[LibreQoS] ookla speedtest results?

[dan]

We're pretty far past that hardware.  Minimum plans at most sites are now
100x20 and we're moving rapidly to 200x200-1Gx1G plans.  The implication
being that we need 4x4 AX wifi radios and that's real tough in an openwrt
compatible hardware right now, not a ton of options.

Flat out though, wisps in general simply do not have the option of
controlling every client side router or even sending out new/compatible
routers for each customer to pull this off.  It's a non-starter for us.
It may be better but it's just not an option that can be pursued.
ingress and egress edge shaping is what we need.

On Mon, Oct 24, 2022 at 11:57 AM Dave Taht <dave.taht at gmail.com> wrote:

> On Mon, Oct 24, 2022 at 10:46 AM dan <dandenson at gmail.com> wrote:
> >
> > Some sort of % drop is better than count from an operator's
> perspective.  I'm actually happy with 'scaling issues' of % on small vs
> large plan because those small plans are the one's we probably want to push
> for upgrades if they're seeing a lot of drops.
> >
> > That said, a 'score' method works really well from a high level.  And
> that can be proportionally adjusted so a slightly higher % in a small plan
> gives the same score as a larger plan would with fewer drops.
> >
> >
> > Also, on the SNMP statement above, we had the same issue with preseem
> being hyper aggressive to pull stats.  Some mikrotik versions have some
> snmp flaws and preseem can peg the CPU.  Hasn't happened in a while but
> that was an issue.
>
> I gave up on mikrotik long before they started working on routeros7,
> and just reflashed everything to openwrt.
>
> I have a nice little custom build for it assembled for the ubnt agws,
> and the mikrotik hap ac lite. Anyone using those?
>
> I would hope they don't have the snmp problem you describe, and if
> they did I could get it fixed in days.
>
> Vastly improved the wifi in particular, see before/after here:
> https://forum.openwrt.org/t/aql-and-the-ath10k-is-lovely/59002/
>
> (After about 40Mbits, the bloat starts moving to the wifi)
>
> Most of my work in the past 9 months has been on improving the
> mediatek mt76 derived products. I left off here:
> https://forum.openwrt.org/t/aql-and-the-ath10k-is-lovely/59002/908
>
> A lot of this stuff is shipping commercially in the openwifi project,
> if you haven't been paying attention to that.
>
> > On Mon, Oct 24, 2022 at 11:42 AM Dave Taht <dave.taht at gmail.com> wrote:
> >>
> >> On Mon, Oct 24, 2022 at 10:35 AM Robert Chacón
> >> <robert.chacon at jackrabbitwireless.com> wrote:
> >> >
> >> > > Presently (in v1.3.), robert (in grafina?) is summing drops + marks.
> >> > which is a good idea, however perhaps drops, marks and drops+marks
> >> > would be better.
> >> >
> >> > Yup, in InfluxDB for the stats by Tin (diffserv4) we display true
> drops as (trueDrops = ecn_mark + drops - ack_drops). But we don't do it by
> sub quite yet.
> >> > And ok, drops, marks and drops+marks can be shown instead - by
> subscriber or by Node (AP/Site).
> >> > Dave - is it ok if we graph this data as a percentage of packets
> sent? So for drops, drop / sent_packets ? I just imagine that  would make
> comparing different subscribers' connections easier.
> >>
> >> Drops and marks are proportional to the plan. You get a heck of a lot
> >> more drops at 5mbits than 50, and the relationship
> >> is not linear, closer to quadratic.
> >>
> >> So comparing customers at their given tiers makes more sense than
> >> across the board.
> >>
> >> As for a percentage, you should see a clear difference in percentages
> >> across tiers.
> >>
> >> I know I'm not quite answering your question well, I'm a big believer
> >> in just getting lots of data, graphing it every which way,
> >> and seeing what's useful.
> >>
> >> > Question for all - is having drops / marks/ drops+marks by
> subscriber/circuit worth the extra modest InfluxDB CPU use? Or should we
> just show those stats by AP/Node?
> >> > If preseem currently offers such stats at a subscriber level we can
> do the same, just wanna be sure I'm not overtaxing CPU.
