PUE is DEAD: The Case for Performance Per Watt

Mark Davidson, Sustainability Officer for JouleX, an innovator in enterprise energy management systems for data centers, distributed office environments and facilities.

Mark Davidson JouleXMARK DAVIDSON

Power Usage Effectiveness (PUE) has been called “The Holy Grail” of data center energy metrics so often that we actually found it impossible to find out who coined the term. As time, technology and sustainability efforts evolve, the PUE metric is no longer the stopping point for energy efficiency measurement, but it has become just one more piece in the larger picture.

What does PUE do? It measures how much of the energy entering a data center facility is used to power the computing devices within, versus the amount used for cooling and overhead of the facility. That’s it.

Realities of Measurement

In an ideal world, the PUE is 1.0, which means that 100 percent of the energy is used by the computing devices in the data center. Since 1.0 is an impossible-to-achieve ideal, the standard goal for most data centers is a PUE of less than 2.0, which means that for every 1 watt of energy used by the computing devices, an additional 1 watt is used for facilities overhead such as air conditioning and lighting.

At first glance, this is a very direct metric. It’s easy to understand and easy to follow. The problem with relying solely on PUE is that it in no way measures the efficiency of the IT devices themselves.

Other Metrics Available

Data center managers today are under pressure to deliver increasingly higher energy efficiency and lower costs. In order to understand the true efficiency levels and progress toward enterprise sustainability goals, managers MUST have access to an accurate performance measurement of each device in the facility. Other available energy metrics include:

Another Measurement – Performance per Watt

There are, however, issues with these other metrics. CADE is a McKinsey metric and is very costly to implement. By the Green Grid’s own admission, DCP and DCeP are very difficult to measure and implement. The measurement recommended by JouleX is Performance per Watt (PPW).

What’s the difference? Like PUE, Performance per Watt is a very straightforward and easy-to-implement solution. The PPW metric measures the actual energy efficiency of every device in the data center and how it is used.

The PPW approach uses a relative performance indicator for each individual asset. This indicator is calculated by the types of hardware and capabilities learned from an asset inventory of that device. This Performance Indicator (PI) is a simple measurement for getting relative performance of the device in question.

At JouleX, we use a derivative of SPECmarks. However, any performance metric can be used as long as it is incorporated into all your measurements. The formula for PPW is as follows:

(PI * Avg Device Utilization / Watts ) * 100

Let’s take a look at two examples:

When the device is at maximum efficiency, the PPW number is higher. The lower the PPW number, the more power that device is wasting. What can the PPW measurement tell you?

  • Are the servers using twice the electricity needed for the jobs they’re doing?
  • Are you wasting electricity powering dead servers?
  • Are the old switches and routers costing you more in power than it would cost you to replace them?
  • Would virtualization of newer, more energy-efficient servers, allow you to retire old servers completely?

When combining PI with live utilization of an asset, along with real-time energy draw of that asset, you get a simple process for measuring PPW. Being able to measure Performance per Watt will help data center managers identify devices that are wasting energy.

And, as we’ve always said: ”You can’t manage what you can’t measure.”

Industry Perspectives is a content channel at Data Center Knowledge highlighting thought leadership in the data center arena. See our guidelines and submission process for information on participating. View previously published Industry Perspectives in our Knowledge Library.

Add Your Comments

  • (will not be published)


  1. Ryan Malayter

    Unfortunately, this scheme has a huge disadvantage in that the benchmarks which matter vary based on application, and therefore numbers will not be comparable across datacenters and workloads. For example, many workloads are IO-bound, so a benchmark like SPC-1 would make sense. But then how do you compare that "efficiency" to a compute-bound workload in another datacenter?

  2. The need to measure "how much work the server is accomplishing" (versus simply how much power it's consuming) echos what Winston Saunders said in proposing Server Usage Effectiveness (SUE): http://communities.intel.com/community/openportit/server/blog/2011/05/19/the-elephant-in-your-data-center-inefficient-servers By "SPECmarks" for your PI, do you mean the original SPEC89 numbers, or just the latest SPEC CPU benchmark (i.e. SPEC CPU2006)? One challenge with using SPEC CPU as the PI is requires conversion between generations of the benchmark. (The larger problem tends to be handling servers that don't have published SPEC results.) An alternative to a single SPEC benchmark is a composite, like RPE2, which gets around the issues of conversion and "missing" benchmarks. Plus, a composite is a better estimate for a broad variety of workloads than a single benchmark that just measures one. (Disclaimer: I work for the company that develops RPE2. I'm not above shameless promotion when it fits!)

