[Starlink] 300ms Telecommunication Latency and FTL Communication

[Colin_Higbie]

Yes, of course it differs for different people. Statistics doesn't care. There will be a mean and median for any sample. Statistical analysis already expects no two people are identical. 

Good point on the interruptions being longer due to longer fix. That makes my original equation a bit more complex, but, to your point, it only intensifies the problem with each ms of added latency being worse than the one before. 

Cheers,
Colin



-----Original Message-----
From: David Lang <david at lang.hm> 
Sent: Friday, June 7, 2024 12:57 PM
To: Colin_Higbie <CHigbie1 at Higbie.name>
Cc: Sebastian Moeller <moeller0 at gmx.de>; David Lang <david at lang.hm>; starlink at lists.bufferbloat.net
Subject: RE: [Starlink] 300ms Telecommunication Latency and FTL Communication

On Fri, 7 Jun 2024, Colin_Higbie wrote:

> GREAT, GREAT point on the increased likelihood of unintentional 
> interruptions and talking over each other the higher the latency. I 
> had not thought of that before, but I think that's a compelling and 
> easily measurable metric. It also gives a clear cost-based selling 
> advantage to the ISP with lower latency, which should make it 
> attractive to business: if an ISP can say to a business (or at-home 
> worker): based on our 20ms lower latency, you will save 3 minutes per 
> day in lost annoyances due to accidental interruptions, that has a 
> value per person in those calls of $X. Therefore, you should switch your business to us.
>
> I do still think, at least for me as a speaker, that video feedback 
> that does not "feel" like it provides instant video feedback in the 
> form of real-time facial expressions to show audience feedback for the 
> speaker is more critical (I can stomach small delays on audio feedback 
> after I ask a question, but not on video feedback, which I need while 
> I'm talking), but I do think the interruption piece is also very real. 
> Further, my point on video is almost binary – either I'm getting close 
> enough to real-time feedback or I'm not and need to plod on without 
> it, which diminishes its value as an argument for lower latency.

the problem is drawing the line at what latency is the problem. This is going to be different for different people and take quite a bit of experimentation to discover (and I would not be surprised if someone who is doing testing is learing how to cope with longer latencies, so you can't just increase/decrease the latency in a predictable manner, you would need to randomize it and ask after each call how good/bad it was)

> In contrast, the increase in accidental interruptions in audio (and 
> video too
> presumably) would scale with latency (not binary means much better for 
> an argument to the ISP). Every 1ms of additional latency would provide 
> a small but measurable % increase in the amount of accidental interruptions.
> Conversely, the lower the latency, the fewer of these. This means that 
> there's not merely a "good enough" level on latency (at least not 
> above human reaction time), but rather the lower the better in a very 
> tangible way. This should be a compelling and objective argument.
>
> I'm not aware of any study, but it stands to reason that every 
> interruption results in X seconds of lost productive talking time. Say 
> X = 5s. While we don't know the slope of the function: 
> #_accidental_interruptions = [some unknown slope] x (latency_in_ms - 
> minimal_human_reaction_threshold_in_ms),
> where minimal_human_reaction_threshold_in_ms probably equals something 
> like 20
> - 50ms. As soon as latency > minimal_human_reaction_threshold_in_ms,
> #_accidental_interruptions becomes a positive number and lost 
> productivity also increases at 5s/interruption (using above assumption).

I also think the interruptions will be longer with higher latency as it will be harder to sync up.

> Even if there is no such study, this is clearly an easily studyable 
> metric to establish the typical human reaction time (that's probably 
> already known), the mean and median time lost per accidental 
> interruption, and the key piece: the slope of the line or shape of the 
> function that says how many additional accidental interruptions occur 
> for every ms of added latency. Anyone here looking for an academic 
> study with a simple metric to provide data to ISP on economic value in 
> reducing latency that is simple enough that even a high-school kid could understand the meaning, this might be it.

or can anyone setup a simple router/pi image that can introduce arbitrary latency and then let's see if we can find some high school students looking for a science projct. That may not get the academia stamp of approval, but if we can get a bunch of people around the world to test things, we would get data back faster than waiting for a more academic process (and it could potentially feed into such a process)

it doesn't even need to be a full blown video conferencing program in use, but with more people on the call, there's more chance of interruptions.

David Lang