> >> >
> >> > On Mon, Oct 24, 2022 at 11:14 AM Dave Taht via LibreQoS <
> libreqos at lists.bufferbloat.net> wrote:
> >> >>
> >> >> Some of the context of my request for libreqos on vs off is from
> here,
> >> >> presentation in the wee hours, wed morning.
> >> >>
> >> >>
> https://www.linkedin.com/feed/update/urn:li:activity:6989281945868283904/
> >> >>
> >> >> On Mon, Oct 24, 2022 at 9:48 AM dan via LibreQoS
> >> >> <libreqos at lists.bufferbloat.net> wrote:
> >> >> >
> >> >> > we also have a bunch of 25/5.  It would be nice to get some more
> detailed numbers from ookla tests.
> >> >>
> >> >> +1! We've argued methods with many a provider, and ookla and samknows
> >> >> are new to this game.
> >> >>
> >> >> As one example I don't know if they use decimal or binary mbits!
> >> >>
> >> >> ...
> >> >>
> >> >> ack-filtering in the libreqos case can and will drop acks in both
> >> >> directions, due to the structure of how htb
> >> >> works in this scenario, at any bandwidth ratio. I *think* it does
> very
> >> >> little harm but I would use the ack-filter
> >> >> not the aggressive ack filter out of caution.
> >> >>
> >> >> Secondly if you are gathering drop statistics from the higher level
> >> >> htb categories, total drops will go WAY up,
> >> >> and it is best to use the cake drop, ack_drop outputs statistics
> directly.
> >> >>
> >> >> ack_drops have nothing to do with congestion control and should be
> >> >> ignored in most cases.
> >> >>
> >> >> Also I keep hoping more will pull out ecn marks separately. Public
> use
> >> >> is way up by e2e transports, and I worry that a high ecn ratio is a
> >> >> e2e network error, which might show up as a big backlog.
> >> >>
> >> >> Presently (in v1.3.), robert (in grafina?) is summing drops + marks.
> >> >> which is a good idea, however perhaps drops, marks and drops+marks
> >> >> would be better.
> >> >>
> >> >> An inverse log scale is needed to plot marks, drops, and ack_drops.
> >> >>
> >> >> I'll always be of the opinion that it's best to run cake on egress at
> >> >> the cpe, first, doing all this work before it hits the wire.
> >> >> Are any of you doing that? Openwrt/dd-wrt/firewalla/mikrotik 7.2?
> >> >>
> >> >> >We also have a lot of Eero units out there, so watching for the
> eero speed test would also be great.
> >> >>
> >> >> eero 5s have cake (and fq_codel on the wifi), eero 6s have a somewhat
> >> >> buggy fq_codel offload:
> >> >>
> >> >>
> https://www.reddit.com/r/eero/comments/u7xm83/gen_2_sqm_vs_gen_3_sqm_stick_with_gen_2_if_you/
> >> >> > On Mon, Oct 24, 2022 at 10:31 AM Herbert Wolverson via LibreQoS <
> libreqos at lists.bufferbloat.net> wrote:
> >> >> >>
> >> >> >> > sniff DNS for those speed test lookups and trigger a packet
> capture for x duration?
> >> >> >>
> >> >> >> That's a good idea (I think there's a service somewhere for
> finding speedtest
> >> >> >> servers?). We're discussing testing the ack-filter feature right
> now, and something
> >> >> >> similar would work well for that. (We have a bunch of 25/5 and
> similarly asymmetric
> >> >> >> connections, in theory it'll help...)
> >> >> >>
> >> >> >> On Mon, Oct 24, 2022 at 10:31 AM dan via LibreQoS <
> libreqos at lists.bufferbloat.net> wrote:
> >> >> >>>
> >> >> >>> sniff DNS for those speed test lookups and trigger a packet
> capture for x duration?
> >> >> >>>
> >> >> >>> On Mon, Oct 24, 2022 at 9:27 AM Dave Taht <dave.taht at gmail.com>
> wrote:
> >> >> >>>>
> >> >> >>>> On Mon, Oct 24, 2022 at 8:20 AM dan <dandenson at gmail.com>
> wrote:
> >> >> >>>> >
> >> >> >>>> > Dave, I think that 'bouncing' is more CPU time in the
> browser.  I don't think it's indicative over what's actually happening.
