Could FVER Take the heat out of PUE by Peter Judge
Measuring data center efficiency is important. It’s not rocket science but there are lots of complexities, so there are in theory plenty of different ways to measure efficiency.
Which one you choose will have consequences on how you design your data center - as well as the goals that the whole industry moves towards.
PUE (power usage effictiveness) the straightforward option put forward by the Green Grid, is very simple - just divide the total power by the power that reaches the IT equipment. It has riegned supreme for a long while, but has plenty of critics.
The British Computer Society (BCS) has put forward a different measure, called FVER(fixed to variable energy ratio) which is different to PUE, but has a strong family resemblance. The BCS’s paper on the subject is here.
The FVER definition is
FVER = 1 + Fixed Energy/Variable Energy
where the fixed energy is that used by the data center when it is not in use, and the variable energy is that used when processors start to do useful work.
As with PUE, the ideal would be 1 - ie, there is no fixed energy at all, and the data center only uses energy when it is working. But in every case, there is always some emergy used keeping the data center up and running.
It’s easy to compare this with PUE, if we write PUE slightly differently. It’s normally expressed as
PUE = Total Facility Power / IT Power
But if we break the total power up in to IT Power and “the rest” which we can call overhead, then we get
PUE = 1 + Overhead/IT Power
This now looks pretty similar (the fact that FVER uses energy, while PUE uses power is not significant. Since both should be averaged over a whole year, they both address energy rather than power in any case).
Now, Liam Newcombe of the BCS Data Centre Specialist Group, believes that PUE has led to a focus on reducing the fixed costs in cooling (the biggest part of the “overhead”) whereas in fact the IT power part of the equation actually includes quite a lot of “overhead” or non-useful work.
In other words, this is a way to get data center people to start distinguishing between how energy is used within the IT systems: what is a fixed cost that could be iminimised, and what is being used for actual processing.
The first obvious thing it does is shift the cooling fans within the IT systems from “work” to “waste”. They are in the IT power part, and they may result in a lower PUE, even though they are less efficient than the larger cooling systems that cover the whole data center.
There are a number of perverse incentives resulting from PUE, and FVER looks a good way to address some of them, without having to go into the kind of subjective territory which asks data center owners to classify work done as “useful” or “not useful”.
I’m interested to see if this idea could be taken further.