Self-adaptive serverless edge computing for edge intelligence

Abstract

The Edge Intelligence application paradigm is an umbrella term for AI applications that cooperate across the edge-cloud continuum. Stringent requirements around low latency, privacy, personalization, and complex interactions are characteristic of them. The edge-cloud continuum is heterogeneous in terms of resources, highly dynamic as users constantly move, and reducing energy consumption is important in edge and cloud deployments. Autonomous application orchestration is crucial for the wide spread adoption of Edge Intelligence. However, current platform paradigms do not fully support Edge Intelligence. Therefore, this thesis presents four contributions that investigate platforms for Edge Intelligence. We conduct a use case study to derive platform requirements and a vision for an Edge Intelligence as a Service platform. Based on these requirements and the necessary support for autonomous application management, we further investigate Serverless Edge Computing. Serverless Edge Computing is a platform paradigm that promises to deliver a fully automated application lifecycle management by abstracting away the heterogeneous edge-cloud continuum and allowing users to simply upload their code as small self-contained functions. We review the current state of Serverless Edge Computing offerings that range from open source to commercial offerings and research. The analysis of Edge Intelligence and the literature review discern that open challenges around mobility and energy-awareness remain. Therefore, we present two novel approaches, which extend Serverless Edge Computing platforms and tackle the challenge of mobile users and energy consumption. The mobility-aware approach proposes a decentralized platform that relies on thresholds to manage functions in response to mobile users (e.g., in Smart Cities). The second approach builds a container scheduler that relies on a graph neural network to estimate the power consumption of devices and reduce overall energy consumption. Based on these two orchestration strategies, we discovered the challenge of properly parameterizing them and propose a self-adaptive platform to tackle this issue. We present a testing framework and a Serverless Edgesimulator that we unify to build a self-adaptive platform prototype. This prototype is fed real-world monitoring data into a simulation that provides insights and on-the-fly optimization of orchestration strategy parameters. In addition, we propose a method to bootstrap prediction-based orchestration strategies by generating synthetic datasets. The former self-adaptive approach targets threshold-based strategies during runtime, while the latter is a prediction model-based approach. Both approaches facilitate a self-adaptive Serverless Edge Computing platform that supports Edge Intelligence applications.