Ovid,

I asked something similar (but different) in the past, and BrowserUk came up with a client/server set of scripts which emphasizes the extraordinary performance improvement using persistent connections for inter-task or inter-machine communications. His scripts are here: Re^7: How to maintain a persistent connection over sockets?. You may want to download them and play with them. The results are quite interesting!

Also BrowserUk's suggestion about the design of the server is right on. I would just enhance it to be 2 machines at two different locations (redundancy and disaster recovery). But you could always improve that later.

Others have already stated some really good ideas, but I'm confused by one thing.

How can the 36 machines get the same input all at once and return 36 similar responses?

That seems like a waste of resources and the inter-machine communications could eat up a lot of your maximum response time!

To scale, you need to have all machines/cores/etc. sharing the work-load not duplicating the process.

On the other hand that was done by the IBM Watson computer ( Jeopardy fame ), but that was to see how many different processes returned similar or exact responses, so that a master task would tally the results and 'speak' with the 'best' answer.

Good Luck, it sounds like a fun project...Ed