Distributed computing at the smart device edge

Overview & Synopsis

With the increasing capacity of smart devices and an incompatibility of privacy/latency sensitive applications with cloud computing, edge computing has emerged as the best deployment solution for such workloads. In this context, this thesis studies the acceleration of heavy workloads in smart device edge networks, by providing observability through filtering of telemetry data, and a pipe-lining framework for throughput acceleration of heavy workloads. This thesis proposes a hybrid approach between cloud-out and edge-in methodologies, which leverages the multiplicity of edge compute by locally offloading computation. The thesis initially focuses on network state observability and fault diagnosis at the edge. A data-driven method to extract intelligible selections of operational features from high-dimensional network telemetry data is introduced, combining data-driven metrics and semantic information contained in meta-data, to produce selections of features which best represent an underlying event. The thesis illustrates the benefits of such a complementary meta-data analysis for data-driven fault diagnosis, highlighting the robustness of the studied approach against variations in the input feature set. With an improved understanding of the state of the edge, this thesis then studies heavy workload distribution in such environments, through the example of DNN partitioning, which consists of distributing inference workloads over several available edge devices, taking into account the edge network properties and the DNN structure, with the objective of maximizing the inference throughput. The thesis describes a process to identify partitionings which maximize the DNN inference throughput while keeping computation on the edge. The analysis of this method has lead to a set of conditions on the link between the edge network and application properties to anticipate the achieved performance and complexity, and effectively size an edge network environment. Finally, the thesis describes a dynamic partitioning framework to improve the system performance and robustness, which leverages the observability of the network to adapt to heterogeneous and dynamic edge networks.