Energy Efficiency Breakthroughs in Data Centres
Recent news that cryptocurrency miners are chartering Boeing 747s to transport masses of GPU processors to their ethereum/bitcoin mining lairs made me reflect on the current state of datacentre power consumption. The economics of cryptocurrency mining are almost entirely based on power prices. As an individual in Australia, you simply cannot compete in the global bitcoin mining market because residential power costs here are so high that you will never get a payback on your $2300 Antminer S9 mining machine. The only theoretical way to make it viable for an Australian crypto miner is to chuck a tri-generation solar array on the roof of their bitcoin mining shed, use the cheap solar power to run the miners during the day and use the heat pump connected to the tri-gen panels to keep the miners from spontaneously combusting from the heat they generate.
Whilst most businesses aren’t thrashing their servers to the extent that a crypto miner would, there’s still a marked asymmetry in a small business’s ability to buy cheap power compared to a multinational giant like Amazon/Google. Economies of scale factor in here like nowhere else. The cloud computing giants can buy the best hardware, hire the best infrastructure engineers, design highly efficient data centres, sign renewable power purchase agreements to lock in cheap electricity prices, write machine learning algorithms to optimise efficiency in real time and even put data centres underwater. You’d need a darn good reason to try to run your own server room these days.
The efficiency gains large data centres have made in recent years is beyond impressive. In 2008, an energy efficiency report to US congress painted a sombre picture of electricity demand from data centres. As usage of Facebook and other major web properties skyrocketed, servers were being thrown into data centres at breakneck speed and the site engineers were struggling to keep up. Their main tactic back then was to blast the entire room of servers with A/C regardless of whether an individual server was hot or not. This was not a very effective mechanism and as a result, electricity consumption from data centres was doubling every five years.
Back then, data centres consumed 2% of the US’ total electricity demand. Nine years later, with another two doubling cycles, the grid would have faced a lot of pressure from these server farms if nothing had changed. However, in an encouraging display of what is possible when innovation is brought to bear, nine years later, data centres still only consume 2% of total electricity demand. You can bet that’s not a result of stagnation in demand for new servers either. As cloud computing has waxed in popularity, data centres have sprung up all around the world. Since 2008, the number of installed servers around the world has risen by at least 40%.
That kind of increase in capability without any increase in electricity usage is exactly what we’d like to see from every business. Arguably we wouldn’t be able to enjoy the joys of web 5.7 (or whatever version we’re up to now) had it not been for these efficiency gains. The cheap computing that powers one’s Facebook feed is only cheap because electricity consumption from the server farms is low. Had Facebook scaled without driving down their energy consumption per server cycle, it’s pretty likely they wouldn’t be the successful company they are today. When you’re operating at that scale, energy bills are not cheap. We’re talking hundreds of millions of dollars a year. You’d have to sell a heck-of-a-lot of ads to pay for that. As such, there’s a pretty strong financial drive to reduce energy consumption.
Tech giants aren’t alone in experiencing big power bills. Aluminium smelters, plastics recyclers, horticultural facilities and every manufacturing business out there face the same challenges around growing output without being savaged by capacity charges. Yet we haven’t seen these type of efficiency miracles from other industries. The reverse has been true. Since July 1st, there have been a number of stories about Australian businesses ceasing to be going concerns due to the price hikes.
We could talk about tactics like water-side economizer systems, hot aisle containment and siting data centres in cold climates but most of these are not relevant outside the industry. What is relevant is the strategy the data centre engineers have used. They collected data…a lot of data. Each server in a data centre provides near real time feeds of its power consumption, utilization level, CPU temperature, fan speed and many other measurements. Having that level of granular data makes it relatively easy to optimise the system. Humans weren’t even involved in optimising some of Google’s server farms.
Most businesses have a dire lack of data on their power usage. Chances are the only insight they get is the total consumption number on their power bill. This single datapoint is impossible to use for any real decision making. About the only conclusion you can make is that you’re paying a lot for power and that you should try and negotiate with your power retailer to shave a cent per kWh off your off-peak rate.
To make better decisions, businesses need to increase the quantity and quality of information available. Getting access to interval consumption data is an important first step. Knowing that your consumption was 50 kWh between 3AM and 3:30AM when there was no-one on site gives you a clear indication that something needs to change. To discern what is causing the high base load requires the next level of instrumentation: sub-metering. By installing sub-meters on the major sets of equipment, you can drill down and see that the high overnight load is due to the HVAC system being left on overnight. Even that is not enough. To truly excel, you need to know how your facility performs compared to similar facilities around the world. Just as Google can compare the PUE (power use efficiency) of its data centre in Georgia with its data centre in New Hampshire, so too can you compare your school with an industry leading school in Arncliffe.
COzero can assist with gathering this type of data. Through EnergyLink, you get access to software similar to what tech giants have used to slash their energy consumption whilst increasing their output. We have submetering solutions and anonymous benchmarking software that shows how your business compares to others in the same industry. To find out more, get in touch.
Originally published at www.cozero.com.au.
