I see, thank you for the reply. Even if it does not make sense for MIG to be supported/documented upstream if someone does come across MIG+OpenStack not backed by virtual GPUs do ping me please :-) I'll be trying to get this working when some cards arrive. Cheers, António On 14/06/21 18:01, Sylvain Bauza wrote:
On Mon, Jun 14, 2021 at 4:37 PM António Paulo <antonio.paulo@cern.ch <mailto:antonio.paulo@cern.ch>> wrote:
Hi!
Has anyone looked into instancing VMs with NVIDIA's Multi-Instance GPU (MIG) devices [1] without having to rely on vGPUs? Unfortunately, NVIDIA vGPUs lack tracing and profiling support that our users need.
I could not find anything specific to MIG in the OpenStack docs but I was wondering if doing PCI passthrough [2] of MIG devices is an option that someone has seen or tested?
Maybe some massaging to expose the MIG as a Linux device is required [3]?
Nividia MIG feature is orthogonal to virtual GPUs and hardware dependent. As the latter, this is not really something we can "support" upstream as our upstream CI can't just verify it.
Some downstream vendors tho have work efforts for trying to test this with their own solutions but again, not something we can discuss it here.
Cheers, António
[1] https://docs.nvidia.com/datacenter/tesla/mig-user-guide/ <https://docs.nvidia.com/datacenter/tesla/mig-user-guide/> [2] https://docs.openstack.org/nova/pike/admin/pci-passthrough.html <https://docs.openstack.org/nova/pike/admin/pci-passthrough.html> [3] https://docs.nvidia.com/datacenter/tesla/mig-user-guide/#device-nodes <https://docs.nvidia.com/datacenter/tesla/mig-user-guide/#device-nodes>