[dpdk-dev] [PATCH v2 0/2] Vhost: unitfy receive paths

[Wang, Zhihong]

> -----Original Message-----
> From: Maxime Coquelin [mailto:maxime.coquelin at redhat.com]
> Sent: Tuesday, May 29, 2018 5:45 PM
> To: dev at dpdk.org; Bie, Tiwei <tiwei.bie at intel.com>; Wang, Zhihong
> <zhihong.wang at intel.com>
> Cc: Maxime Coquelin <maxime.coquelin at redhat.com>
> Subject: [PATCH v2 0/2] Vhost: unitfy receive paths
> 
> Hi,
> 
> This second version fixes the feature bit check in
> rxvq_is_mergeable(), and remove "mergeable" from rx funcs
> names. No difference is seen in the benchmarks
> 
> This series is preliminary work to ease the integration of
> packed ring layout support. But even without packed ring
> layout, the result is positive.
> 
> First patch unify both paths, and second one is a small
> optimization to avoid copying batch_copy_nb_elems VQ field
> to/from the stack.
> 
> With the series applied, I get modest performance gain for
> both mergeable and non-mergeable casesi (, and the gain of
> about 300 LoC is non negligible maintenance-wise.
> 
> Rx-mrg=off benchmarks:
> 
> +------------+-------+-------------+-------------+----------+
> |    Run     |  PVP  | Guest->Host | Host->Guest | Loopback |
> +------------+-------+-------------+-------------+----------+
> | v18.05-rc5 | 14.47 |       16.64 |       17.57 |    13.15 |
> | + series   | 14.87 |       16.86 |       17.70 |    13.30 |
> +------------+-------+-------------+-------------+----------+
> 
> Rx-mrg=on benchmarks:
> 
> +------------+------+-------------+-------------+----------+
> |    Run     | PVP  | Guest->Host | Host->Guest | Loopback |
> +------------+------+-------------+-------------+----------+
> | v18.05-rc5 | 9.38 |       13.78 |       16.70 |    12.79 |
> | + series   | 9.38 |       13.80 |       17.49 |    13.36 |
> +------------+------+-------------+-------------+----------+
> 
> Note: Even without my series, the guest->host benchmark with
> mergeable buffers enabled looks suspicious as it should in
> theory be alsmost identical as when Rx mergeable buffers are
> disabled. To be investigated...
> 
> Maxime Coquelin (2):
>   vhost: unify Rx mergeable and non-mergeable paths
>   vhost: improve batched copies performance
> 
>  lib/librte_vhost/virtio_net.c | 376 +++++-------------------------------------
>  1 file changed, 37 insertions(+), 339 deletions(-)
> 

Acked-by: Zhihong Wang <zhihong.wang at intel.com>

Thanks Maxime! This is really great to see. ;) We probably need the
same improvement for Virtio-pmd.

One comment on Virtio/Vhost performance analysis: No matter what type
of traffic is used (PVP, or Txonly-Rxonly, Loopback...), we need to
be clear on who we're testing, and give the other part excessive CPU
resources, otherwise we'll be testing whoever the slowest.

Since this patch is for Vhost, I suggest to run N (e.g. N = 4) Virtio
threads on N cores, and the corresponding N Vhost threads on a single
core, to do performance comparison. Do you think this makes sense?

For Guest -> Host, in my test I see Rx-mrg=on has negative impact on
Virtio side, probably because Virtio touches something that's not
touched when Rx-mrg=off.

Thanks
-Zhihong