Adaptive Strategies for Speeding Up Sequences of Consensus

The Paxos-MIC protocol allows to solve a sequence of consensus instances in an unreliable asynchronous system. It follows the basic principles of Paxos and uses two optimizations of this protocol: a safe one, S_o that is always activated and a risky one, R_o. The paper focuses on the interest of R_o that has been introduced by Lamport in the Fast Paxos protocol. We study the optimization R_o in favorable and unfavorable scenarios where it may lead to an additional cost. Paxos-MIC is adaptive as it tries to obtain the best performance gain depending on the current context. Indeed, between two consecutive consensus instances, the leader determines if R_o has to be triggered or not. In the particular context of a secure Web architecture whose design relies on a consensus service, we use a trace that contains all the HTTP requests addressed to a real Web site during a period of 16 days. We propose different triggering criteria and we analyze their accuracy to predict collisions when proposed values are produced at the rate observed in this trace.