Characterizing and improving WiFi latency in large-scale operational networks

Summary

A measurement study using round-trip timing over the last, wireless hop of a university campus Wi-Fi network, across more than 47,000 user devices, and of what makes that delay long.

Key findings

• Measured Wi-Fi latency on the Tsinghua University campus network for more than 47,000 devices. The distribution of one-minute average round-trip measurements had a long tail: the median was about 3 ms, the 90th percentile about 20 ms and the 99th percentile about 250 ms.
• Measured Wi-Fi latency on the Tsinghua University campus network for more than 47,000 devices. The ping round-trip delay had a long tail: the median was about 3 ms, the 90th percentile about 20 ms and the 99th percentile about 250 ms.
• 75% of devices had at least one episode in a week in which their average Wi-Fi latency stayed above 20 ms for more than a minute.
• The most informative indicators of Wi-Fi latency were channel utilisation, the number of devices online and the signal-to-noise ratio; received signal strength, commonly used to choose an access point, was less informative.
• On this campus the 5 GHz band had 35% lower average latency than 2.4 GHz, which the authors link to half as many devices using it and lower channel utilisation; devices' default band selection favoured 2.4 GHz.
• Moving devices to access points suggested by the authors' latency model more than halved latency for 72% of the over 1,000 devices that used the service for 2.5 months; adding a 5 GHz-only network and denser access points also helped.
• Section 2.1 defines ping2 as two consecutive pings from a wired server to a user device: the first wakes the device and the second round-trip time supplies the measurement. Six measurements, taken ten seconds apart, are averaged into each one-minute latency record. The authors regard the wired portion as negligible because its 99th percentile was below 1 ms. The reported distribution therefore describes minute-average round-trip measurements, not one-way audio delivery.
• Section 2.1 defines the measured result as the second ping round-trip time (RTT2) after a wake-up ping; the wired portion was treated as negligible (99th percentile under 1 ms). Each stored latency record averages six readings over a minute. These are not one-way audio-packet transit times.

Limits

A general study of one campus data network, not of audio streaming or assistive listening: it estimates the wireless-hop round-trip delay using a second ping after a wake-up ping, not the audio buffering, codecs or phone playback that add to a stream's delay, and reports no audio results. Wi-Fi equipment has changed since 2016. Read in the authors' copy hosted by Microsoft Research; the ACM page is the published version.

Read the original

dl.acm.org/doi/10.1145/2906388.2906393

Link checked 13 October 2026. We link to the original and do not host copies.

Cited in

• How much delay does Wi-Fi streaming add, and when does it matter?Not reliable yet

Cite this source

The accessed date is when this library last checked the link; use the date you read it. To point readers to this summary instead, link to https://assistivelisteningknowledgebase.org/evidence/sui-2016/

Every finding here cites its source, and who funded each source is labelled. This site does not recommend products. Found an error or a missing study? Suggest a correction.