Distributed Diagnosis based on Trellis Processes

Distributed or networked systems become rapidly intractable as their size augments. This is due to the combi- natorial explosion taking place in the state space, but also in the trajectory space. The latter is reinforced by the amount of concurrency in the system (the fact that several non related events can occur simultaneously). We describe a triple strategy to extend standard monitoring approaches to such systems. We first use a true concurrency semantics to represent trajectories of the system. We then describe a very compact data structure to handle sets of trajectories, that we call the trellis of the system. This notion generalizes to distributed systems the standard notion of trellis of an automaton, where one dimension is time, and the other dimension encodes possible states at each time. Finally, we show that trellises have nice factorization properties, that allow the design of distributed monitoring approaches. In other words, a problem like state estimation for example, can be solved by parts, component per component, provided interactions between components are properly handled. We illustrate the approach on a distributed diagnosis problem.