Lessons learned from migrating AI accelerator drivers from the kernel to userspace (asg2024)

The kernel driver is dead; long live the userspace driver! In this talk, we’ll discuss the motivation, challenges and outcomes of migrating drivers for Meta’s AI accelerator chips from the kernel to userspace. Topics include: - Managing systemd units at scale - Experiences of running IPC over D-Bus - Re-writing the driver in Rust - The tooling necessary to support a variety of environments - Overall deprecation challenges and wins Licensed to the public under https://creativecommons.org/licenses/by/4.0/de/ about this event: https://cfp.all-systems-go.io/all-systems-go-2024/talk/77PFFF/