Datacentre energy efficiency: Is the time now for a big switch-off?

Electricity consumption

John Booth, managing director of consultancy Carbon3IT, estimates datacentres, including colocation facilities, account for at least 12% of UK electricity consumption, or 41.11TWh a year.

Cisco has previously forecast global cloud IP traffic to exceed 14.1 zettabytes (ZB) by the end of 2020; IDC’s Seagate sponsored Data Age 2025 report projects overall data growth of 30% a year to hit 175 ZB – with data stored of 7.5ZB, up from 1.1ZB in 2019. Hyperscaler growth alone is expected to continue at a CAGR of 2% a year until 2025, according to ResearchAnd Markets.

Datacentre efficiency: a work in progress

Sustainability innovations of all kinds do continue to increase efficiency in datacentres, but there’s clearly still scope for other approaches.

French researchers Issam Raïs, Anne-Cécile Orgerie and Martin Quinson, writing in peer-reviewed journal Concurrency and computation practice and experience, quantified the likely impact of various shutdown techniques in the datacentre.

Operators have often been reluctant to reduce their number of switched-on servers because of worries about reactivity and hardware failures, and misjudgements about energy gain, they suggest.

The team simulated various production infrastructures and machine configurations under different shutdown policies and workload predictions. These include actual server switch-off and hibernation modes, and suspend-to-disk and suspend-to-RAM techniques, heterogeneous processing, and boot-up costs in time and energy as well as server lifetime, as well as looking at energy-aware algorithms (in related work).

“Shutdown techniques can save at least 84% of energy otherwise lost to power idle nodes. This remains true for prospective energy-proportional hardware and even aggressive shutdown policies do not impact hardware lifetimes,” they write.

SEDC’s Hewkin notes that some forecasts still estimate that datacentre power consumption could devour 20% of UK generation in the next few years if current trends continue. Yet most data generated is only used a few times, sometimes only once, and then stored forever “just in case”, such as bank accounts or a company’s trading accounts. Banks have already started moving older consumer data into “cold” inaccessible storage.

Warm datacentres

Additionally, he anticipates that “warm” datacentres might cost half as much to build as traditional datacentres, with less plant equipment and no backup generators. Existing datacentres could be repurposed, concludes Hewkin, for emerging data-driven use cases from automation to AI.

New technologies and strategies across hardware, media, software and architecture can continue to reduce data traffic and storage requirements.

Chief among these are software solutions, including active archiving, that control access to data – as  a major challenge to a big switch-off at server level has been how to efficiently provision resources dynamically.

On-demand workloads are highly variable, with no well-established taxonomy that defines them as they cross the cloud. Inter- and intra- migration of virtual resources can be a problem even as IaaS, PaaS and containerisation continue to advance alongside prediction methods such as regression or time-series techniques.

Tony Lock, distinguished analyst at Freeform Dynamics, agrees that suitable software solutions are beginning to emerge. Approaches will likely select from and combine the mix of media types available and increasingly cost-effective – from Solid State Disk (SSD) to tape, disk and “perhaps even non-volatile memory in future”.

However, if datacentres are to start switching off customer data, they had better be sure those customers are not going to want to use it or even look at it, he says. A big switch-off “sounds wonderfully simple in theory – the problem is that in practice, it’s an absolute nightmare”.

The whole area of data accessibility is full of pitfalls that are likely to cause massive friction between customers, IT providers, IT teams in the enterprise, and datacentre companies.

“It’s really difficult to actually classify that data. Yes, you can go through it and maybe try and look at when was the last time it was accessed,” says Lock. “Say that was some time ago, so maybe you can move it to a cooler, cheaper form of storage, or eventually switch it to something which doesn’t consume any electricity at all until you want to restore.”

Data discovery investment needed

An investment in data discovery and analysis, which obviously takes time, effort and resources too, is required – and will probably be ongoing. Only after analysis should data be moved offline to a cold, or even warm, storage alternative according to specific business needs and requirements.

“You have to be absolutely certain the user of that data isn’t going to go looking for it very quickly,” says Lock.

“Unless you can start using some sophisticated form of quasi-archiving software, not only knowing what data they’ve got but understanding how valuable it really is to the business, and what are the politics involved in doing something about moving it elsewhere, achieving that is complicated.”

But Lock points to emerging developments in autonomous data management software for multiple storage platforms and cloud by many “very large” incumbents and “extremely well-funded” new entrants, as well as supplier-neutral organisation the Active Archive Alliance (AAA).

The AAA promotes automated policy-based migration that frees up primary storage as data ages – although it focuses on actively archived data that remains retrievable and searchable, somewhat sidestepping the question of changing people’s expectations of data availability.

It believes that at least 60% of the predicted 7.5 ZB stored by 2025 does not need to be on expensive, higher performance tiers and can be archived after 90-120 days of low activity – unchanging and rarely overwritten.

“Not all this data should be archived. But much of it should be, thanks to strong retention policies and the new-found ability to gain value from archived data,” an AAA report states.

Factoring in the user

Lock notes that end users will typically complain when immediate data availability is removed. This is perfectly understandable in behavioural science terms – it is their data, so they want access and control, whenever and however they want it.

Enterprises and users – the politics and policies – must first come to terms with the idea of data not being “always on”. Providers must examine how customers and internal business users will react: have policies actually been explained to the users? Or will the business leave IT to take the blame for any lack of transparency?

“You’ve got to have the right tech in place, but you’ve got to have the right people and policies for the business to be willing to do something about this,” says Lock. “And we’ve been talking about ownership of data for a long time. Businesses want to be able to see their data. They don’t actually want to do anything about it, though – unless it’s really easy, or preferably invisible, for them.”

SEDC’s Hewkin agrees: “The industry has yet to adapt to data that is not available 24x7. It is assumed that everything we need to know will always be available at a touch of a button or click of a mouse.”