On Wed, Jun 10, 2015 at 01:48:49PM -0400, Andy Hill wrote: > When testing force-reload-kmod on XenServer 6.2 running OVS 2.1/2.3 with > bonded LACP configurations, the disruption to data plane connectivity is > much longer than non-bonded configurations. > > While debugging, one area that longer than expected was in the dp-ctl > del-dp calls[1]. The earlier dump-dps call only returned "system@ovs-system > ". > > I straced the calls and was surprised with this call taking over 3s: > > https://gist.github.com/andyhky/b531751f00c5ca94ae24#file-bonded-L197 > > But non bonded <1s: > > https://gist.github.com/andyhky/b531751f00c5ca94ae24#file-non-bonded-L196 > > Repeated testing on different bonded servers all took between 3s-5s. > Testing on non-bonded servers was always ~1s. > > Within 1s of the del-dp command running, traffic drops from the host. Does > anyone have ideas on why this would take so much longer on a bonded > configuration?
At a guess, without looking at code, I'd think that RCU synchronization on removal from a datapath is costly for physical Ethernet devices and that it multiples across the number of them present in the datapath. Pravin, does this make any sense (or do you have another idea)? _______________________________________________ discuss mailing list [email protected] http://openvswitch.org/mailman/listinfo/discuss
