Recently an interesting story hit the news about Jim Gao, a Google engineer who had built an innovative neural network based application to accurately predict the efficiency of the company’s data centers.
Gao, nicknamed Boy Genius, is responsible for tracking the power usage efficiency (PUE) of Google’s data centers. The company has been measuring its PUE every five minutes over the last five years, which has produced an impressive data set. It includes information on 19 different variables from the cooling water temperature to the outside humidity.
Gao decided to apply the power of machine learning to understand how all these variables related to the data center’s PUE. The model he built proved to be extremely accurate at predicting the PUE of any of its data centers at any time. This made it extremely valuable to both try and improve the energy efficiency of the data centers and also to spot any operational issues. The latter is possible because if the prediction and measurement are quite different then there is likely to be some unspecified problem.
enduring popularity
But what is the PUE and how does it relate to energy efficiency? The term was first coined by the Green Grid in 2007 and quite simply is the ratio between the amount of energy drawn by the data center compared to the amount used by the IT equipment. The idea is that the closer one gets to a measure of 1.0, the more efficient the data center is because you are not “wasting” energy on ancillary equipment such as cooling or lighting.
The PUE’s simplicity has been a boon for marketing and many companies boast about how close to 1 their data center PUE has become. Google, for example, publishes its PUE reduction track record along with its current status – which for the record is 1.11 (Q1 2014). It is not alone, Facebook also publishes the PUE of its Prineville data center along with its water usage efficiency.
limitations of PUE
But despite its enduring popularity, PUE has been criticized in some quarters as an inadequate measure of data center efficiency. Tom Raftery in his green blog points out that a data center could actually end up with a higher PUE even if they used more energy efficient IT equipment, just simply from how it is calculated. Even worse, he says, the PUE measure pays no attention to the carbon intensity or water usage of the data center.
Counter-intuitively, virtualization can also increase the PUE of the data center, while actually producing more compute power per watt consumed, as outlined by Jayabalan Subramanian. So although a data center operator is getting more bang per buck, the headline PUE figure could indicate the opposite.
Because it is such as simple measure, PUE doesn’t really measure the energy efficiency of what the data center actually does – i.e. digital processing. Just because the electricity is going to a server doesn’t mean that it is doing anything useful – or doing anything useful efficiently. In addition, PUE is of limited use in comparing the efficiency between different data centers, because there are too many variables. And finally, it is a crude method of measuring the energy efficiency of mixed-use environments, such as if a data center is co-located in an office.
moving forward
Despite these limitations, PUE is likely to continue to play a big part in the push towards greater energy efficiency in the data center. In fact, it will have even more important role through its planned anointment as an ISO standard. There are likely to be several tiers of measurement of PUE that measure it as frequently as every 15 minutes to track changes in energy consumption.
Ultimately, PUE is just one measurement amongst many that can help companies make their data centers more efficient. It continues to be useful as long as you understand and recognize its limitations. To get a more holistic picture of the state of the data center, you need to combine PUE with water usage, carbon intensity of supply and computing power generated over time.
How useful do you think PUE is in measuring data center energy efficiency or would a more sophisticated measurement be a better approach?
Anthony
image : © ristaumedia.de - Fotolia.com
After a Masters in Computer Science, I decided that I preferred writing about IT rather than programming. My 20-year writing career has taken me to Hong Kong and London where I've edited and written for IT, business and electronics publications. In 2002 I co-founded Futurity Media with Stewart Baines where I continue to write about a range of topics such as unified communications, cloud computing and enterprise applications.