Network / 04

Bufferbloat.

Three phases: idle ping baseline (5s), ping under download load (15s), ping under upload load (15s). The latency lift between idle and loaded reveals bufferbloat — the queue-management failure that turns video calls into slideshows the moment someone starts a torrent.

Notes

Grading follows the dslreports.com convention: A+ = ≤5 ms lift, A ≤30, B ≤60, C ≤200, D ≤400, F > 400. A-grade requires modern queue management (fq_codel, CAKE) on your router. Most ISP-supplied modems fail badly.

Latency under load · idle baseline —
Grade
—
—
Idle ping
—
Under DL
—
Under UL
—
DL lift
—
UL lift
—
Worst
—
Frequently asked

What exactly is bufferbloat?

Bufferbloat is the latency spike that appears when a network link is saturated and its queues fill up faster than they drain. Routers and modems buffer packets to avoid drops, but oversized buffers hold thousands of packets in line — meaning every new packet, including your ping, has to wait its turn behind a queue of bulk traffic. The result: your idle ping of 20 ms can balloon to 500 ms or more the instant someone starts a download, killing video calls and competitive gaming.

How does this three-phase test reveal it?

Phase one pings for five seconds on an idle link to establish a baseline. Phase two starts a parallel download to saturate the downlink while continuing to ping every 200 ms. Phase three repeats with an upload flood. The grade compares loaded ping against the idle baseline: under 30 ms of lift earns an A, 30–60 ms a B, 60–200 ms a C, and anything beyond that fails. Most consumer routers without modern queue management land in C or worse.

How do I fix bufferbloat at home?

Enable Smart Queue Management (SQM) — often labelled fq_codel, CAKE, or Adaptive QoS — on your router. OpenWrt, pfSense, and most recent ASUS/Ubiquiti firmwares ship it. Set the upload and download caps to roughly 85–95% of the speeds you measured on /network/speed; that small headroom is what lets the router actually control the queue rather than handing it off to the ISP's modem. Re-run this test afterwards — a properly tuned SQM connection holds latency lift under 20 ms even at full throttle.

Why does upload saturation cause more bufferbloat than download?

Cable and DSL uplinks are typically far slower than downlinks, and the modem's upload buffer is usually the deepest queue in the path. A single cloud backup or video upload can pin the uplink at 100% for minutes, and every TCP ACK from your downloads has to fight through that same congested queue — which is why your downloads also slow down during a big upload. Symmetric fibre helps, but only if SQM is enabled.

How does this relate to your speed and latency tools?

/network/speed measures capacity on an empty link — the marketing number. /network/latency tracks idle ping stability over time. Bufferbloat is the missing third leg: it asks what happens when capacity meets reality. A gigabit connection with severe bufferbloat is worse for real-time apps than a 50 Mb/s connection with good queue management. Always run all three before blaming your ISP.

Does the test work behind a VPN?

It runs, but the numbers reflect the VPN tunnel and the VPN provider's exit-node queues, not your home link. For an honest reading, disconnect the VPN first. If you specifically want to compare VPN-on versus VPN-off latency under load, run the test twice and keep a screenshot of each — the difference is usually striking.