Handley’s simulation shows that, while not as fast as an equivalent ISLL path, long bent-pipe paths would typically have lower latency than terrestrial fiber routes between the same two points.

Mark Handley, a professor at University College London, has made two terrific videos based on runs of his simulation of the first—1,584 satellite—phase of SpaceX’s Starlink Internet-service constellation. I discussed the first video, which assumes that the satellites have inter-satellite laser links (ISLLs), in recent post.

While SpaceX plans to deploy ISLLs in the future, their early satellites do not have them since at 27,000 km/hr they are state-of-the-art technology and may also encounter political problems in some nations. Since it could be a year or so before SpaceX begins launching ISLL-equipped satellites, Handley has made a second video that assumes the phase one satellites do not have ISLLs. This post discusses that video.

Each satellite has four phased-array antennas that can rapidly switch narrowly focused connections to terrestrial antennas falling within a large “footprint” area. The terrestrial antennas might be Internet-connected ground stations or end-user terminals. If there were no ISLLs, long-distance traffic would have to be relayed by bouncing packets up and down between satellites and the ground.

Satellite footprint

Many people—me included—have assumed that these “bent pipe” hops would significantly increase latency on long-distance paths, but Handley’s simulation shows that, while not as fast as an equivalent ISLL path, long bent-pipe paths would typically have lower latency than terrestrial fiber routes between the same two points.

Consider, for example, this six-hop route between Seattle and New York. The bent-pipe route has a round-trip time of 36ms versus 78ms for the current Internet and 38ms for a hypothetical great circle fiber route, which would be impossible because of mountains and other obstructions.

Sample Seattle-New York path

That example was taken from a run in which only six orbital planes had been populated, and it assumed ground stations at popular SpaceX locations plus a few others that Handley assumed would be added at strategic points. With only six orbital planes, global coverage is spotty but it is solid around 53 degrees north (and south). As more satellites are added, coverage becomes wider, and latency improves. By the time all 1,584 satellites are in operation, there is global coverage and latency is consistently better than today’s terrestrial Internet.

While adding satellites improves performance, adding additional ground stations has an even more significant impact. That suggests the possibility of relaying traffic through idle end-user terminals, which also have phased array antennas. Handley ran a simulation assuming relays every 100 km and found that latency across the US was roughly cut in half, and jitter (latency variability) also declined. Still, the number of route changes increased to about one every five seconds. That sounds like a lot of overhead, but Handley feels that it is feasible to handle. It would also require a more expensive user terminal, a little power, and the permission of the user so SpaceX might subsidize the terminals or charge less for service.

User terminals as relays

Handley also considered inter-continental relaying, which would require relay stations on strategically placed ships at sea. (It turns out that no ships would be needed to cross the Northern Pacific, but that would require a relay station in Russia, which might be a political problem). He doesn’t mention the possibility, but couldn’t cargo and cruise ships act as slowly moving relay stations? (They will certainly want to be terminal-users).

The example shown above is for an east-west link, but Handley also looked at long north-south links and found that ground relays actually beat ISLLs in some cases and were always better than fiber, but the best results are achieved by a combination of ISLL and terrestrial links, which we can look forward to once SpaceX and others begin deploying satellites with ISLLs.

Handley concludes by pointing out that since he started making the video, SpaceX had revised their constellation configuration from 24 66-satellite planes to 72 22-satellite planes. It turns out that once the first 1,584-satellite phase is complete, there is pretty much no difference between the new and old configurations. Still, it does require a few more satellites to be deployed before the trans-Atlantic and Pacific relays will work continuously. Note that SpaceX hopes to complete the first phase by early 2021.

I can’t conclude this post without mentioning Handley’s charming disclaimer that he has no inside information, but, based on public statements, has made reasonable assumptions about “what they could do if they wanted to, but probably isn’t what they will actually do.”

Watch the video:

Update Jan 1, 2020:

Handley presented a paper on this research at the 2019 ACM HotNets Conference. You can see a video of his presentation and download a copy of his paper here. (The video of his talk is free, but the paper is behind a paywall).