This issue were created for a better understanding of the scalability of vDPA.

We need to do scalability test to know how well the vDPA can scale. We know vDPA can be used to offload datapath from host to the NIC. Technically it should save host CPU cycles so it would have more CPUs for customer workloads.

We need to know the following things:

1) How many vDPA could be created by a single CX6/BF3 devices, what's the performance for those configurations
2) How many VM/VCPUs could be created in a single host w/ and w/o vDPA (we can assume that each vcpu need a pcpu and each vhost need a pcpu) and what's the performance for those configurations

If we get good numbers for example: vDPA can help for launching more VMs with reasonable performance. It would be a good start for convincing CNV team to use vDPA in OCI.