR/10M_24H_${TIMESTAMP}.json"Raw iperf JSON becomes comparable evidence
flowchart LR Json[24H/*.json] --> Parse[parse intervals<br/>+ end summary] Parse --> Stats[compute stats:<br/>mean · std · loss · jitter] Stats --> Pair[group by<br/>region pair] Pair --> Plots[pair + direction<br/>PNG charts] Stats --> Report[summary_report.md]
- Intervals become time-series arrays for plotting.
- The iperf end summary supplies loss and jitter.
- Pair plots make direction-specific blips visible.
The processor makes every rerun cheap
Once the raw captures are in 24H/, one command regenerates the full comparison set.
- Parsing preserves per-second bandwidth, packet rate, loss, and jitter.
- Pair grouping keeps inbound-to-WUS and outbound-from-WUS separated.
- Zoomed y-axes reveal small variance around the 10 Mbps target.
- Markdown and PNG outputs are ready for a blog, dashboard, or slide deck.
All six directions held the 10 Mbps target
| Direction | Mean Mbps | Loss % | Lost pkts | Jitter ms |
|---|---|---|---|---|
| HK→WUS | 10.0000 | 0.000003 | 2 | 0.0051 |
| WUS→HK | 10.0000 | 0.000184 | 143 | 0.0059 |
| KC→WUS | 10.0000 | 0.000019 | 15 | 0.0066 |
| WUS→KC | 10.0000 | 0.000295 | 230 | 0.0045 |
| UAE→WUS | 10.0000 | 0.002094 | 1,632 | 0.0106 |
| WUS→UAE | 10.0000 | 0.000619 | 482 | 0.0102 |
10 Mbps UDP over 24 hours; totals are from the article's iperf3 summary table.
The signal is loss concentration, not throughput
Throughput stayed flat; the useful difference was that UAE loss was tiny in absolute terms but much higher than HK and KC.
UAE→WUS lost 1,632 packets; WUS→UAE lost 482.
HK and KC stayed close to zero-loss behavior.
Jitter on UAE legs was roughly double the HK/KC range.
Three region pairs, three operational reads
KC ↔ WUS
Also clean, with very low loss in both directions.
- 15 packets lost inbound
- 230 packets lost outbound
UAE ↔ WUS
Still stable, but the first path to investigate.
- 1,632 packets lost inbound
- Wider outbound min/max spread
The plots make small failures visible
Pair charts stack both directions so tiny blips are not hidden by overlapping traces.
- Green lines represent traffic inbound to West US.
- Blue lines represent traffic outbound from West US.
- Separate panels avoid overlap and make direction-specific spikes obvious.
- A zoomed y-axis shows drift around 10 Mbps instead of flattening it away.
HK ↔ WUS stayed visually flat
- Both directions stay tightly centered on 10 Mbps.
- Packet loss is near-zero in the summary table.
- This pair is the visual baseline for the deck.
KC ↔ WUS looked similarly clean
- Mean throughput remains 10.0000 Mbps both ways.
- Loss stays very low: 15 inbound and 230 outbound packets.
- The chart reinforces symmetry across the pair.
UAE ↔ WUS is the path to watch
- Throughput still holds the 10 Mbps target.
- Loss is higher than HK and KC in both directions.
- The outbound leg has the widest min/max spread.
Reproducing the workflow is deliberately boring
Provision endpoints, run both directions, collect JSON, then regenerate diagrams and the report.
- Create iperf3 endpoints and open UDP/5201 in each region.
- Run the 24-hour script once for every direction.
- Collect JSON reports into the 24H/ folder.
- Install Python dependencies: matplotlib and numpy.
- Run process_24h.py to publish tables and PNGs.
Turn the marathon into a regression test
The real value is a repeatable baseline: rerun the same path after routing, region, or workload changes and compare the drift.
The artifacts are deterministic: JSON in, charts and Markdown out.
Direction labels make asymmetry visible without manual cleanup.
A future GitHub Action could publish the same report automatically.
Five lessons from the 24-hour run
- Hold a small UDP stream long enough to expose rare loss and jitter events.
- Name artifacts by direction so analysis can stay automatic.
- Treat charts as part of the test, not a decorative afterthought.
- Use relative differences carefully: UAE was higher-loss, not broken.
- Keep the workflow rerunnable so every future change has a baseline.