  3. I think performance per watt at a data center level is not a useful statistic. As someone else mentioned it depends on the application(and efficiency of that application), but more importantly it doesn't measure the effectiveness of the data center as much as it does the efficiency of the gear running inside of it. And it will be totally useless for any sort of shared hosting environment, colo providers and the like since they would not impose any controls, nor audits to try to measure that sort of thing. It may make SOME sense for those big cloud companies to do it -- but in the grand scheme of things does it really matter what their power efficiency rating is? Short of bragging rights? Does that translate to lower costs passed onto the customer or higher margins by the service provider? I have a friend who runs a data center in Seattle, who has a very good PUE rating (I forgot the exact number this was a while ago), he went into great details about their energy efficiency system how they use the outside air economizers, how they do the hot isle/cold isle isolation, how they do metered power and how that directly translates into lower costs, and higher densities (he was allowing me 12.5kW usable per rack with a pair of 208V 135A 3-phase circuits - their standard is 10kW usable/rack) for the customer. The use of such power circuits I have not yet come across in any other colocation facility, same goes for the use of metered power. I'm about to move into a facility in another state now that I'll be running at about 60% of my peak usable power capacity but until I fill the racks up(which may take a while) I'm paying for as if I am using 100% of my usable capacity = money wasted. And somehow doing some sort of magical power migration changing PDUs out as the power usage goes up and changing circuits out is just too complicated, costly and risky. That same Seattle data center operator went into some details as to why many other PUE ratings aren't really valid - I think it had to do with something like the facilities not taking into N+1 redundancy. Folks like Google and stuff seem to have a strategy where they just shift load away from an entire facility in the event it has an issue, mere mortals don't have that option so we require N+1 redundancy at all levels of the facility (wherever possible anyways). Since I would wager a good 98% of customers out there are running a diverse number of applications and configurations it's not feasible to try to measure a data center at a performance per watt metric.

  4. @Ryan, that is why you can use multiple "performance' metrics to get measurements. It doesn't have to be a process bound metric, perhaps a system metric, or a storage ops metric or an IO metric would be best for your data center and workloads. In some cases ALL of the above, but the simple formula still proves effective. Once you compare the results, even weighted results from multiple types of PPW benchmarks, the most wasteful systems always bubble up to the top. Those are the ones you deal with first. @Nate The SOLE purpose of the data center is to enable computing and network services. Making the facility more efficient is a start at today's efficient datacenter, but you cannot improve over a single watt of power in to a watt of power to IT load. By REDUCING your IT load, then you can have an even larger effect of power reduction in the datacenter, no matter if you measure PUE on an N+1 data center or any other type. Making your IT infrastructure more efficient has the biggest return on your investment. We have often overlooked that as we have been striving for 5 9's of availability. Modern systems and technology is allowing us to reclaim some of those wasted resources and STILL allow 5 9's of availability. @Daniel Thanks for the info. I will look into RPE2

  5. Hi, I've read in JouleX own blog that an agent running in a PC is a waste of energy (while they can't themselves power down the monitors independently from the CPUs and THAT is a waste of energy). I tend to take their statements oriented to people that are lately aware of energy management and not to experts. Well, at the end of the day they are newcomers in the market. ( and yes, this is an interested post) but as I know little about this issues and I find wrong statements, and then when I read something out of my scope I find several corrections... Who knows...)

  6. I remember taking a look at Power Assure's product earlier this year and that's what they were baselining with - the watt. The device would identify/auto discover anything in a facility that drew electricity and then chart the usage. I thought it was a pretty slick solution for anyone that wants to know what they are using, at what workload, etc. but for my pupose (as a pure play colo provider) PUE is fine since it is in my best interest to make sure that the electricity I have is being used to power the greatest number of servers vs. inefficient cafeteria lighting, or old sodium lights in the parking lot.I don't have a material interest in Power Assure, and I did take a look at what they had and for what it did, it seemed very useful.

  7. Hi: re-reading the article I think it makes sense, the results are fair but... do you really need to use it? In your example, a 6 year old server is compared with a current one. The difference in Watts between them is 10.4 W and the performance between a 6 year old machine and a new one is supposed to be huge. Do you really need to be doing much calculations to discover which one is the inefficient when the power draw is so similar? I think that could be more useful comparing a 90W vs a 260W server but this example? I think not. I know nothing about date centers, but I guess the process is to look for underused machines, then to old vs new and make some common sense decisions. I think this deep dive inside where power is used is not important when you worry about the server wasting energy and at the other point the user is connecting his VDI machine to the NFL site watching a match at work hours...

  8. We appreciate you bring attention to the fact that PPW measurements can tell you if there is energy being wasted. Also like that a spotlight was shone on whether old switches and routers are costing more in power than it would cost to replace them. "Going green" is no longer a fad, but a choice that companies must make. At Global Power Supply we believe that business excellence and the health of the environment are unmistakably linked. Accordingly, we understand & accept our responsibility for improving the environment.

  9. Bernie Rob

    PPW is an excellent way to measure one of the aspects of data center efficiency. However, I prefer to use few metrics to see larger picture. For instance, it could be PUE, load balance, utilization and relative performance efficiency of each of the servers http://goo.gl/FIPVA