Section: Scientific Foundations

Management of Quantitative Behavior

Participants : Sandie Balaguer, Benedikt Bollig, Thomas Chatain, Paul Gastin, Stefan Haar, Serge Haddad, Benjamin Monmege.


Besides the logical functionalities of programs, the quantitative aspects of component behavior and interaction play an increasingly important role.

  • Real-time properties cannot be neglected even if time is not an explicit functional issue, since transmission delays, parallelism, etc, can lead to time-outs striking, and thus change even the logical course of processes. Again, this phenomenon arises in telecommunications and web services, but also in transport systems.

  • In the same contexts, probabilities need to be taken into account, for many diverse reasons such as unpredictable functionalities, or because the outcome of a computation may be governed by race conditions.

  • Last but not least, constraints on cost cannot be ignored, be it in terms of money or any other limited resource, such as memory space or available CPU time.

Traditional mainframe systems were proprietary and (essentially) localized; therefore, impact of delays, unforeseen failures, etc. could be considered under the control of the system manager. It was therefore natural, in verification and control of systems, to focus on functional behavior entirely.

With the increase in size of computing system and the growing degree of compositionality and distribution, quantitative factors enter the stage:

  • calling remote services and transmitting data over the web creates delays;

  • remote or non-proprietary components are not “deterministic”, in the sense that their behavior is uncertain.

Time and probability are thus parameters that management of distributed systems must be able to handle; along with both, the cost of operations is often subject to restrictions, or its minimization is at least desired. The mathematical treatment of these features in distributed systems is an important challenge, which MExICo is addressing; the following describes our activities concerning probabilistic and timed systems. Note that cost optimization is not a current activity but enters the picture in several intended activities.

Probabilistic distributed Systems

Participants : Stefan Haar, Serge Haddad.

Non-sequential probabilistic processes.

Practical fault diagnosis requires to select explanations of maximal likelihood; this leads therefore to the question what the probability of a given partially ordered execution is. In Benveniste et al. [69] , [61] , we presented a model of stochastic processes, whose trajectories are partially ordered, based on local branching in Petri net unfoldings; an alternative and complementary model based on Markov fields is developed in [90] , which takes a different view on the semantics and overcomes the first model's restrictions on applicability.

Both approaches abstract away from real time progress and randomize choices in logical time. On the other hand, the relative speed - and thus, indirectly, the real-time behavior of the system's local processes - are crucial factors determining the outcome of probabilistic choices, even if non-determinism is absent from the system.

Recently, we started a new line of research with Anne Bouillard, Sidney Rosario, and Albert Benveniste in the DistribCom team at INRIA Rennes, studying the likelihood of occurrence of non-sequential runs under random durations in a stochastic Petri net setting.

Once the properties of the probability measures thus obtained are understood, it will be interesting to relate them with the two above models in logical time, and understand their differences. Another mid-term goal, in parallel, is the transfer to diagnosis with possible cooperation with René Boel's group in Ghent/Belgium.

Distributed Markov Decision Processes

Distributed systems featuring non-deterministic and probabilistic aspects are usually hard to analyze and, more specifically, to optimize. Furthermore, high complexity theoretical lower bounds have been established for models like partially observed Markovian decision processes and distributed partially observed Markovian decision processes. We believe that these negative results are consequences of the choice of the models rather than the intrinsic complexity of problems to be solved. Thus we plan to introduce new models in which the associated optimization problems can be solved in a more efficient way. More precisely, we start by studying connection protocols weighted by costs and we look for online and offline strategies for optimizing the mean cost to achieve the protocol. We cooperate on this subject with Eric Fabre in the DistribCom team at INRIA Rennes, in the context of the DISC project.

Large scale probabilistic systems

Addressing large-scale probabilistic systems requires to face state explosion, due to both the discrete part and the probabilistic part of the model. In order to deal with such systems, different approaches have been proposed:

  • Restricting the synchronization between the components as in queuing networks allows to express the steady-state distribution of the model by an analytical formula called a product-form  [68] .

  • Some methods that tackle with the combinatory explosion for discrete-event systems can be generalized to stochastic systems using an appropriate theory. For instance symmetry based methods have been generalized to stochastic systems with the help of aggregation theory  [80] .

  • At last simulation, which works as soon as a stochastic operational semantic is defined, has been adapted to perform statistical model checking. Roughly speaking, it consists to produce a confidence interval for the probability that a random path fulfills a formula of some temporal logic  [108] .

We want to contribute to these three axes: (1) we are looking for product-forms related to systems where synchronization are more involved (like in Petri nets), (2) we want to adapt methods for discrete-event systems that require some theoretical developments in the stochastic framework and, (3) we plane to address some important limitations of statistical model checking like the expressiveness of the associated logic and the handling of rare events.

Real time distributed systems

Nowadays, software systems largely depend on complex timing constraints and usually consist of many interacting local components. Among them, railway crossings, traffic control units, mobile phones, computer servers, and many more safety-critical systems are subject to particular quality standards. It is therefore becoming increasingly important to look at networks of timed systems, which allow real-time systems to operate in a distributed manner.

