Why 1 Percent Packet Loss Feels Like Losing the Whole Connection

Archi4 min read

networkingtelecominternetinfrastructure

Contents
  1. A Fast Link That Feels Slow
  2. One percent loss does not sound like much.
  3. Why don't the parallel streams in a speed test catch this?
  4. Where does the loss come from?
  5. Is loss worse than latency?
  6. What does the operator do when they find it?
  7. What can a household check tonight?
  8. The part that surprises people

Speed tests lie in a specific way. They open a dozen parallel streams, run for ten seconds, and report the best result. Real traffic is often one stream that has to survive every queue between your router and the server: a video call, a file sync, a game session.

That gap is where the surprising number comes from. Ask an engineer who handles transport and peering for a regional fiber operator, and the figure they keep returning to is 1 percent. Not 1 percent of bandwidth. One percent packet loss. They asked not to be named because they are not cleared to speak for the company, so what follows is their account of how that single percentage behaves in the field.

One percent loss does not sound like much.#

It is not much until you put it through the TCP throughput rule of thumb. A single flow tops out near the segment size divided by the round trip time, multiplied by the square root of the loss rate. Use ordinary numbers: 1,460 byte segments, a 40 millisecond round trip, and 1 percent loss. The answer lands under 3 megabits per second on a link sold as a gigabit.

At 0.1 percent loss, the same link gives one flow about 9 megabits. At 0.01 percent, roughly 29. Operators see this constantly. A customer on a 1 Gbps plan says uploads crawl while the speed test still shows 940 Mbps.

Why don't the parallel streams in a speed test catch this?#

Because sixteen streams split the loss between them. Each flow runs slow, but the total is the sum, so the test looks healthy. Most applications do not work that way. A backup agent, a single TLS connection to a SaaS app, and a two-way video call each depend on one or two flows.

The engineer's summary: bandwidth is what you buy, and loss is what you actually experience.

Where does the loss come from?#

The usual suspects are unglamorous. Microbursts on a 100 Gbps port with shallow buffers. An uplink that is fine at 10 a.m. and oversubscribed at 8 p.m. A policer on a wholesale handoff that drops bursts instead of queueing them. Wi-Fi retransmissions, which TCP cannot tell apart from real congestion. A duplex mismatch on an old switch. One bad patch cable in a rack nobody has opened in four years.

Their favorite example is a policer set to 1 Gbps on a 10 Gbps handoff. It never buffers, so it drops, and the customer sees 1 percent loss at exactly the moment traffic spikes, which is exactly when it hurts.

Is loss worse than latency?#

For most applications, yes. Adding 30 milliseconds of round trip to a clean path is annoying. Adding 1 percent loss to the same path is worse for a single stream. Latency can be designed around with edge caches and connection reuse. Loss cannot, except by fixing the path or changing the congestion control.

This is also why the same network can feel fine to YouTube and terrible to a self-hosted app. Large content networks run senders that tolerate loss, such as BBR, while plenty of internal applications still use a classic loss-based algorithm that backs off hard.

What does the operator do when they find it?#

Measure first. Per-queue drop counters, TWAMP or iperf3 probes between edge routers, out-of-order and retransmit counts on both ends. The engineer is blunt that most tickets close on evidence, not intuition.

Then fix the shape of the problem. Add uplink capacity where peaks are real. Increase buffer where bursts are the cause. Put fair queueing with CoDel on the access side. Replace policers with shapers, because a queue that delays is survivable and a policer that drops is not. None of these are exotic. They are just unglamorous work that does not photograph well.

What can a household check tonight?#

Start with a cable. Wi-Fi retransmits count as loss to every transport above it. Then look at the router: an older unit forwarding at 100 Mbps or running near full CPU will drop under load, and the symptom looks like a bad line, not a bad router.

On cable, check upstream signal to noise and the modem's error counters. On fiber, check that the ONT is not reporting errors. On any connection, look at the retransmit counter on your own machine (netstat -s on Linux and macOS, Get-NetTCPConnection on Windows) before blaming the provider. A single percent of retransmits is a real finding.

The part that surprises people#

The relationship is a square root, not a straight line. Dropping loss from 1 percent to 0.1 percent buys about three times the throughput on a single flow, not ten. Getting to 0.01 percent buys another three times. Engineers who expect a tenfold jump from a tenfold drop keep getting disappointed.

That curve also explains the support queue. The last fraction of a percent is where most of the calls live, and it is the hardest to measure, the least visible on a dashboard, and the most likely to be dismissed as a customer's imagination.

The engineer puts it flatly: anyone can sell a gigabit. Keeping 99.9 percent of the packets moving is the actual job.

More articles