Proposes a framework to process large distributed systems by parts, through distributed algorithms. We consider distributed (discrete event) systems as the combination of elementary components. Each component defines dynamics on several state variables, and the composition is simply defined by sharing variables. The compound system is asynchronous: each component evolves with its own clock, and exchanges information with its neighbors by means of the shared variables. An interaction graph can be associated to such a compound system: two components are neighbors of each other as soon as they share one (or more) variables. This structure is reminiscent of Bayesian networks, or Markov random fields, which use a graph to display dependencies between random variables. The parallel can actually be pushed quite far. We show that a large family of modular algorithms developed for Markov fields, in order to solve problems like maximum likelihood state estimation, can be translated into distributed algorithms to monitor large distributed dynamic systems.