mbox series

[v5,0/2] one way barrier for split vring idx

Message ID 20200430091437.34958-1-joyce.kong@arm.com (mailing list archive)
Headers
Series one way barrier for split vring idx |

Message

Joyce Kong April 30, 2020, 9:14 a.m. UTC 
  This patchset replaces the two-way barriers with C11 one-way barriers
for split vring idx, when the frontend and backend are implemented
in software.

By doing PVP benchmarking, the test result of 2c1q showed the throughput
increased 20% with the 0.001% of acceptable loss rate on Thunderx2
platform.[1]

By doing vhost-user + virtio-user case benchmarking, 4% performance gain
was measured on Thunderx2 platform by 2c1q RFC2544 test of 0.001% loss,
and 4.7% performance was improved on Dell platform by 1c1q RFC2544 test
of zero packet loss.[2]

[1]https://doc.dpdk.org/guides/howto/pvp_reference_benchmark.html
[2]https://doc.dpdk.org/dts/test_plans/pvp_multi_paths_performance_test_plan.html
   PVP test with virtio 1.0 normal path

v5:
  Fix the compling issue when 'CONFIG_RTE_LIBRTE_VIRTIO_DEBUG_DUMP=y'
  is enabled.

v4:
  Remove some duplicated code comment.

v3:
  Modify some style error.

v2:
  Add test performance statistics.

Joyce Kong (2):
  virtio: one way barrier for split vring used idx
  virtio: one way barrier for split vring avail idx

 drivers/net/virtio/virtio_ethdev.c            |  7 +--
 drivers/net/virtio/virtio_ring.h              |  2 +-
 drivers/net/virtio/virtio_rxtx.c              | 35 ++++-------
 drivers/net/virtio/virtio_rxtx_simple_neon.c  |  5 +-
 drivers/net/virtio/virtio_rxtx_simple_sse.c   |  4 +-
 .../net/virtio/virtio_user/virtio_user_dev.c  |  8 ++-
 drivers/net/virtio/virtqueue.c                |  2 +-
 drivers/net/virtio/virtqueue.h                | 58 ++++++++++++++++---
 lib/librte_vhost/virtio_net.c                 | 19 +++---
 9 files changed, 81 insertions(+), 59 deletions(-)

Comments

Maxime Coquelin April 30, 2020, 8:54 p.m. UTC | #1 
On 4/30/20 11:14 AM, Joyce Kong wrote:
> This patchset replaces the two-way barriers with C11 one-way barriers
> for split vring idx, when the frontend and backend are implemented
> in software.
> 
> By doing PVP benchmarking, the test result of 2c1q showed the throughput
> increased 20% with the 0.001% of acceptable loss rate on Thunderx2
> platform.[1]
> 
> By doing vhost-user + virtio-user case benchmarking, 4% performance gain
> was measured on Thunderx2 platform by 2c1q RFC2544 test of 0.001% loss,
> and 4.7% performance was improved on Dell platform by 1c1q RFC2544 test
> of zero packet loss.[2]
> 
> [1]https://doc.dpdk.org/guides/howto/pvp_reference_benchmark.html
> [2]https://doc.dpdk.org/dts/test_plans/pvp_multi_paths_performance_test_plan.html
>    PVP test with virtio 1.0 normal path
> 
> v5:
>   Fix the compling issue when 'CONFIG_RTE_LIBRTE_VIRTIO_DEBUG_DUMP=y'
>   is enabled.
> 
> v4:
>   Remove some duplicated code comment.
> 
> v3:
>   Modify some style error.
> 
> v2:
>   Add test performance statistics.
> 
> Joyce Kong (2):
>   virtio: one way barrier for split vring used idx
>   virtio: one way barrier for split vring avail idx
> 
>  drivers/net/virtio/virtio_ethdev.c            |  7 +--
>  drivers/net/virtio/virtio_ring.h              |  2 +-
>  drivers/net/virtio/virtio_rxtx.c              | 35 ++++-------
>  drivers/net/virtio/virtio_rxtx_simple_neon.c  |  5 +-
>  drivers/net/virtio/virtio_rxtx_simple_sse.c   |  4 +-
>  .../net/virtio/virtio_user/virtio_user_dev.c  |  8 ++-
>  drivers/net/virtio/virtqueue.c                |  2 +-
>  drivers/net/virtio/virtqueue.h                | 58 ++++++++++++++++---
>  lib/librte_vhost/virtio_net.c                 | 19 +++---
>  9 files changed, 81 insertions(+), 59 deletions(-)
> 


Applied to dpdk-next-virtio/master.

Thanks,
Maxime