> >> >> >>>>
> >> >> >>>> A simultaneous packet capture and wireshark plot of the RTTs
> would
> >> >> >>>> also be helpful.
> >> >> >>>>
> >> >> >>>> >
> >> >> >>>> > On Mon, Oct 24, 2022 at 9:09 AM Dave Taht via LibreQoS <
> libreqos at lists.bufferbloat.net> wrote:
> >> >> >>>> >>
> >> >> >>>> >> On Mon, Oct 24, 2022 at 8:02 AM Robert Chacón via LibreQoS
> >> >> >>>> >> <libreqos at lists.bufferbloat.net> wrote:
> >> >> >>>> >> >
> >> >> >>>> >> > I can run an Ookla test in the middle of the night with
> LibreQoS off for a bit.
> >> >> >>>> >>
> >> >> >>>> >> Groovy.
> >> >> >>>> >>
> >> >> >>>> >> > From what I recall - without Libre we see download bloat
> of 300ms or so. With Libre, it's gone. On waveform we see 0ms added bloat
> each direction for most clients.
> >> >> >>>> >>
> >> >> >>>> >> The numbers report by speedtest bounce around a lot, and
> they tend to
> >> >> >>>> >> pick a lower number as a final result than I'd
> >> >> >>>> >> like. If there's a way to capture a movie of it...
> >> >> >>>> >> >
> >> >> >>>> >> > And I like the "monitor only" mode idea.
> >> >> >>>> >> > Perhaps we could create the HTB tree, but just not attach
> the CAKE qdisc?
> >> >> >>>> >> > Technically even HTB could reduce latency by itself on
> overloaded APs , so it wouldn't be true passive monitoring, but it would
> allow us to compare the before/after of CAKE while still having qdiscs we
> can point cpumap-pping to.
> >> >> >>>> >> >
> >> >> >>>> >> > Alternatively, maybe we could use eBPF PPing just for this
> passive monitoring period?
> >> >> >>>> >> >
> >> >> >>>> >> >
> >> >> >>>> >> > On Mon, Oct 24, 2022 at 8:02 AM Herbert Wolverson via
> LibreQoS <libreqos at lists.bufferbloat.net> wrote:
> >> >> >>>> >> >>
> >> >> >>>> >> >> Are you looking for a shaped result? That'll show
> "whatever bandwidth
> >> >> >>>> >> >> is left on the sector" without Libre, and around the
> customer's paid-for
> >> >> >>>> >> >> target with Libre (hopefully!).
> >> >> >>>> >> >>
> >> >> >>>> >> >> One thing Preseem does well is that their onboarding
> process includes
> >> >> >>>> >> >> "leave it in monitoring-only mode" for a week, gathering
> data before you
> >> >> >>>> >> >> pull the switch (although most ISPs I've talked to just
> pull the switch...)
> >> >> >>>> >> >> It's quite enlightening to see a before/after, even with
> only FQ_CODEL
> >> >> >>>> >> >> as the shaper.
> >> >> >>>> >> >>
> >> >> >>>> >> >> We could probably do something similar with a "monitor
> only" mode that
> >> >> >>>> >> >> doesn't enable any queues (is there a TC equivalent to
> "dummy" on
> >> >> >>>> >> >> FreeBSD?).
> >> >> >>>> >> >>
> >> >> >>>> >> >> On Mon, Oct 24, 2022 at 8:52 AM Dave Taht via LibreQoS <
> libreqos at lists.bufferbloat.net> wrote:
> >> >> >>>> >> >>>
> >> >> >>>> >> >>> I am curious if anyone here has speedtest results with
> libreqos on and
> >> >> >>>> >> >>> off, and can send a screen shot?
