[ovs-discuss] BGP and GRE tunnels

Krzysztof Wojciechowski krzysztof.wojciechowski at tiktalik.com
Mon Nov 23 10:55:38 UTC 2015


Hello

We have two servers with openvswitch 1.7. They are on Centos 6.x with 
2.6 kernel. For now they have 8 bridges with ~20 gre tunnels each. They 
are BGP peers too, so they have ~600k routes in routing table.

Everything works fine, but we would like to update it because of low ovs 
performance (no megaflows and multithreading)...

I installed Debian 8 with openvswitch 2.4 (2.3.1 tested too) and we have 
one problem... When i add even one gre tunnel ovs go up too 70-90% cpu 
usage, when i add more than 3 tunnels cpu is at 100% all the time (one 
core). When i add more than 10 tunnels ovs starts to work veeeeery 
slow... ovs-ofctl dump-flows need ~5 sec to show me empty flow table... 
Adding next gre port takes 5-10 sec etc... And with every next port its 
slow down more and more...

I was searching for problem for long and then... i stopped bgpd.

I understand now that there is a problem with new openvswitch and GRE 
tunnels when there is big routing table. When we have ~50 routes in 
routing table everything works fine. But when bgpd starts, ovs starts 
choking...


Is there anything we can do? or we need to make something to not have so 
much routes on server with OVS?

Regards
Krzysztof Wojciechowski
Tiktalik.com






More information about the discuss mailing list