Pushing the data centre efficiency envelope

by

3 November 2014
Joe Kava of Google

When it comes to making data centres run as efficiently as possible, Google is ahead of the curve. The numbers prove it: Non-computing systems like cooling equipment take up, on average, around 12 per cent of the power channelled into Google’s data centres. Many enterprises are hard pressed to bring that figure down to below 60 per cent.

But Internet giant is apparently concerned. In Singapore recently to deliver a keynote speech, Google’s vice president of data centres, Joe Kava, said that there has been a levelling off in Google’s data centre efficiency scores.

“For the last several quarters, our trailing 12-month PUE has leveled off, to about 1.12 globally. While that’s (a good figure), it is a disturbing trend to me,” he told an audience of technology professionals at the conference organised by DatacenterDynamics.

By PUE, Kava was referring to a widely-used metric used to gauge data centre power consumption efficiency, known as Power Usage Effectiveness. PUE compares data centre facility power consumption, to the amount consumed by the facility’s IT equipment. It highlights how much power is used for non-computing purposes, such as cooling and other overheads.

Google, which lists 12 major data centre locations worldwide on its website, including Singapore and Taiwan, has an average PUE score of 1.12. It is one of the lowest in the industry.

Kava said the industry on the whole has done a pretty good job. “I remember eight years ago we were talking about PUEs in the 2.5 to 3 range. Today, you don’t see any new major data centre designs that aren’t in the sub-1.5 range,” he said.

But the industry is now seeing a levelling off in efficiency numbers partly because of the rising complexities of the modern data centre, he said. The myriad of electrical and mechanical systems that have to work in tandem, coupled with ever-changing operating conditions, are making it extremely difficut for data centre operators to constantly achieve optimal efficiency. In Google’s facilities for instance, operators have to grapple with at least 19 major variables that are critical for efficiency.

“All of us intuitively know that there should be a direct relationship between the cold aisle temperature and PUE, but it is not so simple,” he said.

“Cold aisle temperature is a function of set points like pump speed, flow, differential pressures and airflow management. It is not intuitive for a data centre engineer to look at that vast amounts of data and figure out the right operating condition,” he added.

New approach needed

In his keynote, Kava said that a new approach is needed to add a fresh spark to the efficiency drive. He pointed to an idea Google first mentioned in its whitepaper released in May this year. Authored by Google engineer Jim Gao, the whitepaper proposed the use of machine learning algorithms to predict and optimise data center operations.

Machine learning is the way to go because today’s complex data centres produce too much data and have too many operating permutations, to be effectively tweaked by human operators in the quest for optimal power usage.

“If we could apply machine learning models to PUE, we could figure out what the hidden story is, which could help us identify how to break through that plateau,” said Kava.

“This is a good example of data-driven innovation and the next wave of, not just data centres, but the industry as a whole. The ability to take all of the data that we are deriving everyday and make sense of it all is going to help all of us drive forward into the next level of efficiency and move our industry forward,” he added.

Enterprises need to do better

At the sidelines of the DatacenterDynamics conference, Kava told ConvergenceAsia in an interview that many enterprises are not putting enough focus on data centre efficiency.

“They are worried about just keeping the data centre up and running. No one wants to be the person that puts the data centre at risk because of the potential of business interruption,” he explained.

He reckoned that enterprises on the whole are operating at PUE figures of above 1.6, a “huge disparity” compared to figures of around 1.2 that big Internet companies now achieve. In Singapore, a 2013 study recorded an average PUE of 2.6 for data centres here. This figure was cited in the “Green Data Centre Technology Roadmap” by the Singapore government, released in July this year. The report also gave the average PUE ratings of data centres in the US and Europe at 2.2 and 2.0, respectively.

“You know what is the biggest problem in most small companies to even large enterprises? The people using the electricity aren’t the ones paying the bills. The IT department is using the electricity, but the finance or real estate department is paying the bill,” said Kava.

“They don’t see or they don’t share. That’s why I think you have to motivate them, by showing them that, ‘this is how much you are spending and here is how much your budget is’.”