One of the first things I did outside of work at Google was to find out how many computers the company has. It’s a fairly secret number; it’s not quite a topic that people in the Googz like to talk about.
It took me a week to piece together the answer; and a few months to come to terms with my discovery. It’s hard to talk to people outside of the big G about the kind of stuff they pull off there, and I’m not talking about making ball pits out of director’s offices.
I can finally talk about this, now that this information is explicitly public, published in an article by MapReduce Gods Jeff Dean and Sanjay Ghemawat (bloggy synopsis here). In the paper, they talk of 11,081 machine years of computation used in Sept 2007 alone, for a subset of their MapReduce work. That’s 132972 machine months of CPU used in one month. Assuming all the computers were running at 100% capacity, without failure, without any break for the entire month, that’s almost a hundred and fifty thousand machines worth of computing used in September Oh Seven.
In other words, Google has about one hundred and fifty thousand computers that are reported here.
But does that account for ALL the computers at Google?
To find out, go ask a Google employee to violate his NDA today!
for your information, this may not be the right number. it should be obvious why. for example, they never said anything about not using hamsters. hamsters are 10x faster than computers, which would mean they could just have 10,000 hamsters and it would be fine.