Moreover, the idea of using user-terminals/gateways as ground-relays to bounce signals up-and-down is quite impractical, since you would be greatly reducing the capacity available to satellites on those "intermediate" satellite.
Having said that, this kind of wide-area low-latency bounced routing is never going to be used for you or me to watch Netflix. It will be reserved for high paying customers who really really care about latency. For you and me, we'll be dumped into the terestrial network at the nearest possible location that isn't already saturated.
The second generation of satellites should have optical inter-satellite links, and then you would only use ground relays rarely.
Thanks for your reply. I agree with you, using user-terminals is extremely challenging, especially from a link-budget perspective. You cannot pump enough data to make it worth it. For the gateways, my main concern is that the Ka-band spectrum would have to be shared between user-data and "inter-satellite" data.
Finally, I think that the use-case your described for those latency-sensitive customers is going to be hard to pull off, mainly because of link-availability concerns. There are too many "passes" through the atmosphere to guarantee the availability numbers that a user of such a service requires (99.5%?). Rain in any of these links might cause an outage or a re-route (causing too high jitter). Plus, it would be extremely difficult to have signals traveling from one continent to another.
If you can't eliminate all queuing delays, you can't factor all delays into your decision of when to change routes, and so you will get some jitter and hence reordering. If so, you need a reorder queue in the final receiving groundstation, so as to avoid confusing TCP. This removes jitter at the expense of adding some latency. How much latency depends on the queuing delays you're trying to smooth out, so it all really comes down to avoiding queues in the satellites.
The user terminal as a gateway idea is also not practical due to link budget (EIRP and G/T), but also the much lower availability they'll be dealing with.
https://youtu.be/m05abdGSOxY?t=428
I'm running Dijkstra across this mesh at 30fps in real-time on my laptop while also doing the 3d animation. My laptop fan does spin a bit, but it's not crazily optimized code, and for the video I was also recording to H.264 simultaneously. Doing routing for all customers simultaneously is certainly feasible if their groundstations do the computation, based on routing state supplied in real-time by the constellation. Other solutions are probably possible too, but this seems simplest to me, and scales linearly with customers.
https://www.youtube.com/watch?v=QEIUdMiColU
And here without ISLs:
https://www.youtube.com/watch?v=m05abdGSOxY
Now, these videos assume no queuing delay. It's really hard to guarantee low queuing delay in a traditional IP network, but there's been a lot of research over the years, and we have a range of ideas that can be used to deliver minimal queuing with reasonable utilization. In fact doing this for Starlink is something I'm actively researching at the moment. I'm convinced it can be done, but it won't look like a conventional IP network internally. I don't know what SpaceX will actually do, so if the FCC has doubts, this is perhaps where they are concerned. Other delays might be in the ground segment, but that's more or less the same for SpaceX's terestrial competitors, and the FCC seems to think they can do it.