Section: New Results

Specific studies: network and service diagnosis

Participants : Eric Fabre, Carole Hounkonnou.

This work represents part of our activities within the research group “High Manageability,” supported by the common lab of Alcatel-Lucent Bell Labs (ALBLF) and Inria. It is also supported by the UniverSelf EU integrated project, and conducted in cooperation with Orange Labs.

The objective is to develop a framework for the joint diagnosis of networks and of the supported services. We are aiming at a model-based approach, in order to tailor the methods to a given network instance and to follow its evolution. We also aim at active diagnosis methods, that collect and reason on alarms provided by the network, but that can also trigger tests or the collection of new observations in order to refine a current diagnosis.

Since 2011, an important effort was dedicated to a key and difficult part of this approach: the definition of a methodology for self-modeling. This consists in automatically building a model of the monitored system, by instantiating generic network elements. There are several difficulties to address:

• The model must capture several layers, from the physical architecture up to the service architecture and its protocols. As a case-study, we have chosen VoIP services on an IMS network, deployed over a wired IP network.

• The model should be hierarchical, to allow for multiscale reasoning, and to reflect the intrinsic hierarchical nature of the managed network.

• The model should be generic, i.e. obtained by assembling component instances coming from a reduced set of patterns, just like a text is obtained by assembling words.

• The model should be adaptive, to capture the evolving part of the network (e.g. introduction of new elements) but also its intrinsically dynamic nature (e.g. opened/closed connections).

• The model should display the hierarchical dependency of resources, specifically the fact that lower-level resources are assembled to provide a support to a higher level resource or functionality.

• The model should allow progressive discovery and refinement: for a matter of size, it is not possible to first build a model of the complete network and then monitor it; one must adopt an approach where the model is build on-line, and where the construction is guided by the progress of the diagnosis algorithms.

Elements of methodology achieving these goals were proposed in 2011, and further refined in 2012. Besides, we have also worked on the definition of generic Bayesian networks, that could translate into mathematical terms the dependency relations between network resources, in order to reason about them for failure diagnosis. A methodology was then designed to reason on such models. The idea is that one should first consider a subset of network resources (at a given granularity), in order to localize the origin of a given malfunction. The natural start point is the graph of all resources involved in the delivery of the malfunctioning service. As the fault localization is statistical, the model is then progressively expanded to capture more network elements and thus more observations, and thus refine the diagnosis. This model expansion is performed by introducing first the most informative network elements, using information theory criteria. The result is a fault localization algorithm that explores only part of the network, and builds at runtime the necessary part of the model it should use to explain a malfunction [28] . The current efforts aim at extending these ideas to allow for the refinement of the model of some component (multiresolution reasoning).