Gigabit Broadband Tests Fast, So Why Does Backing Up Photos Make Everyone’s Video Calls Stutter?

Fast speed tests, laggy calls during photo backups: a Linux queue experiment explains bufferbloat, SQM trade-offs and practical home-network diagnosis.

Picture a familiar evening: a speed test on your computer comes close to the rate advertised by your broadband plan, and your phone shows full signal strength. You start uploading the day’s photos to cloud storage, and a family member’s video call begins to break up. Pause the backup, and the call recovers.

Buying a more expensive wireless device might seem like the obvious fix. But if the problem is a queue at the broadband connection’s exit, replacing an AP may not help at all.

The question worth asking is: If the network is still transferring data at high speed, why does it feel slow to the people using it?

The answer often lies in two different measures: how much data can be sent in a second, and how long a small packet must wait before it can leave. The first determines when the backup finishes. The second directly affects whether calls, games, and remote controls respond promptly.

This article uses a reproducible, isolated Linux network experiment to explain the difference, then translates it into checks and trade-offs you can apply to a home network. The experiment is not a test in a real home and does not represent the performance of any particular router.

The recommendation up front: if “uploading makes everything stutter” is reliably reproducible over a wired connection, first check upload capacity and latency under load. Once you confirm queuing at the broadband exit, compare limiting the backup’s upload rate with using SQM. If the problem occurs only in a particular room, or only when accessing a local NAS, investigate that local network path first.

Gigabit in Which Direction?

“Gigabit broadband” often leaves out a crucial detail: the separate download and upload rates.

An excellent download speed-test result does not mean the upload capacity is the same. Check the plan and measure the actual rates; do not assume every home broadband connection is asymmetric. Photo backups, security-camera uploads, and remote access to a NAS at home primarily consume upload capacity.

Suppose a connection has an actual usable upload rate of about 20 Mbit/s. One background task could saturate it even while plenty of download capacity remains.

Once it is saturated, new packets have to queue. A video call may use far less data than a photo backup, but its packets must arrive on time. The file data ahead of them can wait; the call data cannot.

An idealized calculation gives a sense of the scale:

Queuing delay ≈ amount of data queued ahead × 8 ÷ bottleneck transmission rate

1,000,000 bytes × 8 ÷ 20,000,000 bit/s = 0.4 seconds

Here, 1 MB is calculated in decimal units. The 400 milliseconds is only the delay contributed by this queue. It excludes propagation, other network nodes, and application processing, and it is not the full round-trip time. Actual results also depend on scheduling, packet sizes, and bursts.

This also explains why “no packet loss” does not necessarily mean the network is healthy. If a device tries to retain all the data in a long queue, the file may eventually arrive while real-time communication has already missed its playback deadline.

Buffers are not inherently bad; they absorb short bursts. The problem is a queue that stays backed up, turning brief buffering into sustained waiting. What people call bufferbloat is associated with this excessive queuing delay.[1][2]

Before Changing the Router, Find Where the Data Is Queuing

A home network does not have just one “network speed.” The same backup action can encounter entirely different bottlenecks depending on its destination.

Action and symptom Path to check first Could SQM on the WAN help?
Uploading to cloud storage also disrupts an internet call on a wired computer Router exit, optical modem, and ISP upload access Possibly, provided you can control the queue ahead of the relevant bottleneck
A phone transfers photos to a NAS on the same LAN, and nearby wireless devices slow down Wireless airtime, the AP’s wired interface, and switching links Usually not; the traffic may never traverse the WAN
A group of upstairs devices gets busy and the whole floor slows down Floor switch uplink and negotiated interface speed Usually not for a bottleneck within that floor
Only one room has problems, and a wired connection works normally Coverage, interference, retransmissions, and wireless backhaul Do not attribute it to the broadband exit first
Multiple internet destinations remain problematic even when the network is idle Access line, routing, device load, and remote endpoints Diagnose first; “stuttering” alone does not establish bufferbloat

In particular, distinguish access to a local NAS from remote access through a public address, proxy, or relay. An application saying it is “connected to the NAS at home” does not automatically mean its traffic stays on the LAN.

