An ideal class of problems is Monte-Carlo simulation. You distribute identical programs to the machines, and different random-number seeds. Then each machine computes any desired amount of time (1 hour ? 1 day ? 1 week ? 1 year ?) without ANY communication. At the end each machine communicates just one or a few numbers to central, which takes the averages - that's the result. I had Cray machines working at theoretical max performance (all processors) that way. John G. Zabolitzky