Hacker News Reader

Comment by foobiekr | Hacker News Reader

The article is 100% rubbish. Just complete trash.

Each submarine cable is 12+ strands which are DWDM multiplexed to 20-40 channels at 100-400gbps.

The entire _planned_ starlink constellation has less aggregate switching capacity than a pair of current generation core routers. The market for terrestrial ethernet _routed_ 400Gbps ports is 10k+ annually and 800Gbps is just getting warmed up.

Satellite is not going to even make the slightest dent in terrestrial networking. Starlink isn't even viable without most of the traffic doing two leg ground-satellite-ground for SP-network connected ground stations. 100% of that traffic is landing on the terrestrial internet.

View on HN

That Starlink estimate must be low, they currently have 1.5 million customers and a significant percentage of total bandwidth is over ocean and thus mostly wasted.

If you assume a pessimistic 100x over subscription and 50% waste over oceans and non serviced countries * 50mbps (upload + download) * 1.5 million customers that’s well over 1.5 TBPs current capacity for customers which then doubles when you include their base stations. And they that’s today when the current 3887 Satilites which mostly v1.0 and v1.5 which have dramatically less individual bandwidth than v2.0.

PS: Also your port estimate overstates total bandwidth as each point to point link requires a port on each side, and packets generally make multiple hops through the internet.

Current generation 12 slot chassis with 51.2Tbps per slot is 612Tbps per chassis, and so a pair is 1228Tbps which is roughly your estimate. Companies like Juniper and Cisco sell tens of thousands of the previous generation per year _today_. Starlink is not even a drop in the bucket in terms of forwarding.

Moreover, with a very tiny exception over antarctica or open ocean, every single packet forwarded by starlink is hitting a terrestrial link after a single ground-space-ground hop.

Nobody is going to be making backbones out of Starlink or anything that looks like Starlink. Only the people who know nothing about modern networking performance think otherwise.

If you followed the thread, my estimate was ~3,000 TBps, welterde gave “1.7 - 2.3 Pbit/s” = 1700 to 2300 TBps. 1228 TBps is well below both of our estimates. Last I checked global IP traffic it’s still well under 2,000 TBps looking at the rate traffic is increasing Starlink really can act as a major internet backbone.

Also, Starlink customer to Starlink customer is very compelling for may applications. Undersea cables for example can create much higher latency between areas than their physical separation for example Australia to Madagascar. Similarly there’s still quite a bit of point to point microwave tower connections etc that can now be cheaply replaced.

You are not comparing the same thing here though. If you applied the same theoretical maximum method to all networking hardware (or even just core router class hardware) sold today you would arrive at numbers very very much in excess of a couple Pbit/s.

For it to be really viable as a backbone it would need to be able to do several 100 Gbit/s from one location and that Starlink cannot do (or is built to do). A single satellite has less than 100Gbit/s of RF capacity and you will only have a couple visible above you at any one time. And after that bottleneck comes the inter-satellite link capacity bottleneck which is less than half of that if the information I have found is correct (4 links à 10Gbit/s). I don't think they will be able to sell more than 10 GBit/s service to anyone and that only in not so populated areas.

A single datacenter can reach low Tbit/s of connectivity _alone_ _today_. I don't see how it could make any dent whatsoever in the backbone market.

The other usecases are a different discussion and not as dependent on very high bandwidth numbers.

Core routing hardware isn’t sold at multiple Pbit/s because no location on earth has that much traffic. Nobody is wasting that kind of money buying hardware when they don’t have fiber to make use of it. Starlink is aiming physically to eventually have 10+ Pbit/s worth of actual hardware because customers aren’t spread evenly across the globe so there’s inefficiencies involved. The only unfairness is comparing future equipment to current equipment.

First data centers very wildly in size and have connections to multiple networks for redundancy, Starlink would be very appealing from that standpoint. Also you’re underestimating bandwidth here, V2 mini are ~80Gbps full V2 weigh 2.5x as much and presumably the bandwidth difference is even higher, but they aren’t launching hardware yet so it’s unclear. Any given location several satellites above it at the current density, at 42,000 you’re talking 50+. Plus if it’s data center to Starlink customer then you’re effectively replacing a Starlink ground station so they might make such installations very cheap.

Every packet traverses multiple such core routers (most likely from multiple internet providers), so 1 Tbit/s from A to B requires many more Tbit/s of actual hardware capacity sold. Or take the edge capacity of small FTTH providers and you get to quite large theoretical numbers too. Hundreds of Tbit/s easily if one is offering Gbit/s or faster service (my ISP has around 1Tbit/s of equipment installed just to service around a 300m radius of residential homes). Both are true for Starlink as well: In case both source and target are not visible to the same satellite they have to be routed via multiple hops or sent upstream (ground station) to the rest of the internet. The edge capacity (RF up/downlink) boasts insane numbers, but the actual internal backbone is much more limited. For consumer it is less of an issue since they won't mostly fill up that 1Gbit/s service they have, but businesses will absolutely fill that up (so you can oversell much less).

Starlink doesn't have nice overview of actual specs of future satellites, so I can only use what people have figured out or what they have published. But you are right about the number of satellites, I looked at the wrong column (it was for phase 1 and not 3), so density will indeed be much higher, so from that perspective it will be somewhat more capable (depending on fast/slow their rollout is). And it's not clear if it would actually be cheaper for SpaceX to have such customers (since some of the traffic would go to the rest of the internet - requiring ground station resources) or go elsewhere on the planet (filling up inter-satellite capacity), but that's not too much different from how regular ISPs have to plan (some customers are cheaper for them than others).

> Every packet traverses multiple such core routers (most likely from multiple internet providers), so 1 Tbit/s from A to B requires many more Tbit/s of actual hardware capacity sold.

