J. Bowen at No Watermelons found this paper (PDF format) that details the architecture of Google.
Very cool. I knew that they had somewhat graduated from their shelves of cheap PCs but their direction is interesting. Racks and racks of cheap 1U Single Proc boxes.
And speaking of scalability...
I worked for MSFT for some time both as permatemp and blue-badge and managed this lab:
Click on Image for full size
This is one row of the 1200 'client' machines -- these simulated a large volume of traffic coming in to a web site and database.
Click on Image for full size
A very gorgeous box (with a stack of Compaq servers and a disk array on the left). This was the Unisys ES-7000, they were a mainframe company and decided to evolve PC technology to mainframe specs. 32-proc, 96 PCI bus slots, gobs and gobs and gobs of RAM, three separate power inputs (you would be on three separate power sources and two could fail and it would still run). Excellent engineering and one of the most intelligent people I have had the pleasure of meeting.
Unfortunately, Windows does not scale that well. Neither do may other OS's so the Google technique is the way to go. Expensive lesson for all. Distributed systems of small boxes work best. You do need to keep your ducks in a row on this but the technology is straightforward and easy to scale.
Finally, a pr0n shot: we got a different Compaq disk array and 700+ drives.
Each drive came in it's own box.
Here is the mess that we left for the cleaning crew.
(I did give them two days notice)
(Marc and I broke down the boxes after the photos)
(Recycle yes -- I'm still using a bunch of them)
Click on Image for full size
