Data Centre Productivity Will Be Measured In Kittens Per KWh
There’s a new measure of data centre efficiency. Peter Judge says it might as well measure kittens per kiloWatt hour
No one really knows how to measure data centre efficiency. There are problems with the Green Grid’s PUE, but the proposed new measure DCeP (data centre energy productivity) looks so vague, it’s hard to imagine anyone using it.
It’s important to measure how efficient data centres are: electricity is now a major cost, and data centre owners want to use less of it. PUE (power usage effectiveness) simply takes the amount of power a data centre uses, and divides it by the amount that reaches the IT kit. It’s a simple idea and has been very successful – arguably too successful.
The Green Grid originally said that PUE should not be used as a way to compare data centres, but that is exactly how it is being used. After all, if companies are pitching for business based on how green their service is, they need a figure to quote, don’t they?
Whose PUE is better?
Whose PUE is better?PUE is being used as a benchmark. But like all benchmarks it has problems, and creates perverse incentives to meet non-real demands. Because it divides total power by IT power, you can get a great result (close to 1) when the IT kit is burning lots of power, as long as not much is used by other systems such as cooling.
But what are the IT systems up to? PUE doesn’t care how effective the IT systems themselves are at delivering work, or even processor cycles. So an “efficient” low-PUE data centre might be simply idling and completely wasting all its IT power.
There is a very good replacement, FVER, suggested by British group the BCS, and developed by Liam Newcombe of Romonet. It compares fixed energy use with variable energy use, so it encourages data centre owners to reduce waste in both IT and the overhead of the centre. Very neat, but to displace PUE would take a huge shift in the data centre community, who have really got behind quoting PUE at ever opportunity.
In any case, both measures are about making sure that computer cycles are provided with as little energy input as possible, and neither looks at how effectively those cycles are used. What we need is a measure of data centre productivity, not data centre efficiency.
What about productivity?
Last week, it looks like we might have got just that. DCeP (data cente energy productivity), was proposed by a global taskforce, consisting of the Green Grid, along with two US government agencies (the Department of Energy, and the EPA’s Energy Star intiative), Europe’s Joint Research Centre (source of the EU code of conduct) and two Japanese government bodies.
What is DCeP? It’s very simple – and like PUE it’s possibly too simple. It’s the “useful work” done by the data centre in a given time, divided by the energy the data centre used in that time.
What does “useful work” mean? The taskforce says “DCeP allows each organisation to define ‘useful work’ as is applicable to its business”. So for a retail outfit, the DCeP would measure number of sales per kWh, while for a search engine, it would measure number of searches.
I assume for Youtube, DCeP is the number of kittens per kWh,and for Facebook, it would be some aggregate of the number of Buzzfeed links, unwelcome adverts and LOLs.
The taskforce press release says this makes DCeP a “custom and meaningful metric”. In my view it’s more likely to make it a non-metric. There’s no way to compare Amazon’s sales-per-kWh with Youtube’s kittens-per-kWh and Facebook’s LOLs-per-kWh. And there’s no way to make day-to-day comparisons in a single company, as the make-up of the business will change. On top of this, companies will adjust their own personal DCeP to suit them (for instance, maybe counting the volume of sales rather than the number).
There are hints in the Green Grid’s White Paper that generic DCeP variations might evolve, such as one for e-commerce. At the moment, it’s difficult to say whether the taskforce may has showed up the difficulty of the search for a generic measure of data center efficiency, or demonstrated its futility.
A version of this article appeared on Green Data Center News.
