Lag when someone uploads? Bufferbloat explained

Waveform points to phone photo apps that back up photos and videos to the cloud as soon as they're taken, and says upload is the direction that catches people out: a single cloud backup can add hundreds of milliseconds to every packet in the house.

An example, not a real test: a video call freezes on and off while a phone uploads a weekend's videos, yet a speed test afterwards looks fine. Nothing is broken; the router let a long queue build. Our guide on how to see what's using your broadband helps you find the device responsible.

Our rule of thumb, not a source's: bufferbloat in your own router usually tracks what your household is doing and eases when the big transfer ends. Lag at peak times when nobody at home is busy is more likely beyond your home; see why broadband slows down in the evening. A wired test with the house quiet settles it.

Waveform's test measures latency idle, then during a download and an upload, and grades the bigger increase. It suggests reading the grade first, then latency, then speed.

On a laptop or desktop it also shows jitter (how much the delay varies) and 95th percentile latency (the delay that 95 in every 100 measurements stay under) next to each chart. Waveform describes a stable connection as jitter under 30 ms, a 95th percentile close to the median, and a grade of B or better.

Because it runs in a browser, it doesn't count dropped packets, but Waveform says packet loss still leaves clues: higher jitter, a 95th percentile well above the median, and speeds below your plan. Speeds can also vary between browsers, though the grade should be fairly consistent.

Cloudflare's current documentation (updated April 2026) says its Aggregated Internet Measurement (AIM) turns latency, packet loss, speeds, loaded latency and jitter into a score from Bad to Great for streaming, gaming and webchat/RTC (real-time communication, such as video calls). A 2023 Cloudflare blog post said the video-call score used the difference between loaded and unloaded latency; the current documentation doesn't say how each score is weighted, so check the loaded latency figure yourself.

Apple's test measures how well a network copes when several devices and apps share it. On a Mac running macOS Monterey or later, open Terminal, type networkQuality and press Return. On an iPhone or iPad (iOS or iPadOS 15 or later) you first install a "WiFi Performance Diagnostics" profile from Apple's developer website (which may ask for an Apple Developer login), then run Responsiveness under Settings > Wi-Fi > Info > Diagnostics. Apple says these steps are meant mainly for developers.

The result is in round-trips per minute (RPM): higher is better, and bufferbloat.net describes the range as roughly a hundred (poor) to a few thousand (good). Apple rates the network Low (calls or gaming may be unreliable while another device downloads a film or backs up photos), Medium (momentary pauses when several devices share it) or High.

1. Pause everything else. Waveform's advice is to pause backups, cloud sync and large downloads on every device, because the test can't get a clean reading if the line is already busy.
2. Start with a cable. Waveform recommends testing over Ethernet first. That takes the Wi-Fi's own buffers out of the picture. Waveform adds that on a very fast line Wi-Fi is usually the slower link, while the test needs to fill your internet connection, not your Wi-Fi. See wired Ethernet at home if you need a cable run.
3. Then test on Wi-Fi from the problem room, on the device that suffers.
4. Repeat a few times, at a quiet time and at a busy one. Our guide on why speed test results differ explains why results wobble.
5. Note each result with the time, the tool and wired or Wi-Fi, so you can show your provider or router maker.

For a plain speed check alongside these, use our broadband speed test.

Waveform grades the number of milliseconds (ms) latency rises under load, using what it calls a slightly modified version of the rubric the DSLReports test used:

Bufferbloat.net's fixes page says that if a test shows latency below 15 to 25 ms, bufferbloat is under control. Its FAQ frames it as an increase under load of less than 15 to 25 ms. Its tests page suggests acting on a result above about 50 ms, or a grade lower than B.

• Good on Ethernet, poor on Wi-Fi (Waveform's example is an A and a C): the Wi-Fi is the weak link, so look at coverage first.
• Poor on both: Waveform says that points to your router or your provider, and a new Wi-Fi access point won't fix it. This is where SQM can help.
• Bad only while uploading, or only while downloading: the queue is building in that direction.
• Jitter low when idle but jumping under load points to bufferbloat, Waveform says; high jitter even when idle usually means Wi-Fi trouble.

Quality of service (QoS) settings let you rank traffic so that chosen packets go first. Bufferbloat.net says that helps but won't solve bufferbloat completely. The rest of the queue still has to be sent, and QoS has no way to tell the big senders they're sending too much, so their packets keep piling up. Every rule also needs adjusting when your traffic, router or speed changes.

OpenWrt describes SQM as an integrated system that combines per-flow scheduling, active queue management, traffic shaping, rate limiting and QoS prioritisation, whereas "classic" QoS only prioritises. So SQM's main job is keeping the queue short for every flow, rather than asking you to rank devices. For providers' own priority features, such as game and work modes, see our guide to QoS, work mode and game mode.

SQM runs on the router's processor, so slower routers may not keep up at the full speed of the line, and it doesn't work with hardware flow offloading (a speed-up feature on some routers), though software offloading is fine.

OpenWrt's guide reports one user's results on a cable connection (their figures, not a test of ours; other lines and routers will differ):

In that example, a little less raw speed bought no rise in latency when the line was busy.

eero, which makes mesh routers, says that with its SQM on every device still gets bandwidth, but you may see slower performance on certain devices. Waveform says SQM is CPU-intensive, so faster connections need faster router processors, and Apple says routers with SQM generally need some expert configuration.

SQM works best when it knows your line's speed, which is a problem on connections that vary. Bufferbloat.net notes that mobile (4G and 5G) links, cable and Starlink can change speed from morning to evening, or even minute to minute, and suggests cake-autorate, which keeps watching traffic and latency and adjusts cake's limits up and down. Starlink users can find more about the service on our Starlink broadband page.

Bufferbloat.net calls this a very common solution. Its steps:

1. Turn off the Wi-Fi on the provider's router.
2. Connect the new router's WAN (internet) port to one of the LAN ports on the provider's router.
3. Set up the new router's SQM.
4. Connect all your equipment to the new router, leaving nothing on the old one, which bufferbloat.net says would throw off the algorithms. Then test again.

Our reading is that SQM belongs on the box doing the routing: eero's isn't available in bridge mode, and OpenWrt says to test the upstream router first if a device works as an extender, repeater or bridge. Our guide to using your own router covers the settings each type of line needs and what you can lose by not using the provider's hub.

Meanwhile, Cloudflare's tips are to use a wired connection, move closer to the router and use a router capable of smarter queueing. Moving cloud backups and big updates to overnight also helps.