Timed automata are a well-studied formalism to describe reactive systems that come with timing constraints. For modeling distributed real-time systems, networks of timed automata have been considered, where the local clocks of the processes usually evolve at the same rate [98] [76] . It is, however, not always adequate to assume that distributed components of a system obey a global time. Actually, there is generally no reason to assume that different timed systems in the networks refer to the same time or evolve at the same rate. Any component is rather determined by local influences such as temperature and workload.

Distributed timed systems with independently evolving clocks

Participants : Benedikt Bollig, Paul Gastin.

A first step towards formal models of distributed timed systems with independently evolving clocks was done in [62] . As the precise evolution of local clock rates is often too complex or even unknown, the authors study different semantics of a given system: The existential semantics exhibits all those behaviors that are possible under some time evolution. The universal semantics captures only those behaviors that are possible under all time evolutions. While emptiness and universality of the universal semantics are in general undecidable, the existential semantics is always regular and offers a way to check a given system against safety properties. A decidable under-approximation of the universal semantics, called reactive semantics, is introduced to check a system for liveness properties. It assumes the existence of a global controller that allows the system to react upon local time evolutions. A short term goal is to investigate a distributed reactive semantics where controllers are located at processes and only have local views of the system behaviors.

Several questions, however, have not yet been tackled in this previous work or remain open. In particular, we plan to exploit the power of synchronization via local clocks and to investigate the synthesis problem: For which (global) specifications 𝒮 can we generate a distributed timed system with independently evolving clocks 𝒜 (over some given system architecture) such that both the reactive and the existential semantics of 𝒜 are precisely (the semantics of) 𝒮? In this context, it will be favorable to have partial-order based specification languages and a partial-order semantics for distributed timed systems. The fact that clocks are not shared may allow us to apply partial-order–reduction techniques.

If, on the other hand, a system is already given and complemented with a specification, then one is usually interested in controlling the system in such a way that it meets its specification. The interaction between the actual system and the environment (i.e., the local time evolution) can now be understood as a 2-player game: the system's goal is to guarantee a behavior that conforms with the specification, while the environment aims at violating the specification. Thus, building a controller of a system actually amounts to computing winning strategies in imperfect-information games with infinitely many states where the unknown or unpredictable evolution of time reflects an imperfect information of the environment. Only few efforts have been made to tackle those kinds of games. One reason might be that, in the presence of imperfect information and infinitely many states, one is quickly confronted with undecidability of basic decision problems.

Implementation of Real-Time Concurrent Systems

Participants : Sandie Balaguer, Thomas Chatain, Stefan Haar, Serge Haddad.

This is one of the tasks of the ANR ImpRo.

The objective is to provide formal guarantees on the implementation of real-time distributed systems, despite the semantic differences between the model and the code. We consider two kinds of timed models: time Petri nets  [99] and networks of timed automata  [63] .

Time Petri Nets allow the designer to explicit concurrent parts of the system, but without having decided yet to localize the different actions on the different components. In that sense, TPNs are more abstract than networks of timed automata, which can be seen as possible (ideal) distributed implementations. This raises the question of semantical comparison of these two models in the light of preserving the maximum of concurrency.

In order to implement our models on distributed architectures, we need a way to evaluate how much the implementation preserves the concurrency that is described in the model. For this we must be able to identify concurrency in the behavior of the models. This is done by equipping the models with a concurrent semantics (unfoldings), allowing us to consider the behaviors as partial orders.

For instance, we would like to be able to transform a time Petri net into a network of timed automata, which is closer to the implementation since the processes are well identified. But we require that this transformation preserves concurrency. Yet the first works about formal comparisons of the expressivity of these models [78] , [75] , [77] , [79] , [106] do not consider preservation of concurrency.

In contrast, we aim at formalizing and automating translations that preserve both the timed semantics and the concurrent semantics. This effort is crucial for extending concurrency-oriented methods for logical time, in particular for exploiting partial order properties. In fact, validation and management - in a broad sense - of distributed systems is not realistic in general without understanding and control of their real-time dependent features; the link between real-time and logical-time behaviors is thus crucial for many aspects of MExICo's work.

Weighted Automata and Weighted Logics

Participants : Benedikt Bollig, Paul Gastin, Benjamin Monmege, Marc Zeitoun.

Time and probability are only two facets of quantitative phenomena. A generic concept of adding weights to qualitative systems is provided by the theory of weighted automata [60] . They allow one to treat probabilistic or also reward models in a unified framework. Unlike finite automata, which are based on the Boolean semiring, weighted automata build on more general structures such as the natural or real numbers (equipped with the usual addition and multiplication) or the probabilistic semiring. Hence, a weighted automaton associates with any possible behavior a weight beyond the usual Boolean classification of “acceptance” or “non-acceptance”. Automata with weights have produced a well-established theory and come, e.g., with a characterization in terms of rational expressions, which generalizes the famous theorem of Kleene in the unweighted setting. Equipped with a solid theoretical basis, weighted automata finally found their way into numerous application areas such as natural language processing and speech recognition, or digital image compression.

What is still missing in the theory of weighted automata are satisfactory connections with verification-related issues such as (temporal) logic and bisimulation that could lead to a general approach to corresponding satisfiability and model-checking problems. A first step towards a more satisfactory theory of weighted systems was done in [11] . That paper, however does not give final solutions to all the aforementioned problems. It identifies directions for future research that we will be tackling.