Network Infrastructures for Highly Distributed Cloud-Computing

Software-Defined-Network (SDN) is emerging as a solid opportunity for the Network Service Providers (NSP) to reduce costs while at the same time providing better and/or new services. The possibility to flexibly manage and configure highly-available and scalable network services through data model abstractions and easy-to-consume APIs is attractive and the adoption of such technologies is gaining momentum. At the same time, NSPs are planning to innovate their infrastructures through a process of network softwarisation and programmability. The SDN paradigm aims at improving the design, configuration, maintenance and service provisioning agility of the network through a centralised software control. This can be easily achievable in local area networks, typical of data-centers, where the benefits of having programmable access to the entire network is not restricted by latency between the network devices and the SDN controller which is reasonably located in the same LAN of the data path nodes. In Wide Area Networks (WAN), instead, a centralised control plane limits the speed of responsiveness in reaction to time-constrained network events due to unavoidable latencies caused by physical distances. Moreover, an end-to-end control shall involve the participation of multiple, domain-specific, controllers: access devices, data-center fabrics and backbone networks have very different characteristics and their control-plane could hardly coexist in a single centralised entity, unless of very complex solutions which inevitably lead to software bugs, inconsistent states and performance issues. In recent years, the idea to exploit SDN for WAN infrastructures to connect multiple sites together has spread in both the scientific community and the industry. The former has produced interesting results in terms of framework proposals, complexity and performance analysis for network resource allocation schemes and open-source proof of concept prototypes targeting SDN architectures spanning multiple technological and administrative domains. On the other hand, much of the work still remains confined to the academy mainly because based on pure Openflow prototype implementation, networks emulated on a single general-purpose machine or on simulations proving algorithms effectiveness. The industry has made SDN a reality via closed-source systems, running on single administrative domain networks with little if no diversification of access and backbone devices. In this dissertation we present our contributions to the design and the implementation of SDN architectures for the control plane of WAN infrastructures. In particular, we studied and prototyped two SDN platforms to build a programmable, intent-based, control-plane suitable for the today highly distributed cloud infrastructures. Our main contributions are: (i) an holistic and architectural description of a distributed SDN control-plane for end-end QoS provisioning; we compare the legacy IntServ RSVP protocol with a novel approach for prioritising application-sensitive flows via centralised vantage points. It is based on a peer-to-peer architecture and could so be suitable for the inter-authoritative domains scenario. (ii) An open-source platform based on a two-layer hierarchy of network controllers designed to provision end-to-end connectivity in real networks composed by heterogeneous devices and links within a single authoritative domain. This platform has been integrated in CORD, an open-source project whose goal is to bring data-center economics and cloud agility to the NSP central office infrastructures, combining NFV (Network Function Virtualization), SDN and the elasticity of commodity clouds. Our platform enables the provisioning of connectivity services between multiple CORD sites, up to the customer premises. Thus our system and software contributions in SDN has been combined with a NFV infrastructure for network service automation and orchestration.