[Cake] Re: [Codel] [Rpm] Re: [Bloat] Re: "Fi-Wi is a new forwarding plane for wireless" - Bob McMahon

[Sebastian Moeller <moeller0@gmx.de>]

Hi Frank,

(I shortened the CC list, I believe the list sghould suffice, no?)

> On May 25, 2026, at 14:45, Frantisek Borsik <frantisek.borsik@gmail.com> wrote:
> 
> Not saying that modern Wi-Fi has no issues or that these issues will be mitigated at all (let alone any time soon), but compensated with FQ-CoDel and CAKE, coupled with QoE middle-box and proper peering, it's more than just good enough.

I respectfully disagree, and I am not saying this to spite you, but because I am currently running some WiFi experiments in my home network, and to my taste it is not good enough. And that is with OpenWrt APs that already support:

[ TXQS ]: FQ-CoDel-enabled intermediate TXQs 
[ AIRTIME_FAIRNESS ]: airtime fairness scheduling
and, where applicable
[ AQL ]: Airtime Queue Limits (AQL)

with the following aql twaeks
# AQL Tweaks
aql_txq_limit_l=2500
aql_txq_limit_h=8500
for ac in 0 1 2 3; do echo $ac $aql_txq_limit_l $aql_txq_limit_h > /sys/kernel/debug/ieee80211/phy0/aql_txq_limit; done
for ac in 0 1 2 3; do echo $ac $aql_txq_limit_l $aql_txq_limit_h > /sys/kernel/debug/ieee80211/phy1/aql_txq_limit; done

so from the wifi side this is close to as good as it currently gets for Linux/OpenWrt with open source drivers (for WiFi4/5/6 APs)

and still under WiFi-limited scenarios h++ps://test.libreqos.com shows latency excursions into thte 500-900ms range. For my taste that certainly leaves ample room for improvements. Sure, when I get closer to the AP so WiFi throughput exceeds WAN throughput my WAN shaper kicks in and I get decent test results (with no noticeable increase in latency-under-load), but that is a work-around only.

> And it's here, at the our disposal. This knowledge it's just not evenly distributed and used at scale.

Again, knowledge is fine, but not sufficient, if I run my APs under-capacity (that is if my WAN shaper limits traffic rates to well below WiFi capacity) things look "solved" but the moment the AP is further away and the signal gets weaker things go pear shaped fast. IMHO this is not a roaming issue (sure it would be nice if stations would always pick the best AP) but really a "performance under saturating conditions" issue that ideally should be fixed, as "never let the WiFi become the bottleneck" is not a real solution and also IMHO not what "make WiFi fast" was all about.

> 
> I'm not claiming any special knowledge of Jonathan's work on his new generation of queuing discipline(s), but I remember discussions with Dave re: getting into micro-seconds territory. I still do believe that super smart people from these lists can dig up something helpful from Dave's scoping of CAKEMQ:
> https://docs.google.com/document/d/1tTYBPeaRdCO9AGTGQCpoiuLORQzN_bG3TAkEolJPh28/edit?tab=t.0
> 
> As for 3GPP playbook, not at all. Wi-Fi is getting objectively better, so all I'm saying is that we can reasonably expect some improvements on it with Wi-Fi 8. 
> This is nothing close to 3GPP B(ritish) S(tandarts) playbook :)

From my perspective you just drunk the IEEE cool-aid ;)

> 
> Let me share a story. Back then in 2023, Dave sent me to Wi-Fi Alliance meeting here in Prague with a list of pleas, asks and questions. There was a few great people from the bufferbloat mailing lists and LibreQoS customers helping to put it together. And one great person from Qualcomm got me in, made introductions to everybody and their mother: from Broadcom, to Intel, to MediaTek, to HPE, to Juniper, to Cisco, to his Qualcomm colleagues. What I've learned? They are not giving a flying F about really fixing bufferbloat. For the most part, most of them, with REAL deal, like baking FQ-CoDel and CAKE into Wi-Fi. All they will do is some virtue signalling at best, like introducing a "Wi-Fi QoS management" and other things, which are marginal, potential improvements, at best.

That seems close to the 3GPPP playbook to me, over-promise, under-deliver. That is not to say that 3GPP or IEEE do not advance their respective technologies, but not in the specific way that we are discussing here, I feel we are just a bullet point in their marketing slides, not a core engineering issue (and that is sort of OK, I just wish marketing would soften the latency claims until engineering is actually "on it").

But see above, the Linux WiFI stack already includes fq-codel and still performance under saturating conditions is less than ideal.

> 
> We will have more vendors adopting FQ-CoDel and CAKE (and it's iterations) over time, but other than that, it will be QoE middle-boxes that will really manage to deliver these improvements at scale, in more than just ISP networks out there. University campuses, venues, airports, manufacturing plants, mines, cruise ships, aircraft...

