Section: Research Program

Management of large distributed systems

The generic terms of “supervision" or “management" of distributed systems cover problems like control (and controler synthesis), diagnosis, sensor placement, planning, optimization, (state) estimation, parameter identification, testing, etc. These questions have both an offline and an online facet. The literature is abundant for discrete event systems (DES), even in the distributed case, and for some quantitative aspects of DES in the centralized case (for example partially observed Markov decision processes (POMDP), probabilistic diagnosis/diagnosers, (max,+) approaches to timed automata). And there is a strong trend driving formal methods approaches towards quantitative models and questions like the most likely diagnosis, control for best average reward or for best QoS, optimal sensor placement, computing the probability of failure (un)detection, estimating the average impact of some failure or of a decision, etc. This second research axis focuses on these issues, and aims at developing new concepts and tools to master some already existing large scale systems, as telecommunication networks, cloud infrastructures, web-services, etc. (see the Application Domains section).

The objective being to address large systems, our work will be driven by two considerations: how to take advantage of the modularity of systems, and how to best approximate/abstract too complex systems by more tractable ones. We mention below several objectives that we consider as more important.

• Approximate management methods: this consists in exploring the relevance of ideas developed for large scale stochastic systems, as turbo-algorithms for example, in the setting of modular dynamic systems.

• Self-modeling: consists in managing large scale systems that are known by their building rules, but which specific managed instance is only discovered at runtime, and on the fly. The model of the managed system is built on-line, following the needs of the management algorithms.

• Distributed control: explores issues related to asynchronous communications between local controllers, and abstraction techniques to address large systems.

• Test and enforcement: considers coverage issues for the test of large systems, and the test and enforcement of properties for timed models, or for systems handling data.