Dependable Decentralized Cooperation with the Help of Reliability Estimation.

Internet supercomputing aims to solve large partitionable computational problems by using vast numbers of computers. Here we consider the abstract version of the problem, where n processors perform t independent tasks, with n <= t, and each processor learns the results of all tasks. An adversary may cause a processor to return incorrect results, and to crash. Prior solutions limited the adversary by either (i) assuming the average probability of returning incorrect results to always be inferior to 1/2, or (ii) letting each processor know such probabilities for all other processors. This paper presents a new randomized synchronous algorithm that deals with stronger adversaries while achieving efficiency comparable to the weaker solutions. The adversary is constrained in two ways. (1) The set of non-crashed processors must contain a hardened subset H of the initial set of processors P, for which the average probability of returning a bogus result is inferior to Notably, crashes may increase the average probability of processor misbehavior. (2) The adversary may crash a set of processors F, provided vertical bar P - F vertical bar is bounded from below. We analyse the algorithm for three bounds on vertical bar P - F vertical bar : (a) when the bound is linear in n the algorithm takes Theta(t/n log n) communication rounds, has work complexity Theta(t log n), and message complexity O(n log(2) n); (b) when the bound is polynomial (vertical bar P - F vertical bar = Omega(n(a)), for a constant a epsilon (0, 1)), the algorithm takes O(t/n(a) log n log log n) rounds, with work O(t log n log log n), and message complexity O(n log(2) n log log n); (c) when the bound is polylog in n, it takes O(t) rounds, has work O(t.n(a)), and message complexity O(n(1+a)), for a epsilon (0, 1).