06 Jan 2021

PUE - The data centre measure of efficiency that is still going strong...

I've spent almost three decades working within the data centre industry and around this time of year it's usual to see blogs, articles and commentary on "industry predictions" outlining new ways of operating, cooling and monitoring data centre capacity.

Despite all this annual change, one measure of efficiency that has always steadfastly remained is the measure of power usage effectiveness (PUE) which shows no signs of tiring.

This is because energy efficiency remains a crucial topic for data centre operators and PUE is still the best way to track performance, both from an economic and sustainability perspective. So, how do you make the most of it – and what do you do when PUE can’t be improved any further?

Data centres can be power-hungry, especially ones like our Kao Data campus in Harlow that looks after high performance and supercomputing deployments. If you're running megawatts of compute, high density server configurations and NVIDIA GPU-supported clusters you want to make sure your data centre's power profile is as lean and mean as possible.

In 2018 data centres consumed one per cent of all electricity used that year, worldwide. That’s a lot, but then data centres power businesses, economies, entertainment networks and much more. In 2020, we’ve depended on them more than ever. But how do we ensure that they use energy efficiently and responsibly?

The interesting news is that data centres used one per cent of the world’s electricity in 2010 as well, but although the amount of computing handled in data centres more than quintupled to 2018, the proportion of energy used remained the same. Commercial colocation data centres - especially latest generation facilities and those at industrial and hyperscale - typically do a better job of this than smaller on-premise legacy data centres because any energy they save is profit and modern design, architecture and scale can really squeeze PUE. However, sustainability agendas, energy availability and costs are giving every business a reason to cut energy use.

Driving down PUE

Doing that requires data centres to work out how they can make better use of energy, which is where power usage effectiveness (PUE) comes in. It’s calculated by dividing the power coming into the data centre by the power used to run the computing equipment in the data centre. A PUE of one would mean that all the power coming in was used by computing equipment, but this is unrealistic. Cooling, lighting, backup power supplies, security and other services all require power, but the aim is to get as close to one as possible.

At Kao Data we've worked hard to get our PUE to an excellent, market-leading 1.2, even at partial loads. However, there are still areas - design, configuration, cooling, etc - we are focusing on to see if we can tweak to reduce this even further. Google says it averages a PUE of 1.11 across its large-scale data centres, with some sites coming in under 1.06. The average across data centres in general has fallen from 2.5 in the mid-2000s to around 1.6 in 2020. However, Uptime Institute says the average PUE hasn’t really changed since 2013.

That’s partly because there are limits to what can be done to improve PUE at an existing data centre. Choosing a location with a low ambient temperature will reduce the need for cooling, for example, and designing the building to let in natural light will cut lighting costs. However, these measures need to be applied when the data centre is being built, and obviously not all data centres can be located in areas like the Arctic Circle. If you're running compute in hot and humid climates your PUE task becomes much harder, but operators are still pulling out all the stops to get their facilities to be as efficient as possible.

Other measures can only be implemented once. Painting the outer walls with a weather shield coating can help with cooling, as can a roof terrace with vegetation. The layout of server racks can be optimised too: positioning alternate rows so that all cold air intakes face one way and hot air exhausts face the other creates a hot aisle/cold aisle configuration that helps to optimise cooling. Switching to LED lighting will consume less energy, too.

Developing a range of measures

Day-to-day management of the data centre can also help to optimise PUE. Monitoring the efficiency of servers and how much power they consume under pressure or while on standby can help to keep them running with optimal efficiency. Likewise, the UPS load should be matched to the system load to help keep PUE at the lowest level.

Important though it is, PUE is just part of the picture. As mentioned above, you can’t always move your data centre to another location, some compute will always need to be close to the end-user in urban environments and this is a big factor in effectiveness. A very efficient data centre in a tropical climate might have a higher PUE than a less efficient, legacy data centre in a very cold location.

Operators that have done all they can on-site can still reduce their data centre’s footprint by other means, such as by using renewable energy or finding ways to reuse the heat generated by the data centre. At Kao Data we're doing both - using 100% certified green energy, and exploring ways we can help heat buildings across Kao Park.

PUE is by no means a perfect solution, but it remains a valuable tool that operators can use to measure their energy effectiveness at a time when pressure is growing to be as efficient as possible. PUE is here to stay and my 'New Year prediction' is it won't be going anywhere, anytime soon!

If your compute's PUE is as important to you as it is to us, then get in touch and have a virtual tour of the Kao Data campus and we'll be delighted to show you what we're doing to ensure we stay at the front of the PUE race.

Spencer Lamb

Spencer Lamb is Vice President at Kao Data and an experienced figure within the industry. He has a keen interest in HPC, AI and hyperscale computing.



Share

Other articles

13 Jul 2021

DGX-Ready - But, what does that actually mean?

Read more
21 Jun 2021

Reinforcement Learning and its position within the AI family tree

Read more
13 Jun 2021

Turbo-charging the engine of innovation in Cambridge

Read more
Get a quote