Wednesday 7 April 2010

New Scientist fails maths.

A recent article in New Scientist online (I hope it doesn't/didn't get past the editors into the print version - oh wait, it says Magazine issue 2754, oops!) has the title "Search engine's dirty secret". In it physicist James Clarage explains to us the laws of thermodynamics and does some simple calculations to conclude that every Google search uses 100 Watt hours of energy,
so one search has the same energy cost as turning on a 100-watt light bulb for an hour.
But does it? Really? Does that number sound way too high to anybody else?

Ignoring the clunky use of the laws of thermodynamics to explain in an extremely round-a-bout way that computers use electricity (duh?), I'd like to pick over his calculation, which contains three serious errors, even though it only has three terms. His simple (very simple; too simple) argument basically goes like this: take the amount of power Google's servers use at any one time and divide this by the number of searches every hour to give an energy cost per search. This is the calculation he makes...
(1,000,000 servers)×(1,000 Watts of power per server)/(10,000,000 searches/hour)=(100 Watt hours/search)
Pretty straight forward, but where do these numbers come from? The first term, the number of servers Google has, is an estimate from Gartner research, and I have no reason, or need, to dispute it.

The second states that each server constantly uses 1 kW of power. This is a conveniently round number, but is it true? A simple search (with the lights off, I promise) reveals the following about Google's servers...
The company also revealed for the first time that since 2005, its data centers have been composed of standard shipping containers--each with 1,160 servers and a power consumption that can reach 250 kilowatts.
Er, so that would be 216 W per server, at maximum power. So already we've gotten our cost per search down to about 21 watt hours, or about 12 minutes of leaving the light on, assuming that every one of those 1 million severs is running at full capacity 24 hours a day - not a very convincing assumption.

The 10 million searches/hour figure is an estimate based on the number of searches in the USA in February 2009 measured by comScore, and once again I have no gripe with this number. Actually, yes I do. It's probably accurate, in and of itself, but it is the number of searches in the USA, and Google services the entire world. A better figure, also from comScore comes from December 2009, when Google served 88 billion searches worldwide, approximately ONE HUNDRED MILLION *inserts little finger into side of mouth with great comedic effect* per hour. So now our figure drops to 2.1 watt hours, or about 1 minute of power of your 100-watt light globe.

But that's not all. Unbeknowst to some, apparently, Google's servers do more than just process searches. In fact, that is probably only a small fraction of what they do, given that they first have to index every website on the planet, as well as running Gmail, Google Docs, Analytics, Adsense, Blogger, YouTube etc. etc.

To conclude, I really cannot understand why New Scientist (a generally high quality science magazine, that I have enjoyed for many years) would publish an opinion piece featuring a back-of-the-envelope calculation based on figures that are wrong by orders of magnitude and using assumptions that border on the ridiculous, when they could have just, you know, asked Google how much energy there web searches use.
How efficient is our infrastructure? Google-designed data centers use about half the energy of a typical data center. As a result, the energy used per Google search is very small; to be precise, we currently use about 1kJ (0.0003 kWh) of energy to answer the average query.
Which would power a 100-watt light globe for approximately 10 seconds, or, as Google puts it...
In the time it takes to do a Google search, your own personal computer will likely use more energy than we will use to answer your query
Touché.