Fix the four conditions
| Condition | Why it must be fixed |
|---|---|
| Hour | Peak and off-peak are different networks in practice |
| Target and file | Different hosts and sizes produce incomparable numbers |
| Client and rules | A rule change mid-test invalidates the comparison |
| Device and network | Wi-Fi, Ethernet and mobile are not interchangeable |
What to measure
- Sustained transfer. A file large enough to run for tens of seconds, not a 1 MB sample.
- Variance. Run it three times and record all three. The spread predicts stalling better than the average.
- Interactive behaviour. If your use is calls, note whether audio drops during the same window.
- The control. Same hour, proxy off. This is the measurement people skip and then misinterpret.
The control matters more than the test
If direct traffic also degrades at 21:00, the result describes your line, not the provider. Without the control you will attribute a local peak to the node, change nodes, and see no improvement — then conclude the provider is bad when the constraint was never on their side.
Reading the comparison
| Peak result | Control at the same hour | Conclusion |
|---|---|---|
| Degrades badly | Also degrades | Local or ISP peak; not a node problem |
| Degrades badly | Fine | Route or provider capacity at peak |
| Fine | Fine | No peak problem for this target today |
| Fine | Degrades | The proxy route is better than direct at this hour |
How many runs
One evening is an anecdote. Three evenings, on different weekdays, at the same hour, with the same target, is enough to say something you can act on. Record the dates; a provider cannot investigate a single timestamp that might have been an outlier.
What to send a provider
- Dates, hour and time zone.
- Node label, protocol, observed exit.
- Fixed target and file size.
- Three runs at peak with all values, plus the control.
- Whether other nodes in the same region behaved the same way.
Why a single number cannot describe peak behaviour
A throughput figure is one sample of a distribution. Peak-hour problems are distribution problems: the average may look acceptable while the tail — the slowest 10% of seconds — is where streaming stalls and calls drop. That is why three runs and their spread matter more than one impressive result.
| What you record | What it predicts | What it hides |
|---|---|---|
| Best of three runs | Ceiling under good conditions | Everything about consistency |
| Average of three runs | Typical throughput | The stalling tail |
| Worst of three runs | The floor you will actually experience | How often you hit it |
| All three values | Spread, and therefore predictability | Nothing important at this sample size |
Report all three. A provider can act on “21:00, three runs: 42, 8, 6 Mbps, direct control 45 Mbps” and cannot act on “slow at night”.
What peak-hour data cannot tell you
- Why it slowed. The measurement locates the hour and the magnitude, not the mechanism.
- Whether tomorrow will be the same. Three evenings give a pattern; one gives a coincidence.
- Whether another provider is better. Only the same test, same target, same hour, on their service answers that.
- Whether the service is suitable for streaming. Bitrate requirements and platform detection are separate questions from raw throughput.
Scope
This page describes a measurement protocol. It publishes no results for any service and makes no claim about evening performance on any particular network — that depends on region, carrier and month.