> I am interested to learn what people think about whether equal-cost > multi-path routes are needed in homenet. Given the previous discussion > about parallel wireless links - which I know I have in my house and can't > use - I've been wondering if these have been considered.
As Toerless noted, source-specific routing combined with something like MP-TCP gives similar effects to ECMP. Clauz from the Ninux project gave a talk at Battlemesh with his experiences with such a setup, however, and the conclusion was that in his particular network he couldn't get it to work well enough to be useful for his particular application (VPNs to a central server). https://www.youtube.com/watch?v=ztVogLI_ZJs (That was a great talk, by the way. We need more talks about what people were unable to get to work, talks about successes are boring.) As to ECMP itself -- in a wireless environment, it must be done with care. If you have two parallel links at the same radio frequency, you will get collisions between the two subflows, which will get you worse performance than what you'd get by using a single subflow. So your routing protocol needs to be able to distinguish between interfering and non-interfering routes, and only do ECMP on non-interfering subflows. In the particular case of Babel, this means that in order to efficiently do ECMP over wireless you need to use the "Babel-Z" extension [1], which propagates interference information in order to make better routing decisions in multi-radio meshes. I've been refraining from mentioning this extension on this list, since I've repeatedly been told that meshy things are out of scope for Homenet. [1] https://tools.ietf.org/html/draft-chroboczek-babel-diversity-routing Note that this extension is not frozen yet, I am planning a backwards-incompatible change before I submit it for publication as an RFC. (The TLV format will stay, but the interpretation of some values will change; existing implementations will not be impacted.) Babel does not do ECMP yet (it does do feasible successors, of course). If there's interest, and somebody willing to do some testing for me (empirical data only, please), then I think I could conjure something up pretty quickly. -- Juliusz _______________________________________________ homenet mailing list homenet@ietf.org https://www.ietf.org/mailman/listinfo/homenet