Orbital Data Centers connecting directly to Starlink via laser

[Robotbeat]

If your users are accessing the data set using Starlink, however, putting the data on Starlink itself cuts down the latency by half. It also eliminates the bandwidth to the Gateways if you cache stuff on Starlink itself.

[DanClemmensen]

If your users are accessing the data set using Starlink, however, putting the data on Starlink itself cuts down the latency by half. It also eliminates the bandwidth to the Gateways if you cache stuff on Starlink itself.

Sorry but that does not work. The satellites move, so your data would be orbiting the earth unless it was being continuously moved from satellite to satellite. Such movement of any appreciable amount of data is infeasible as it would require too much ISL bandwidth and power.

[Robotbeat]

If your users are accessing the data set using Starlink, however, putting the data on Starlink itself cuts down the latency by half. It also eliminates the bandwidth to the Gateways if you cache stuff on Starlink itself.

Sorry but that does not work. The satellites move, so your data would be orbiting the earth unless it was being continuously moved from satellite to satellite. Such movement of any appreciable amount of data is infeasible as it would require too much ISL bandwidth and power.

It works just fine for super common data you can just put on each Starlink satellite. Like Netflix cache, for instance.

“Infeasible” is a function of capability. Just handwaving without analysis doesn’t help anything.

The Netflix catalogue isn’t that big (can easily fit on small solid state devices) but it uses a large portion of Starlink bandwidth. Similar for other streaming video catalogues.
We explored this quantitatively I believe on this thread.

Might not make sense for current Starlink satellites, but once they become multiple tons, it may be worth doing.

Near-line storage like this (where commonly used data is copied dozens of times so data doesn’t have to traverse the whole internet) is super common for other ISPs, and I don’t see why it won’t eventually happen for Internet ISPs like Starlink.

Since many satellites will be in view at any one time, it may even be possible to interleave the satellites. Starlink sat A has Netflix catalogue, Starlink sat B has a Disney+ catalogue, etc.

[DanClemmensen]

If your users are accessing the data set using Starlink, however, putting the data on Starlink itself cuts down the latency by half. It also eliminates the bandwidth to the Gateways if you cache stuff on Starlink itself.

Sorry but that does not work. The satellites move, so your data would be orbiting the earth unless it was being continuously moved from satellite to satellite. Such movement of any appreciable amount of data is infeasible as it would require too much ISL bandwidth and power.

It works just fine for super common data you can just put on each Starlink satellite. Like Netflix cache, for instance.

“Infeasible” is a function of capability. Just handwaving without analysis doesn’t help anything.

The Netflix catalogue isn’t that big (can easily fit on small solid state devices) but it uses a large portion of Starlink bandwidth. Similar for other streaming video catalogues.
We explored this quantitatively I believe on this thread.

Might not make sense for current Starlink satellites, but once they become multiple tons, it may be worth doing.

Sorry, but the term "data center" usually implies something like Amazon AWS, not just a global cache of static data, at least to me. When we looked at a Netflix-type application, the best place for it was at the user terminal on the ground. Latency is zero, update was by continuous low-priority data stream in every beam. Its basically easier to store on millions of UTs than on tens of thousands of satellites.

[Robotbeat]

“You’re not allowed to pick the low hanging fruit of data center functionality!” is not a very persuasive argument to me.

And while the Netflix catalogue may be relatively small, it’s still way too big to put on all the user terminals:

(I already showed this picture, but you ignored it?)

[launchwatcher]

“You’re not allowed to pick the low hanging fruit of data center functionality!” is not a very persuasive argument to me.

And while the Netflix catalogue may be relatively small, it’s still way too big to put on all the user terminals:

(I already showed this picture, but you ignored it?)

I've looked over https://openconnect.netflix.com/en/appliances/

The entire Netflix catalog (~3000 TB in 2013 according to one source) cannot fit on one of these devices (360TB raw capacity on the largest one through a mix of traditional rotating media and solid state storage, but useful capacity will be lower).   It's a cache.   If it serves users in one particular locality it would likely have a higher hit rate than if it had to serve users all over the globe.

A couple TB of cache in each ground station would likely be more cost effective for netflix-like workloads than 300TB in each satellite.   

[Robotbeat]

300TB is only like $30,000 worth of enterprise SSD. It’s not that much. If Starlink satellites get larger as some have speculated, that’ll be pretty doable (with some basic shielding!).

[abaddon]

Netflix is waaaaay bigger now with far more content in 4k and far more content in general.  I’d be shocked if it’s not at least an order of magnitude bigger now.  The idea of caching the entire catalogue in a satellite is a non-starter.  A smart cache that keeps the top N% of shows, where N is a small number?  Sure.

[Robotbeat]

Netflix is waaaaay bigger now with far more content in 4k and far more content in general.  I’d be shocked if it’s not at least an order of magnitude bigger now.  The idea of caching the entire catalogue in a satellite is a non-starter.  A smart cache that keeps the top N% of shows, where N is a small number?  Sure.

Give me a number for its size. Also, I bet that 90% of watch time is from 10% or less of their catalogue.

EDIT: (see my post below… this became a lot more than just an edit LOL)

[launchwatcher]

Netflix is waaaaay bigger now with far more content in 4k and far more content in general.  I’d be shocked if it’s not at least an order of magnitude bigger now.  The idea of caching the entire catalogue in a satellite is a non-starter.  A smart cache that keeps the top N% of shows, where N is a small number?  Sure.

The cache size vs cache hit rate curves for each service would be interesting to see (and likely both highly proprietary, and very different for netflix vs youtube vs ...).

Given how existing caches are operated and the inability to visit satellites to swap out cache appliance, the system architecture of on-board caches would likely have to resemble a virtual machine hosting model which would have lower efficiency than the bare iron appliance model currently used in caches.

