#1 -- Évalution de l’impact énergétique du scheduling pour l’entraînement d’IA sur des CPU hétérogènes.
COMPAS
Système
Nix
RAPL
Andrew Mary Huet de Barochez, Stéphan Plassart, Sébastien Monnet
Papier Description ArchiveThe growth of artificial intelligence raises concerns about electricity consumption, especially since the rise of genera- tive models. At the same time, hardware is becoming increasingly heterogeneous, which is often an obstacle to efficient and fru- gal usage of computer resources. CPUs are a great example, as heterogeneous topologies are getting more common. Their cores with different frequencies, cache size or hierarchy, and thread number make it challenging for developers to efficiently parallelize tasks. In this context, we investigate how task-based programming can be used to adapt to such constraints. The ability to schedule tasks based on runtime informations, such as hardware topology, and worker load can solve these problems. To this end, we built DAHL, a task-based programming machine learning framework using the StarPU runtime system. As a use case, we study an image classification task, using datasets with increasingly larger images. We perform our experiments over machines with heterogeneous and homogeneous CPUs. A reference is established by comparing different Pytorch configurations against DAHL. Then we evaluate the runtime and energy gains of several StarPU schedulers. Results show that, specifically on heterogeneous CPUs, the default Pytorch configuration can be 7.5× slower and consume 17× more at worst. Additionally, we demonstrate that the scheduler choice on StarPU can impact the runtime with up to 22% speedup, and energy consumption with up to 15% energy gains. Last, we draw guidelines to choose scheduler families depending on context.