Supervision : Lúcia Maria DE ASSUMPÇÃO DRUMMOND
Co-supervision : SENS Pierre

Multi-FedLS : A Scheduler of Federated Learning Applications in a Multi-Cloud Environment

Federated Learning (FL) is a new area of distributed Machine Learning (ML) where learning ensures data privacy. Each client has access only to its own local and private dataset. This approach is attractive in various domains of knowledge because it allows different institutions to collaborate without sharing their confidential data. As the amount of data required for training has grown significantly in recent years, most institutions cannot afford physical data centers to store and manipulate all their data. A viable option is to utilize cloud storage services offered by providers with different data privacy and availability guarantees. The user is responsible for choosing the regions where their data is stored and controlling access to it.Additionally, cloud providers offer various services to execute an application. They provide users with the ability to create Virtual Machines (VMs) with different configurations, where users have full control over them. This type of service is known as Infrastructure-as-a-Service (IaaS). Thus, a multi-cloud environment is conducive to the collaboration of different institutions in creating a Machine Learning model through Federated Learning.In this thesis, we propose extit{Multi-FedLS}, a robust framework designed to execute FL applications in a multi-cloud environment. The framework considers the current location of each client's datasets, communication delay, and cost of utilization in the clouds, focusing on cost and runtime reduction. Moreover, Multi-FedLS utilizes cheaper instances whenever possible to reduce costs, even though they may be revoked at any time by the cloud provider. Thus, to ensure the successful execution of FL applications, the framework employs fault-tolerance techniques such as checkpoints and work migration to resume training on another VM after a revocation. Multi-FedLS comprises four modules: Pre-Scheduling, Initial Mapping, Fault Tolerance, and Dynamic Scheduler. The obtained results demonstrate the feasibility of executing applications in multi-cloud environments using low-cost VMs, employing mathematical formulation, fault-tolerance techniques, and simple heuristics for selecting new VMs. The framework achieved a cost reduction of 56.92% compared to application runtime using more expensive VMs, with only a 5.44% increase in runtime on commercial cloud providers.

Phd defence : 11/29/2023

Jury members :

Pierre Sens,
Luciana Arantes
Maria Clicia Castro (UERJ)
Cristina Boeres (UERJ)
Aline Paes (UFF)
Christophe Cerin (Paris 13)
Guillaume Pierre (Univ. Rennes/INRIA).

2023 Publications

• 2023
  • R. Correia Brum : “Multi-FedLS : un ordonnanceur d’applications d’apprentissage fédéré dans un environnement Multi-Cloud”, thesis, phd defence 11/29/2023, supervision Maria de assumpção drummond, Lúcia, co-supervision : Sens, Pierre (2023)