On Tue, May 10, 2016 at 02:49:43PM -0700, Rich Lane wrote:
> I see a significant performance improvement with these patches, around 5% at
> 64
> bytes.
Thanks for testing.
>
> The one patch that didn't give any performance boost for me was "vhost:
> arrange
> virtio_net fields for better cache sharing".
Yeah, same here from my test. I mean, in theory, it should give us a
tiny boost, it doesn't in real life though. And since it (should) do
no harm, I would still include this patch in this set. Maybe I should
have noted at first that no real perf gain from the 3rd patch.
--yliu
>
> Tested-by: Rich Lane <rich.lane at bigswitch.com>
>
> On Mon, May 2, 2016 at 5:46 PM, Yuanhan Liu <yuanhan.liu at linux.intel.com>
> wrote:
>
> Here is a small patch set does the micro optimization, which brings about
> 10% performance boost in my 64B packet testing, with the following topo:
>
> ? ? pkt generator <----> NIC <-----> Virtio NIC
>
> Patch 1 pre updates the used ring and update them in batch. It should be
> feasible from my understanding: there will be no issue, guest driver will
> not start processing them as far as we haven't updated the "used->idx"
> yet. I could miss something though.
>
> Patch 2 saves one check for small packets (that can be hold in one desc
> buf and mbuf).
>
> Patch 3 moves several frequently used fields into one cache line, for
> better cache sharing.
>
> Note that this patch set is based on my latest vhost ABI refactoring
> patchset.
>
>
> ---
> Yuanhan Liu (3):
> ? vhost: pre update used ring for Tx and Rx
> ? vhost: optimize dequeue for small packets
> ? vhost: arrange virtio_net fields for better cache sharing
>
> ?lib/librte_vhost/vhost-net.h? |? ?8 +--
> ?lib/librte_vhost/vhost_rxtx.c | 110
> ++++++++++++++++++++++++------------------
> ?2 files changed, 68 insertions(+), 50 deletions(-)
>
> --
> 1.9.0
>
>
>
