energy proportionality

Think of a car that burns nearly as much fuel idling at a red light as it does cruising on the highway. That would be absurd and wasteful — you would want fuel use to track how hard the engine is actually working, dropping to almost nothing when you are stopped. Energy proportionality is the wish that a computer behave like a well-designed engine: draw power in proportion to the work it is doing, so an idle or lightly-loaded server sips power and a busy one drinks a lot, with everything in between scaling smoothly. The trouble is that real servers historically have not behaved this way at all.

The precise problem is idle power. A typical server at 0% utilization — doing nothing useful — still draws a large fraction, often half or more, of the power it would draw at 100%. So the relationship between load and power is not a line through the origin; it is a line with a big constant offset. Why does this matter so much in a warehouse-scale computer? Because servers there spend most of their time at low-to-moderate utilization (often 10–50%), rarely flat out and rarely fully idle, and that is exactly the region where a non-proportional server is least efficient — it is paying near-peak power to deliver a fraction of peak work. Energy proportionality asks designers to flatten that offset so power falls steeply as load falls: ideally a server at 20% load uses about 20% of peak power.

Energy proportionality matters because it directly attacks total cost of ownership and carbon footprint at exactly the operating point WSCs actually live in. Techniques that push toward it include dynamic voltage and frequency scaling (slowing and lowering the voltage of cores when demand is low), powering down idle components, and consolidating work onto fewer machines so others can be put into deep low-power states. Note the contrast with PUE: PUE is about the building's overhead, while energy proportionality is about the servers themselves wasting power when underused — they are complementary, and you want both.

An honest caveat: perfect proportionality is hard, because much of a server's hardware (memory, the chip's leakage current, fans, power supplies) has a floor that does not scale down with load, and waking from deep low-power states adds latency that can hurt tail latency. Idle servers also represent stranded capital — you paid for hardware that sits unused — so the deepest fix is often not just lowering idle power but raising utilization (packing more work onto each machine) so servers spend less time idle in the first place. Proportionality and high utilization are two sides of the same efficiency coin.