> -Original Message-
> From: Flavio Leitner
> Sent: Thursday 9 January 2020 14:45
> To: d...@openvswitch.org
> Cc: Stokes, Ian ; Loftus, Ciara
> ; Ilya Maximets ;
> yangy...@inspur.com; Flavio Leitner
> Subject: [PATCH v3 0/3] Add support for TSO with DPDK
>
> Abbreviated as TSO, TCP Segmentation Offload is a feature which enables
> the network stack to delegate the TCP segmentation to the NIC reducing
> the per packet CPU overhead.
>
> A guest using vhost-user interface with TSO enabled can send TCP packets
> much bigger than the MTU, which saves CPU cycles normally used to break
> the packets down to MTU size and to calculate checksums.
>
> It also saves CPU cycles used to parse multiple packets/headers during
> the packet processing inside virtual switch.
>
> If the destination of the packet is another guest in the same host, then
> the same big packet can be sent through a vhost-user interface skipping
> the segmentation completely. However, if the destination is not local,
> the NIC hardware is instructed to do the TCP segmentation and checksum
> calculation.
>
> The first 2 patches are not really part of TSO support, but they are
> required to make sure everything works.
>
> There are good improvements sending to or receiving from veth pairs or
> tap devices as well. See the iperf3 results below:
>
> [*] veth with ethtool tx off.
>
> VM sending to: Default Enabled
>Local BR 859 Mbits/sec 9.23 Gbits/sec
>Net NS (veth) 965 Mbits/sec[*] 9.74 Gbits/sec
>VM (same host)2.54 Gbits/sec 22.4 Gbits/sec
>Ext Host 10.3 Gbits/sec 35.0 Gbits/sec
I performed some similar tests. I recorded the following improvements in
throughput:
VM -> VM (same host): +5.4x
VM -> Ext Host: +3.8x
I tested VM -> Ext Host for both MT27800 (ConnectX-5) and a XL710 (Fortville)
NICs.
The result above was measured for the XL710.
Two things to note when testing with an i40e NIC:
1. The following patch is required for DPDK, which fixes an issue on the TSO
path:
http://git.dpdk.org/next/dpdk-next-net/commit/?id=b2a4dc260139409c539fb8e7f1b9d0a5182cfd2b
2. For optimal performance, ensure the IRQs of the queue being used by the
iperf server are pinned to their own core, as opposed to the same core as the
server process, which appears to be the default behavior.
I intend on submitting a follow up patch which documents the above.
Tested-by: Ciara Loftus
Thanks,
Ciara
>Ext Host (vxlan) 8.77 Gbits/sec (not supported)
>
> Using VLAN:
>Local BR 877 Mbits/sec 9.49 Gbits/sec
>VM (same host)2.35 Gbits/sec 23.3 Gbits/sec
>Ext Host 5.84 Gbits/sec 34.6 Gbits/sec
>
> Using IPv6:
>Net NS (veth) 937 Mbits/sec[*] 9.32 Gbits/sec
>VM (same host)2.53 Gbits/sec 21.1 Gbits/sec
>Ext Host 8.66 Gbits/sec 37.7 Gbits/sec
>
> Conntrack:
>No packet changes: 1.41 Gbits/sec33.1 Gbits/sec
>
> VM receiving from:
>Local BR 221 Mbits/sec 220 Mbits/sec
>Net NS (veth) 221 Mbits/sec[*] 5.91 Gbits/sec
>VM (same host)4.79 Gbits/sec 22.2 Gbits/sec
>Ext Host 10.6 Gbits/sec 10.7 Gbits/sec
>Ext Host (vxlan) 5.82 Gbits/sec (not supported)
>
> Using VLAN:
>Local BR 223 Mbits/sec 219 Mbits/sec
>VM (same host)4.21 Gbits/sec 24.1 Gbits/sec
>Ext Host 10.3 Gbits/sec 10.2 Gbits/sec
>
> Using IPv6:
>Net NS (veth) 217 Mbits/sec[*] 9.32 Gbits/sec
>VM (same host)4.26 Gbits/sec 23.3 Gbits/sec
>Ext Host 9.99 Gbits/sec 10.1 Gbits/sec
>
> Used iperf3 -u to test UDP traffic limited at default 1Mbits/sec
> and noticed no change with the exception for tunneled packets (not
> supported).
>
> Travis, AppVeyor, and Cirrus-ci passed.
>
> Flavio Leitner (3):
> dp-packet: preserve headroom when cloning a pkt batch
> vhost: Disable multi-segmented buffers
> netdev-dpdk: Add TCP Segmentation Offload support
>
> Documentation/automake.mk | 1 +
> Documentation/topics/dpdk/index.rst | 1 +
> Documentation/topics/dpdk/tso.rst | 96 +
> NEWS| 1 +
> lib/automake.mk | 2 +
> lib/conntrack.c | 29 ++-
> lib/dp-packet.h | 158 +-
> lib/ipf.c | 32 +--
> lib/netdev-dpdk.c | 318
> lib/netdev-linux-private.h | 4 +
> lib/netdev-linux.c | 296 +++---
> lib/netdev-provider.h | 10 +
> lib/netdev.c| 66 +-
> lib/tso.c | 54 +
> lib/tso.h | 23 ++
> vswitchd/bridge.c | 2 +
> vswitchd/vswitch.xml| 12 ++
> 17 files changed, 1013 insertions(+), 92 deletio
