Dear all,
I'm currently debugging a (testing) deployment with neutron/OVN, currently most ports in external flat network.
So far I'm happy everything can communicate, but we see a strong performance degradation with increasing number of VMs/ports
hosted on a hypervisor.
The setup is similar to the one described in the routed provider networks docs [1].
For routing between the different segments, we have physical routers, that are not managed by neutron.
Debugging the issue showed traffic originating from the DC router to VMs go perfectly as unicast traffic through OVN.
Communicating from the VM to other networks via the DC router, however, will broadcast the traffic to all attached ports
on the internal bridge.
Is there a way to tell neutron/OVN to use a specific MAC address for the router and forward the traffic through the patch
port instead of broadcasting it?
I could not find any documentation of whether I am missing a neutron port or some other configuration on the hypervisor.
Just to make sure: The VM and DC routers have the right ARP information of itself, the traffic that is duplicated is from
VM to router/external network. (via default route to router)
As far as I could see, this issue does not appear with private networks, since all MAC addresses are known to OVN/Neutron.
Versions:
Nova/Neutron: Yoga
ovs-vswitchd (Open vSwitch) 3.1.4
ovn-northd 22.12.1
kernel 5.14.0-284.30.1.el9_2.x86_64
Best,
Daniel