[Cerowrt-devel] CeroWRT and "FTTN" 50/10 VDSL2 (aka "FIBE")

[Dave Taht]

On Mon, Jun 30, 2014 at 8:35 AM, Michael Richardson <mcr at sandelman.ca> wrote:
>
> Dave Taht <dave.taht at gmail.com> wrote:
>     > On Fri, Jun 27, 2014 at 12:37 PM, Dave Taht <dave.taht at gmail.com> wrote:
>     >> On Tue, Jun 24, 2014 at 3:48 PM, Michael Richardson <mcr at sandelman.ca> wrote:
>     >>>
>     >>> On Monday I had VDSL2 installed at my home office.
>     >>>
>     >>> 50Mb/s down, 10Mb/s up from storm.ca. Native IPv6, but as far as I can tell,
>     >>> they aren't speaking DHCPv6; anyway I kept my /56 from my previous
>     >>
>     >> Cool. I note that the wndr3800 runs out of horsepower at about 50mbit/10.
>
>     > And when I say that, I mean it. HTB eats all the cpu, leaving nothing
>     > for wifi , wifi crypto, the web interface, etc, at these rates. I have
>     > long been looking for something more efficient than htb without much
>     > luck.
>
> Right now, with it still set at 50000, I see 49.86Mb/s from my ISPs' speedtest.storm.ca.
> I'll take your advice and turn off download policing.
>
>     > So assuming your downstream bloat isn't "bad", and you stick with cero
>     > on this interface, I'd disable the downstream rate limiter entirely
>     > (set it to 0), and just use sqm on the upstream. It is increasingly
>     > hard to hit the higher bandwidth limits with normal traffic.
>
>     > (but I'd love a rrul measurement all the same... )
>
> I'll try to get it setup.... anecdotally, my ssh is really fast all the time.
> I spend a lot of time ssh'ed into my SOHO desktop from my office, running
> emacsclient in "-nw" mode.  In the pre-CODEL days, this resulted regularly in
> me getting 20-40seconds ahead of the screen updates.  Post cerowrt/CODEL,
> there were occasional periods of time when codel couldn't protected my typing
> SSH from the ssh that was running sshfs/rsync.
> (sshfs/rsync needs to invoke ssh with some option to make it apply DSCP to
> the packets)

I switched to mosh for terminal sessions 2+ years ago and I'm never
going back. Also, I'd supplied a patch to mark those packets as AF42,
as well, which moves it up in priority in the SQM system. This was in
mainline mosh for a while, until it was found to mess up two systems
(out of 10s of thousands, 2 would stop forwarding AF42 marked packets
after 10 seconds), and rather than switch it to tos immediate, they
merely reverted the patch.

This probably won't re-apply but would be simple to add back in.

http://snapon.lab.bufferbloat.net/~d/0001-Set-AF42-on-mosh-by-default.patch


Secondly, I made patches available for rsync's native protocol to use
another tcp algorithm and mark the packets as CS1 (background), also
two years ago. (I don't know how to do the right thing to ssh, but
should probably get around to it one day)

These two patches should (in addition to fq_codel) reduce your
interruptions of interactive service from pretty darn good with
fq_codel, to nearly never, with sqm.

https://git.samba.org/cifs-utils.git/?p=rsync-patches.git;a=history;f=congestion.diff;h=5fb8a79edd399b6e8cdefb251dce0e53c0951c65;hb=7578744b8595da0d9ee785c96c74fff810e721ea

I don't think this landed in rsync head, either.

>
> I'm think that with 10Mb/s up, I just won't care :-)

Measure...
>
> --
> ]               Never tell me the odds!                 | ipv6 mesh networks [
> ]   Michael Richardson, Sandelman Software Works        | network architect  [
> ]     mcr at sandelman.ca  http://www.sandelman.ca/        |   ruby on rails    [
>



-- 
Dave Täht

NSFW: https://w2.eff.org/Censorship/Internet_censorship_bills/russell_0296_indecent.article