On Mon, Nov 23, 2015 at 10:56:34AM -0500, Russell Bryant wrote: > On 11/23/2015 10:49 AM, Ben Pfaff wrote: > > I think we'll need more rationale than that. BFD is expensive without a > > good reason. > > It was on the ovn/TODO list, at least. :-) > > > * Use BFD as tunnel monitor.
I think I wrote that. It came from an NVP mindset. In NVP, the hypervisors use a set of "service nodes" to implement broadcast: a hypervisor sends a broadcast packet to a service node, then the service node replicates it and sends it to all of the destination hypervisors. In that design, it's really important for the hypervisors to have BFD set up between the hypervisors and the service nodes, because if one service node goes down the hypervisors need to shift to using a different service node to ensure availability. The same goes for the "gateway nodes" that provide access to physical networks. For OVN, we're not using service nodes (which simplifies the system a great deal; service nodes were never well-motivated), so we don't need BFD for that reason. We'll need something for the OVN gateways when we start supporting HA for gateways, and that something will probably involve BFD. For hypervisor-to-hypervisor tunnels, it's much less important to have tunnel monitoring (such as BFD), because when the tunnel goes down there's no alternative to switch to. If there's someone working on some kind of monitoring system for OVN, then it might make sense to turn on BFD (probably at a slow rate) to allow monitoring of the physical network. But absent that I don't think it's really worth the cost. Anyway, that's my story and I'm sticking to it ;-) _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