The earlier article Villa Wi-Fi Roaming and Cabling Plan covers coverage, roaming, and backhaul. This article addresses a different layer: even when you stay in one place, have a strong signal, and have working wired backhaul, queuing at the broadband exit can still slow everyone’s interactive traffic.

An Isolated Experiment: The Same Bottleneck, a Different Queue

To avoid drawing conclusions from an analogy alone, this experiment created two separate network namespaces in the WSL2 Linux kernel and connected them with a pair of virtual network interfaces. The experimental network had no default route and was not connected to a home broadband line.

On the sending side, HTB fixed the outbound rate at 10 Mbit/s. Two queue configurations were then attached in turn:

  • FIFO: first in, first out, with a queue limit of 1,000 packets.
  • FQ-CoDel: also with a 1,000-packet limit, using per-flow scheduling and active queue management; the experiment used a 5 ms target and a 100 ms interval.

A continuously sending TCP flow generated the load, while a ping was sent every 20 milliseconds for a total of 300 pings. An idle baseline was recorded first. Each of the two loaded configurations was then run three times, with the order varied between runs. TSO, GSO, and GRO were disabled on the virtual interfaces to reduce the effect of aggregated large packets on the experiment.

The measured results follow, in milliseconds. P95 was calculated separately from the successful responses in each run; the three runs were not combined into a single “best” result.

Configuration Run Median RTT P95 RTT Maximum RTT Pings received / sent
Idle baseline 1 0.066 0.099 0.427 300 / 300
FIFO + TCP load 1 140.0 164.0 210.0 300 / 300
FIFO + TCP load 2 141.0 164.0 206.0 300 / 300
FIFO + TCP load 3 141.0 171.0 204.0 300 / 300
FQ-CoDel + TCP load 1 1.220 3.640 5.710 300 / 300
FQ-CoDel + TCP load 2 1.705 3.970 5.950 300 / 300
FQ-CoDel + TCP load 3 1.425 4.580 6.210 300 / 300

The most instructive point is that both configurations received every ping response, yet their waiting times differed greatly. The FQ-CoDel queue statistics still showed packet drops, so “zero ping loss” must not be presented as “zero loss across all traffic.” This experiment did not measure effective throughput at the receiver, so it cannot establish that the latency improvement came at no throughput cost.

The comparison shows whether, under the same rate limit, bulk transfer must make a separate, sparse flow wait a long time.

It must not be extended into a claim that “turning on SQM will make your home video-call latency match the numbers in the table.” This setup had no real wireless retransmissions, ISP queues, complex traffic mix, or video encoding and decoding. Ping is not equivalent to an actual audio/video stream. The FIFO limit of 1,000 packets was an intentionally chosen comparison condition, not evidence that your router uses that queue.

The experiment also did not isolate the respective contributions of fair scheduling and CoDel, or measure the maximum processing rate of different routers. Although the offloads listed above were disabled, the queue logs still contained kernel packet lengths larger than a typical Ethernet MTU. The setup therefore cannot be treated as a frame-by-frame recreation of transmission over a physical cable. It supports a distinction between queuing mechanisms, not a ranking of devices to buy.

What Does SQM Actually Do? Separate Three Actions

Router interfaces use many names, including “Smart QoS,” “game acceleration,” and “bandwidth management.” To assess a feature, it is better to examine its specific mechanism.[1][2]

The first action is shaping: controlling how quickly data is sent.

Suppose the actual bottleneck is in the ISP’s equipment, while your router keeps feeding it data at a higher rate. The backlog may then form somewhere you cannot control. Setting the router’s sending rate slightly below the stable usable capacity aims to move the controllable bottleneck to your side.

That is why deliberately sending a little less can sometimes produce a substantial improvement in responsiveness. But how much less requires testing: set the rate too low and you waste capacity; set it too high and you may fail to control the downstream backlog.

The second action is scheduling: deciding whose packet goes next.

FIFO serves packets in arrival order. Per-flow fair scheduling instead gives different flows separate attention, preventing one continuous download or upload flow from keeping a sparse flow waiting behind it for a long time.

Per-flow fairness, however, is not the same as fairness between household members. A task that opens many connections may receive a different share from a single-connection task. VPN encapsulation can also make multiple inner flows appear as a small number of outer flows at a given observation point. Whether host isolation is available, and whether the queue sees traffic inside or outside the tunnel, can both affect the result.[2]

The third action is active queue management: not waiting for a buffer to fill before addressing congestion.

