One division and three arguments
412,000 kWh through the facility meter, 286,000 kWh through the IT meter, thirty days. The ratio is 1.441 and the overhead is 126,000 kWh, which is 175 kW running continuously to support 397 kW of servers. At 14.2 cents that overhead is $217,686 a year, and that is the number worth putting in front of anyone who thinks efficiency work is a hobby.
The three arguments are all about definitions. What is inside the facility boundary. Where the IT meter sits. Whether the period was representative. None of them are settled by arithmetic, and all of them move the answer more than a year of tuning would.
Moving the IT meter is not an improvement
A UPS running at partial load can lose several percent of what passes through it. Read the IT figure at the UPS output and those losses land in the overhead, where they belong. Read it at the UPS input and they get counted as IT, the numerator stays put, the denominator grows, and the ratio falls. Nothing in the room changed. Some of the most impressive PUE improvements on record are exactly this, and it is why the boundary note matters more than the number.
The tightest honest boundary is the rack PDU total, which puts UPS losses and distribution losses both in the overhead. It gives the worst-looking number of the three and it is the one that tracks real change.
What a hundredth is worth
At the defaults, one hundredth of a PUE point is $4,941 a year. That is a usable unit for judging a project: a containment retrofit that takes the ratio from 1.44 to 1.35 is worth about $44,000 a year at this load and this price, and it is a fixed thing rather than a percentage that shrinks as the room grows.
Working from 1.441 to a target of 1.30 is $69,452 a year at the same IT load. Whether the room can get there is a question about the equipment and the building, not about the arithmetic — a facility with an air cooled plant in a hot climate and one with a water side economiser in a cold one are not competing in the same event.
Why the trend lies in both directions
Decommission a rack and PUE gets worse. The lighting, the controls, the standing losses and the minimum flow through the plant are all still there, spread over less IT load, so the ratio climbs while the building draws less power. Add a rack into an under-loaded room and PUE improves for the same reason in reverse. Neither is an efficiency change.
Seasonal swing does the same thing on a shorter cycle. A room with mechanical cooling reads best in February and worst in August, and comparing a July figure with a January one tells you about the weather. Twelve rolling months compared with the twelve before is the smallest comparison that carries information, and even then the IT load needs to be roughly flat for it to mean what it looks like.
Questions people ask
What is a good PUE?
This page will not name one, because the published figures are survey averages and marketing claims rather than requirements, and the populations behind them look nothing like a room with three cooling units and a UPS in the corner. The comparison that carries information is your own room against itself, measured at the same meters, over the same length of period, at a similar IT load and a similar outside temperature.
How do I calculate PUE without sub-meters?
You mostly cannot, and it is better to say so than to estimate. The one case that works is a room whose entire load, cooling included, sits behind a single meter, with the IT total available from the UPS or PDU displays. Anything else means apportioning shared plant, and an apportioned PUE is an opinion. If the aim is to track improvement rather than publish a figure, a clamp meter on the mechanical feeds for a fortnight is more use than a guess for a year.
Why did my PUE get worse after we removed servers?
Because the overhead is largely fixed and the denominator shrank. Lighting, controls, standing transformer losses and the minimum flow the plant will run at do not fall when a rack leaves, so the same overhead now sits on top of less IT load and the ratio climbs. The building is using less energy and the metric reads worse. This is the best known reason not to manage a room by PUE alone.
Is DCiE just PUE upside down?
Yes, exactly that. DCiE is 1 divided by PUE, expressed as a percentage, so a PUE of 1.441 is a DCiE of 69.4 percent. It says the same thing in a form some people find easier: 69 cents in every energy dollar reaches a server. There is no extra information in it and no reason to compute both except that different audiences prefer different shapes.
Should I include the office lighting?
That is a boundary decision, and the only wrong answer is making it differently each time. If the office is on the same switchboard and cannot be separated, either include it and note that you did, or fit a sub-meter. What matters is that the boundary drawn this quarter is the boundary drawn last quarter, because otherwise the trend is measuring your accounting rather than your building.