AtlasFL: A Federated Learning Workload Generator with Energy and Carbon Emission Support

Federated Learning (FL) is prevailing as the dominating service paradigm facilitating the training of Machine Learning (ML) models among a set of collaborating entities in a distributed fashion. Although FL settings are ideal for scalable model training and can limit data privacy exposure for the collaborating entities, evaluating the impact of FL on the underlying infrastructure is challenging due to several different configuration knobs. To address this, we introduce AtlasFL, a framework designed to simplify the benchmarking of FL deployments and enable users to generate realistic workloads that can be used to assess what-if scenarios. A key feature of AtlasFL is that on top of FL-level metrics, AtlasFL also exposes execution traces for the underlying infrastructure utilization, such as compute and memory usage as well as energy consumption. Moreover, AtlasFL can integrate with services that expose energy grid data and energy availability from direct access to renewables to provide carbon footprint estimations for requested FL experiment configurations. To illustrate the utility of AtlasFL, a benchmarking process embracing 5 FL implementations was conducted over an edge micro-DC testbed and afterwards, a use-case scenario is designed to introduce the impact of time-shifting when running together a total of 15 FL experiments.