On Tue, Sep 25, 2018 at 2:27 PM Juliusz Chroboczek <[email protected]> wrote: > > >> B) you run out of cpu - babeld uses linked lists, > > Not quite -- the route table is a flat array, with per-destination > linked-lists. We don't walk the linked lists often, we mostly walk the > flat arary. > > >> and tries to recalc bellman-ford every 4 seconds also. > > No, it doesn't. It only does the Bellman-Ford dance when it receives > a new update -- and the computation is incremental.
ok. I stand corrected on both counts. > >> my rtod tests showed babeld typically falling over for any one of > >> these four reasons in well under 4k routes on low end mips and arm > >> hardware. Even the low end apu2 eats a whole cpu with about that many > >> (ipv6) routes. > > Dave, these are interesting results -- please send me your profiles. The > one known major inefficiency is in parsing xroutes, if you've got others, > I'm interested. A big part of the dance was duking it out with odhcpd. I'm not in a position to not run that. All this data - and my interpretation of it - was old, I merely tried to make things "crash" last night. I did. I think a key to all that damage I did with less routes last time was that I had a lot more routers on the same link and an infinite mcast queue on some of them. I made the rtod tool available for those that wanted to do more 2 years ago. The core part of the gprofile - where I stopped work after ripping out enough of memcmp to finally see it - was after figuring out that a goodly part of the problem was in the non-qsort thing. both you and I hoped someone else would tackle it. You can't see it on the gprofile without putting in all the xor stuff I did. Aside from that... what I'm trying to do right now is upgrade my ~6 core routers to lede 18.06.1 - which I can use as a base for future work. I put in the patches I was using to keep it stabler before (just posted here earlier today), and I just got off the roof with a nanostation. Nearly all my core routers date from about this era..... Linux roof-to-office.lodge.lupinlodge.org 3.18.7 #1 Thu Feb 19 15:51:50 PST 2015 mips GNU/Linux so... assuming I don't fall off the roofs and can get this upgrade online and stable I'll be able to look better at real data again. :) Next week. > > >> I think making some version of babel (be it bird or frr ) scale well > >> to at least 64k routes would be a very good idea, > > > I agree. > > 64k sounds like an achievable goal. > > > We are just starting another test network for a city-wide mesh which > > will be based on babeld. > > Excellent. Could you please send me some more detailed info? (With > permission to include it in my next activity report -- my boss is cool > with my hacking Babel, but he understandingly feels more comfortable when > I can provide him with hard data to put into the activity report.) > > Thanks, > > -- Juliusz > > _______________________________________________ > Babel-users mailing list > [email protected] > https://alioth-lists.debian.net/cgi-bin/mailman/listinfo/babel-users -- Dave Täht CEO, TekLibre, LLC http://www.teklibre.com Tel: 1-669-226-2619 _______________________________________________ Babel-users mailing list [email protected] https://alioth-lists.debian.net/cgi-bin/mailman/listinfo/babel-users
