Thank you for your reply, Ben. Some heads up, we changed our flow idle_timeout to 300, then ran the test again, and got around 6000 packet-in per second. After some investigation, it turns out that most of the ovs-vswitchd CPU cycles are consumed by add flow, like classifier_rule_overlaps(), miniflow_equal_in_minimask(), miniflow_get__ ()...
On Tue, May 27, 2014 at 11:02 PM, Ben Pfaff <b...@nicira.com> wrote: > On Mon, May 26, 2014 at 09:55:33PM +0800, lin shaopeng wrote: > > I am currently running tests to test packet-in performance of OVS > > (ovs-2.0.1). > > By starting up a large number of different connections simultaneously, > > causing a bunch of packet-in messages sent to the controller. It turns > > out that OVS can only handle about 3000 packet-in messages per second, > > then the CPU usage of ovs-vswitchd reaches up to 100%, while the thread > > dispatcher and miss_handler are still almost idle.(my CPU is intel E2640 > > 2.5GHZ) > > > > Is this normal? Or something wrong with my configuration. > > > > I notice that, openflow messages, iterations for updating all the > bridges, > > ports, > > and connections configurations are all handled by one single > > thread(ovs-vswtichd), is there any plans to implement multi-core support > to > > improve the performance? > > I don't know of anyone working on that. Feel free to submit patches to > improve packet-in performance. >
_______________________________________________ discuss mailing list discuss@openvswitch.org http://openvswitch.org/mailman/listinfo/discuss
