On 21 Jul 2026 at 12:56:33 PM, Felix Huettner via dev <[email protected]> 
wrote:

> Previously we guarded expectations using an rwlock. While there are
> generally only a few or none expectation entries, taking and releasing
> the rwlock in read mode still takes more time than using a cmap.
>
> When doing a loadtest with OVN and natting across two LRs we increased
> our packet rate from 2.5 Mio pps to 12 Mio pps with only 40% of the

nit: Mpps is more grep friendly

> previous load. This mainly comes from the UNSNAT and UNDNAT stages of OVN that
> generally do not match anything as long as we are in not specific cases.
>

numbers are interesting. Moving this to cmap was for sure something
beneficial, so thanks for taking care of it.

Can you please elaborate a little on your test setup, e.g. threads,
rxqs, type of traffic and so forth?

> Co-Authored-by: Florian Werner <[email protected]>
> Signed-off-by: Florian Werner <[email protected]>
> Co-Authored-by: Sebastian Riese <[email protected]>
> Signed-off-by: Sebastian Riese <[email protected]>
> Signed-off-by: Felix Huettner <[email protected]>
> ---
>
> Notes:
>     v3->v4: smaller changes
>
>  lib/conntrack-private.h |   9 ++--
>  lib/conntrack.c         | 112 +++++++++++++++++++---------------------
>  lib/conntrack.h         |   5 +-
>  3 files changed, 60 insertions(+), 66 deletions(-)

_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to