Sustainable Data Centres, Part II
JUNE 6, 2021
The first step in achieving a sustainable energy system is to reduce the energy demand. However, data centres are only one-third of the ICT system, the others being the network/transport/access and the energy consumed at the point of consumption and data generation – dominated by 'us' and our mobile phones.
In many energy systems that are aiming for sustainability, a demand reduction is a behavioural change by the users, such as the dress-code of occupiers in a low energy house – putting on a jumper when it gets cooler instead of immediately adjusting the heating thermostat, and vice-versa with the air-conditioning when it gets hotter. Unfortunately for us, the lid of Pandora's Box has been ripped open, and the demand for ICT services has grown exponentially – and data growth drives higher energy consumption. The fastest increase in demand comes from society and is enabled and encouraged by ever-faster connectivity through ever-broader bandwidth. Since the dot-bomb crash of the new millennium, data centres have only responded to demand and never created it. In this respect, 5G and UHD/4K will be bad for climate change as we build to meet the demand but later will have to pay for the consequences of streaming video, social networking, and gaming, etc. I always remind myself that 10% of the world live in abject poverty on less than US$1/day and that 3Bn people do not have access to an internet connection – and probably never will have, due to a lack of disposable income. Those people are often in places where climate change will strike hardest, and yet they will have played no part in causing it.
In the network, the behavioural change should be using the internet less and even restricting ourselves to LD monochrome video on mobile phones, but I think that consumption will only rise unless regulation intervenes under the guise of energy reduction, as described by The Coal Question (1865) author William Stanley Jevons who wrote, 'It is a confusion of ideas to suppose that the economical use of fuel is equivalent to diminished consumption. The very contrary is the truth, and he has, through the Jevons Paradox, also known as the Rebound Effect, been proven right ever since.
For our purposes, we will concentrate on 'reduction in demand' outside of the data centre, although inside data centres, this could include removing unneeded ICT hardware and comatose servers.
How big is the problem?
Since I was first involved with data centres in 1989, I have seen more estimates of how much energy data centres consume from our grid, both nationally and globally, than I can remember. However, I am tempted to say that no one really knows. The reasons for this lack of accuracy are numerous and varied, but the hyperscale facilities are often the exception to the rule as they tend to be loaded quickly, and their actual energy consumption reaches 75-85% of their utility connection capacity within the first 18 months after opening. However, many data centres are not hyperscale. They are existing or new builds that fail to attract publicity and often run at partial loads down at the 35-45% range.
Humans appear to love top ten lists, and off-the-wall comparisons between diverse objects – the classic comparison between chalk-and-cheese – and the data centre world has indulged in a few of them:
Over twenty years ago, someone wrote that data centres consumed more energy than the airline industry. I never found a calculation or the data that backed that up and never really understood the comparison either. However, it still forms the introduction to many publications on the impact of ICT on the environment. Of course, it makes no sense now, especially if it were true at the time. Commercial airline fuel consumption increased 50% over that period until 2019 (to 96Bn US-Gallons), with increased fuel efficiency balancing the growth in air-miles although the pandemic has knocked it back to the 2000 level (to 51Bn US-Gallons). However, data centre energy has been consistently growing globally with around a 10% CAGR. If the airline comparison was valid in the year 2000, data centres would now account for nearly 5x the energy consumption of the airline industry (excluding the impact of the COVID pandemic during 2020). At 15% CAGR, it would be closer to 10x.
But how much electrical energy would 96Bn US-gallons of aviation kerosene produce in a combined cycle turbine? About 1.16TWh, or 25% of the total annual electricity consumption of the USA, would only be the data centres' share of the ICT infrastructure. The developed world wants both air travel and ICT services, so trying to compare the energy consumption of one with the other doesn't seem to be productive, and the 20-year-old prediction can't be right now if it was then.
