Would you please elaborate on this good-idea a bit further? Specifically, the part about the UDP transmission of the word/count pairs. Somehow the words have to be parceled-out among the various machines who are counting them. How would you approach that piece of your scenario, such that the distribution is fair among the “hundreds or thousands of” nodes and yet we can be sure (although using UDP) that every pair is in fact counted or considered by someone. Also, what sort of UDP/TCP network bandwidth is this scenario relying upon?