Despite the benefits of the power usage effectiveness metric, it's not always sufficient to fully optimize data...
center efficiency. IT teams need more detailed data center metrics that encompass both the power and cooling infrastructure, as well as the computing systems, to truly optimize the data center.
As a step in this direction, the Environmental Protection Agency in 2007 published a report on data center efficiency that stated the intention of developing an ENERGY STAR rating for servers -- and they did, albeit years later than expected. The Green Grid published the water usage effectiveness and carbon usage effectiveness metrics in 2010, although they have gotten relatively little attention. In June 2016, Green Grid announced the Performance Indicator (PI), a metric that looks at cooling criteria and integrates it with energy efficiency.
We're being inundated with data center metrics on performance and efficiency, but for most IT teams -- even those regularly tracking their power usage effectiveness (PUE) -- taking steps beyond that metric requires more advanced data gathering and analytics skills. For those who are ready to use more sophisticated optimization tools, here are some key considerations.
The purpose of extended data center metrics
It's common to save energy by shutting down comatose servers, consolidating and virtualizing applications or purchasing ENERGY STAR-rated hardware. But unless the power and cooling infrastructures are also optimized -- which can be difficult with existing systems -- the reduced load will not significantly improve energy efficiency. This means that the PUE will get worse. By using PUE as a benchmark, rather than a tracking metric, it could appear as though energy was not saved, and that the investment and effort were not worthwhile. Management expects their investment to produce a lower PUE number with no comprehension of what the number actually means.
That's where other data center metrics -- such as The Green Grid's data center energy productivity (DCeP) -- are valuable. DCeP quantifies the amount of useful work the data center actually produces per watt of energy it consumes. The DCeP metric lets the user establish their definition of useful work. For example, an online search company might define that as the number of searches completed. For a retailer, it may be the number of sales. PUE is still necessary, but the DCeP number may be better understood by those with less experience.
Although DCeP is not a scientifically accurate metric, it provides a way to quantify what you're actually accomplishing with the energy you use. If a bank of servers does little work and runs idle most of the time, it draws minimal energy, requires minimal cooling and won't significantly affect the PUE. But it's still drawing power to accomplish almost nothing. The DCeP would show that; its goal is to minimize energy consumption and maximize useful work. For organizations at the leading edge, that want to squeeze every ounce of productive computing out of every watt of power they use, sophisticated servers can provide operational data far beyond CPU utilization, and more sophisticated data center metrics can track the results.
But PUE and DCeP are still all about energy efficiency and energy reduction. Using them can have unrecognized consequences because they don't reveal compromises made to save energy that negatively affect cooling and reliability. The new PI metric does, and is most useful for administrators who have better-than-average data gathering skills, and the ability to optimize every facet of their operation.
Four levels of PI measurement
There are four levels of PI measurement. Level 1 is basic, and does not need sophisticated equipment to use. Level 2 requires more thorough and accurate measurements. Levels 3 and 4 add computational fluid dynamics (CFD) air flow modeling to provide a visual picture of performance, and to enable what-if scenarios that look at future capacities and failure modes along with energy efficiency. Level 3 is normal modeling. Level 4 uses actual, detailed measurements to calibrate the CFD model as an accurate baseline for other examinations.
There are three requirements to use extended data center metrics like PI and DCeP:
1) Track the PUE;
2) Define what constitutes useful work in a computing operation; and
3) Obtain detailed measurements of power draw and temperature at every rack.
PI complements existing approaches by combining PUE, IT thermal compliance and IT thermal resilience. The latter two are based, respectively, on the ASHRAE recommended and allowable thermal envelopes. Thermal compliance and resilience examine how well redundant cooling works under both normal and abnormal conditions. If computer room air conditioning air temperature has to be lowered to meet the thermal compliance goal, the PUE may increase. The point of the PI metric is to know how reliably hardware is being cooled, as well as how energy efficient the facility is, and how one condition affects the other. Decide how close to the maximums you want to operate, as well as where you want to target energy efficiency or PUE. Then, measure actual conditions and plot them on a triangular diagram, also known as a spider diagram, to see how close they are to your goals.
There are now several data center metrics available to maximize overall data center performance in terms of energy efficiency, work output per unit of energy expended and reliability relative to energy efficiency. For most operations, adhering to fundamentals and tracking power usage effectiveness should still be priority. The rest can follow, but PUE remains the basis. And even if you're not ready for more esoteric measures, it's still good to know what the industry is proposing -- if only to have a target to aim for.
Ways to save money and be energy efficient in the data center
PUE keeps data center energy use in check
How to measure total energy use in the data center