On Thu, Mar 06, 2014 at 05:20:23PM -0800, Joe Stringer wrote: > This series implements a cache for xlate_actions() so that full flow > translation does not need to be performed for long-lived flows; instead, > references are kept to the modules affected by it, and this cache can be used > to attribute stats and effect other side-effects of xlation. > > This patch series improves CRR performance mildly in the typical case; > modestly > for cases where revalidation is constantly occurring; and makes the work for > revalidator threads much more efficient, allowing more flows to be maintained > in the datapath. Note that patch #8 causes revalidators to remove any flows > that are not experiencing high throughput. > > The "modules" which logic is implemented for are: > * Stats attribution: > + Rules > + Netdevs > + Bonds > + Mirrors > * Side effects: > + Learn actions - hard timeout refresh > + Normal actions - mac learning > > There are minor testsuite failures in the later patches, but as a whole this > is > fairly complete. I am looking for feedback on the approach, specifically > whether I may have missed some modules, if there are any holes in the logic > where typical operation may cause failures, and if there is perhaps areas that > I do not go far enough. I do plan to look into shifting ukey creation into > handler threads, but that is a later topic.
I applied patches 1 through 4 to master. I think I'm happy with patches 5 through 7, but I want Ethan's feedback on them. You noted that the later patches cause testsuite failures. We'll need to resolve those, one way or another, before we can apply them. Any particular feedback you're looking for in the meantime? _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
