Re: GPU passthrough memory bandwidth performance degration

[Paul Menzel]

Dear Jiatong,


Thank you for your message.


Am 04.07.23 um 03:58 schrieb Jiatong Shen:

>    I am testing a passtrhough gpu performance by measuring device to host &
> host to device memory copy bandwidth. The tested GPU is nvidia t4. The
> benchmarking script I am using is
> https://developer.nvidia.com/blog/how-optimize-data-transfers-cuda-cc/.

Please share your QEMU version and the command line for QEMU.

> On the baremetal machine, the result is
> [image: image.png]

Instead of making screenshots, please copy (and paste) the output to the 
message.

> in the virtual machine, the result is,
>
> [image: image.png]
>
> My question is what could be the reason for the degradation and is there
> anything I can do to improve it? Thank you very much for the help.

Did you compare the Linux kernel messages from both setups, and for 
example `lspci` output? Are you using the latest versions?

With more information, others can hopefully help you.


Kind regards,

Paul


PS: If you care:

> --
>
> Best Regards,
>
> Jiatong Shen

Your signature delimiter is missing a trailing space [1].

PPS: For mailing lists, plain text messages without HTML part are 
preferred [2].


[1]: https://en.wikipedia.org/wiki/Signature_block#Standard_delimiter
[2]: https://useplaintext.email/