BTW, I really appreicate your efforts on reviewing this patchset.
It would be great if you could take some time to review my another
patchset :)
[PATCH 0/7] vhost: vhost-cuse removal and code path refactoring
It touchs a large of code base, that I wish I could apply it ASAP.
So that the chance a later patch will introduce conflicts is small.
--yliu
On Tue, Aug 23, 2016 at 10:42:11PM +0800, Yuanhan Liu wrote:
> On Tue, Aug 23, 2016 at 04:18:40PM +0200, Maxime Coquelin wrote:
> >
> >
> > On 08/23/2016 10:10 AM, Yuanhan Liu wrote:
> > >This patch set enables vhost Tx zero copy. The majority work goes to
> > >patch 4: vhost: add Tx zero copy.
> > >
> > >The basic idea of Tx zero copy is, instead of copying data from the
> > >desc buf, here we let the mbuf reference the desc buf addr directly.
> > >
> > >The major issue behind that is how and when to update the used ring.
> > >You could check the commit log of patch 4 for more details.
> > >
> > >Patch 5 introduces a new flag, RTE_VHOST_USER_TX_ZERO_COPY, to enable
> > >Tx zero copy, which is disabled by default.
> > >
> > >Few more TODOs are left, including handling a desc buf that is across
> > >two physical pages, updating release note, etc. Those will be fixed
> > >in later version. For now, here is a simple one that hopefully it
> > >shows the idea clearly.
> > >
> > >I did some quick tests, the performance gain is quite impressive.
> > >
> > >For a simple dequeue workload (running rxonly in vhost-pmd and runnin
> > >txonly in guest testpmd), it yields 40+% performance boost for packet
> > >size 1400B.
> > >
> > >For VM2VM iperf test case, it's even better: about 70% boost.
> >
> > This is indeed impressive.
> > Somewhere else, you mention that there is a small regression with small
> > packets. Do you have some figures to share?
>
> It could be 15% drop for PVP case with 64B packet size. The test topo is:
>
> nic 0 --> VM Rx --> VM Tx --> nic 0
>
> Put simply, I run vhost-switch example in the host and run testpmd in
> the guest.
>
> Though the number looks big, I don't think it's an issue. First of all,
> it's disabled by default. Secondly, if you want to enable it, you should
> be certain that the packet size is normally big, otherwise, you should
> not bother to try with zero copy.
>
> > Also, with this feature OFF, do you see some regressions for both small
> > and bigger packets?
>
> Good question. I didn't check it on purpose, but I did try when it's
> disabled, the number I got is pretty the same as the one I got without
> this feature. So, I would say I don't see regressions. Anyway, I could
> do more tests to make sure.
>
> --yliu
