The Internet of Things (IoT) consists of low-cost, highly-specialized, resource constrained devices deployed in technologically limited environments to collect data, perform computations and actuate over the results. These deployments co-locate computers and storage systems at the "edge" of the network or as close to the environment they actuate over. Together these "edge deployments" localize communication, computation, and storage for security, with increased efficiencies (e.g. lower latency response), and reliability. To keep these deployments low-cost and sustainable, edge clusters increasingly rely on commodity hardware and shared power, network infrastructures.
Increasingly, advancements in applications like Machine Learning (ML), Artificial Intelligence (AI), and device hardware, have allowed these systems to go beyond just sensing. Applications that used to require offloading heavy computations to compute, network-rich environments like the cloud, can now sufficiently operate locally at the edge. However, unlike the cloud, with its abundant and stable compute, network and power resources, resource-constrained edge environments exhibit dynamic network and compute conditions, leading to fluctuating workload performance. This body of work showcases the impact of such computations, communications under realistic edge deployment conditions. Specially, by modeling and designing edge deployments with a power first approach, the dissertation shows how to account for these novel power-compute interactions on real edge clusters, and finally how to both efficiently and effectively utilize real edge clusters under these dynamic conditions for ML/AI applications.