I agree, but look up global IP traffic numbers it’s just crossed 1 Pbit recently and that’s across the globe. No individual router needs to handle anything that close to 1 Pbit today, though modular designs would let you manufacture such a router.

You are underestimating the capacity of modern core routers. A single PTX10016 router could route the entire starlink customer base with no over-subscription at all (230 Tbit/s vs 75 Tbit/s from 50 Mbit/s * 1.5M). Actual estimates of the total capacity of Starlink puts it more in the 20Tbit/s range (and that's without taking into account coverage over less populated areas).

First my 3 Tbit/s estimate was wildly pessimistic and only covered current network utilization. Adding missing countries plus ships more than doubles land area, 42,000 satellites is more than 10x current satellites, v2 is ~10x current individual satellite bandwidth. So current goals are at least 200x current bandwidth. (Satellites also so point to point links which we are ignoring in terms of what the total routing capacity of the network has.)

Total estimate for a “fully deployed” Starlink network is therefore 600 Tbit/s from my pessimistic estimate (or half that if your thinking in terms of customer bandwidth vs traffic handled by the network)

A more realistic estimate is more like 3,000 TBPs as lower density areas can offer higher bandwidth per customer a village of 100 in the Sahara gets the same bandwidth as a town of 100,000 in the US. How much of that actually gets used is a different story.

PS: Individual switches can handle a surprisingly large percentage of total internet bandwidth as packets travel through many of them. 100 switches / average of 4 hops per packet has 25x the bandwidth of a single switch.

I was only talking about the currently deployed constellation, which of course is much smaller, but for which we have actual numbers from measurements. For the completed Phase 3 constellation I have read numbers in the range from 1.7 - 2.3 Pbit/s as theoretical maximum. So would need around 8-12 PTX boxes to reach the same total routing capacity as the whole future constellation. The current single-fiber bandwidth record is 1.53 Pbit/s.

And nevermind that Starlink will run out of ground station / inter-satellite link capacity long before reaching the theoretical maximum.

Ah, you said planned here so I assumed that was what you where talking about:

> The entire _planned_ starlink constellation has less aggregate switching capacity than a pair of current generation core routers.

I generally agree with those estimates based on how they’re defining limits though those definitions really depend on what you care about. Aka predicting actual utilization at one end vs just considering what the satellites could physically do in ideal conditions etc. Thus the theoretical impact of direct customer to customer links for a satellite over the Atlantic for example can vary wildly.

Does it count as a waste if their main selling point is to deliver internet access to places that are hard or impossible to reach otherwise?

Well you're calculating bandwidth so yes

But Starlink would have no reason to exist in the first place if it didn't offer service in those remote places.

Most Starlink customers are in areas “easy” to serve by high speed internet like the eastern US. It’s really only economical because of poor internet access across most of the US compared to similar density areas in other developed counties.

Isn’t the latency of satellite internet always going to be significantly worse? Unless, of course, you could make larger hops, avoiding unnecessary switching delays. But I don’t think that will happen, except some rare circumstances.

Not a clear cut, and one intended application of Starlink was . You need less detours and light in fiber is slower than light in vacuum.

Fiber slows down Berlin <-> NYC by a lot. I usually use Wolfram Alpha for this and it's like 30ms fiber vs. 21ms vacuum.

So in theory starling could offer less hops and better latency. The detour to space isn't that bad either, ~1100km on a 6500km route, and fiber will need some extra hops and detours too.

> Each submarine cable is 12+ strands which are DWDM multiplexed to 20-40 channels at 100-400gbps.

Being able to pump 40 terabits per second up to a satellite from a single uplink, as they hope, is significant-- exceeding most subocean cables.

Preface: Day job is owning a low latency telco.

a single cable system under construction(which we are buying part of) (there are over 30) has over 120t of capacity @400G waves, but we are already discussing doing 800G waves, which will double that. We currently operate on over 7 submarine cables, one of the landing stations we operate out of has over 10 cable landing which is over 850T of capacity across all cables.

Most of my metro DF backhaul can push 16-19T down a single pair, and most of my builds are 1728 core fiber.

Starlink is very cool, it will connect ALOT of under served or simply not served people across the world, but the notion that it will even have the slightest chance at dinting the Subsea cable market is laughable.

Most of us are not talking about Starlink. I'm talking about this optical uplink technology, which can push tens of terabits per second through the atmosphere between each uplink and satellite pair.

No, a single uplink can't replace a subsea cable.

Any free space optical approach you can imagine will be slower and noisier than the equivalent technology in a fiber. You are not going to get DWDM channel density in free space or throughput.

This is not real.

> Any free space optical approach you can imagine will be slower and noisier than the equivalent technology in a fiber.

Granted. More path loss, dispersion, potential for interference. The question is how well can you do?

> This is not real.

I believe ETHZ; of course there's more to do to make a practical system.

Which will go oh so well when an adversary appears with a laser uplink of their own and starts fiddlefarting the Rx Sat to ruin your day.

Lines have one biiiiig advantage. They're very easy to monitor(liveness), very hard to imperceptibly tap anymore, they don't move once you put them down (generally), and if they do, you know just where to start looking.

Most infrastructure is vulnerable in one way or another... just a few miles from me someone cut an important SONET ring in two places.

The telescopes involved in this are strongly directional, so whomever is tampering would need to be rather close to the real uplink.

Since you seem to have some experience I want to ask:

How much latency is added by routing and switching on a typical transatlantic journey?

And a followup, are there any promising developments for moving the networking infrastructure fully to light as a carrier signal without any electron-based hardware in between?

Reply on news.ycombinator.com