[Intel-wired-lan] VF/SRIOV question

[Alexander Duyck]

On Thu, Oct 17, 2019 at 11:06 AM JD <jdtxs00 at gmail.com> wrote:
>
> Hello,
>
> I couldn't find anything in the spec/doc for the Intel 82599 series
> NIC's regarding performance penalties with a higher number of VF's.
>
> Currently I'm using 16 VF's with SR-IOV/QEMU, but the 82599 NIC
> supports up to 63 VF's. From a driver/NIC perspective, are there any
> performance considerations or penalties of enabling/using all of the
> available VF's on a NIC?
>
> If there isn't, is this the same case for other models (besides 82599) as well?
>
> Thank you!

One hardware limitation I am aware of is that as you spread the work
over more queues, or in this case more VFs you may not be able to
achieve 64B line rate with small packets. The issue is as you add more
queues the descriptor fetching becomes more interleaved between the
queues which will reduce the performance. So instead of being able to
hit 14.88Mpps you may only see about 10 or 11Mpps. If you are working
with packets larger than 128B or larger you should see little to no
impact.

Another limitation I can think of is the number of queues per PF/VF.
The hardware only has a certain number of queues. For 82599 that
number is 128. So when you have 31 or fewer VFs you get 4 queues for
the PF, when you have 32 or more the PF has to drop to 2 queues.
Depending on your workload this may mean more stress on the CPUs
handling PF traffic as the number of queues is reduced.

Hope that helps.

- Alex