Advanced Computing in the Age of AI | Monday, August 8, 2022

Pfeiffer Dishes on Data Center Metrics and Trends 

Power Assure’s CTO Clemens Pfeiffer spoke with Green Computing Report this week about ten trends he sees controlling the green computing world in the coming years. The discussion inevitably turned to the state of green IT as a whole, where he noted that too much emphasis was being placed on metrics such that they are often misreported. In some cases, the emphasis on PUE in particular comes at the expense of actual efficiency.

As mentioned on GCR this week, the sentiment that PUE is a valuable metric that simply gets over- and misused is growing. There exist other valuable metrics, such as total energy usage and The Green Grid’s new recycling metric that we covered yesterday, EDE.

Power Assure’s CTO Clemens Pfeiffer spoke with Green Computing Report this week about ten trends he sees controlling the green computing world in the coming years. The discussion inevitably turned to the state of green IT as a whole, where he noted that too much emphasis was being placed on metrics such that they are often misreported. In some cases, the emphasis on PUE in particular comes at the expense of actual efficiency.

Along with challenging the implementation of certain data center metrics, we also discussed Pfeiffer’s ten trends that he considers instrumental to green computing moving forward.

“(Trend number two) is one that’s becoming more and more critical,” Pfeiffer said. His second trend mentions how the power consumed by servers will become an issue of increasing importance, especially as programs like the U.S. Environmental Protection Agency’s ENERGY STAR come more into effect.

What it comes down to for Pfeiffer is a simple turning off of machines when they’re idle. This principle may seem intuitive, but there exists a significant fear, especially high-power HPC systems, that machines that turn off may never turn back on.

“When you get out of your car to go into a grocery store, you don’t leave your car running,” Pfeiffer said as he analogized high performance computers to cars. “Maybe you did thirty years ago when cars were less reliable and took a longer time to start but you don’t anymore.” For Pfeiffer, the machines have advanced to the point where it’s okay to turn them off when they’re not in use. Unfortunately, for the most part, logic has yet to catch up with conventional wisdom.

“Keeping everything running 24-7 will not be the norm long-term.”

Conventional wisdom has taken shape a little too rapidly around a certain TGG-developed metric in PUE.

For example, Pfeiffer told a story of a facility claiming to have a PUE of 0.99. Remember that the simple version measures total IT energy over total facility energy, with the margins coming in the form of, for example, heat generated to maintain computing and the energy loss in cooling the excess heat. In essence, that facility was then claiming that their IT power usage was larger than the total power usage of the entire center.

“People either report what the designed minimum PUE is or they report what the best PUE was during the year.” For those that are followers of Star Trek, breaking below a 1.0 PUE is like breaking the Warp 10 barrier: impossible by definition.

The problem is reporting for comparison’s sake as opposed to measuring to actually improve one’s efficiency. Almost every new data center announced will report a PUE of around 1.3. For Pfeiffer, these measurements are largely shambolic as those numbers are difficult to achieve. “To get to a PUE of less than 1.3 is very challenging. You can get to 1.3 and to some of these numbers on some ideal days over the year. The challenge is to do it on an average over 12 months…If you can’t use outside air, no one will have a PUE of less than 1.5 or 1.6. It’s just not possible.”

For those with the ability to use outside air, like Google’s Finnish data center and the MGHPCC in Massachusetts, Pfeiffer joked that it’s so cold up there that no one really wants to spend time there. But in seriousness, part of the problem lies in just how, well, seriously organizations take that one metric.

“PUE is one measurement. It’s an important one but isn’t the only one. The absolute power consumption should also be a factor and the efficiency of the equipment should be a factor. The less IT power consumption you have, the higher your PUE; usually because your cooling is increasing with lower utilization on the IT side. So it’s kind of a counter-productive measure from that extent.”

For example, he related an amusing tale of a data center manager who installed more efficient IT equipment…only to see his PUE go down? The explanation as to how that happened was fairly simple: newer equipment runs more efficiently, meaning it doesn’t have to use as much power. When both the numerator and the denominator of a fraction decrease by the same amount (and again, PUE is in its simplest form just a fraction), the total number will go up.

And yet, the manager was upset because, as predicted, that number went up. From Pfeiffer: “I was in a data center the other day, they had a PUE of 1.3 they said. And then they started replacing old servers with new servers. They shot up to 1.5, the guy was so frustrated. He said, ‘I replaced all my old servers with new servers and my PUE went up.’ I said, ‘why, that’s good, your IT is using less power.’ And he said, ‘Yeah but my PUE went up. Management doesn’t like that.’ But he saved 50 percent power. That’s the kind of discussion they had.”

It should be noted that Pfeiffer’s comments do not represent a call to dispatch the metric altogether. By all means, the measurement of Power Usage Efficiency should help a data center compare itself with past versions of itself. That was the purpose of its introduction. The Green Grid, who developed PUE several years ago, stated clearly that it was not meant for cross-facility comparison. Yet inevitably this sort of comparison happens.

This discussion informs a great deal of Pfeiffer’s full list of ten trends, which can be found here.

Related Articles

Heat-Trapping and NREL's Green Datacenter Leadership Effort

The Green Grid Forum 2013 Tackles Metrics, Standards

Groundwater Flows through New Datacenters

Add a Comment

EnterpriseAI