See, I am in this long enough to understand that this is really really sub-optimal for variable-rate links. Will FiWi solve all of these issues? Likely not, but it addresses (some of) them at their root, which seems the right thing to do. One of the open questions is, will FiWi actually make live better for the billions of existing WiFi devices in the field, if it does, more power to it, as IEEE's approach will likely not help there at all.

> 
> So QoE middle-boxes and bufferbloat, Open Source, communities, will be saving the day.

I like the spirit!

> Can Fi-Wi find some niche use case or help Wi-Fi to fix some of its inherent issues? Absolutely.

Not sure this and the sentence before are mutually exclusive though.

> 
> All the best,
> 
> FrankFrantisek (Frank) Borsik
> 
> In loving memory of Dave Täht: 1965-2025
> https://libreqos.io/2025/04/01/in-loving-memory-of-dave/
> 
> https://www.linkedin.com/in/frantisekborsik
> Signal, Telegram, WhatsApp: +421919416714 
> iMessage, mobile: +420775230885
> Skype: casioa5302ca
> frantisek.borsik@gmail.com
> 
> 
> On Mon, May 25, 2026 at 8:59 AM Sebastian Moeller <moeller0@gmx.de> wrote:
> 
> 
> On May 25, 2026 7:43:01 AM GMT+02:00, Frantisek Borsik <frantisek.borsik@gmail.com> wrote:
> >Hey Bob,
> >
> >This is all nice and all, but it really doesn't matter. The only measurable
> >term IMO we really need is this - is the setup we are using good enough and
> >delivering all we need? And the answer for ISP like Dan and many that will
> >come "to Jesus" is sounding YES. It's here, it's cheap, it's getting better.
> 
> Not sure I fully buy this, if I understand correctly, Dan uses traffic shaping to "neuter" WiFi from its worst self congestion tendencies... that is a clever way to work around clear short comings in modern WiFi not evidence that modern WiFi has no issues...
> 
> >
> >You will get your evidence and then what? When (and big iF) Fi-Wi will
> >materialize, we will have Wi-Fi 8
> ><https://www.linkedin.com/feed/update/urn:li:activity:7460998580456116224/>
> >for those interested in chasing marginal improvements and the latest thing.
> 
> That is the 3GPP playbook, always overpromise for the coming generation N+1 and once that materilalizes and typically under-delivers roll-over the unfulfilled promises simply into N+2, maybe add a few more claims... Sure fitting for the latexstage capitalusm we find ourselves in, but let's not get fooled by marketing promises and judge WiFi8 once it exists by its merits/performance, please.
> 
> >Jonathan Morton will deliver his improvement on CAKE, we might even get
> >into microsecond territory.
> 
> Last time I researched his deltic approach it looked very much not going into the microsecond territory, I remember 25ms as latency target somehow... 25ms is certainly not bad, and not high-latency, but seems different from ultra-low-latency approaches. So I might have missed some interesting development, if you have a link you could share to get me up to speed again, I would be grateful.
> 
> > There will be FQ-PIE, PURPLE CAKE, HTB-MQ (to
> >supplement CAKE-MQ)...
> >
> >Anyway, you will like this:
> >https://arstechnica.com/gadgets/2017/02/going-hands-on-and-behind-the-scenes-at-the-plume-wi-fi-hq/
> >
> >Plume used to do this, you will have your Fi-Wi house test bed up and
> >running, soon. Let's have fun and do another round of the good ole
> >Wi-Fi/mesh router rumble:
> >
> >https://arstechnica.com/gadgets/2016/09/the-router-rumble-ars-diy-build-faces-better-tests-tougher-competition/
> >https://arstechnica.com/gadgets/2016/12/review-comparing-google-wifi-to-other-mesh-networking-heavyweights/
> >
> >Maybe Jim and Ars Technica would be interested in it.
> >
> >
> >All the best,
> >
> >Frank
> >
> >Frantisek (Frank) Borsik
> >
> >
> >*In loving memory of Dave Täht: *1965-2025
> >
> >https://libreqos.io/2025/04/01/in-loving-memory-of-dave/
> >
> >
> >https://www.linkedin.com/in/frantisekborsik
> >
> >Signal, Telegram, WhatsApp: +421919416714
> >
> >iMessage, mobile: +420775230885
> >
> >Skype: casioa5302ca
> >
> >frantisek.borsik@gmail.com
> >
> >
> >On Sun, May 24, 2026 at 11:57 PM <bob.mcmahon@umbernetworks.com> wrote:
> >
> >> Hi Frank,
> >>
> >> The question needs to be framed in measurable terms.
> >>
> >> The systems you named operate at different layers than the 802.11 MAC/PHY
> >> layer where TXOP allocation, retries, EDCA contention, AMPDU behavior, and
> >> airtime utilization occur throughout the building. batman-adv routes
