A theory of system behaviour in the presence of node and link failure

Abstract

We develop a behavioural theory of distributed programs in the presence of failures such as nodes crashing and links breaking. The framework we use is that of Dπ, a language in which located processes, or agents, may migrate between dynamically created locations. In our extended framework, these processes run on a distributed network, in which individual nodes may crash in fail-stop fashion or the links between these nodes may become permanently broken. The original language, Dπ, is also extended by a ping construct for detecting and reacting to these failures. We define a bisimulation equivalence between these systems, based on labelled actions which record, in addition to the effect actions have on the processes, the effect on the actual state of the underlying network and the view of this state known to observers. We prove that the equivalence is fully abstract, in the sense that two systems will be differentiated if and only if, in some sense, there is a computational context, consisting of a surrounding network and an observer, which can see the difference.