Skip to content
BenchDevice

What bufferbloat is and how to tame it

Why a connection lags the moment something downloads, how the test measures it, and what fixes it.

The problem

You start a large download, or someone in the house does, and everything else goes to sleep. Pages take an age to open, a video call freezes, a game stutters. The speed test said the line was fine, and it was: the line is not slow. It is queuing.

Every router and modem has a buffer, a small store where packets wait their turn when more arrive than the line can send. A buffer is useful: it smooths out bursts. The trouble starts when the buffer is far too large. A download fills it to the brim and keeps it full, so every other packet, including the tiny ones a web page or a call sends, has to wait at the back of the queue before it leaves your house. That waiting is bufferbloat. The line's speed is unchanged; its delay has grown by the time it takes to drain the queue.

Why speed tests miss it

A speed test measures how many bits pass in a second, and a bloated buffer does not reduce that. If anything it helps, since the line never sits idle. What suffers is latency, the time one small request takes to go out and its answer to come back, and only while the line is busy. A test that times round trips on an idle connection sees nothing wrong.

How the test measures it

The bufferbloat tester times round trips to BenchDevice's server in three phases: at rest, while the page downloads, and while it uploads. It sits at rest for five seconds, then downloads with four streams for ten seconds, then uploads for ten seconds, sending a tiny request every tenth of a second the whole way through; a request with no answer after two seconds counts as lost. The upload sends chunks that double in size while each one finishes quickly, so the phase loads a fast line as hard as a slow one.

The result is one table. For each phase you see the typical (median) round trip, the slow (95th percentile) one, how many requests were lost, and the speed the load reached in Mbit/s. The row that matters most is the rise over rest: the loaded phase's median minus the median at rest. That is how much longer every small request waits while your connection is busy.

Two caveats. The requests share the browser's connection with the load itself, so the rise is what a web page feels on this device, not a pure measure of the line's queue. And BenchDevice's server caps the speeds, so a very fast line may read lower here than on a speed test.

What is normal

The tool gives numbers and no verdict: there is no threshold in it that says pass or fail. Compare the rise with the round trip at rest, and compare the figure with a second run after you change something in the router: the difference between the two runs is what tells you whether the change worked.

It is common to see a rise under upload but not under download. Home connections often have far less upload capacity, so a small upload queue fills in moments. Video calls and cloud backups both live on upload, which is why they are the first to suffer.

What helps

Smart Queue Management. Many routers can run a queue manager, often listed as SQM, fq_codel or CAKE in the settings. You tell it the speed of your line and it keeps the queue short by design, dropping or marking packets a little early so no single download can hog the buffer. This fixes bufferbloat on most lines. Set the speeds a little below what the line actually delivers; the manager needs to be the bottleneck for it to work.

Quality of Service. Older or simpler routers offer QoS instead: rules that put calls and games ahead of downloads. It helps, but it relies on you naming the right traffic, and it does nothing about the queue itself.

A different router. Many providers' boxes have neither setting. A modest router that supports SQM, placed behind the provider's modem, usually does the job.

Fewer things at once. Pausing a cloud backup or a game update while you are on a call is not a fix, but it helps for the evening.

What to do next

Run the test with the house quiet, note the rise, then change one thing in the router and run it again. The number tells you plainly whether the change worked. If you are on Wi-Fi, run the connection stability tester as well, once on Wi-Fi and once on a cable: if the Wi-Fi trace is the spiky one, the radio is adding its own delay on top, and SQM will not cure that part.

All guidesBack to the network tools