Using AI to cut humans out of the datacentre sustainability equation

As the idea of using artificial intelligence to reduce energy use by datacentres gathers momentum, does this spell the end of human intervention on sustainability matters?

The potential for artificial intelligence (AI) to cut the power consumption of datacentres is an area of growing interest in the industry, as operators seek new ways to reduce costs and drive up the performance of their facilities.

Google is among the tech giants leading the way on this at the moment, having spent the past two years experimenting with machine learning and AI technologies in its hyperscale datacentres.

Earlier this year, Google published details of how tapping into the AI expertise of its Deepmind division had led to the creation of a machine learning system that could help optimise the cooling of its datacentres and, in turn, cut their energy consumption.

The system comprises a series of neural networks that are “trained” using data collected from the thousands of datacentre sensors used by the internet giant to track changes both inside and outside the datacentre environment, and predict how some of these conditions – the temperature and pressure – might change within the hour.

The company deployed the system at a live datacentre site and claims the experiment saw a 40% reduction in the energy used to cool the facility, leading to the lowest power usage effectiveness (PUE) score the site had ever achieved – 1.06.

Francois Sterin, Google’s director of global infrastructure, told delegates at November’s Datacenter Dynamics Zettastructure conference in London that the quoted 40% savings figure had been repeated consistently as the firm replicated its trials across other sites in its datacentre portfolio.

What prompted the company to experiment with AI and machine learning in the first place was the realisation that, after several years of using “human engineering” to drive down the PUE scores for each of its facilities, its progress had started to slow.

“We used human engineering to get down to 1.1 or 1.11 but then, clearly, we were kind of reaching a plateau,” said Sterin.

The end is nigh

It is impossible to say whether Google would ever have gone on to achieve a PUE score of 1.06 (or even lower) without using AI. Even so, news of its advances have prompted industry watchers to ponder whether human intervention will be required to help datacentres achieve even greater efficiency improvements in future.

Roel Castelein, membership services director at datacentre energy efficiency consortium The Green Grid, told Computer Weekly that Google’s work had certainly raised questions about the ongoing value of relying on humans to improve the operational efficiency of server farms.

“Google’s work makes the incredible point that the human mind can only go so far, and then you need statistical analysis to improve your process,” he said.

“They are using a statistical model that uses a lot of data and statistical analysis to get these fantastic improvements, and go beyond the point where the human mind cannot follow any more.”

Given that PUE is calculated by dividing the total power consumption of a datacentre by the amount of energy used by the IT equipment inside it, a score of 1.1, which Google quoted before its AI intervention, suggests the operator needs to address the energy use of its IT kit to achieve a lower score.

Simon Brady, datacentre optimisation programme manager at Emerson Power Network, told Computer Weekly: “If we look at the technology we deploy in datacentres now – from a UPS or cooling perspective – we are at the cutting edge and at the limits of the efficiencies we can reach.

“We as an industry are doing everything we can to improve efficiency, but now it is mostly falling on the IT suppliers to do more.”

Pooling resources

Earlier this month, the University of Lancaster revealed details of a self-assembling form of AI software it is developing that could be used to drive down the energy consumption of datacentre servers.

The system, dubbed REx, comprises a set of software components (such as memory caches or search and sort algorithms) that can independently reassemble themselves into the most efficient form of the software, depending on the type of workloads the server needs to run. 

According to the four-strong team working on the project, the software has the potential to eradicate the need for human involvement in some datacentre management tasks.

Barry Porter, a lecturer at the university’s school of computing and communications, works on the project. He said it was a big ask to expect human beings to deal with the complexity of monitoring, managing and tweaking concurrent datacentre workloads in real time.

“These facilities are extremely complex compositions of lots of hardware and software, all working together, and they are subjected to different workloads from users over the course of the day,” he told Computer Weekly. “For humans to understand that complexity in real time is just not possible.”

Read more about datacentre efficiency and sustainability

Machines, however, are much better equipped to collect, analyse and act on the information generated by day-to-day datacentre processing tasks, said Porter.

“We are all now looking at how we can get machines to consume all this data about how user requests are handled every day, how datacentre systems respond to these requests, analyse them and, in turn, find out the most efficient way to do this,” he said.

Many of the hyperscale datacentre providers, including Amazon, Facebook and Microsoft, are experimenting with AI systems, said Porter. Some are doing so for energy efficiency reasons, while others are using AI to help them pinpoint issues within their complex datacentre infrastructures.

Given how many organisations are devoting research and development (R&D) resources to this area, Porter said even greater efficiencies could be achieved in future by combining their approaches.

“The Google stuff is focused on how do you supply exactly the right amount of cooling over the course of a day, based on predicting likely use, and Facebook is really interested in making its five massive datacentres around the world more efficient,” he said.

“We need a fully connected solution that considers all the factors [each company is targeting] together and makes the best use of everything, because we absolutely should combine these things.”

Back to basics

In the meantime (and for the majority of operators), there are some simpler steps they can take to improve the PUE of their datacentres, said Emerson’s Brady, who is also chairman of the Green Grid’s EMEA-wide liaison committee.

“Everyone talks about what the big hyperscale datacentres are doing and all the new datacentres that are being built with the newest technology, but a lot of legacy datacentre sites are 10 or 20 years old and have a lot of older technology in them,” he said.

“While at the high end there is some really great technology, the vast majority of datacentre estates have got a long way to go with regard to getting the basics [to improve PUE] right.”

Brady suggests, for example, taking steps to ensure any holes in a datacentre’s raised floors are sealed up and implementing other pieces of best practice recommended by The Green Grid.

Not necessarily bad

It is also worth bearing in mind that a PUE score of two or three is not necessarily a bad thing, because each facility is different.

“A colocation facility can be half empty for years and, as a result, have a high PUE for years,” said Brady. “And they shouldn’t be ashamed of that. A high number just means there is room for improvement or your datacentre isn’t full, which a lot of them aren’t.

“If you’re a bank, insurance fund or stock exchange, the [datacentres] need to be available 24/7, 365 days of the year, so you’re going to have highly available systems, which may affect how energy-efficient the site is.

“In that instance, the needs of the business come first and you just have to try to be as efficient as possible.”

Read more on Datacentre energy efficiency and green IT