Experimental evaluation of the unavailability induced by a group membership protocol

Group communication is an important paradigm for building highly available distributed systems. However, group membership operations often require the system to block message traffic, causing system services to become unavailable. This makes it important to quantify the unavailability induced by membership operations. This paper experimentally evaluates the blocking behavior of the group membership protocol of the Ensemble group communication system using a novel global-state-based fault injection technique. In doing so, we demonstrate how a layered distributed protocol such as the Ensemble group membership protocol can be modeled in terms of a state machine abstraction, and show how the resulting global state space can be used to specify fault triggers and define important measures on the system. Using this approach, we evaluate the cost associated with important states of the protocol under varying workload and group size. We also evaluate the sensitivity of the protocol to the occurrence of a second correlated crash failure during its operation.