> >> >> >>>> >> >>>
> >> >> >>>> >> >>> --
> >> >> >>>> >> >>> This song goes out to all the folk that thought Stadia
> would work:
> >> >> >>>> >> >>>
> https://www.linkedin.com/posts/dtaht_the-mushroom-song-activity-6981366665607352320-FXtz
> >> >> >>>> >> >>> Dave Täht CEO, TekLibre, LLC
> >> >> >>>> >> >>> _______________________________________________
> >> >> >>>> >> >>> LibreQoS mailing list
> >> >> >>>> >> >>> LibreQoS at lists.bufferbloat.net
> >> >> >>>> >> >>> https://lists.bufferbloat.net/listinfo/libreqos
> >> >> >>>> >> >>
> >> >> >>>> >> >> _______________________________________________
> >> >> >>>> >> >> LibreQoS mailing list
> >> >> >>>> >> >> LibreQoS at lists.bufferbloat.net
> >> >> >>>> >> >> https://lists.bufferbloat.net/listinfo/libreqos
> >> >> >>>> >> >
> >> >> >>>> >> >
> >> >> >>>> >> >
> >> >> >>>> >> > --
> >> >> >>>> >> > Robert Chacón
> >> >> >>>> >> > CEO | JackRabbit Wireless LLC
> >> >> >>>> >> > _______________________________________________
> >> >> >>>> >> > LibreQoS mailing list
> >> >> >>>> >> > LibreQoS at lists.bufferbloat.net
> >> >> >>>> >> > https://lists.bufferbloat.net/listinfo/libreqos
> >> >> >>>> >>
> >> >> >>>> >>
> >> >> >>>> >>
> >> >> >>>> >> --
> >> >> >>>> >> This song goes out to all the folk that thought Stadia would
> work:
> >> >> >>>> >>
> https://www.linkedin.com/posts/dtaht_the-mushroom-song-activity-6981366665607352320-FXtz
> >> >> >>>> >> Dave Täht CEO, TekLibre, LLC
> >> >> >>>> >> _______________________________________________
> >> >> >>>> >> LibreQoS mailing list
> >> >> >>>> >> LibreQoS at lists.bufferbloat.net
> >> >> >>>> >> https://lists.bufferbloat.net/listinfo/libreqos
> >> >> >>>>
> >> >> >>>>
> >> >> >>>>
> >> >> >>>> --
> >> >> >>>> This song goes out to all the folk that thought Stadia would
> work:
> >> >> >>>>
> https://www.linkedin.com/posts/dtaht_the-mushroom-song-activity-6981366665607352320-FXtz
> >> >> >>>> Dave Täht CEO, TekLibre, LLC
> >> >> >>>
> >> >> >>> _______________________________________________
> >> >> >>> LibreQoS mailing list
> >> >> >>> LibreQoS at lists.bufferbloat.net
> >> >> >>> https://lists.bufferbloat.net/listinfo/libreqos
> >> >> >>
> >> >> >> _______________________________________________
> >> >> >> LibreQoS mailing list
> >> >> >> LibreQoS at lists.bufferbloat.net
> >> >> >> https://lists.bufferbloat.net/listinfo/libreqos
> >> >> >
> >> >> > _______________________________________________
> >> >> > LibreQoS mailing list
> >> >> > LibreQoS at lists.bufferbloat.net
> >> >> > https://lists.bufferbloat.net/listinfo/libreqos
> >> >>
> >> >>
> >> >>
> >> >> --
> >> >> This song goes out to all the folk that thought Stadia would work:
> >> >>
> https://www.linkedin.com/posts/dtaht_the-mushroom-song-activity-6981366665607352320-FXtz
> >> >> Dave Täht CEO, TekLibre, LLC
> >> >> _______________________________________________
> >> >> LibreQoS mailing list
> >> >> LibreQoS at lists.bufferbloat.net
> >> >> https://lists.bufferbloat.net/listinfo/libreqos
> >> >
> >> >
> >> >
> >> > --
> >> > Robert Chacón
> >> > CEO | JackRabbit Wireless LLC
> >>
> >>
> >>
> >> --
> >> This song goes out to all the folk that thought Stadia would work:
> >>
> https://www.linkedin.com/posts/dtaht_the-mushroom-song-activity-6981366665607352320-FXtz
> >> Dave Täht CEO, TekLibre, LLC
>
>
>
> --
> This song goes out to all the folk that thought Stadia would work:
>
> https://www.linkedin.com/posts/dtaht_the-mushroom-song-activity-6981366665607352320-FXtz
> Dave Täht CEO, TekLibre, LLC
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.bufferbloat.net/pipermail/libreqos/attachments/20221024/5c37ea34/attachment-0001.html>