Centos7 RDO VXLAN TCP segment losses
Hi everyone!
I have a problem with multinode Juno installation (although I have had the same problems with Kilo and switched to Juno to work with RHEL OSP6 platform). The problem is simple: after few transmitted TCP packets encapsulated in VXLAN segments there is a 6 seconds interruption; there are three strange sympthoms:
- NO interruption for ICMP packets even in flood mode.
- VXLAN IF recieves all packets, but br-tun Table-0 flow counter doesn't increase for lost packets.
- small packets are also lost (not an MTU issue).
I disabled iptables, but it didn't help. Any thoughts about this issue are highly appreciated.
UPD1. I' ve tried following workarounds (with no success):
- Downgraded openvswitch and neutron modules.
- Installed openvswitch 2.3.2 from source.
- Insert Accept rules in all filtering chains on compute node and netns router instance.
- Played with systctl parameters (rp_filter, net.bridge.bridge-nf-call-iptables).
- Played with neutron parameters (ovs l2pop, fragmentation, arp processing).
- Replaced OVS br-tun flow-processing rules with drop actions to forward rules (I noticed that lost packets doesn't increment counter in table 0, so I don't think that it will help)
- Switched to GRE tunnels (same result, some TCP segments are lost in transmission over VXLAN).
- Lowered MTU value (by default I use Jumbo frames).
All unsuccessfull, so again, any thoughts on this issue are highly appreciated.
UPD2. Tried linuxbridge driver. It is quite messy and doesn't work correctly for my network setup.