Re: [Qemu-devel] [virtio-dev] Re: [virtio-dev] Re: [PATCH v2 00/16] Vhost-pci for inter-VM communication

["Michael S. Tsirkin" <mst@redhat.com>]

On Fri, May 19, 2017 at 11:10:33AM +0800, Jason Wang wrote:
> 
> 
> On 2017年05月18日 11:03, Wei Wang wrote:
> > On 05/17/2017 02:22 PM, Jason Wang wrote:
> > > 
> > > 
> > > On 2017年05月17日 14:16, Jason Wang wrote:
> > > > 
> > > > 
> > > > On 2017年05月16日 15:12, Wei Wang wrote:
> > > > > > > 
> > > > > > 
> > > > > > Hi:
> > > > > > 
> > > > > > Care to post the driver codes too?
> > > > > > 
> > > > > OK. It may take some time to clean up the driver code before
> > > > > post it out. You can first
> > > > > have a check of the draft at the repo here:
> > > > > https://github.com/wei-w-wang/vhost-pci-driver
> > > > > 
> > > > > Best,
> > > > > Wei
> > > > 
> > > > Interesting, looks like there's one copy on tx side. We used to
> > > > have zerocopy support for tun for VM2VM traffic. Could you
> > > > please try to compare it with your vhost-pci-net by:
> > > > 
> > We can analyze from the whole data path - from VM1's network stack to
> > send packets -> VM2's
> > network stack to receive packets. The number of copies are actually the
> > same for both.
> 
> That's why I'm asking you to compare the performance. The only reason for
> vhost-pci is performance. You should prove it.
> 
> > 
> > vhost-pci: 1-copy happen in VM1's driver xmit(), which copes packets
> > from its network stack to VM2's
> > RX ring buffer. (we call it "zerocopy" because there is no intermediate
> > copy between VMs)
> > zerocopy enabled vhost-net: 1-copy happen in tun's recvmsg, which copies
> > packets from VM1's TX ring
> > buffer to VM2's RX ring buffer.
> 
> Actually, there's a major difference here. You do copy in guest which
> consumes time slice of vcpu thread on host. Vhost_net do this in its own
> thread. So I feel vhost_net is even faster here, maybe I was wrong.

Yes but only if you have enough CPUs. The point of vhost-pci
is to put the switch in a VM and scale better with # of VMs.

> > 
> > That being said, we compared to vhost-user, instead of vhost_net,
> > because vhost-user is the one
> > that is used in NFV, which we think is a major use case for vhost-pci.
> 
> If this is true, why not draft a pmd driver instead of a kernel one? And do
> you use virtio-net kernel driver to compare the performance? If yes, has OVS
> dpdk optimized for kernel driver (I think not)?
> 
> What's more important, if vhost-pci is faster, I think its kernel driver
> should be also faster than virtio-net, no?

If you have a vhost CPU per VCPU and can give a host CPU to each using
that will be faster.  But not everyone has so many host CPUs.


> > 
> > 
> > > > - make sure zerocopy is enabled for vhost_net
> > > > - comment skb_orphan_frags() in tun_net_xmit()
> > > > 
> > > > Thanks
> > > > 
> > > 
> > > You can even enable tx batching for tun by ethtool -C tap0 rx-frames
> > > N. This will greatly improve the performance according to my test.
> > > 
> > 
> > Thanks, but would this hurt latency?
> > 
> > Best,
> > Wei
> 
> I don't see this in my test.
> 
> Thanks