Home Broadband Under Load: Four Connections Tested for Video Work

Four home connections measured between 38 and 210 Mbps on a speed test, and all four carried the same video call equally badly during the evening peak. Download speed was the least predictive number in the set.

The tests ran for three weeks in four households doing similar work: two video calls a day, a shared drive and occasional large uploads. Each household recorded latency and call quality rather than headline speed.

What the speed test measures

A speed test measures a short burst to a nearby server, which is the best case the network will ever offer. It says nothing about how the connection behaves when a neighbour is streaming, and it says nothing at all about upload.

Two of the four connections advertised upload speeds that were not reached during the day, let alone at eight in the evening.

The four connections

Connection A measured 210 Mbps down and 21 up, cable, city centre. Connection B measured 74 down and 18 up, fibre to the building. Connection C measured 38 down and 9 up, older copper with a long line. Connection D measured 120 down and 12 up, fixed wireless.

On paper A was best and C was worst, which turned out to be true for downloads and irrelevant for calls.

Latency under load

Idle latency was between 9 and 24 milliseconds. Under a simultaneous upload, which is what happens when a shared folder syncs during a call, three of the four connections rose above 200 milliseconds, and one reached 780.

At 200 milliseconds conversation becomes a series of interruptions. At 780 the call is unusable regardless of resolution.

Bufferbloat, plainly

The pattern is caused by queues filling at the router when upload capacity is saturated. The technical fix is a queue management setting that most consumer routers hide or omit. The practical fix is to stop the upload.

Connection B’s router exposed a setting labelled something like smart queue, which held latency to 60 milliseconds under load. The other three had no equivalent.

What each household changed

Household A moved the shared folder sync to a scheduled window after 22:00. Household C turned off automatic photo backup during working hours, which had been using 9 Mbps of upload capacity without anyone noticing.

Household D added a second connection on a different technology for calls only, at 15 euros a month, which is cheaper than any upgrade offered by the primary provider.

The upload figure nobody quotes

Video calls need about 3 Mbps up for a good picture and 1.5 for an acceptable one. Every connection in the test exceeded that at idle. None of them delivered it while a sync was running, because sync processes take whatever is available.

Upload capacity is not a speed problem. It is a scheduling problem, and scheduling is free.

What the team concluded

Three changes to ask about before signing: upload speed, latency under a saturated upload, and whether the router exposes a queue setting. None of the three appear in advertising, and two of the four providers could not answer the second question at all.

The device count nobody budgeted for

Household A had 14 devices on the network, of which nine were idle most of the day and still kept connections open. The router handled them without complaint, and the calls were unaffected, which the household checked by unplugging four of them for a week and comparing.

The number of devices is the figure providers ask about and the one least related to call quality. Bandwidth per active call mattered; the idle count did not.

What a wired connection changed

One household ran a cable to the desk for a fortnight. Latency under load fell by about 30 milliseconds and the call quality was unchanged, because the bottleneck was upstream rather than on the local network.

The cable stayed for the stability rather than the speed, which is a habit rather than a measurement.

Two hundred and ten megabits once looked like the best connection in the group and behaved like the worst during a call. Thirty-eight megabits with scheduled syncs behaved better, which is the only result here worth acting on.