Date of Award
Doctor of Philosophy (PhD)
Recent years have witnessed an explosive growth of Internet of Things (IoT) devices, which collect or generate huge amounts of data. Given diverse device capabilities and application requirements, data processing takes place across a range of settings, from on-device to a nearby edge server/cloud and remote cloud. Consequently, edge-cloud coordination has been studied extensively from the perspectives of job placement, scheduling and joint optimization. Typical approaches focus on performance optimization for individual applications. This often requires domain knowledge of the applications, but also leads to application-specific solutions. Application development and deployment over diverse scenarios thus incur repetitive manual efforts. There are two overarching challenges to provide system-level support for application development at the edge. First, there is inherent heterogeneity at the device hardware level. The execution settings may range from a small cluster as an edge cloud to on-device inference on embedded devices, differing in hardware capability and programming environments. Further, application performance requirements vary significantly, making it even more difficult to map different applications to already heterogeneous hardware. Second, there are trends towards incorporating edge and cloud and multi-modal data. Together, these add further dimensions to the design space and increase the complexity significantly. In this thesis, we propose a novel framework to simplify application development and deployment over a continuum of edge to cloud. Our framework provides key connections between different dimensions of design considerations, corresponding to the application abstraction, data abstraction and resource management abstraction respectively. First, our framework masks hardware heterogeneity with abstract resource types through containerization, and abstracts away the application processing pipelines into generic flow graphs. Further, our framework further supports a notion of degradable computing for application scenarios at the edge that are driven by multimodal sensory input. Next, as video analytics is the killer app of edge computing, we include a generic data management service between video query systems and a video store to organize video data at the edge. We propose a video data unit abstraction based on a notion of distance between objects in the video, quantifying the semantic similarity among video data. Last, considering concurrent application execution, our framework supports multi-application offloading with device-centric control, with a userspace scheduler service that wraps over the operating system scheduler.
Hu, Bo, "System Abstractions for Scalable Application Development at the Edge" (2022). Yale Graduate School of Arts and Sciences Dissertations. 609.