The emergence of new digital services with applications such as sub-second data analytics, high-speed data feed delivery or high-performance computing is imposing stringent performance requirements on the networking infrastructures that support these services. This project is looking at how to take advantage of recent advances in dataplane programmability, with initiatives like P4, to offload some application computational tasks to the network, making in essence the network part of the application's workload compute cycle. The project is particularly interested in the convergence in terms of resource management between the network and the compute domains of a datacenter.

This project aimed at closing the loop in software-based networking environments between high-level business objectives that dictate what is expected for the applications and services running on top of the network infrastructure and low-level commands driving the configuration of network resources. In this project, we were particularly interested in understanding how to design better abstraction(s) for network and management functions so as to reduce the complexity of today's network management practices.

This project focused on trade-offs for large-scale and efficient network monitoring. A key question that our work aimed to answer is: how to achieve accurate and timely network monitoring report collection while performing scalable and low-overhead statistics extraction? We specifically focused on two scenarios currently receiving a strong interest in the community, namely software-defined networks and software packet processing. Our main objective was to investigate how to take advantage of adaptive monitoring mechanisms to satisfy hardware constraints, while reducing potential accuracy degradation.

This project investigated from an architectural and functional point of view, the applicability of SDN-based approaches to network resource management. A key contribution of our research was the development of a novel SDN-based management and control framework to support both static and adaptive resource management applications, as well centralised/decentralised deployment. Its main feature lies in its modular design enabling a clear separation of concerns between control and management. In addition, to support communication in decentralised settings, we also developed a new signalling approach building on top well defined interfaces and re-usable functions. In parallel, we also investigated how to use the OpenFlow technology for realising flexible traffic management.

This research focused on the development of scalable and lightweight cache/content management approaches for ISP-operated content delivery services. A key contribution that resulted from our research efforts is a hybrid cache management approach for virtualised ISP networks, combining periodical proactive cache reconfiguration with distributed reactive cache replacement. We also developed a distributed approach based on the parallelisation of the decision-making process and the use of network partitioning to cluster the distributed decision-making points.

This concerns the work carried out during my Ph.D. The objective of my thesis was to address, through realistic and specific in-network self-management applications, some of the main engineering challenges raised by the design and implementation of self-managed decentralised and adaptive networks. In this context, I proposed a novel framework for decentralised network resource management and demonstrated its functionality on three realistic use case scenarios for dynamic traffic engineering, energy efficiency and cache management in Internet Service Providers.