Freitag Jan 16, 2009

Seven Grams Revisited

After the "7 gram story" not only did hit Techcrunch, but also my favourite german-language news site, let's revisit my original calculation:
• In 2005, Prof. Paul A. Strassmann George Mason University, held a presentation “Google - A Model for the Systems Architecture of the Future”, pdf here. Slide 11 says that in those days, 31654 2-socket x86 servers worked in parallel to serve 40 million searches a day.
• So how much energy does a 2-socket x86 server at Google consume? Google engineer Luiz André Barroso gave an answer in his 2007 paper "Power Provisioning for a Warehouse-sized Computer". They measured 143 Watts.
• These 143 Watts have to be multiplied with a power factor. In 2007 I used a factor of 2 (the industry standard, meaning: "For every Watt fed into the computer you feed another Watt in Cooling etc.). In 2008 Google reported a PUE factor of 1.21. With this, every server actually draws 173 watts for power and cooling combined.
• So 31654 servers times 173 watts times 24 hours gives us 131427 KWh energy a day. Divided by 40mln searches, that's 3.3 Wh or 2 grams CO2 per search, using 600 grams of CO2 for every KWh generated with the german energy mix of nuclear, carbon-based and renewable sources
• In 2005, a dual socket x86 server meant two cores. Today we have eight cores and roughly four times the throughput per watt, bringing us down to 0.8 Wh or 0.5 grams CO2 per search
• So the difference between the "7 grams" from 2007 and this calculation today is caused by: Quad Core CPUs, lower measured power draw of the servers, PUE of 1.21, carbon factor calculated with energy mix instead of coal only. Net result: A factor of 12
• And the real number Google disclosed this week is even lower: 0.3Wh or 0.2 grams CO2
• That is really not much: If you start a search every 30 seconds on a typical Laptop (30 Watts), your client uses more energy triggering than Google executing the search.

• Montag Jan 12, 2009

Wrong by a factor of 35

Sorry Google: In 2007, I guestimated the carbon dioxide emission of one Google search to be roughly 7 grams. That number was never intended to frighten people from doing internet searches, because I always assumed that Google would implement Search as efficient as possible. Just think of the advantage you get when you do search 10% more efficient than your competitors while doing billions of searches.

After the Sunday Times had published last week that Alex Wissner-Gross estimates the same number Google now officially has answered.
I was wrong. Very wrong. Wrong by a factor of 35. Wrong even when you take into account that Moore's Law and Google engineers had 20 months to increase efficiency since my first guestimate.

So now we have it: One Google Search produces as much CO2 as 10 seconds of breathing!

Update: Could be that the Sunday Times directly took my number (after generating 0.2 grams of CO2 and finding this blog with a Google search) and connected it to other work done by Alex Wissner-Gross. Funny how an innocent number can make big news 20 months later!
At least I got my 2 seconds of fame on Techcrunch: "This obscure blog post" ... "Rolf Kersten’s Weblog (who?)" Cheers!