Regular reports of the latest high-efficiency data center could easily lead one to believe that data centers generally are becoming more efficient, even if overall power consumption is rising. If Facebook, Google or Yahoo can build a facility with a PUE of 1.2 or under, surely other companies aren’t too far behind. But have the headlines swayed the industry’s view of itself, or could the problem be something else entirely?
Average Data Center PUE
According to results from a recent Digital Realty Trust survey of selected large companies (greater than $1 billion in annual revenue or more than 5,000 employees) in North America, the average power usage effectiveness (PUE) is 2.9. In other words, on average, only about one in three watts of power consumed by data centers actually goes to the IT equipment. In addition, the survey found that only 20% of participants reported a PUE below 2.0. These results suggest that some previous estimates of average PUE—which have been as low as about 1.9—miss the mark.
A number of salient points arise from a cursory look at the data. First, the sample involves large companies with significant resources. Thus, the average PUE estimate isn’t swayed by smaller companies that may be unable to implement technologies to reduce energy consumption. Furthermore, if energy-saving upgrades really do offer reasonably fast returns on investment, even smaller companies would have an incentive to implement them. According to the Digital Realty Trust report on the study, the average IT load for participants was 2.6MW. Again, the average in this case is a sizable data center, not just a data closet.
Second, the survey indicated that 81% of participants measure power usage in their facilities (the majority doing so at three or more locations), and 80% have implemented hot-aisle/cold-aisle containment. Furthermore, five out of six report reliance on data center infrastructure management (DCIM) software. Although these measures don’t represent the totality of possible energy-efficiency improvements, they demonstrate serious reliance on common practices for reducing data center power consumption. Yet the average PUE is still 2.9.
Or maybe The New York Times was right about data centers as energy wastrels?
PUE: A Valid Comparison?
The problems with using a single number to evaluate efficiency in the context of a single data center are manifold. But is cross-comparison of multiple data centers using a single number—specifically, PUE—even more problematic?
Even if PUE isn’t a good absolute indicator of energy efficiency, it can, using the right approach, be a decent measure of a given data center’s improvement (or lack thereof) over time. Yes, it requires care to avoid pitfalls like an increase in PUE owing to an upgrade of IT equipment to more-efficient models. Internally, however, having a formulaic approach to calculating efficiency can be beneficial in identifying areas that need improvement or adjustments that have proven to be worthwhile.
The problem with PUE may well be its use beyond the company as a means of comparison with other data centers. Gabe Andrews of Greater St. Louis said at DatacenterDynamics, “Energy efficiency metrics should continue to remain in-house and be utilized as benchmarks of performance and not for publication in any magazine or online forum. As an enterprise, it may be acceptable to have a PUE of 3.3 based on the multiple levels of redundancy built into the facility to ensure a lack of downtime. In the market this single number is outlandish and inefficient.”
Difficulties in comparing PUE among several data centers range from climate to reliability and so on. For example, an obvious problem arises when one data center is located in a cooler climate than another. In the former case, cooling is less energy intensive, so one would naturally expect that facility to have an inherent efficiency edge. Similarly, some data centers implement more redundant infrastructure, which may crimp efficiency. Nevertheless, that redundancy may be a business necessity to maximize availability; other data centers do not require the same high availability, so their efficiencies may be higher from the start.
Unfortunately, once companies are segregated by climate, availability, facility size and on and on, the value of PUE as a single number summarizing the industry fades. A weighted average is one possibility, but this approach invites controversy over the weights for each characteristic. And the problem isn’t PUE specifically—it’s any single number as a means of summarizing efficiency. Even in a single data center facility, PUE varies depending, among other things, on usage levels. Idle equipment thanks to less-than-peak utilization can drag the PUE down. Ian Bitterlin noted, “Any such ‘comparison’ system cannot adapt itself to partial load. For example, Facebook in Prineville is 1.07 [PUE] now when full, but was surely (it was) over 10 the day they turned it on.”
Is the Picture Good or Bad?
So, are data centers efficient or not? The answer is it’s difficult to know. A given facility can be compared with its own performance last year, in peak-usage times over idle times and so on, but comparing one data center with another invites trouble. Even ignoring variations in methods of calculating and reporting PUE, the values cannot be fairly judged relative to one another without numerous (potentially confusing) qualifications.
Yes, data centers use a large and growing amount of power. But users demand these services—including The New York Times, which might take issue with the regular outages or service problems that could occur if data centers only maintained enough equipment to operate near full utilization, for example. The question, however, is how much of that power is going to waste. Answering that question is difficult, at least in any numerically precise sense. Whether the real average PUE is 1.9 or 2.9 makes little difference if the underlying sample data is suspect.
Efficiency is an aspect of data center operation that will always leave some room for improvement. The industry could quite easily implement an array of improvements that cuts PUE across the board—but the concomitant rise in service prices might be more than consumers (inside or outside of the company) are willing to bear.
An widely applicable, single-value measure of data center efficiency will remain elusive. A single number simply cannot convey the variables that affect data center efficiency, from climate to tier and so on. This isn’t to say the efficiency of the industry at large is unknowable, so we should just ignore it—but care is needed whenever a single number purports to represent a number of data centers, to say nothing about its meaning for just a single facility.
Photo courtesy of Darlingguest14