Decomposing monolitic computations for execution on the Edge for fun and profit

• Main Author: Gomes, Diogo João de Paiva
• Publication Date: 2023
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10362/160380
• topic: Cloud; Edge; IoT; Computation; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Cloud computing is emerging in the IT industry and is being established as the standard way of providing high quality applications and services, being able to scale when neces- sary and maintaining low latency. This high availability allows new users and devices to connect which may possibly cause some saturation in the network. The edge computing paradigm has recently emerged to address raising concerns on the ability of cloud computing infrastructures (and data links connecting end user devices to such infrastructures) to receive, process, and reply to operations generated by edge devices (e.g., sensors or other more sophisticated devices) in an expedited way. One way to mitigate this challenge, put forward by the edge computing paradigm, is to decompose the computations executed by applications in smaller components that can then be pushed into the devices (either on the cloud or edge) that are in close vicinity to the location of data generation and consumption. This allows to remove pressure from network links (by avoiding to move data) and can, in particular cases, speedup the production of results to be consumed in the edge. However, achieving this intuitive vision is cumbered by several challenges, in particular: 1. How to decompose the logic of a (large) computation into smaller units; 2. Understanding the requirements for the correct processing of such smaller units; 3. Define mechanisms to compose the results of the smaller computations; 4. Effectively pushing the computations to the locations where they will be conducted and ensuring their execution despite failures. In this thesis we will start to pave the way towards such a vision by studying how to decompose computations, classifying relevant properties of a subset of classes of these computations, and exploring the design space of the runtime support to execute smaller units in an integrated way across the cloud and edge. We will focus on settings with multiple cloud datacenters potentially with a few edge locations.