On May 9, 2013, at 9:41 AM, Anoob Soman <anoob.so...@citrix.com> wrote:
> There are few patches submitted for mega flows in last couple of weeks. I > have a vague idea on what mega flows could be. Is it possible to share some > more detail on how exactly we would be doing the mega flow match in the > kernel. Couple of months ago I did an experiment with ovs, which tried to > reduce the granularity of flows in kernel in order to reduce the number of > packets queued to userspace (flows miss). In order to make this work I sent > only layer 2 flow information (src-mac, dst-mac, next-proto) from kernel > space to userspace. Thus only layer-2 based flows were created in kernel > space. This worked well for situation where there were large number of short > lived flows. Is mega flow similar to this ? Yes. Basically, the plan is to support wildcarding in the kernel. There is some complication in userspace, since it needs to understand when it can push a wildcard entry versus exact-match. I'm working on some patches that are based on a prototype by Ethan that will allow us to determine how much wildcarding we can do with a kernel rule. You can expect to see that going out for review within the next week. --Justin _______________________________________________ discuss mailing list discuss@openvswitch.org http://openvswitch.org/mailman/listinfo/discuss
