Lab 09: MIG: profiles, mig-parted, Slurm MIG GRES, Kubernetes mixed strategy
240 minutes suggested, up to 360 in one session. Price: 65 credits per minute. Includes Check my work.
Hosted, this lab covers Parts A, B, C, D. The other Parts need your own machine (Home Lab Kit) and are shown collapsed in the lab.
The MIG tier is not available on this site yet.
What you will be able to do
- Enable MIG, read profiles and placements, and create and destroy GPU and compute instances by hand.
- Explain why some combinations of profiles do not fit, and create layouts largest first.
- Keep layouts in a
nvidia-mig-partedfile, and apply, assert and export them. - Configure Slurm to schedule MIG devices as typed GRES, and request one with
--gres=gpu:<profile>:1. - Install the GPU Operator with the
mixedstrategy, repartition a node with one label, and run a Job that requests one MIG slice. - Hand a GPU from one owner to another safely, and recover when an owner's change fails.