> >> between mesh nodes. CAKE and FQ-CoDel operate on IP-layer queues. QoE
> >> middle-boxes operate from the WAN/IP side.
> >>
> >> So the question is whether IP-layer AQMs or WAN-side QoE systems affect
> >> measurable 802.11 MAC/PHY behavior under load.
> >>
> >> If they do, the predicted MAC/PHY effects would include:
> >>
> >>    - lower failed TXOP percentage
> >>    - lower retransmission airtime fraction
> >>    - improved AMPDU completion efficiency
> >>    - higher payload delivered per TXOP
> >>    - lower AMPDU truncation frequency
> >>    - reduced EDCA wait distributions
> >>    - improved airtime utilization
> >>    - reduced service-time variance
> >>    - changed MCS distributions across stations under realistic spatial
> >>    topology, particularly at coverage edges and where multiple APs or RRHs are
> >>    candidates for node selection
> >>
> >> That is the measurement gap I am pointing at. The industry has extensive
> >> 802.11 feature lists and marketing claims, with comparatively little open
> >> tooling to falsify what those actually do to building wide 802.11 MAC/PHY
> >> behavior under real multi-client, multi-AP load.
> >>
> >> I developed and still maintain iperf2 because throughput alone has proven
> >> an insufficient measurement model for network behavior, e.g. latencies and
> >> responsiveness under load. Features added to iperf2 include:
> >>
> >>    - trip-times
> >>    - bounceback testing
> >>    - one-way IP packet timing measurements
> >>    - packet-level latency histograms
> >>    - message-level (TCP write/read completion) latency histograms
> >>    - responsiveness under load measurements
> >>    - enhanced per-interval reporting
> >>    - CWND reporting
> >>    - RTT reporting
> >>    - bytes/packets in-flight reporting
> >>    - TCP retransmission reporting
> >>    - TCP congestion-control selection (CUBIC, BBR, Prague, etc.)
> >>    - udp-l4s including CE counters and durations
> >>    - isochronous traffic generation
> >>    - socket-layer pacing via SO_MAX_PACING_RATE
> >>    - token-bucket write-rate control
> >>    - Markov-chain packet-size generation
> >>    - UDP latency/jitter analysis
> >>    - multicast testing
> >>
> >> Those measurements operate above the 802.11 MAC/PHY. The missing layer
> >> remains open 802.11 MAC/PHY telemetry, which is what I'm building now.
> >>
> >> This measurement standard should apply equally to any claim from any
> >> architecture, including Fi-Wi.
> >>
> >> The question becomes: does centralized MAC telemetry and coordinated
> >> airtime control affect measurable 802.11 MAC/PHY behavior under load
> >> relative to autonomous AP operation?
> >>
> >> Measure the same MAC/PHY metrics under equivalent topology, offered load,
> >> and client mix. If those measurements do not improve relative to baseline,
> >> the architectural claim of Fi-Wi fails. We won't know until Fi-Wi is
> >> designed, built, deployed and measured at many different locations. The
> >> theoretical case is strong.  Centralized scheduling, joint optimization,
> >> and observability all argue for it. The design and measurements are the
> >> next steps. Then we'll see based on evidence.
> >>
> >> Bob
> >>
> >> Technically, Betamax was superior to VHS, and yet...
> >>
> >> If we would be talking about FiWi pre-jump from Wi-Fi 5 to Wi-Fi6/E/7, I
> >> would even try to do my best and ignore https://fastgood.cheap.
> >> But we are not there, for better or for worse.
> >>
> >> ISPs like Dan that use Wi-Fi 6 (with mesh, especially based on batman-adv
> >> and FQ-CoDel / CAKE) and above at customer premises and Quality of
> >> Experience middle-box on their last-mile, while providing good, personal
> >> customer service (not some crappy outsourcing of it to India or "AI") will
> >> be hard to beat, even for a big telco/ISP offering cheaper price. Yes, some
> >> people will leave but the most of them will be coming back.
> >>
> >> This is, believe it or not, not a high bar to jump over, even though not
> >> many ISPs are getting it already. We all know that more or less: "Bandwidth
> >> is a lie, Bandwidth is dead," so it's coming. They will be throwing more
> >> bandwidth on it for some time, but we will be getting from "innovators" to
> >> "early adopters" soon and then, all we need is just that "crossing the
> >> chasm."
> >>
> >> Despite all the difficulties, it will be faster and cheaper - also "good
> >> enough," than FiWi. Or rather, it's here already, it's just not evenly
