Re: HVM performance once again

["Roger Pau Monné" <roger.pau@citrix.com>]

On Wed, May 31, 2023 at 09:49:32AM +0200, Marek Marczykowski-Górecki wrote:
> Hi,
> 
> I returned to HVM performance once again, this time looking at PCI
> passthrough impact evaluating network throughput.
> The setup:
>  - Xen 4.17
>  - Linux 6.3.2 in all domU
>  - iperf -c running in a PVH (call it "client")
>  - iperf -s running in a HVM (call it "server")
>  - client's netfront has a backend directly in server
>  - frontend's "trusted" is set to 0
>  - HVM have qemu in a stubdomain in all the cases
>  - no intentional differences about HVM besides presence of a PCI device
>    (it is a network card, but it was not involved in the traffic)
> 
> And now the results:
>  - server is plain HVM: ~6Gbps
>  - server is HVM and has some PCI passthrough: ~3Gbps
> 
> Any idea why such huge difference?

Just a wild guess, when domains have a PCI device assigned the memory
cache types from the guest are enforced, otherwise it all defaults to
write-back (see epte_get_entry_emt()).

If you are not using the PCI device you might want to play with
epte_get_entry_emt() and see if that makes a difference.

Do you see the same performance regression when testing on AMD?

Thanks, Roger.