Towards an Adaptive Supervision of Distributed Systems C´ edric Herpson, Vincent Corruble, Amal El Fallah Seghrouchni Abstract The traditional, centralized, approach to supervision is challenged when communications between supervision and supervised systems become either slow, disrupted or too costly. To obtain a supervision system that is able to dynamically adapt itself to the communications’ state, we propose to distribute the supervision process through several autonomous agents. To evaluate our approach, we made experiments on a simulator for distributed systems using three different supervision approaches. Results show that our agent’s decision model does lead to a relevant autonomous supervision in distributed systems where a short response time prevails over a limited repair extra-cost. 1 Introduction The advent of physically distributed systems and the need to minimize the down- time of services and production processes call for more efficient supervision sys- tems. Indeed, as the complexity of systems increase, humans can no longer process the flow of information arriving at each instant. So, the improvement of system effectiveness requires the delegation of more tasks to the supervision system, and more automation of its decisions/actions. The supervision of distributed systems such as ones found in services, produc- tion, larges infrastructures and organizations is so far mainly centralized. As a re- sult of which, even though a number of malfunctions have predefined repairs, the unbounded communication time between the supervision system and the geograph- ically distributed regions of the supervised system delay the repairs. This – some- times unnecessary – communication time reduces the effectiveness of the supervi- sion and so increases the supervised system’s down-time. C´ edric Herpson, Vincent Corruble, Amal El Fallah Seghrouchni LIP6, Universit´ e Pierre et Marie Curie, 75225 Paris, France. e-mail: FirstName.LastName@lip6.fr 1