Treating machine learning workflows as microservices
keywords ALGORITHMS, EXPERIMENTAL DESIGN, TESTBED
Reference persons CLAUDIO ETTORE CASETTI, CARLA FABIANA CHIASSERINI, PAOLO GIACCONE
External reference persons Prof. Dr. Holger Karl, Hasso Plattner Institute, Potzdam University (Berlin), Germany
Research Groups Telecommunication Networks Group
Thesis type EXPERIMENTAL AND MODELING
Description Machine-learning workflows often comprise individual components, acting separately on their input data. In that sense, they are similar to conventional microservices (e.g., three-tier web applications). In other senses, they are quite different in their data flows, computational patterns, etc.
It is the goal of this thesis to investigate if and to what degree ML workflows can be managed (at runtime) similar to microservice chains. An ideal outcome would be to identify relevant examples, cast them as MS chains, and show that they can be orchestrated by an existing orchestrator like Open-Source MANO (OSM). A possible approach could be to automatically generated description files used by such an orchestrator.
Further, the thesis work shall extend orchestration logic to better deal with ML workflows and showing performance benefits.
See also https://hpi.de/karl/teaching/bachelor-and-master-theses.html
Required skills Familiar with software-engineering concepts like microservices
Familiar with machine-learning concepts
Good practical development skills
Deadline 02/08/2023 PROPONI LA TUA CANDIDATURA