Note · 2026-01-28 · shelf server, 6.6 kernel, schedutil

Taking four cores offline buys almost nothing

It is one of the first ideas everyone has about an idle server: if half the cores are doing nothing, switch them off and save their share of the power. The trouble is that a core doing nothing has already switched most of itself off, and you are left paying to delete something that was already asleep.

Why it is tempting

The arithmetic looks obvious. Eight cores, half of them idle, so remove four and remove their power. It survives exactly until you look at what an idle core actually draws, which on this part in its deepest state is a tenth of a watt. The active power the core could draw is not being drawn, because the core is not active; there is nothing there to reclaim.

The measurement

I took the shelf server down from eight online cores to six to four, idle, five ten-minute windows at each step, nobody logged in. Deepest-state residency is read as a delta, not a single total, for the reasons in the idle-loop note.

Three tenths of a watt for cutting the core count in half. The meter resolves 0.1 W, but the window-to-window scatter on this machine is about 0.2, so 0.3 W is a result I will state and will not defend to a second decimal. Residency barely moves because the online cores were already spending ninety-one per cent of their time asleep; offlining their neighbours does not change how the survivors rest.

What it costs you

The power you save is the deep-idle power of the cores you removed - almost nothing. The thing you lose is that the cores are no longer there when work arrives.

Even that understates the loss, because the synthetic load is light and fits in four cores with room to spare. A real burst that wanted all eight would find four of them gone and would have to wait while they came back, and bringing a core back online is itself a flurry of work the scheduler has to absorb.

On top of which there is the part no table captures: a monitoring system that suddenly sees four cores missing will tell you the machine is broken, every rebalance now steps around a hole in the topology, and the next person to log in has to be told why. That is a standing operational cost paid forever for a one-off third of a watt.

Verdict

Not worth it, and not close. If the goal is a quieter machine, the watt is in the wakeups and the governor, not in the core count - a core that is genuinely asleep is already giving you almost everything offlining it would, without taking itself out of the building. The place this idea does earn its keep is thermal or acoustic, on a fanless box fighting to stay cool, and that is a different measurement than the one on this page.

Machines and method on the notes page. The companion argument - that keeping idle cheap is worth doing for the instrument it leaves you, not the watts - is the last section of the wakeups note.