On thinking about it some more I suspect you're better off with larger caches in/near ground stations, each provided and operated by each streaming service, placed into colocation racks in the equipment huts.   With power, cooling, and mass much cheaper on the ground you could provision much more capacity and caches that stay near their clients would benefit from any geographic effects on cache locality.

[Robotbeat]

Ah, but maybe those things wouldn’t be much cheaper indefinitely. If you start with that as a given, it will force a certain result.

SpaceX uses pretty cheap solar panels, the cells of which are manufactured for terrestrial applications (I believe). And they get much better capacity factor than on the ground.

[Naito]

......

clearly none of you guys have ever run a datacenter

[Robotbeat]

......

clearly none of you guys have ever run a datacenter

I did, in a previous life.

[Nomadd]

......
clearly none of you guys have ever run a datacenter

I did, in a previous life.

I sort of did, when 9600bps was considered high speed and an interstate circuit at that rate ran about 2 grand a month. 
 I miss being able to tell data speed by listening to it.

[Robotbeat]

Netflix’s catalogue varies in size (and may actually have gotten smaller in terms of hours as usage rights vary over time). The most recent articles from 2020-2021 say that Netflix has about 13,000-15,000 titles (but at most around 5000 available per country), translating to about 36,000 hours.

Netflix streams “4K” at just 7GB/hour, HD at 3GB/hour, so you could fit their entire worldwide catalogue in both 4K and HD in just 360TB.

The largest enterprise solid state drives in the 3.5” form factor is 100TB, so you could easily fit those in an EXISTING Starlink satellite.

(Although you wouldn’t use that exact form factor and you probably want to wait until the Starlink satellites grow significantly.)

And it’s not even that much data to serve. A Starlink satellite serves, what, 20Gbps?

Netflix’s OpenConnect appliances were serving 100Gbps of 4K content back in 2017.

I don’t think this will happen right away. Starlink satellites need to probably be about 10 times their current size so the overhead of hosting CDN boxes (plus a few inches of shielding) isn’t a problem. They also need millions of customers. The terminals probably need to talk to multiple satellites simultaneously so you don’t have to put the same cont nt on every box.

But it’s not nearly as absurd as what you all are saying.

Again, in 2017, 100Gbps for a box this size containing the whole Netflix catalogue, which is much lighter than an existing Starlink satellite let alone whatever size they eventually grow to. About 30 liters volume, 30kg mass, and 650W peak power. Could probably knock that volume and mass by a factor of 4 or better, and the power possibly as well. But for a larger Starlink satellite, not required.

Probably only SpaceX has the experience and the data to know if or when this would be worth it. Only they have experience with using lots and lots of COTS electronics in orbit. But I think this is worth considering eventually.

[Robotbeat]

Starlink has about 6500W worth of solar power peak. One of these appliances uses about a tenth of that to serve 5 times the throughput. So from an energy perspective, it doesn’t look absurd.

Small data center rental near each gateway would cost about $17,000 annually (large data centers would be much cheaper, but can’t always find a data center over 5000 sqft nearby) for a 650W appliance like that. A Starlink satellite costs ~$250k to build, lasts 4-10 years. So we’re in the right ballpark here for something using a tenth its capacity.

[Robotbeat]

Netflix is waaaaay bigger now with far more content in 4k and far more content in general.  I’d be shocked if it’s not at least an order of magnitude bigger now.  The idea of caching the entire catalogue in a satellite is a non-starter.  A smart cache that keeps the top N% of shows, where N is a small number?  Sure.

oh really? Non-starter? Did you even try to calculate it?

As I proved above, this is just false. Netflix has 36,000 hours, equivalent to 360TB (if stored at both 4K & HD streaming rates). You can get 100TB SSDs that weigh 538 grams each, so 4 of those would weigh just 2kg (and 360TB worth of microSD cards weighs just 100 grams… although you’d want to double that to get any kind of useful life). (A full server may weigh 10 times that, but still reasonable… and realistically, SpaceX would probably just make their own CDN server-on-a-board with a bunch of NAND chips soldered to it that sits flush with the rest of the Starlink electronics with thermal interface and behind some shielding.)

It’s just not that much data anymore.

But I get it, what you’re saying SOUNDS true, whereas the idea you can cache all of Netflix in orbit SOUNDS false… But the numbers just don’t look absurd at all.

[DreamyPickle]

This is absolutely ridiculous.

What does this get you compared to placing Starlink ground stations next to existing data centers? This is a sensible thing that SpaceX is already doing.

You get maybe 20ms less latency towards Starlink customers at the expense of having to spread computing resources equally around the globe (because they move).

[DanClemmensen]

This is absolutely ridiculous.

What does this get you compared to placing Starlink ground stations next to existing data centers? This is a sensible thing that SpaceX is already doing.

You get maybe 20ms less latency towards Starlink customers at the expense of having to spread computing resources equally around the globe (because they move).

Also, for the only use case being proposed (Netflix), latency is not important.

[Robotbeat]

This is absolutely ridiculous.

What does this get you compared to placing Starlink ground stations next to existing data centers? This is a sensible thing that SpaceX is already doing.

You get maybe 20ms less latency towards Starlink customers at the expense of having to spread computing resources equally around the globe (because they move).

It saves you Gateway bandwidth.

For every Gbps served from on-board CDN, that’s one less Gbps that has to traverse the atmosphere. Since streaming video traffic is most of today’s bandwidth usage, that’s not a small effect.

And this makes an even bigger difference once you add ISLs because it means people far from a gateway don’t need to wait. And you’re not using up all that ISL bandwidth just for streaming video.

Also, renting a small data center near a gateway isn’t free, either.