On Fri, Aug 16, 2013 at 07:25:18AM +0900, YAMAMOTO Takashi wrote: > > On Thu, Aug 15, 2013 at 08:25:44PM +0900, YAMAMOTO Takashi wrote: > >> move per-backer wait calls (dpif_wait/udpif_wait) from ofproto_wait > >> to ofproto_type_wait. > >> > >> this fixes O(N^2) behaviour (eg. poll slot consumption) for > >> dpif-netdev where N is the number of bridges. > >> > >> Signed-off-by: YAMAMOTO Takashi <y...@mwd.biglobe.ne.jp> > > > > Applied, thanks. > > > > I couldn't see how this was O(N^2), although I agree that it is > > excessive, so I changed the second paragraph above it: > > > > This eliminates excessive poll slot consumption when there is more > > than one bridge. > > a backer for N bridges would have M ports. > M >= N because of local ports. > > each bridges was doing dpif_wait, which does rx wait calls for each ports. > it costs N*M >= N^2.
OK. Anyway, I applied this with the explanation above. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
