On Thu, Aug 01, 2013 at 07:08:34PM -0700, Ethan Jackson wrote: > This patch retires a venerable library whose inception dates before > the first patch of the current repository: tags. They have served us > well, but their time has come for the reasons listed below. > > 1) They don't actually help much. > In theory, tags had been used to reduce revalidation necessary when > using bonds, mac-learning, and frequently changing flow tables. With > bonds and mac-learning, things change happen so rarely that tagging > isn't worth it. That leaves flow table changes. With the complex flow > tables in my testing, the revalidate_set gets so overwhelmed with > tags, that we end up revalidating every facet every time through the > run loop. In other words, they tags are giving us no benefit. > > 2) They complicate the code. > This patch simplifies the code and removes a couple of rather ugly > kludges. > > 3) They complicated locking once threading hits. > Because of the calculate_flow_tag() function, the table_dpif structure > would require locking in a multi-threaded OVS. Though this problem > isn't insurmountable, it's annoying and probably would cause lock > contention. > > Of course, we could try to work around these problems with a more > advanced tagging infrastructure, but this moves in the opposite of the > direction we should be. Ideally we'll have a more-or-less stateless > ofproto-dpif supporting a massive number of datapath flows. Tags (or > facets for that matter) aren't going to work in this new world. > > Signed-off-by: Ethan Jackson <et...@nicira.com>
I think that we should keep ofoperation_get_victim() and the related comment in ofproto_class, because it seems completely reasonable that some ofproto implementation would need this information. I am surprised that ofproto-dpif no longer needs it. Otherwise: Acked-by: Ben Pfaff <b...@nicira.com> _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
