I sat down with this patch series and ran a case of stamping out 800 of the following patterns (each in its own openstack project) to see how this patch set improves things within lflow_run:
n1 -- r1 -- x1 n1 is an openstack network with a dhcp subnet r1 is an openstack router x1 is an openstack provider network. The result is 800 networks with 2 ports each and one network with 800 ports on it. At the end of creating 800 of these patterns, this patch series results in a 2% improvement in the number of flows processed, a 10% improvement in the number of matches processed, and about a 16% improvement in the execution time of lflow_run. While this is a good first step and I'm in favor of merging this set, these patches can only be viewed as a short term fix as the number of processed flows and matches still grow supra-linearly as scale increases. This tells me that we still need some sort of incremental processing in this routine. Ryan IRC Handle: regXboi _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
