[Starlink] FQ for rocket scientists

Thu Feb 24 09:23:17 EST 2022

To bring this back to starlink more coherently, FQ would provide
stabler behavior for voip, videoconferencing and gaming traffic, even
if the characteristics of the underlying link mutate from 2Mbit to
300Mbit on any interval under 20ms and the underlying buffering kept
short. In the now ancient data we had gathered, even the sat-sat
switching delay was under 40ms on roughly a 2m interval, which is
easily compensated for in a jitter buffer (though I might tune one to
adjust more quickly), and I imagine they can switch sats faster in the
future.

I wish we knew more about the underlying multiplexing of the uplink.
We did some high resolution irtt based measurements in the early days,
having *two* starlink terminals fairly close together and repeating
that L3 measurement with well synced clocks would reveal much (as
would analog measurements). My guess is for subscriber density they
are more bound by the uplinks than downlinks (please correct me if I'm
wrong)! as a downlink transmission could contain data for all
terminals all the time, separately decoded....

I hope the dishy provides a GPLS locked ntp clock one day.

That leaves the backhaul (skyhaul?) questions still wide open. I would
be trying to regulate bandwidth (and buffering) on the ground stations
to keep the queues in the sky relatively empty, doing ack-filtering in
both directions (because that works well with bursty, blocked, traffic
and reduces pressure on packet fifos), but when the laser links go
up... my head still explodes. Too many unknown unknowns.

On Mon, Feb 14, 2022 at 2:21 PM Bjørn Ivar Teigen <bjorn at domos.no> wrote:
>
> FQ clearly gets us very good behavior, so no argument there.
> Also agree the cost of giving up work conservation might be too high, and I'm not saying we should necessarily do that. I just wanted to point out that insisting on work conservation also comes at a cost (which we are free to choose to pay of course!). It's a choice between adding some middlebox complexity or accepting the amount of jitter induced by the amount of queuing needed to maintain high utilization (which depends on the jitteriness of the interface working to empty the queue).
>
> Regards,
> Bjørn Ivar
>
>
> On Mon, 14 Feb 2022 at 18:39, Sebastian Moeller <moeller0 at gmx.de> wrote:
>>
>> Regarding work conservation, FQ as so often appears to be a decent solution that while not optimal will also not be pessimal. Which IMHO is as much as we can hope for unless we want to burden all middleboxes with deducing relative importance of packets....
>>
>> Regards
>> Sebastian
>>
>>
>>
>> On 14 February 2022 18:52:29 CET, "Bjørn Ivar Teigen" <bjorn at domos.no> wrote:
>>>
>>> Sebastian,
>>>
>>> On Mon, 14 Feb 2022 at 17:15, Sebastian Moeller <moeller0 at gmx.de> wrote:
>>>>
>>>> Hi Bjørn,
>>>>
>>>>
>>>> I guess I should have started with the obvious. Nice short article!
>>>
>>> Thank you!
>>>>
>>>>
>>>> > "See the problem? Buffers produce jitter!"
>>>> >
>>>> > I would rephrase that as "over-sized and under-managed" buffers increase jitter unduly. Interestingly, bound jitter can be converted into static latency by using, wait for it, (delay) buffers and a scheduler.
>>>> >
>>>> > You are absolutely right. I was hoping someone would spot that! There is one problem with that approach though. To actually remove the jitter the scheduler can no longer be work-conserving (needs delay as you also point out), and that increases TCP ramp-up times (among other things). Can be a hard sell. I think the benefits outweigh the costs though, so I would do it your way.
>>>>
>>>>         But that is what end-points already do, on-line games do this to equalize internet access quality between players (allowing them to make matches over larger populations), as do DASH type video streaming applications, where the isochronous play-out takes the role of the scheduler and the race-to-fill-the-play-out-buffers serves to keep the buffers filled so the scheduler never runs "dry".
>>>
>>>
>>> Good points. The application clients and servers can choose to use their resources in a way that is not work-conserving and thus achieve sharing of resources without introducing jitter. I would argue it's a different story with the queues "in the network", in routers, switches, access points, etc. There seems to be an unwritten law that those schedulers must be work conserving, presumably to minimize round-trip times, and this makes it harder to achieve well-behaved sharing. So in network devices configured to forward packets as quickly as possible, I think it's mostly true to say that buffers produce jitter.
>>>
>>>>         Given Pete's data at https://github.com/heistp/l4s-tests I am cautious to hold my breath... or to put it differently, I am certain they will not deliver on their promises, the bigger question is whether the incremental improvement they offer (over the default FIFO) is decent enough to retroactively justify the disruption they will have caused...
>>>
>>>
>>> Ohh, that work has been updated a lot since the IETF showdown! Thanks for reminding me, I'll have a look at it again.
>>>
>>> - Bjørn Ivar
>>>
>>>>
>>>> Regards
>>>>         Sebastian
>>>>
>>>>
>>>> > >
>>>> > >
>>>> > > On Mon, Feb 14, 2022 at 7:07 AM Bjørn Ivar Teigen <bjorn at domos.no> wrote:
>>>> > >>
>>>> > >> Hi everyone,
>>>> > >>
>>>> > >> I was inspired by the latest Starship presentation to write a piece on network quality in the language of rocket science. The blog can be found here: https://www.domos.no/news-updates/network-quality-for-rocket-scientists
>>>> > >>
>>>> > >> Cheers,
>>>> > >> Bjørn Ivar Teigen
>>>> > >>
>>>> > >> --
>>>> > >> Bjørn Ivar Teigen
>>>> > >> Head of Research
>>>> > >> +47 47335952 | bjorn at domos.no | www.domos.no
>>>> > >> WiFi Slicing by Domos
>>>> > >> _______________________________________________
>>>> > >> Starlink mailing list
>>>> > >> Starlink at lists.bufferbloat.net
>>>> > >> https://lists.bufferbloat.net/listinfo/starlink
>>>> > >
>>>> > >
>>>> > >
>>>> > > --
>>>> > > I tried to build a better future, a few times:
>>>> > > https://wayforward.archive.org/?site=https%3A%2F%2Fwww.icei.org
>>>> > >
>>>> > > Dave Täht CEO, TekLibre, LLC
>>>> > > _______________________________________________
>>>> > > Starlink mailing list
>>>> > > Starlink at lists.bufferbloat.net
>>>> > > https://lists.bufferbloat.net/listinfo/starlink
>>>> >
>>>> >
>>>> >
>>>> > --
>>>> > Bjørn Ivar Teigen
>>>> > Head of Research
>>>> > +47 47335952 | bjorn at domos.no | www.domos.no
>>>> > WiFi Slicing by Domos
>>>>
>>>
>> --
>> Sent from my Android device with K-9 Mail. Please excuse my brevity.
>
>
>
> --
> Bjørn Ivar Teigen
> Head of Research
> +47 47335952 | bjorn at domos.no | www.domos.no
> WiFi Slicing by Domos

-- 
I tried to build a better future, a few times:
https://wayforward.archive.org/?site=https%3A%2F%2Fwww.icei.org

Dave Täht CEO, TekLibre, LLC