[OpenStack][GPU] OpenStack對於GPU的支援

目前OpenStack對於GPU的支援作法

• PCI passthrough
  – Nova VM-based compute (e.g. Libvirt+KVM) with PCI passthrough
• Ironic
  – 透過ironic設定含有GPU的compute node
• CPU pinning、NUMA
• vGPU
• RDMA(Remote Direct Memory Access)
  – 透過一台主機上的網路卡API存取到另一台主機上的記憶體之技術

vGPU

目前若cpu有支援intel GVT-g，就可以支援全虛擬化的vGPU, 雖然效能略低於PCI passthrough，但可以多個VM間做sharing (up to 15)

[vGPU的啟動方法]
Nova
1. 於nova-compute啟動GPU type

[devices]
enabled_vgpu_types = nvidia-35

2. 於nova controller設定flavor

Configure a flavor to request one virtual GPU:

$ openstack flavor set vgpu_1 --property "resources:VGPU=1"

RDMA(Remote Direct Memory Access) 

GPU on K8S

• 於K8S 1.10開始支援nv-docker
• 一個pod會吃滿一張GPU卡，無法切分

[參考資源]

• https://wiki.openstack.org/wiki/GPUs
• https://gist.github.com/claudiok/890ab6dfe76fa45b30081e58038a9215
• https://specs.openstack.org/openstack/nova-specs/specs/queens/approved/virt-add-support-for-vgpu.html
• https://www.openstack.org/assets/presentation-media/Enable-GPU-virtualization-in-OpenStack.pdf

• https://blueprints.launchpad.net/nova/+spec/add-support-for-vgpu
• https://docs.nvidia.com/grid/5.0/pdf/grid-vgpu-user-guide.pdf
• https://docs.openstack.org/nova/queens/admin/virtual-gpu.html
• https://gist.github.com/claudiok/890ab6dfe76fa45b30081e58038a9215
•