CoDel tracks how long packets spend in the queue. During sustained congestion, it signals congestion to senders by dropping packets or, for eligible traffic, marking them with ECN. Transport protocols that respond to those signals adjust their sending behavior. This does not guarantee that all traffic will cooperate, nor does it guarantee zero packet loss.

FQ-CoDel combines scheduling with active queue management. But attaching an fq_codel queue on its own does not automatically shape the broadband exit to the connection’s actual rate. HTB supplied the rate limit in this experiment. CAKE integrates additional capabilities, including shaping; OpenWrt’s SQM scripts organize these mechanisms into a working configuration.[1]

So “I already checked the QoS box” is not the end of the investigation. Ask which interface and direction it controls, whether it has an appropriate shaping rate, and whether the traffic actually passes through that queue.

Limit the Backup First, or Go Straight to SQM?

If one known task causes the problem, the simplest remedy is often effective: lower that application’s upload limit, or schedule large backups for times when nobody needs a video call.

For example, if the stable actual upload rate is about 20 Mbit/s, you could start by trying a backup limit of 12–15 Mbit/s, then see whether the call quality and backup completion time are acceptable. This is a starting example, not a guaranteed setting. If the application displays MB/s, convert the units first. Limits set separately on multiple devices may still add up to enough traffic to saturate the exit.

With many devices and unpredictable upload tasks, managing limits one by one becomes difficult. That is when applying SQM at the exit becomes more valuable.

Method When it fits Cost and limits
Application rate limits and scheduled backups The responsible task is known, and there are few devices Simple, but another task may saturate the exit
Traffic priority Important traffic can be identified and classified reliably Classification can be wrong; priority alone may not help if the backlog is elsewhere
Shaping with fair scheduling and AQM Multiple devices share the connection, and loaded-latency problems are reproducible Requires tuning and CPU resources, and usually sacrifices some peak throughput
Upgrade genuinely insufficient upload capacity Total workload consistently exceeds capacity More capacity helps, but does not replace checking behavior under full load

One distinction is particularly important: if the call and the upload tasks that must keep running together need close to or more than the available upload capacity, even excellent queuing cannot create bandwidth. It can improve how capacity is allocated; it cannot turn 20 Mbit/s into 100 Mbit/s.

Four Easy Mistakes to Make on OpenWrt

First, entering the plan’s advertised rate directly. OpenWrt’s official introductory guide recommends starting at about 90% of measured upload and download speeds, then testing and adjusting.[1] “90%” is a starting point, not a law that fits every connection. Measure each direction separately, identify fluctuations, account for encapsulation overhead such as PPPoE, and check the units used by the software interface.

If usable capacity drops substantially during busy periods, shaping configured against an idle-time peak may stop working. Fixed settings have inherent limits on access networks whose capacity varies greatly.

Second, choosing the wrong interface or letting traffic bypass the queue. Do not blindly copy an interface name from someone else’s screenshot. Check the actual WAN path and the upload and download directions, then verify that the queue counters change under load.

OpenWrt’s current guidance explicitly warns that hardware flow offloading is incompatible with its SQM because it bypasses some kernel processing. The same guidance says software flow offloading can be used with it.[1] It would therefore be wrong to claim that “all traffic offloading must be disabled.” Verify other firmware or vendor acceleration paths according to their specific implementations.

Third, insufficient router processing power. A router capable of ordinary gigabit forwarding may not maintain the same throughput with SQM enabled. Watch CPU use under full load, particularly individual cores and soft interrupt load, while recording throughput. A modest overall CPU percentage can hide a saturated core that handles critical processing.

Fourth, assuming download shaping can undo delay that has already happened. On upload, you can schedule packets before they leave the home. On download, packets have already crossed the access link by the time they reach your home. Ingress shaping can influence senders through subsequent congestion feedback and other mechanisms, but it cannot remove time the packets have already spent waiting upstream. Nor is it guaranteed to handle all traffic that fails to respond to congestion.

These limits mean SQM should be treated as a configuration change you can verify and roll back, not a switch that is permanently correct once enabled.

