Uploaded image for project: 'Fast Datapath Product'
  1. Fast Datapath Product
  2. FDP-1897

Improve userspace datapath TSO with tunnel checksum performance for mlx5 nics

XMLWordPrintable

    • Icon: Story Story
    • Resolution: Unresolved
    • Icon: Undefined Undefined
    • None
    • None
    • openvswitch3.5
    • None
    • 3
    • False
    • False
    • rhel-9
    • None
    • rhel-net-ovs-dpdk
    • ssg_networking
    • OVS/DPDK - Sprint 11 - East, OVS/DPDK - Sprint 12 - East, OVS/DPDK - Sprint 13 - East
    • 3

      It was noticed that TSO over Geneve tunnels (with tunnel checksum enabled) with mlx5 nics gives lower performance than the kernel datapath.

      This comes from the fact that mlx5 (and other nics, like i40e) do not support outer UDP checksum.
      Because of this, all TSO requests are done by OVS itself.
      This can be seen by looking at coverage counters:

      # ovs-appctl coverage/show | grep seg
netdev_soft_seg_good       0.0/sec     0.033/sec        0.0861/sec   total: 169593

      After some brainstorming, I realised that the outer UDP checksum can be computed before letting the nic segment the packets as it keeps the outer L4 header untouched.
      So it is one option to optimise such operation.

              rhn-support-dmarchan David Marchand
              rhn-support-dmarchan David Marchand
              Votes:
              0 Vote for this issue
              Watchers:
              1 Start watching this issue

                Created:
                Updated: