Accelerating orchestration with in-network offloading

The demand for low-latency Internet applications has pushed functionality that was originally placed in commodity hardware into the network. Either in the form of binaries for the programmable data plane or virtualised network functions, services are implemented within the network fabric with the aim of improving their performance and placing them close to the end user. Training of machine learning algorithms, aggregation of networking traffic, virtualised radio access components, are just some of the functions that have been deployed within the network. Therefore, as the network fabric becomes the accelerator for various applications, it is imperative that the orchestration of their components is also adapted to the constraints and capabilities of the deployment environment. This work identifies performance limitations of in-network compute use cases for both cloud and edge environments and makes suitable adaptations. Within cloud infrastructure, this thesis proposes a platform that relies on programmable switches to accelerate the performance of data replication. It then proceeds to discuss design adaptations of an orchestrator that will allow in-network data offloading and enable accelerated service deployment. At the edge, the topic of inefficient orchestration of virtualised network functions is explored, mainly with respect to energy usage and resource contention. An orchestrator is adapted to schedule requests by taking into account edge constraints in order to minimise resource contention and accelerate service processing times. With data transfers consuming valuable resources at the edge, an efficient data representation mechanism is implemented to provide statistical insight on the provenance of data at the edge and enable smart query allocation to nodes with relevant data. Taking into account the previous state of the art, the proposed data plane replication method appears to be the most computationally efficient and scalable in-network data replication platform available, with significant improvements in throughput and up to an order of magnitude decrease in latency. The orchestrator of virtual network functions at the edge was shown to reduce event rejections, total processing time, and energy consumption imbalances over the default orchestrator, thus proving more efficient use of the infrastructure. Lastly, computational cost at the edge was further reduced with the use of the proposed query allocation mechanism which minimised redundant engagement of nodes