The fault span of crash failures

Abstract

A crashing network protocol is an asynchronous protocol whose memory does not survive crashes. We show that a crashing network protocol that works over unreliable links can be driven to arbitrary global states, where each node is in a state reached in some (possibly different) execution, and each link has an arbitrary mixture of packets sent in (possibly different) executions. Our theorem considerably generalizes an earlier result, due to Fekete et al., which states that there is no correct crashing Data Link Protocol. For example, we prove that there is no correct crashing protocol for token passing and for many other resource allocation protocols such as k -exclusion, and the drinking and dining philosophers problems. We further characterize the reachable states caused by crash failures using reliable non-FIFO and reliable FIFO links. We show that with reliable non-FIFO links any acyclic subset of nodes and links can be driven to arbitrary states. We show that with reliable FIFO links, only nodes can be driven to arbitrary states. Overall, we show a strict hierarchy in terms of the set of states reachable by crash failures in the three link models.