Re: [EToSat] Packet lost on last mile network?

[Roland Zink <[e-mail]>]

Am 20.05.2020 um 20:09 schrieb Christian Huitema:
>
>
> On 5/20/2020 9:56 AM, Carsten Bormann wrote:
>>> And, if they are only deployed in a
>>> fraction of the Wi-Fi network, then the transport folks will still need
>>> an end-to-end solution.
>> They sure will.  But that never can be as good as local recovery.
>
> In a nutshell, that's the entire discussion of end-to-end vs middle-boxes.
>
> It is always possible to do a local optimization by adding some 
> intelligence in the network. That's the case for pretty much all 
> middle boxes, and it seems that the proxies envisaged in Loop fall 
> into that category. But there are always downsides, such as 
> management, flexibility, and most importantly ossification by old 
> proxies prone to bit rot.
>
yes
>
> Given effort, it is almost always possible to find end-to-end 
> solutions that are "good enough", getting you within 90% of the 
> optimum or some such. FEC as tail probe would be one example of that. 
> The end-to-end solutions also tend to get better over time, as 
> implementers devise more clever algorithms and update the code. All it 
> takes is some motivation that the effort is important.
>
This is the question. Is there enough motivation for the 0.x% case?
>
> An example of that would be support for cellular networks. Twenty 
> years ago, this was a niche, and TCP was not optimized for it. Hence 
> the deployment of middle-boxes in the cellular networks. But as this 
> became an ever more important scenario, TCP improved, and QUIC will 
> improve over that. The cellular middle-boxes become harder to justify, 
> and will start moving from the "optimization" to the "hindrance" category.
>
The cellular boxes are handling radio, mobility and things like local 
retransmits and FEC. So is the situation now better because of improved 
TCP or because of improved mobile network boxes? Probably both, but you 
are probably only talking about PEPs. Those may go away but they are 
also built into those mobile network boxes doing the other stuff like 
eNodeBs for example.
>
> Satellite networks are a peculiar case. They remain very much a niche 
> application to this day, something like a communication media of last 
> resort in places that fiber or cellular cannot reach. Given the niche 
> status, it is unclear whether end-to-end transports will bother 
> implementing optimized algorithms, even when these are available. But 
> it is also unclear whether "mainline" networks will bother deploying 
> middle-boxes solely for the benefit of the satellite niche, which 
> kinds of rules out solutions like 'deploy a box in every Wi-Fi 
> network'. 'Deploy a proxy at the ground station' might work, because 
> the satellite operator has the right business incentives, but of 
> course that requires nodes to still do something special for 
> satellites, i.e., discover the proxy and use it. Not obvious.
>
> Hence my effort to look at end-to-end solutions covering not just the 
> satellite case, but also other long distance networks such as 
> intercontinental connections. These are still niches today, given VPN, 
> but they are still useful for P2P applications. Given enough niches, 
> there might be enough motivation for doing the effort.
>
My fear is that niches are served late or not at all. The local network 
operator has a high interest in improving the situations but at global 
scale the end-2-end solution for it may not provide enough motivation. 
Assuming not all nices faces the same problem which can be solved then 
for all niches. Also solutions may need to be applied to all 
applications, e.g. if it is fixed in application a it still may need 
attention by application b. Applications may bit rot as well.

Anyway if you manage to build an end-to-end solution which does work for 
all sort of niches than that would be great. Please continue to work on 
an end-to-end solution.

> -- Christian Huitema
>
>
Regards,

Roland