On the andvantages of free choice

Abstract

It is shown that distributed systems of probabilistic processors are essentially more powerful than distributed systems of deterministic processors, i.e., there are certain useful behaviors that can be realized only by the former. This is demonstrated on the dining philosophers problem. It is shown that, under certain natural hypotheses, there is no way the philosophers can be programmed (in a deterministic fashion) so as to guarantee the absence of deadlock (general starvation). On the other hand, if the philosophers are given some freedom of choice one may program them to guarantee that every hungry philosopher will eat (with probability one) under any circumstances (even an adversary scheduling). The solution proposed here is fully distributed and does not involve any central memory or any process with which every philosopher can communicate.