On Thu, Feb 07, 2013 at 10:04:14PM +0000, Rajahalme, Jarno (NSN - FI/Espoo) wrote: > > On Feb 7, 2013, at 22:36 , ext Ben Pfaff wrote: > > > With complicated flow tables and multiple levels of resubmit, I see > > flow setup performance improvements of up to 5-6% with this patch. > > > > I played around with the same issue last week and came up with a > little bit different solution. I think I spotted couple more locations where > the search can be stopped based on the max priority. > > I'll send my patches right away. There are two parts: > > 1. Maintain max priority of each table. > 2. Maintain tables in descending priority order. > > The latter should be beneficial e.g. for longest prefix match tables (such > as an IP routing table that is implemented by assigning higher priorities > for rules with longer prefixes) as tables with shorter prefixes need not be > searched after a match is found. > > I'd be interested to know how these perform with your test case, if you > care to test. Anyway, please feel free use these as you see fit.
Your patch gives me close to a 15% improvement (versus the 5-6% I see with my patch) and, furthermore, they move the classifier a direction I was thinking about going anyway. I'll review them with an eye toward application. Thanks a lot. Ben _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
