Kate's Comment

Thoughts on British ICT, energy & environment, cloud computing and security from Memset's MD

Energy-efficient cloud computing: Jevons Paradox vs. Moore’s Law

I am in my final year of a part-time PhD attempting to answer whether Moore’s Law (which has accurately predicted that computational power will roughly double per unit of cost or unit of energy every two years for some time) will mitigate the Jevons Paradox (see inset) in relation to energy efficiency of ICT thus allowing ICT to be an enabler of the green economy.

The Jevons Paradox, and more specifically the Khazzoom-Brookes postulate (a modern interpretation), states that increased energy efficiency paradoxically tends to lead to increased energy consumption due commoditisation. The Khazzoom-Brookes postulate has been shown to be true under the neo-classical growth theory for a range of scenarios and Jevons Paradox in the wider context has numerous historical precedents, most notably around fossil fuel utilisation. The current explosive growth of cloud computing could well be an example of the paradox in action.

ICT is undoubtedly part of the solution to climate change, but many commentators have pointed out how our sector’s power consumption is increasing at a dramatic rate. My research is not theoretical – I have been working with Surrey University to examine the efficiency of our Miniserver Virtual machines, among other approaches.

Data gathering

This part of the research is an ongoing project examining the load levels and power consumption of our Miniserver VM host server estate which has been running for almost three years now.

The first element of the project has been lots of measurement! As well as measuring the power the servers draw we monitor a number of load characteristics on each host and each individual virtual machine, including:

  • CPU utilisation
  • RAM utilisation
  • Disk storage utilisation
  • Disk input/output transactions per second
  • Disk bandwidth
  • Network bandwidth

We take samples every hour, and have been doing so for some time, so we now have a couple of billion data points.

That data has allowed us to see where the bottlenecks are in the infrastructure. One of the first things we noticed was that CPU is hardly touched on the host machines, but that disk I/O was limiting. We therefore disabled swap (virtual memory) and made RAM cheaper instead. That helped a lot!

The second generation came when still customers were not really using the CPU, so we put even more RAM into servers and more disks.

We are now onto our fourth generation of host since Summer 2006, and there are many iterations in between where we have tried different architectures, different types of disks, different ways of distributing the virtual machines (eg. mixing ones of the same size vs. putting ones of the same size on the same host) and so forth.

This is necessary since the server technology is constantly evolving too! We also refit old host machines where practical and have an analytical approach to what to replace.

Beating Moore’s Law

By iteratively refining and fine-tuning the host hardware to optimise it for the actual load profile, we’ve been able to reduce the average power consumption of a one Miniserver Compute Unit (MCU – see inset) from 26 watts in mid-2009 to 4.7 watts by the end of 2011. The chart below shows this progress:

Our “MCU” or “Memset Compute Unit” is based on our VM2000 which has 1 x 0.4GHz X34xx series Xeon core or equivalent, 1GB RAM, 80GB RAID1 HDD. Our latest generation of host servers are Dell R310’s with 2.93GHz X3740 Xeon CPUs, 32GB of RAM and 4 x 2TB hard disks in a dual RAID1 array.

Moore’s law is a good approximation for the rate at which energy efficiency of computers improves. Even taking the more aggressive figure of a doubling of efficiency every 18 months (see inset) Moore’s Law predicts those VMs should be using 8.41 Watts – 79% more than Memset’s power usage. Even if we apply Moore’s law to the lower extreme of my error range for the figure in 2009 we are still beating its prediction by 47% (6.92 Watts/MCU).

Moore’s Law technically refers to a doubling of CPU transistor density every 24 months, however when computers are viewed overall including other advances such as multi-core, higher clock speeds and storage density their performance per £pound or per watt doubles roughly every 18 months.

This is no mean feat because in the above I am not merely showing the efficiency of our latest generation of Miniserver VM host, but the average power cost across our entire estate. We do have a programme of replacement but that happens after 3-4 years – even with a lot of older, inefficient hardware we are still beating the odds by a lot!

Embedded & data centre energy Please note that this paper does not examine the embedded energy or the cooling and other data centre overheads. To get those figures add about 20% for the embedded energy (see my article on the embedded energy of servers and about 50% for cooling and other overheads including the network.

The Jevons Paradox / Khazzoom-Brookes postulate

However, we are also fighting a battle. As you can see from the red line on the chart above as virtual machines have become cheaper – more of a commodity – their average size has increased. My operations manager despairs at the profligate wastage of resources by our customers but when the resources are so cheap they don’t really care, so logs are not rotated, software bloats in terms of CPU and RAM requirements and Web pages become ever-larger yet our customers are still largely doing the same things they were doing ten years ago.

This is the Jevons Paradox in action: as a resource becomes more efficient the market forces move (mainly via price elasticity) to paradoxically increase overall utilisation. But is that really the case? Well I’m relieved to report that it is not – at least at Memset. The chart to the right shows our average power consumption per logical server (dedicated or virtual) and as you can see we are turning the tide – just.

The effect is pretty minimal however and as we move up the food chain with big name customers for whom our fees are miniscule we are again fighting a battle to convince them that they should ditch wasteful dedicated servers and use efficient, scalable cloud solutions.

Further, we are arguably the undisputed leader in the green hosting space in Britain. If we as exemplars are only just managing to reign in our power consumption per logical server (partly by encouraging customers to go from dedicated to virtual) then what of our less environmentally-inclined peers?

Conculsion

In short, we’re doing great stuff by making our Miniservers more efficient, and delivering the following benefits:

  • Lower prices (we’re currently cheaper than Amazon Web Services)
  • Environmental responsibility (UK’s first accredited Carbon Neutral ISP and still trail blazing!)
  • Better VM performance (by fine-tuning we are eliminating bottlenecks)

However, the question of whether we as an industry (rather than just one efficient corner) can employ Moore’s Law to combat the Jevons Paradox and deliver on our promise of being part of the solution to climate change without becoming part of the problem remains unclear – but that’s why I’m doing my PhD and I hope to have the answer quite soon!

9 comments