Put energy use beside PUE
When an efficiency ratio falls, I would like to see the two numbers that existed before the division.
A report with good news
Suppose a data center efficiency report arrives. Its first page says PUE fell from 1.30 to 1.17. That looks like progress. On the next page, total energy use has increased.
Which page did we misread?
The US Department of Energy’s explanation gives a clear calculation: total annual facility energy divided by annual IT equipment energy. Dividing kWh by kWh leaves a dimensionless ratio.
One number is convenient. Before the division, though, there were two.
IEA 4E’s public overview of its liquid-cooling report also identifies a limitation: PUE can understate efficiency gains from liquid cooling. I have read the overview, and am not evaluating individual savings estimates or the operating conditions of a particular installation here.
It made me want to put the original quantities back beside the ratio.
Before the division
Take small, invented quantities unrelated to any real installation. In the first year, IT equipment uses 1,000kWh and cooling, power infrastructure, and other non-IT loads use 300kWh. Total energy is 1,300kWh. PUE is 1.30.
In the next year, suppose IT uses 1,200kWh and the rest uses 200kWh. The total is 1,400kWh. Divide 1,400 by 1,200 and the result rounds to 1.17.
PUE has fallen. Total energy has risen by 100kWh. The calculations agree; the ratio and the total answer different questions.
Non-IT energy in this example fell by 100kWh while IT energy rose by 200kWh. Keeping only the quotient makes those movements harder to revisit.
We cannot attach a cause such as “the cooling change raised server energy” to these numbers. They were invented for arithmetic and contain no information about what changed. They supply no basis for recommending a coolant temperature.
We also do not know how much work was completed. Doing substantially more computation would provide context for increased energy use. Using more energy for the same work would raise another question. Completed work is not the denominator of PUE.
I would first add two fields to the report: total energy and IT energy, using matching periods and measurement boundaries. Beside them, I would want a way to examine what work was completed.
Room for another number
I do not think an efficiency ratio is inevitably a poor summary. Comparing supporting infrastructure energy with IT energy gives it a useful job. Asking that one ratio to explain energy savings and computational performance creates the difficulty.
The reader can be more specific too. Did the facility use less energy? Did the overhead relative to IT fall? Did the same computation finish using less energy? Each question needs different evidence.
The reporting periods need to match. Comparing a year on one side with a brief high-load interval on the other makes the change hard to interpret. The report also needs to say which non-IT equipment was included.
A wider table seems a reasonable price. A large ratio on its own might let a meeting move quickly, but someone asking about the total would send us back to the underlying records.
Return to the imaginary report. Its first page now shows 1,400kWh beside 1.17. There is no reason to remove either. Both came from the example’s arithmetic.
I would leave the final space for the work completed with that energy. Before turning the page, I would ask about that field.