On Thu, 30 Oct 2025 at 09:24, David Marchand <[email protected]> wrote: > > Hello, > > On Wed, Jul 16, 2025 at 5:44 PM Ilya Maximets <[email protected]> wrote: > > > > On 7/7/25 2:25 PM, David Marchand via dev wrote: > > > A dp_packet object comes from either dp_packet_new() (which resets all > > > fields to good default values) or from a DPDK mempool. > > > > > > Reset offsets in netdev-dpdk so that inner offsets are in a clean state > > > before calling flow extraction for the first time. > > > > > > On recirculation, inner offsets can be left untouched when extracting > > > flow informations again. > > > > > > Reported-at: https://issues.redhat.com/browse/FDP-1517 > > > Signed-off-by: David Marchand <[email protected]> > > > --- > > > lib/flow.c | 9 +++------ > > > lib/netdev-dpdk.c | 1 + > > > 2 files changed, 4 insertions(+), 6 deletions(-) > > > > Hi, David. Thanks for the patch! > > > > I see some performance difference with this change. Interestingly, I see > > a performance decrease in a basic V2V case with EMC, but an increase in a > > Simple Match case. Have you noticed anything similar in your testing? > > IIRC, I saw no change in my own v2v setup. > I'll rebuild the setup.
I see similar number w/ and w/o the patch. Before patch: - simple: 5965611 - normal: 5499010 After patch: - simple: 5981719 - normal: 5497476 In all cases, pmd 4 (polling vhost1, which is the "emitter") is at 100%. I also tested resetting only the outer offsets in parse_tcp_flags() (which would make sense as part of this change), and I see no difference. -- David Marchand _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