> >> distributed.
> >> There might be some good niche use case for FiWi, though, once it will
> >> mature and get there.
> >>
> >> All the best,
> >>
> >>
> >> Frank
> >>
> >>
> >>
> >> Frantisek (Frank) Borsik
> >>
> >>
> >>
> >> *In loving memory of Dave Täht: *1965-2025
> >>
> >> https://libreqos.io/2025/04/01/in-loving-memory-of-dave/
> >>
> >>
> >>
> >> https://www.linkedin.com/in/frantisekborsik
> >>
> >> Signal, Telegram, WhatsApp: +421919416714
> >>
> >> iMessage, mobile: +420775230885
> >>
> >> Skype: casioa5302ca
> >>
> >> frantisek.borsik@gmail.com
> >>
> >> On Sun, May 24, 2026 at 2:12 AM dan <dandenson@gmail.com> wrote:
> >>
> >>
> >>
> >> On Sat, May 23, 2026 at 10:36 AM <bob.mcmahon@umbernetworks.com> wrote:
> >>
> >> > I don't have arguments on the technical bits of your reply.
> >>
> >> The industry hasn't provided open source tools to analyze 802.11 properly.
> >> The proprietary ones are tightly coupled to chip firmware, owned by 802.11
> >> vendors, and not for sale.
> >>
> >> You've built solid networks, especially given how little 802.11 MAC-layer
> >> and MCS observability current tooling exposes. I hope to make open source
> >> 802.11 MAC telemetry tools available sooner rather than later. These will
> >> run on both an ESP32-C5 and an RPi5. The inexpensive ESP32 allows monitors
> >> to be placed throughout a venue at low cost.
> >>
> >> A key issue is what's being measured. Your deployment data is a capacity
> >> analysis. Once baseline capacity is adequate, user experience is driven by
> >> tail latency and service-time variance rather than throughput. A network
> >> can saturate aggregate throughput while still showing large service-time
> >> variance and long per-flow tails under contention. These are different
> >> measurements.
> >>
> >> And CODEL and CAKE are IP-layer mechanisms. They don't operate on 802.11
> >> TXOPs or airtime sojourn, which is where the contention and service-time
> >> variance actually live. AQL is closer. It operates on airtime at the
> >> driver/firmware boundary, but it's still per-AP. There's no coordination
> >> across APs and no MAC telemetry exposed upward.
> >>
> >> Slide 6 of my DPDK Summit Stockholm talk lays out the layering:
> >> https://www.umbernetworks.com/DPDK_WiFi_Stockholm_Pres.html
> >>
> >> Check out iperf2's advanced features around --bounceback, --trip-times,
> >> and --histograms. These are available as open source. Man page:
> >> https://iperf2.sourceforge.io/iperf-manpage.html
> >>
> >> I'll try to respond with some metrics soon. My rig is down at the moment,
> >> so give me a few days.
> >>
> >> Bob
> >>
> >>
> >> I agree.  However, I'm not actually just measuring throughput, but running
> >> latency sensitive applications without issues.  Frank might attest to my
> >> efforts to reduce latency and I routinely share data on lqos (and another
> >> QoE product's) TCP measurements for latency and among QoE users, I believe
> >> my network is top 1% in latency and jitter and in the WISP space.
> >> Obviously I'm not going to compete with XGSPON operators.. (though I'm only
> >> measuring TCP because that's the 'easiest' to measure passively
> >> practically).  While this is not an engineering adequate benchmark, if my
> >> VoIP handsets work over wireless then the wireless is good is a very
> >> reasonable latency, jitter, loss argument.  And I run a few brands of WiFi
> >> based VoIP phones (Fanvill Wxxx series, Yealink Wxx and AXxx series).
> >> Similarly, if the sonos works and is in sync, the wifi is good.
> >>
> >> As a service company and as a user, I don't really care so much where the
> >> goodness is in the system, mac layer or IP having cake work it's magic.
> >> Basically running an AP on OFMDA (as much as possible) and well under the
> >> 'red line' capacity delivers great results on WiFi7 radios (and some WiFi6
> >> radios).  I have no doubt that FiWi could get more of the theoretical
> >> throughput delivered at the mac layer.
> >>
> >> Deliverables are all that matter to me and I think to buyers and users.
> >> Benchmarks test deliverables which is great and I'm a routing iperf* user.
> >> It's engineering's problem to develop the next tech *AND* solicit funding
> >> to do so.
> >>
> >>
> >>
> >_______________________________________________
> >Codel mailing list -- codel@lists.bufferbloat.net
> >To unsubscribe send an email to codel-leave@lists.bufferbloat.net
> 
> -- 
> Sent from my Android device with K-9 Mail. Please excuse my brevity.