Due to OVN's usage of learn flows for ECMP symmetric reply handling, benchmark tools that start many short-lived connections can overload OVS.

Instead, OVN can use the kernel's CT functionality to determine the reply next-hop without adding learn flows. This greatly reduces OVS CPU usage in this scenario and allows higher throughput.

Upstream OVN patch:
https://patchwork.ozlabs.org/project/ovn/patch/20230901105557.970938-1-dceara@redhat.com/

These downstream builds and later include the patch:
ovn23.09-23.09.0-beta.31.el9fdp
ovn23.06-23.06.1-8.el9fdp
ovn23.03-23.03.1-10.el9fdp
ovn22.12-22.12.1-8.el8fdp