How Can You Check at Home Without Relying on “It Feels Faster”?

For the first round, connect the test computer by cable, keep the device, server, and testing period consistent, and record the current configuration. Exclude wireless variation first, then repeat the checks at the Wi-Fi location where the problem occurs.

Test the following four states for at least one minute each and repeat each three times. This is a starting procedure for home troubleshooting, not a universal certification standard.

State Load to generate What to watch
Idle Pause backups and downloads; check traffic from other devices Whether baseline round-trip latency is stable
Upload saturated Upload continuously to a reliable external destination Whether uploading causes latency to rise
Download saturated Continuously download enough data Whether the download direction has a similar problem
Bidirectional load Upload and download simultaneously Competition in both directions, CPU, and actual application behavior

Do not ping only the router. The gateway can respond normally while a WAN queue is backed up. Observing both the gateway and several stable external targets makes it easier to narrow down the location. Conversely, a remote endpoint may rate-limit or deprioritize ICMP, so a poor ping result to one target alone does not prove that application traffic has a problem.

Some speed-test tools already report latency under load. The issue is not that “a speed test can never reveal this,” but whether you look only at its largest Mbps number.

For each run, it is useful to keep a record like this:

Time / wired or wireless / test target / configuration version
Load direction / duration / actual upload and download throughput
Median RTT / P95 RTT / packets sent and lost
CPU and queue counters / video-call or remote-control behavior

P95 means that about 95% of successful response samples are at or below that value, helping reveal the slow tail. It cannot replace the packet-loss rate: a timeout produces no RTT, and a P95 calculated while ignoring timeouts may look better as a result. Jitter statistics in applications have their own definitions, too; do not casually treat “maximum RTT minus minimum RTT” as audio/video jitter.

Then change just one thing: try limiting the backup, or enable one clearly documented set of SQM parameters. Retest under the same conditions without selecting only the best run. Finally, while the load is still present, make an actual video call or use the remote-control application to confirm that the improvement reaches the service that matters.

There is no “passing P95” that applies to every home. More useful questions are how much latency rises from idle to full load, whether the slow tail consistently comes down, whether packet loss is acceptable, and whether the throughput sacrificed is worth it.

If throughput falls markedly but loaded latency does not improve, do not keep lowering the rate blindly. Check the interface, bottleneck location, CPU, and test targets before tuning further or restoring the previous configuration. After a router restart, confirm that the rules still apply and recheck the key load scenarios.

Leave Room for the Home Network to Respond on Time

Back in the opening scenario, the most useful next step is not necessarily to buy new equipment.

First confirm where the photos are going. Compare idle and upload-saturated latency over a wired connection, then use a backup rate limit as a comparison. If the evidence points to queuing at the broadband exit, SQM has a clear role. If the bottleneck is wireless airtime or a floor uplink, repair that part of the path instead.

We pay for gigabit service to make the internet easier for the whole family to use. An acceptable outcome may be a backup that finishes a few minutes later, while every word of a family member’s call reaches the other end on time.

That moment should be part of how a network solution is judged: when the background is busiest, can the people in the foreground still use it normally?

Sources and Experiment Notes

  1. OpenWrt: SQM (Smart Queue Management). Used to check starting configuration guidance, CPU costs, offloading compatibility, and CAKE/FQ-CoDel choices. Checked: 2026-11-16.
  2. RFC 8290: The Flow Queue CoDel Packet Scheduler and Active Queue Management Algorithm. Used to explain scheduling, AQM, per-flow fairness, and encapsulation boundaries. This document is categorized as Experimental; it should not be described as a mandatory standard that all home devices must implement.

The 20 Mbit/s upload rate, 1 MB queue calculation, and backup rate-limit range in this article are illustrative examples. The Linux experiment’s script, raw logs and results, and reproduction instructions are stored with the experiment materials in the same directory. No tests were conducted on a physical home broadband line, video-call quality, or different router models.

Download the complete experiment materials (ZIP, including script, raw logs, and results).

AI-generated scene illustration, not a photograph of this project or a wiring instruction.

Discussion

Share a question or your experience. Comments appear after moderation. Your email is never displayed publicly. Comment privacy

Loading comments…