Note · 2026-10-06 · desk machine, 6.6 kernel, schedutil
What an idle login costs
The cheapest way to spoil a power measurement is to stay and watch it. A terminal left open on the machine you are measuring draws real, repeatable watts of its own, and they land on top of whatever you were trying to see.
The number, and why it is awkward
The governor note put the cost of a stray login at 1.1 W on the shelf server, in a single parenthesis, as the reason I log out before logging. This note is the measurement behind that parenthesis, taken on the desk machine, where it comes to the same figure by a slightly different route.
The awkward part is that 1.1 W is not «ssh». A bare shell sitting at a prompt is almost free. The watt arrives with everything people hang off the prompt - the status line that re-renders every second, the clock, the thing that runs a version-control query each time the cursor moves. Most of the cost of being logged in is the cost of the login being comfortable.
Where the watt goes
| Session state | Draw | Δ |
|---|---|---|
| nobody logged in | 19.3 W | - |
| one bare shell, no prompt dressing | 19.6 W | +0.3 |
| a prompt that redraws once a second | 20.4 W | +1.1 |
| plus a refresh loop left running at two seconds | 21.1 W | +1.8 |
The baseline is the schedutil figure from the governor note, so the rows are comparable to everything else on the site. The 1.1 W I keep quoting is the third row - the ordinary case, a prompt doing a little work every second, which is what almost everyone actually has open.
Counting the wakeups behind it
The watts are easier to believe once the wakeups are accounted for, using the same discipline as the wakeups note: count the total, attribute it, check that the parts add up to the whole.
| Source | Logged out | Prompt open |
|---|---|---|
| scheduler tick housekeeping | 14 | 14 |
| network card link and arp | 5 | 5 |
| filesystem commit | 3 | 3 |
| pseudo-terminal and line discipline | 0 | 8 |
| the prompt and the processes it spawns | 0 | 39 |
| total | 22 | 69 |
Forty-seven extra wakeups a second for 1.1 W, which is a touch over twenty milliwatts apiece on this part - in the same range as the daemon and the agent in the wakeups note, which is the sort of agreement between two unrelated measurements that makes me trust both a little more.
The one source I left off the table on purpose is the keepalive probe the connection sends to notice a dead link. At one packet every thirty seconds it is 0.03 wakeups a second and well under the meter’s error; putting it in the table would imply I could measure it, and I cannot.
Why it still matters at two euros a year
As a saving it does not. At what I pay for electricity, 1.1 W is about two euros a year, and nobody should spend an afternoon on it for its own sake.
It matters because it is bigger than the things I am usually trying to measure. The whole spread between the best and worst governor on this machine is 6.2 W; the step from schedutil to powersave is 0.9 W; the disk policy in another note is worth less than a watt in real duty. A stray login is larger than any of those, which means a measurement taken while I am logged in cannot see them at all - the thing I came to measure is hiding under the thing I am doing to measure it.
So the rule from the first note is not fussiness. Start the logger, leave, come back. If you cannot leave - a laptop on battery, a machine you can only reach through one session - at least strip the prompt to nothing for the duration, which the table says recovers most of the watt. And write down that you were logged in, because a number you cannot reproduce without yourself in the room is not a measurement, it is an anecdote.
Method and the two machines are described once on the notes page. The wakeups note uses the same accounting on a worse offender.