In a context where computing and communications infrastructures consume over 4% of the world's energy, the aim is to design approaches that offer a better compromise between Quality of Service (QoS) and the overall energy efficiency of today's infrastructures (Cloud/Fog/Edge).

Allocating dynamic services, managing their lifecycle, monitoring the Quality of Service (QoS) delivered to users and the energy consumption generated have become recurring and very important issues since the advent of on-demand services. In today's innovative architectures, all these issues are all the more topical because their use, possibilities and complexity have increased. What's more, we're moving towards architectures that are increasingly distributed, flexible and reconfigurable (both in terms of applications and topology). This is the case, for example, with the Fog computing paradigm, which, by its very definition, accepts the fact of hosting a number of devices deployed at the four corners of the architecture (and based on data generated by the lower IoT layer). From an energy point of view, this represents a major departure from the architectures studied a decade ago (Cloud Computing, for example), as resources must now be powered independently and no longer depend on a single data or computing center (data-center). As a result, energy efficiency issues are changing due to this first architectural reason. What's more, a change of vision in the way these systems are powered will have to be closely studied. It's no longer a question of trying to minimize the consumption of a large number of wired resources, but of making the best use (the term "optimization" is often misused to describe this) of the amount of energy available for each (or a set) of devices.

Our work, which is being carried out as part of the ANR DELIGHT project, aims to finely assess and reduce the energy consumption of federated learning using various levers (gradient compression, data summarization, speed-scaling, etc.). Given the heterogeneity of the data, another objective will be to study the negotiation and coalition-building process between nodes, in order to understand the extent to which a node has an interest in collaborating with others. The techniques developed will be validated empirically on computer vision and NLP tasks using the Flower toolkit.