You can measure your data center’s present PUE and set a PUE goal to work toward, but how do you know you’ve set the right goal? How do you know your goal is achievable yet represents the best efficiency rating you can get out of the facility?
That’s one of the questions a new metric proposed by a former Amazon data center designer can help answer. Engineering Operations Ratio, or EOR, is a way to assess and express the gap between a computing facility’s actual performance and the performance it was designed (or claimed to have been designed) to deliver.
“My goal with EOR is to provide a simple but effective way to drive data center operational effectiveness,” Osvaldo Morales wrote in a recently published paper describing the concept. “Having EOR data would allow us to double down on the good and fix the bad.”
Morales spent the period between 2000 and 2016 designing the data centers that underpin Amazon’s vast ecommerce, online video, and cloud computing empire. He left the company this past March after three years as VP, Global Data Center Platform for Amazon Web Services, according to his LinkedIn profile.
His EOR paper, released earlier this month, is the first publication by Infrastructure Masons, a data center design think tank of sorts, started earlier this year by Dean Nelson, the former eBay data center chief who now oversees all things computing infrastructure at Uber. IM’s advisory council members oversee some of the biggest global data center networks – Microsoft, Google, Facebook, eBay – and the operations of data center provider Switch, which lists most of them as customers.
The first draft of the proposed metric describes it as a master ratio between data center design and operational performance levels that combines multiple similar ratios for various data center subsystems, primarily mechanical and electrical infrastructure. Amazon's infrastructure team did not use a metric similar to EOR while he was there, but the company may decide to adopt the idea today, Morales told Data Center Knowledge.
Here’s an example of EOR calculations for a hypothetical 20MW data center’s top-level PUE and its subsystems Morales included in the paper:
Each ratio is derived by dividing an observed performance value by the number the design intended to achieve in cases where a higher value indicates better performance and vice versa (design divided by observed) in cases where a lower value indicates better performance, such as PUE, where lower PUE is better than higher PUE.
As you can see, EOR dictates that you take into account both underperforming and overperforming systems. It’s important to keep tabs on both, according to Morales. (more details in the paper itself)
Like PUE, the proposed “unit-less” metric is meant to be simple to communicate a data center’s performance to non-technical team members.
The paper is a first draft, and Infrastructure Masons invites others involved in data center design and operation to comment and make suggestions for improvement.