Re: [virtio-dev] Re: [RFC] Virtio RDMA

[Yongji Xie <xieyongji@bytedance.com>]

On Wed, Feb 16, 2022 at 10:21 PM Hannes Reinecke <hare@suse.de> wrote:
>
> On 2/16/22 15:08, Yongji Xie wrote:
> > On Wed, Feb 16, 2022 at 8:06 PM Leon Romanovsky <leon@kernel.org> wrote:
> >>
> >> On Wed, Feb 16, 2022 at 06:03:29PM +0800, Junji Wei wrote:
> >>>
> >>>> On Feb 16, 2022, at 5:48 PM, Leon Romanovsky <leon@kernel.org> wrote:
> >>>>
> >>>> On Wed, Feb 16, 2022 at 04:00:53PM +0800, Junji Wei wrote:
> >>>>
> >>>> <...>
> >>>>
> >>>>>>
> >>>>>> What is the use case for this virtio-rdma? Especially in context of RXE.
> >>>>>
> >>>>> Hmm... yes, we didnât find one. In passthrough case we can use RXE directly.
> >>>>
> >>>> It doesn't sound like a good sales pitch.
> >>>
> >>> Maybe I misunderstanded what you mean. We mean we didnât find a user case
> >>> for virtio-rdma with passthrough net device. Do you want to know the user
> >>> case for our virtio-rdma(RoCE) proposal?
> >>
> >> Yes, please.
> >>
> >
> > I think one point is: when running RDMA accelerated applications in
> > VM, the virtio-rdma solution should get better performance than RXE
> > since it has a shorter data path (guest app -> host dpdk, bypass guest
> > kernel).
> >
> And you can (potentially) implement RDMA functionality in the hypervisor
> by 'simply' to a memcpy between guests.
> (I know it's not simple. But you don't actually need RDMA hardware here
> and still should be able to provide RDMA functionality to guests.)
>

Yes, and I think we're able to achieve that with virtio-rdma when one
DPDK serves multiple VMs.

> And if you look at the current RoCE CNAs they operate in exactly the
> same way; there's a network interface, and _additionally_ an rdma
> device. So for the OS both will be detected as distinct devices, who
> just happen to run on the same PCI device.
>

Got it.

Thanks,
Yongji