Combining Different Failure Detectors for Solving a Large-Scale Consensus Problem
Serge Haddad, François Nguilla · 1999
Dependable services in distributed systems rely on some kind of agreement. Such an agreement can be obtained by solving the consensus problem. Most of the proposed consensus' algorithms are based on mutual knowledge of the participants and thus inadequate to wide area networks (WANs). In previous papers, we proposed protocols which deal with WANs constituted of interconnected physical groups of machines (LANs/domains). These protocols rely on Chandra and Toueg' unreliable failure detectors model for asynchronous systems extended to handle broadcast addresses and safeness of a whole subnet. Nevertheless, we assumed the same characteristics for local and distant failure detectors. In the present paper, we propose a new algorithm which takes into account a different behavior for the detectors. More precisely, local detectors behave as the S class of Chandra and Toueg and so are more reliable than the distant detectors which behave as the \\PiS class. Moreover, the distant failure detector we dene, has the ability to test the failure of a local network via a broadcast address. We prove the correctness of the new algorithm and give some implementations hints in the Internet context.