I have basic queries regarding QPs in GDA-KI,
Diagram in below links shows that QP is created by CPU
But looking at the pics at ( Magnum IO GPUDirect, NCCL, NVSHMEM, and GDA-KI on Grace
Hopper and Hopper systems -S61368)
specially slide 20 , Workqueue is present inside GPU.
So my question are
1.Who creates QPs CPU or GPU? In general who manages QP ( creation, destroy etc )
2.How is GPU Memory used for MR, its pre-registered or dynamic
3.How is NIC Bar exposed to GPU?