On May 31, 2012, at 6:09 PM, Dong Jin wrote: > I have not yet seen any real openflow switch hardware, and I guess there is a > single ovs daemon running side, right? If so, just curious why the ovs daemon > is designed to create and manager a group of bridges (each bridge can > represent an Ethernet switch), rather than a single Ethernet switch? Can > someone pls advise or correct me if I am wrong? > > It looks that emulation testbed like mininet is running a single ovs daemon > in the hypervisor and create/manage a group of Ethernet switches in the > hypervisor. A follow-up question would be if in the real hardware one ovs > manages one ethernet switch, would it be more realistic for the emulation > testbed to model every Ethernet switch in a VM and run an ovs daemon inside > the VM?
This is not entirely an OVS question, but I thought I'd comment anyway. In a VM server environment, it is quite useful for a variety of reasons (isolation, administration, security, priority, other policies, slicing, etc.) to be able to support multiple switches with different sets of VMs attached to different switches. You may note that the Linux bridge supports multiple bridges as well. I expect this is why OVS can create multiple switches. Although it may be possible to run multiple OVS daemons, I expect it would create additional overhead and provide little benefit in most cases. (Though perhaps it would help with supporting mulitiple administrative domains?) As you note, supporting multiple virtual switches with a single kernel instance is also very useful for emulating networks. It's not nearly as efficient to create a full VM for each switch, which reduces the scale and speed of the networks that you can emulate. Regarding "realism," OVS is a a software switch rather than hardware, but as long as it can keep up with the traffic which is being presented to it, it should be fairly realistic. Currently Mininet can place OpenFlow reference user switches in separate namespaces and connect them by an emulated control network. This works, but it's slower than making multiple kernel switches with OVS. Also I haven't been able to get OVS (in user or kernel mode) to attach to an interface that is not in the root namespace, so the approach doesn't currently work with OVS. -Bob > > kevin > > _______________________________________________ > discuss mailing list > [email protected] > http://openvswitch.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list [email protected] http://openvswitch.org/mailman/listinfo/discuss