Then I remember from the early days a claim that a Google search used the same energy as boiling a kettle. Well, my kettle boils a litre of water using 0.135kWh which, for 5.4Bn searches per day (courtesy of a Google search), equates to a continuous load of 30GW, not far away from the average demand of the UK grid (66m people) and probably more than 20 times larger than Google now consumes in its 'search' provision. So, if it was correct around the millennium, which I doubt, it is a testament to Moore's Law and its derivatives in continuously decreasing the Watts per packet of ICT load. It certainly is not true today, yet it still appeared in an EU document only a couple of years ago and still in numerous dissertations.
The growth in monthly European data traffic (suggested to be a proxy for power demand) since 2001 is shown below, compared to the typical ICT hardware capacity development curve represented by the 40% CAGR of Moore's Law.
I also remember the 1/3rd rule that I am sure came from Johnathan Koomey in the late 90s. If not, I apologise to the originator and would be pleased to be corrected. The original version was that the total national energy consumption of the share was 1/3rd equally between users, network, and data centres. The forecast for the USA ICT was 12% of the utility for ICT in total, but that quickly became a generic one for developed countries with a stock-market and data centre estate. That was transposed to the UK as 9% for ICT and 3% for data centres, with almost no proof. In recent years, the total ICT figure was revised upwards. In 2017, 15-20% was quoted for the USA, but I don't recall any reference to the network and point-of-use. However, Rabih Bashroush (Global Head of IT at Uptime Institute) recently wrote a paper on energy effectiveness that addressed the topic well. For countries without hyperscale or social networking behemoths, the 1/3rd rule is still the best guess. However, for hubs like Dublin, Amsterdam, and Singapore, that number breaks down as the facilities serve a much larger region than the country in which they are located.
Lastly, in this review of claims, I should mention the current pandemic. In many quarters, the data centre infrastructure has been lauded as an enabler of home working, provisioning, and teaching in the essential isolation. This is clearly true, and internet shopping has expanded hugely, with Amazon ramping up to cope. We have hard evidence in Europe that in March 2020, when the first wave and lockdown began, internet traffic jumped 15-20%. For the first couple of weeks (in the UK at least), bandwidth appeared inadequate, but apps like Zoom quickly adapted and arranged more bandwidth from the core network, and the connectivity improved. However, an inspection of the data traffic profile over every 24 hours was fascinating: The trough remained at 4am, and the peak remained at 9pm. The shape of the curve stayed identical and just shifted upwards. So, at the 9pm daily peak, the traffic was, and still is, not related to 'work' but to home/family/teenager streaming HD video, internet TV, gaming, social networking, etc. It will be interesting to see what happens in the coming months when (hopefully!) we all begin to return to 'normal' – if that ever happens.
The profile in hourly traffic profiles are compared from 2016, May 2020, and April 2021, which clearly shows that the impact (of streaming video by domestic consumers?) has remained unchanged through the pandemic – not 'working from home' as many commentators suggest.
What is mitigating the growth in energy?
Over the past 25 years, the speed and capacity of ICT hardware has increased at least at the 40% CAGR of Moore's Law, and this has slowed down the growth in power draw. Let's take into account the gradually improving idle-power and utilisation enabled by virtualisation and a reducing interval for refresh rates that keeps the performance per Watt increasing year after year. We can see that Moore's Law itself (doubling of transistor count on a microprocessor every 18 months) has been exceeded by the technology that surrounds it.
The difference between growth in load demand and growth in hardware capacity looks to be continuing to result in 10% CAGR data centre capacity growth, although we need to get a handle on the impact of partial load in enterprise and colocation facilities.
With the consumptive nature of society and the vast untapped market of those poorer nations with low internet penetration, we would be too optimistic to assume that we can curb or even limit the demand for ICT connection and services. Jevons suggested that the only way of constraining consumption was through the price, either cost or taxation, and therein lies the most effective mechanism.
We are about to witness the initial roll-out of 5G in the most affluent and population-dense urban areas, and this will drive consumption up unless the cost of the microcell wireless infrastructure is not met by mobile-phone users who are satisfied with the price-performance point of 4G.
So, to continue with the search for sustainable ICT, the data centre industry is best advised to concentrate on the data centre itself and hope that the load itself continues to be our 'holy cow' and is addressed by society as climate change takes a firmer grip.
In the next part of this series, we address the second step in sustainability; improving the process.