> Based on the discussion, I don't object to sending a gratuitous ARP > when we rebalance. But naively dropping traffic that comes in on the > wrong slave when rebalancing is going to cause trouble, because it > will cause some packets to get dropped during the transition. I think > that we would need a "grace period" of a second or two during which we > accept traffic on the old and new link, or some other heuristic to > figure out when the next-hop switch has switched links.
Are we chasing a long tail here? We have equivalent behavior to Xenserver running the bridge. I think generally things are less painful for users to migrate the more we mimic them. If people care that much about the possibility of the occasional DUP on failover they can switch to a LACP bond. If we decide to drop traffic, I think a reasonable strategy would be to send a gratuitous arp on the first packet which comes in the wrong slave, wait for a 1 or 2 second grace period, and then drop traffic. Ethan _______________________________________________ dev mailing list [email protected] http://openvswitch.org/mailman/listinfo/dev
