The ongoing quest for energy efficiency in data centers, driven in part by the looming potential for regulations and taxes, motivates a desire for a simple means of benchmarking a facility over time and against other facilities. Since 2007, The Green Grid’s PUE (power usage effectiveness) metric has been the staple of companies in this regard, but it’s not an approach that pleases everyone. With all the alternative, albeit less popular, standards in use or under development, should PUE remain the top choice?
Problems With PUE
Picking on PUE is easy. Any number of scenarios illustrate the shortcomings of this metric. Perhaps the most egregious example is where a company updates the IT equipment in its data center to more-energy-efficient systems. Because PUE is simply a ratio of the total power consumed by the facility to the amount consumed by the IT equipment (meaning a value closer to unity is better), an increase in IT efficiency can cause PUE to likewise increase. At face value, this would seem to indicate the data center is less efficient! Other scenarios can and have been cited.
Quite apart from the difficulties inherent to PUE are manipulations of the system by companies looking for a PR edge. PUE, rather than a tool to gauge efficiency improvement, has become a marketing statistic; thus, companies often feel pressure to minimize their PUE ratings—possibly by fudging some numbers. (For instance, perhaps lighting is omitted from total facility power consumption—after all, that’s just for the people, it’s not for the data center itself—or PUE is measured during the winter when cooling power is minimal.) The companies may know what their PUEs really are, but no one else, whether customers or other companies looking for points of comparison, can know whether the numbers are legitimate.
Nevertheless, despite its problems, PUE remains the chief metric for measuring data center energy efficiency. Most companies touting the sophistication of their data centers will cite PUE rather than one of the myriad competing metrics. These other approaches suffer from the disadvantage both of being later to the scene, thus lacking momentum, and of being largely grouped in a mass from which no single alternative stands out. In other words, the incumbency of PUE means it is unlikely to be toppled by an alternative for some time.
The Problem With All of Them
Think of how a scientist might rate an individual’s intelligence. Intelligence quotient (IQ) probably comes to mind as the chief metric developed thus far. If you think further, you might come up with means of refining IQ. Ultimately, however, IQ suffers from a serious problem: it is a single number that tries to rate a nonsingular characteristic.
To illustrate, imagine the scientist must rate two individuals using the IQ system: a brilliant musician and a talented mathematician. What should the scientist do? Ask questions about music? Test musical capability? Ask questions about mathematics? Pose a series of quantitative problems? No matter how he conducted his test, the results could easily put both the musician and mathematician on the same level as, say, a renaissance man who just happens to be adequate in both specialties.
PUE is similar: it is a single number that tries to encapsulate a number of different factors, such as the efficiency of cooling infrastructure and the efficiency of IT infrastructure—the problem is that these two can offset each other, causing unintuitive changes in PUE. But let’s not pick on just PUE; the same can be said of the majority of competing metrics. Some try to solve the problem of increasing IT efficiency leading to higher PUE by measuring a broader characteristic, such as the delivered performance per watt of energy consumed.
Such metrics still run into problems: what about data centers that use measurements from the middle of winter instead of summer? One could argue that this is just manipulation of the system, but when should measurements be taken? Is a measurement taken at maximum cooling load better than one taken at minimum load? Perhaps measurements should be taken year round, with some form of average serving to remove the extremes. But the weather one year may differ greatly from that of another; furthermore, data centers in cooler geographies will have an inherent advantage over those in warmer geographies.
You can probably conceive of various ways to incorporate these various factors into a single number: maybe divide by the average temperature of the data center’s surroundings, take an average of regular readings of the metric over different times of year and different times of day (to try to balance the effect of workload), and so on. But the problem remains: excellence in one area can sway the number to indicate mediocre results overall. Attempting to use a single number to measure data center efficiency is just like trying to use IQ to rate intelligence: it may have some broad applicability, but it lacks accuracy.
No Easy Solution
The nice thing about PUE and similar metrics is their simplicity: when you’re dealing with a single number per data center, you can quickly make a comparison. Data center A is more efficient than data center B (at least according to the metric). This simplicity also provides a convenient marketing tool. A data center with a measured PUE of 1.10, for instance, can quickly and effectively shout that number in the media and garner attention. Imagine if, instead, the company had to provide and explain a table of values, a graph or two, and some other information to accurately describe the efficiency of its facility; the press would yawn and move on, because few readers would have the patience to assimilate all that information. A number is faster, cleaner and punchier.
As a result, The Green Grid has undertaken efforts to refine PUE, and others have developed similar metrics that try to get around PUE’s problems, but they all fall prey to the difficulty of reducing a complex data center characteristic (energy efficiency) to a single number. Each alternative will thus have its problems and will be open to abuse by unscrupulous companies looking for a marketing edge. Debates over which is better will continue.
Although PUE has its downsides, its incumbency as the metric of choice (if for no other reason than it’s what “everyone” knows) means competitors are at a major disadvantage. Toppling PUE means gaining a following among data centers—particularly among those built by large companies that can afford the types of innovations that yield extraordinarily high efficiency ratings. If one metric emerges that minimizes the problems associated with using a single number (or, maybe, it can neatly present efficiency in two or three numbers), it might have a chance of gaining some traction against PUE. In the meantime, however, PUE—despite its difficulties—will remain the metric of choice. As such, it is still a useful tool, as long as those who use it or interpret measurements keep its shortcomings in mind.
Photo courtesy of simon.carr