Google Server Set Up and Neural Nets

| | Comments (0)

The keynote at the CIC Tech conference by Charles Severance was very informative if you are interested in Google. There were several factoids that I found interesting, but what I noticed the most is that the Google server architecture is sort of like a neural net.

When I was a grad student, I was introduced to the concept of a "neural network". The extremely simplstic idea is that individual neurons do one limited task and then send data to another neuron to do the next task.

I think the best example of a neural network I know is modeling color vision. For instance, you can begin with sensor neurons which detect different levels of light (e.g. detect level of red light vs detect level of green light). The sensor neurons can then send their input to another set of neurons which only have the job of this light level data into a simple calculation called "hue identification" (e.g. 100% red + 100& green = yellow), and then sends that data to another system (e.g. the general visual system then object recognition then word recognition etc).

Each neuron is fairly limited in function, but the architecture is set up to perform complicated tasks very quickly.

Getting back to Google, Severance showed a video about how a search query works. Between load balancers, data storage, querying and instant HTML publishing, a typical query can actually hit 1000 servers in less than 2 seconds. Holy You Know What!

More interestingly, Google apparently uses cheap servers. Apparently they use truck trailers worth (as in they plug in entire trucks of servers into a server farm). And apparently, they must have data centers all over the world. This apparently explains how Google mail can be efficient south of the equator when otherwise the Internet tends to slow down (at least between the hemispheres).

Severance called this "building a brain", and for once I don't think it's hype.

Leave a comment