Saturday 22nd September’s article in the New York Times, “Power, Pollution and the Internet“, which highlights the huge amount of energy wasted by data centres in general and internet data centres in particular, has already attracted a great deal of blog space. Many IT industry pundits are pointing out that, whilst the NYT has commissioned McKinsey to do the research, the biggest internet companies such as Google and Facebook don’t release their server utilisation figures.
The McKinsey stats quoted show that only 6 – 12% of energy used by data centres actually does any computing and this is not very different from the figures which have been widely accepted in the IT industry for decades. I’ve seen calculations that demonstrate the power that actually gets used doing any “computing” is as low as 2% and you can debate whether it’s now improved to 6% or 12%.
What we all know for sure is that it is low. We all know that most of the power that goes into a data centre gets used along the way, in UPS and air conditioning systems, running the server fans and spinning the disks, and very little is actually used by the processors themselves.
The NYT article ends with a brief comment on “the cloud” and the fact that this will further increase the demand for large, power hungry, data centres, stating: “Some industry experts believe a solution lies in the cloud: centralizing computing among large and well-operated data centres.”
So the question this raises is: “will using the cloud waste less energy than operating your own data centre?”
The advocates of “the cloud” claim, as you would expect, that it will reduce energy use. This claim seems to be based on two premises: firstly, that the large data centres that host your cloud service will be more energy efficient (more efficient servers and better PUE) than your own small one and secondly, that the effect of being able to spread the load among a number of data centres reduces the amount of spare capacity required and therefore of energy wasted.
So let’s take the first of these, is a large data centre necessarily more energy efficient than a small one? The answer is: no.
As far as the servers go, the servers you buy to install in your own data centre are just the same as the ones the cloud providers buy. As for PUE, it is quite possible to build a small data centre with a very good PUE just as it is possible to build a large one with a very poor PUE. The factor which is most likely to affect the energy efficiency of a data centre is not its size but its age and the age of the servers. New servers are much more energy efficient than old ones and any properly designed data centre constructed in the last 3 – 5 years will (or at least should) be vastly more efficient than one built say 8 or 10 years ago.
This brings us to one of the difficulties which the cloud presents for the user, you don’t actually know which data centre is processing your data. So, whereas with old fashioned “hosting” you could audit the data centres you intend to use and verify their energy efficiency, resilience, security etc, now you can’t, or at least not if the cloud works the way its proponents say it does.
On the second claim for energy efficiency, that spreading the load reduces the need for spare capacity: this could be true if the times of peak load were staggered.
For global internet use this is quite possibly the case: private internet use peaks at certain times of the day and so different time zones around the world will peak at different times. This means data centres serving consumers across the whole planet can expect their load to be relatively smooth.
For Google, Facebook, Amazon, ebay, etc this may well be so. But what about for corporate and institutional users, some will run processes over night while others have their maximum loads during the business day, others may have a fairly constant demand 24 hours a day and any spare capacity they require is only for redundancy. Still others may have other peaks such as monthly payroll runs or, for universities for instance, the annual intake of new students. What happens when all these peaks coincide or when the user takes the cloud provider up on the promise of being able to increase capacity at any time?
It seems to me that for cloud providers to be able to keep their promises they will need more spare capacity, not less, than if the same users operated their own facilities.
To take a simple example: if there are 10 small data centres each with a maximum capacity of 100kW and an average load of 60kW, that’s 1 megawatt of data centre capacity operating at 60% utilisation. If that is going to go into 2 large data centres, each able to take the full load, that is now 2 megawatts of capacity operating at 30% utilisation. Not only do we now have double the capacity but this also impacts on the second biggest factor affecting data centre energy efficiency, after age of the facility, and that’s its load utilisation. While it is relatively easy to get good PUE figures at full load this is much more difficult to achieve them at low loads. If the cloud is going to be able to provide the elasticity it claims then it will mostly be running at low load and consequently not at optimum efficiency.
However there is one way in which the use of cloud computing or indeed other types of hosting service can save energy although the cloud providers have not yet, so far as am aware, picked up on this. It is by users making use of the cloud to allow for different levels of uptime required for different applications. Most users will have some applications which are not business critical and others that are, but their data centres are not large enough to be able to build part of it to a “Tier 4” standard, for the really critical stuff, and part with a lower level of redundancy for the less critical.
This means they often run everything with a higher level of redundancy than is required. Why does this matter?
Because, as stated above, one of the major factors affecting energy efficiency is power utilisation (operating load compared to the full load capacity) and this is greatly affected by the level of redundancy. Put simply, a Tier4 data centre with 2N power systems is running everything at 50% load, even if the data centre is at 100% load. So with a typical data centre running at around 60% capacity each UPS is running at only 30% load (or possibly even less, depending on the exact arrangement). This is very inefficient. Indeed right at the very top of the list of best practices in the European Code of Conduct for Data Centres is to ensure that you don’t run services with a higher level of resilience than is justified by the business requirements.
So how can outsourcing help? Data centres offering hosting will, at least one would hope, be of Tier 4 standard and therefore users can make use of these for applications that really need it and operate their own tier 2/3 data centre for the rest of their applications, giving them the benefits of lower whole of life costs and convenience of having their servers in their own building.
So I’m not saying the cloud is bad. As with hosting and collocation, “software as a service” and all the other methods of out-sourcing your IT that have come and gone over the years, the cloud will suit some users and it will be right for some applications. Whether or not it will save energy will depend on how intelligently it is used.
It is perhaps worth noting that a tier 3 data centre is built so as all maintenance can be carried out without the need to shut down the facility and IT services are not disrupted by power outages or component failures. In short, for many organisations tier 4 is overkill, being unnecessarily expensive both to build and to operate.
Bio. James Wilman is the Sales & Marketing Manager at Future-tech EMEA. He sits on the British Standards TCT/7/3 expert panel which is working on the new data centre design, construction and operation standard. James has a passion for resilient energy efficient data centre design and operation. Linkedin profile