Most of us take for granted the inner workings of a Google search and all we're concerned about is receiving near instantaneous results to our query. And just as well, because Google hasn't been one to pony up much specifics on the hardware it uses to sift through oodles of web pages. That is, until Google Fellow Jeff Dean gave a keynote talk at ESDM 2009.
According to Dean, a typical Google search might consume the processing power of 1000 machines. Talking about Google's growth in the past decade, Dean said search queries have gone up by x1000, and so too has the company's processing power (# machines * speed of the machines). In addition, Dean claims query latency has dropped from under 1000ms to normally under 200ms, as well as dropping the update latency by a staggering x10000, so that crawler updates have been reduced from several months down to just a few minutes.
Dean says the improved performance is a result of switching the holding indexes to now being completely in memory. The result is that it takes thousands of machines to process a search query, but it has made near instantaneous searches possible.