Anonymity-preserving failure detectors

Abstract

The paper investigates the consensus problem in anonymous, failures prone and asynchronous shared memory systems. It introduces a new class of failure detectors, called anonymity-preserving failure detectors suited to anonymous systems. As its name indicates, a failure detector in this class cannot be relied upon to break anonymity. For example, the anonymous perfect detector AP, which gives at each process an estimation of the number of processes that have failed belongs to this class. The paper then determines the weakest failure detector among this class for consensus. This failure detector, called C, may be seen as a loose failures counter: (1) after a failure occurs, the counter is eventually incremented, and (2) if two or more processes are non-faulty, it eventually stabilizes.