On Thu, Feb 19, 2015 at 12:14 PM, Dave Taht <dave.t...@gmail.com> wrote: > On Thu, Feb 19, 2015 at 11:18 AM, Ted Lemon <mel...@fugue.com> wrote: >> On Feb 19, 2015, at 2:11 PM, Mikael Abrahamsson <swm...@swm.pp.se> wrote: >>> I'd imagine it's easier to do AQM on routed ports instead of switched ports >>> as well, that's where I can imagine CeroWRT choosing this approach. >> >> I don't think it is easier to do AQM on routed ports. If you do the easy >> version of AQM for switched ports, that's easy; if you do per-port AQM I >> think it's equally hard in both cases. > > In either case AQM and FQ has to be per port in the switch hardware > itself. I have been patiently awaiting for that sort of hardware to > arrive. > > On most consumer routers the cpu forwarding path isnt capable of gbit > rates in the first place. > > If you are doing software bridging (say between wired and wireless), > you apply fq_codel to the underlying interfaces, not the bridge, and > you are golden, except for dealing with multicast... > >>The plumbing for the L2 solution would look different than for the L3 >>solution, and that might make it trickier, though. >> >> CeroWRT did this, as far as I understand it, because Dave wanted to shake >> out cross-subnet issues, not because he felt it was technically preferable >> in the long run, but perhaps I am misrepresenting his position on the topic. > > You are correct, I primarily chose routing between tons of interfaces > precisely to expose what cross-subnet issues existed in the real > world, and to make it be soooo painful as to inspire those writing > code that overused multicast to find unicast solutions! > > The core things that break are things we long ago identified - mdns > and local service discovery, notably.
Also firewalling became a log(nports) problem without the special solutions in cerowrt. That became AMAZINGLY slow when we got past 4 interfaces. It still is quite slow even with the pattern match we use there. >(I have generally found that > every android app lets me plug in an ip address for a given service, > so for me it hasnt been a huge problem - except that not having a > dynamic ipv6 to name mapping makes dynamic ipv6 unusable in such > cases) > > The net result is a lot of cero people said to hell with that and went > back to bridging everything, instead of working harder on mdns-proxy, > cidr, etc. > > This microcosm of induced routing pain was also intended to show how > multicast didnt scale at all into larger wifi networks in the the > small business, or educational campus, that are bridged to wifi, here > the problems induced by multicast are so crippling as to leading to > the vast establishment of things like AP isolation, which disconnects > everyone from everyone else, on the same AP, which to me, is a > terrible solution. > > I do note that not bridging ethernet and wifi together in cero has led > to making wifi quite a bit more pleasant and less jittery on apps like > webrtc, at least for me, and there is work in play on improving wifi´s > aggregation handling and multicast queue management coming up soon. > >>> It's good that we're having this discussions since I seem to not be the >>> only one thinking that there should be one port per subnet. > > I do like occasionally having more than one vlan, but in the general case, > routing is far more expensive than switching. > > The no-nat case here shows the impact of a current 44 entry routing > table on downloads, in the present linux 3.18 fib lookup system. > > http://snapon.lab.bufferbloat.net/~d/openwrt-3.18.7/openwrt-wndr3800-3.18.7.svg > > and also shows the performance difference between ipv6 and ipv4 on > this hardware. > > There has been some GREAT work on improving linux´s fib lookup system > of late, but hardware bridging will always outperform software > bridging which will always outperform routing, IMHO. > > Anyway I cant imagine a homeowner wanting more than 2-3 vlans at most, > and most, just one. There is no problem scaling an ethernet broadcast > domain to 4096 devices. Using up 16 ports on 16 different /64 subnets > is kind of crazy, especially since at least one provider (comcast) is > only supplying a /60 in the first place. > > wifi on the other hand, barely scales to 30 active stations. > > >> Indeed, there are two of you! >> >> _______________________________________________ >> homenet mailing list >> homenet@ietf.org >> https://www.ietf.org/mailman/listinfo/homenet > > > > -- > Dave Täht > > thttp://www.bufferbloat.net/projects/bloat/wiki/Upcoming_Talks -- Dave Täht thttp://www.bufferbloat.net/projects/bloat/wiki/Upcoming_Talks _______________________________________________ homenet mailing list homenet@ietf.org https://www.ietf.org/mailman/listinfo/homenet
