On Mon, Jun 08, 2015 at 10:54:55AM -0700, Jarno Rajahalme wrote: > > > On Jun 8, 2015, at 9:36 AM, Ben Pfaff <b...@nicira.com> wrote: > > > > This replaces the very specialized and difficult-to-maintain miniflow > > assembly macros by a simpler data structure named "miniflow_builder" that > > is more general and easier to use. > > > > It's not complete--it needs polishing and it mysteriously fails a few > > tests--but I'd like some early feedback. In particular, I know that the > > transformation to the current form was done for performance, so I'd like > > to make sure that this new code will not cause a regression. Therefore: > > how was the performance measured last time, so I can measure it the same > > way? > > > > Yes, the current form is pretty much optimal when looking at the > generated assembly, so I’d be surprised if anything more general is > faster. > > The only direct and really performance critical use for > miniflow_extract() is in the userspace datapath > (dpif-netdev.c). miniflow_extract() is called right before exact match > cache lookup, so any significant performance difference should be > visible in the simplest (single-flow) DPDK performance tests.
I did some simple measurements and the miniflow-builder version was significantly slower. I couldn't fix that simply, so I'm shelving this for now. I pushed the current version to my "ovs-reviews" repo as the branch "miniflow-builder". _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
