Total data centre energy use - it hasn't doubled - yet

photo by freefotouk on Flickr

The EPA report to congress on server and data center efficiency was one of the key documents that existed when the EU Data Centre Code of Conduct was created. An often cited part of the document was the estimate that data centres consumed  around 2% of all electricity production. Jonathan Koomey has recently updated that study and his new findings suggest that consumption has only seen half the predicted growth between 2005 and 2010. With everyone from Greenpeace to governments picking on data centre energy consumption this update makes some interesting reading, particularly the suggested causes of the lower than expected growth in Western markets.

You should go and read his own post on this topic, or download the updated report to read.

Energy use has not doubled

The findings are quite significant for the data centre industry; our energy consumption has not doubled as was expected at the time of the 2007 report to Congress on data centre energy use; instead growth has been roughly half of that originally predicted (56% worldwide, 36% in the US) for a number of reasons. It is important to note here that most of the projection is based on IDC data for number of servers shipped and not on actual metering of data centres.

Reasons, #1 the economy collapsed

The most concerning reason put forward is that the lack of growth is due to the economic crisis slowing down new IT equipment purchasing. This is concerning for a number of reasons, firstly that is has produced a substantial quantity of pent-up demand which is now being fulfilled. There is no shortage of articles or vendor press releases about server sales growth in recent quarters, indeed Gartner claims 17.3% for Q1 2011. If this is the case then we may be in for another period of substantial increases in installed IT equipment and therefore data centre energy consumption growth in the US and Europe.

Reasons, #2 we got more efficient but what about demand growth?

Improved data centre efficiency (PUE) and increased use of virtualisation are also cited as reasons for the lower than expected forecasts but these are also likely to be subject to ‘wisdom in hindsight’.

When we look back in 2015 there may well be substantial demand increases due to price elasticity as both of these improvements serve to reduce the marginal cost of each new IT service delivered into an ever-hungry market. The increasing popularity of cloud-based services (social networking, dropbox, iCloud etc.) and many High Performance Computing (HPC) applications which are largely creating new consumption rather than presenting a more efficient “alternative” to an existing consumption pattern is a particular risk here. See my post on price elasticity for a better description of this effect.

Problems, #1 Power consumption of idle equipment

Other areas for concern are the rising energy consumption of storage and network equipment, this is a particular issue because, despite the progress made in volume servers, most storage and network equipment still consumes the major part of its peak power when idle or at low load.

This is frequently compounded in virtualised environments where the improvements in power consumption at low load for newer servers through power management are blocked by virtualisation environments which do not properly support power management and therefore leave most of the servers consuming most of their power even when idle. Yes, there are tools which are able to mitigate some of this through shutting down entire server chassis, but these are only in use for a small part of the market.

Another issue for virtualised, cloud and HPC type environments is that people are buying bigger servers with more processors and memory now that they are able to properly utilise them, and this is reflected in the observed shift from volume to mid range and high end servers in Koomey's analysis.

These issues combine to deliver data centres full of higher specification equipment, much of which is consuming the major part of its peak load power even when doing nothing. This is supported by observed data from real data centres showing that the IT load is still virtually constant when it should have started to move significantly based on server churn.

Problems, #2 Growth markets – not our usual vendors

Whilst the results indicate that energy growth has slowed down considerably in Europe and the US there are a number of emerging market areas where new builds of data centres are continuing at a much greater rate; South America, the Middle East and, of course China. I won’t moan on about the carbon intensity of the utility grids in these areas but I will raise the issue that many of these data centres are not being filled with equipment from the vendors we recognise. In the emerging markets many of the servers come from ODMs and realistically the only way to count how many are being shipped direct to these markets from Korea, China and Taiwan would be to look at the number of processors Intel shipped in the period. When we look at this indicator we see something more like what we expected, a fifth consecutive quarter of record revenue with a substantial part of that revenue and the growth in processors destined for data centres, whoever makes the box they are shipped in.