To share some concrete numbers: we tested with 500 L2 networks and 4500 pods(two logical ports each) across 24 nodes. Average pod ready latency dropped from 21s to 7s and p99 dropped from 41s to 22s. There's likely room to tune the p99 further and bring it closer to the average.
On Mon, Jun 8, 2026 at 2:20 PM Patryk Diak <[email protected]> wrote: > Hi Dumitru, > > Testing this in ovn-kubernetes with layer2 primary UDNs, which use remote > ports extensively, showed amazing improvements at scale. > Average time for a pod to become ready dropped by over 60%, and p99 > dropped by approximately 45%. It would be great if it could be backported. > > Patryk > > On Fri, Jun 5, 2026 at 4:33 PM Dumitru Ceara <[email protected]> wrote: > >> Hi Ales, Mark, >> >> On 6/5/26 9:58 AM, Ales Musil wrote: >> > On Thu, Jun 4, 2026 at 10:57 PM Mark Michelson via dev < >> > [email protected]> wrote: >> > >> >> Thanks Dumitru! >> >> >> >> Acked-by: Mark Michelson <[email protected]> >> >> >> >> .... >> >> >> >> >> >> > Thank you Dumitru and Mark, >> > >> > applied to main for now. Dumitru, would you mind doing a bit more >> testing >> > before backporting? >> > >> >> Thanks for the review and for merging this! >> >> I ran the full ovn-kubernetes upstream CI with this change and >> everything is green: >> >> https://github.com/dceara/ovn-kubernetes/actions/runs/27004054146 >> >> I'm also going to work with Patryk next week to do some more downstream >> validation of this change at scale. I'll reply once I have more >> information. >> >> Thanks, >> Dumitru >> >> _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
