Skip to content

Move to the nvidia gpu operator, add time-slicing support

Ricardo Rocha requested to merge nvidiaoperator into master

Drop our custom nvidia gpu setup and rely on the upstream gpu operator instead.

  • Download nvidia images from cern magnum registry repo
  • Add nvidia license gridd server

Added daemonset to set selinux permissive on GPU nodes

Enable nvidia's kubernetes GPU timesharing

  • shared GPU's as a new resource
  • 2 time slicing profiles available [slice-4 and slice-10]

OS-12528 and OS-12888

Closes: https://gitlab.cern.ch/kubernetes/project/-/issues/122 Closes: https://gitlab.cern.ch/kubernetes/project/-/issues/194 Closes: https://gitlab.cern.ch/kubernetes/project/-/issues/142

Edited by Diogo Filipe Tomas Guerra

Merge request reports