A 10 Gbps campus with 2 Gbps of traffic does not need QoS. A 200 Mbps circuit carrying a voice call, a backup, and forty Chromebooks does.
QoS is not a performance feature you turn on. It is a set of decisions about who waits when something is full, and the first job is finding out what is actually full.
The four problems
| Problem | Measured as | Felt as |
|---|---|---|
| Bandwidth | bits per second | Slow downloads, long page loads |
| Delay | one-way latency | Conversation lag |
| Jitter | variation in delay | Robotic voice, frozen video |
| Loss | dropped packets | Clicks, retransmits, stalls |
QoS can influence all four, but only when there is contention. On an idle link it changes nothing except CPU load.
Bufferbloat: why a fast link feels slow
A router with a large buffer accepts a burst, holds it, and delivers it late. Latency climbs to seconds while throughput looks excellent. This is bufferbloat, and it is the reason a 500 Mbps connection can feel worse than a 50 Mbps one.
The fix is a shaper that keeps the queue short (see the WAN series), plus a scheduler that puts latency- sensitive traffic first.
Throughput is not responsiveness
A speed test measures the pipe when nothing else is using it. Users experience the pipe when everything is using it. The number that predicts complaints is latency under load, not peak throughput.
When QoS is the right answer
- A WAN circuit smaller than the traffic that wants to use it.
- Voice or video sharing a link with bulk transfers.
- A wireless network where airtime is the bottleneck.
- Any link where a single flow can starve others.
When it is not
- Congestion caused by a failing link, a duplex mismatch, or a loop.
- A LAN with 10× headroom.
- A capacity problem that needs a bigger circuit.
- An application problem that looks like network latency.
QoS moves the queue; it does not create capacity.
Worked example: 100 Mbps to a small campus
Traffic at 09:00: 60 Mbps of web, 15 Mbps of backup, 2 Mbps of voice.
At 77 Mbps, nothing waits. At 10:00 a cloud backup saturates the link, latency goes from 12 ms to 400 ms, and voice breaks up.
QoS does not add bandwidth. It ensures the 2 Mbps of voice is transmitted first, at the cost of the backup finishing later — which nobody notices.
What breaks in production
- QoS enabled everywhere with no contention, adding complexity and no benefit.
- QoS as a substitute for capacity, so the link stays saturated and everyone is equally slow.
- No measurement, so nobody knows whether it worked.
- Policy that classifies everything as priority, which is the same as no policy.
- Ignoring the return path, so inbound congestion still ruins calls.
Checklist
- You can name the link that is actually congested.
- You know the traffic mix on it at peak.
- QoS is applied where the queue forms.
- Success is measured as latency and jitter under load.
- Capacity has been considered before policy.
Next
Decisions require labels. Continue with Classification and Marking.