Diploma and Master Theses (authored and supervised):
"Serverless Edge Analytics";
Supervisor: S. Dustdar;
Institute of Information Systems Engineering, Distributed Systems Group,
final examination: 2020-01-20.
Due to increasing amount of information that gets available at client devices, the centralized layout of the internet begins to be outdated. By now, we have tremendous amounts of data that is only available at end-devices from where data can, due to connection limitations and missing capacity at cloud data centers, not be sent to central
locations for analysis. Machine learning training requires enormous amounts of input data to find meaningful patterns and would benefit by the availability of the data that can currently not be sent to centralized clouds. To reduce the resource overhead of cloud services, the Function as a Service (FaaS) paradigm was recently proposed as a possible improvement for cloud data centers. Novel frameworks bring serverless functions on edge computing devices too and could dramatically improve analytical services.
The aim of this work is the analysis of serverless edge computing under the context of applying machine learning at those edge devices in order to use data that can not easily be sent to the cloud. In order to do this, beside analysis of existing frameworks and
literature, it was a task to investigate on the feasibility of implementing an extension for AWS Greengrass that allows dynamic forwarding of serverless function triggers to edge or cloud devices based on the utilization each AWS Greengrass edge node.
The findings of this thesis reveal that monitoring the processing time and utilization of AWS Greengrass nodes is not meaningful with today´s solutions. As it was necessary for this thesis to research various serverless function frameworks and their monitoring approaches, an extensive study about their capabilities is given. Beside finding that
monitoring of AWS Greengrass delivers vacuous results, it was found that forwarding triggers of serverless functions from AWS Greengrass to the cloud-based AWS Lambda results in extremely slow response times. As it was necessary to test machine learning workload on AWS Lambda, various insights and limitations were found during this process. In general it was found that serverless edge computing, edge computing hardware and AWS IoT are currently limited in many ways.
Non the less, while serverless computing has proven to be limited for low-latency applications, various possible further research fields, like Federated Learning, have been identified as promising ways to tackle the challenges of analyzing data within serverless
edge computing and edge computing in the future.
Created from the Publication Database of the Vienna University of Technology.