Some NICs like Mellanox allow you to create "virtual functions" (basically a hardware-assisted virtual NIC). These are meant to be passed directly through to VMs instead of generating "tap" interfaces tied to a bridge.
Supposedly this sort of thing completely hardware-offloads all communication from VM to VM freeing up resources on the host and reducing latency.
You can assign a VLAN per virtual function, or use vlan 4095 to pass through *all* vlans available:
https://community.mellanox.com/docs/DOC-1123
It looks like if you create an LACP LAG/bond on the host, it will auto route traffic to the "up" port even though the VF is tied to one port, which is a nice feature so you don't have to do some bonding hack in the guest:
https://community.mellanox.com/docs/DOC-2612
It would seem like Proxmox could add management interfaces to assign and configure VFs dynamically like it does with tap interfaces. I'm not really seeing any limitations to this vs bonding, but I do see a huge potential upside, especially if using high performance (10/25/40G) nics.
Are there any plans for doing something like this?
Thanks!
-Brad
Supposedly this sort of thing completely hardware-offloads all communication from VM to VM freeing up resources on the host and reducing latency.
You can assign a VLAN per virtual function, or use vlan 4095 to pass through *all* vlans available:
https://community.mellanox.com/docs/DOC-1123
It looks like if you create an LACP LAG/bond on the host, it will auto route traffic to the "up" port even though the VF is tied to one port, which is a nice feature so you don't have to do some bonding hack in the guest:
https://community.mellanox.com/docs/DOC-2612
It would seem like Proxmox could add management interfaces to assign and configure VFs dynamically like it does with tap interfaces. I'm not really seeing any limitations to this vs bonding, but I do see a huge potential upside, especially if using high performance (10/25/40G) nics.
Are there any plans for doing something like this?
Thanks!
-Brad