Bufferbloat in 2026: Why a Gigabit Connection Still Stutters on Video Calls, What Latency Under Load and L4S Actually Fix, and How to Tame the Queue in Your Router
- Internet Pros Team
- October 14, 2026
- Networking & Security
Your speed test says 900 megabits. Then someone in the office starts uploading a large video to the cloud, and every Teams call freezes, voices turn robotic and the web feels like 2005. The connection did not get slower. It got later. The problem has a name, bufferbloat, and in 2026 it is still the most common reason a fast internet line feels slow.
Speed and Delay Are Different Things
Bandwidth is how much data a link can carry per second. Latency is how long a single packet takes to get there and back. A video call uses very little bandwidth, often 2 to 4 Mbps, but it needs low and steady latency. A delay of 30 milliseconds is invisible. A delay of 600 milliseconds makes people talk over each other.
Most speed tests measure latency while the line is idle, and then measure throughput. That hides the real question: what happens to latency while the line is busy? That figure is called latency under load or working latency, and it is where bufferbloat shows up.
Where the Delay Comes From
Every router, modem and Wi-Fi access point holds a buffer: a waiting line for packets that arrive faster than they can be sent. A small buffer absorbs short bursts and is useful. The trouble started when memory became cheap and manufacturers gave devices huge buffers, believing that dropping a packet was always bad.
TCP, the protocol behind most downloads and uploads, speeds up until it sees a sign of congestion, usually a lost packet. A huge buffer delays that signal. So a single large upload keeps filling the queue until it holds hundreds of milliseconds, sometimes whole seconds, of data. Your small, urgent call packets then wait at the back of that line. Researcher Jim Gettys named the effect “bufferbloat” in 2010, and it was usually worst on the upload side, because upload links are slower and fill first.
A bigger pipe does not fix a long queue. If the line in front of the gate is a second long, a faster gate behind it changes nothing for the packet at the back.
How to Test for It
You need a test that measures latency while the line is saturated. Good options:
- Waveform Bufferbloat Test runs a download and an upload and grades the latency increase from A+ to F.
- Cloudflare Speed Test and Ookla Speedtest both now show “loaded” latency during download and upload, next to the idle figure.
- networkQuality, built into macOS, reports responsiveness in RPM (round trips per minute). Higher is better; a few hundred RPM is poor, over 1,000 is good. The IETF is standardising this metric.
Run the test wired, then on Wi-Fi, at a busy time of day. If idle latency is 15 ms and loaded upload latency jumps to 300 ms or more, you have bufferbloat. An increase under about 30 ms is healthy.
The Fix: Smarter Queues, Not Bigger Ones
The cure is active queue management (AQM): the router watches how long packets have been waiting and signals senders to slow down before the queue grows. Several algorithms are now standard:
| Technique | What it does | Where you find it |
|---|---|---|
| CoDel / PIE | Drops or marks packets when queue delay stays above a small target (about 5 ms for CoDel) | PIE is required in DOCSIS 3.1 cable modems |
| fq_codel (RFC 8290) | Gives each flow its own short queue, so one big upload cannot starve a call | Linux default; OpenWrt, many Wi-Fi drivers, firewalls such as OPNsense and pfSense |
| CAKE | fq_codel plus a built-in shaper, per-host fairness and overhead compensation | OpenWrt SQM, Linux since 4.19, several gaming routers |
| L4S (RFC 9330-9332) | Marks packets with ECN instead of dropping them, so apps back off early and keep queues near 1 ms | Rolling out on some cable networks and supported in recent Apple, Android and cloud gaming stacks |
Why the shaper matters
Your router can only manage a queue it controls. If your ISP modem is the bottleneck, the bloated buffer sits there, out of reach. Smart queue management (SQM) solves this by shaping traffic in your router to slightly below the line speed, typically 85 to 95 percent of the measured rate. The queue then forms in your router, where fq_codel or CAKE keeps it short. You trade a few percent of peak throughput for a large drop in delay.
What L4S Changes
Classic AQM still works by dropping packets, which costs a little throughput and only signals congestion after a queue has begun to form. L4S (Low Latency, Low Loss, Scalable Throughput) uses Explicit Congestion Notification bits to tell a sender “slow down slightly” many times per second. Senders that understand L4S, using congestion controls such as TCP Prague or modern versions of BBR and real-time video stacks, react in small steps. The result is full speed with queues of about a millisecond.
The catch is that both ends and the bottleneck must support it. Some cable operators began L4S deployments in 2025, and support is growing in phone and computer operating systems and cloud gaming services. For most businesses in 2026, L4S is a welcome bonus where available, not a replacement for good queue management on your own router.
A Practical Plan for an Office
Cut latency under load in an afternoon
- Measure first. Record idle and loaded latency, wired and on Wi-Fi, so you can prove the fix worked.
- Turn on SQM or Smart Queues on the edge router or firewall, using CAKE or fq_codel, and set it to roughly 90 percent of your measured download and upload speeds.
- Put the ISP modem in bridge mode where possible, so your own router is the device in charge of queuing.
- Check the Wi-Fi. Choose access points with airtime fairness and per-station queuing; one slow, distant laptop should not stall everyone.
- Schedule big transfers. Run cloud backups and large sync jobs after hours, or cap their rate.
- Re-test after changes and whenever your ISP upgrades your plan, because the shaper rate must follow the new line speed.
The Bottom Line
When calls break up on a fast connection, buying more bandwidth is usually the wrong fix. The problem is a queue that is too long, not a pipe that is too narrow. A router with SQM, set up correctly, typically brings loaded latency from hundreds of milliseconds down to a few dozen. L4S will make the internet more responsive over the next few years, but you can fix most of the problem today with your own router.
Are video calls dropping when the office is busy?
Internet Pros tests latency under load, configures smart queue management on business routers and firewalls, and tunes Wi-Fi so calls stay clear while backups run. Contact our team for a network performance review.
