Toward auto-configuration in software networks

Software networks have the potential to take the network infrastructure to a more advanced level, a level that can make the configuration autonomic. This ability can overcome the rapidly growing complexity of current networks, and allow management entities to enable an effective behavior in the network for overall performance improvement without any human intervention. Configuration parameters can be automatically selected for network resources to cope with various situations that networks encounter like errors and performance degradation. Unfortunately, some challenges need to be tackled to reach that advanced level of networks. Currently, the configuration is still often generated manually by domain experts in huge semi-structured files written in XML, JSON, and YAML. This is a complex, error-prone, and tedious task to do by humans. Also, there is no formal strategy except experience and best practices of domain experts to design the configuration files. Different experts may choose different configurations for the same performance goal. This situation makes it harder to extract features from the configuration files and learn models that could generate or recommend automatic configuration. Moreover, there is still no consensus on a common configuration data model in software networks, which resulted in heterogeneous solutions, such as TOSCA, YANG, Hot, etc. that make the end-to-end network management difficult. In this thesis, we present our contributions that tackle the aforementioned challenges related to automating the configuration in software networks. To tack the problem of heterogeneity between the configuration files we propose a semantic framework based ontologies that can federate common elements from different configuration files. And, to tackle the problem of generating automatically the configuration, we propose two contributions, one contribution that considers deep neural networks to learn from configuration files models for recommending the configuration and another contribution based on a model-driven approach to assist automatically the design of the configuration files.