Software networks: orchestration, resilience, and programmability concerns

Network softwarization constitutes a paradigm shift that disrupts computer networking. It has emerged as a fundamental approach to enable flexibility, automation, and network as a service. This dissertation is a summary of part of the research activities I have been conducting and developing over the past 5-6 years on software networks. It is the result of many collaborations, on projects and with juniors, Ph.D. students, postdoctoral researchers, and research engineers. The first part of this dissertation addresses the question of the logical separation of the control plane from the data plane. This approach transforms the way network elements are conceived and piloted as it aims at enabling software-based network control. It brings new challenges, especially along the questions of the centralization/distribution of the control logic, and the level of abstraction of the network. We addressed resilience concerns at the control layer and data plane programmability concerns. We motivated the need to have a resilient and scalable control plane and then proposed a distributed Software-Defined Networking (SDN) controllers architecture that enables to geo-partition the network, and thus distribute control load, while ensuring end-to-end connectivity and supporting failures. In this perspective, we studied the problem of the controller placement and tackled the failure of an SDN controller. Finally, we addressed network programmability through two advanced network functions that enable in-switch primitive functions: programmable packet generation and Distributed Denial of Service attack detection. The second part of this dissertation presents a number of contributions on orchestration in Network Functions Virtualization (NFV) systems. One of the key challenges to realize flexible networks and network services is to efficiently place and chain Virtual Network Functions (VNFs) and service chains. We explored several complementary research facets of this question, addressing different optimization objectives: reduce the deployment cost, consider a maximum latency, reach an availability target, and find a trade-off between the deployment cost and the number of network reconfigurations when the demands change over time. We proposed for each of them a mathematical optimization formulation, providing bounds, and practical algorithms that aim at being close to optimal solutions in an operation-compliant time.

1 Docteur 2020