Terabit-Per-Second Class Connections over FTTH

Big Fat Dave writes "Thanks to research from Japan's Tohoku University, an article at Tech.co.uk wonders if someday the megabit and gigabit classes of net connections will join kilobits in the 'antique tech' bin. By doing some advanced mathematics and 'tweaking' existing network protocols, researchers may be able to enable standard fiber-optic cables to carry data at hundreds of terabits per second. 'At that speed, full movies could be downloaded almost instantaneously in their hundreds. At the heart of the development is a technique already used in some digital TV tuners and wireless data connections called quadrature amplitude modulation (QAM). One glance at the Wikipedia explanation shows that it's no easy science, but the basics of QAM in this scenario require a stable wavelength for data transmission. As the radio spectrum provides this, QAM-based methods work fine for some wireless protocols, however the nature of the optical spectrum means this has not been the case for fibre-optic cables ... until now.'"

Comments

ya but..

[mastershake_phd]

At that speed, full movies could be downloaded almost instantaneously in their hundreds

Not until the PC buses catch up..

Re:ya but..

[gad_zuki!]

>No "PC"s on "backbones" I think.

Then no terabit connection for you. I dont care how fast the backbone is. Where I live the last-mile technology is DSL which for my location maxes out at 1.5mbps.

I think the "OMG LOOK HOW FAST TIS IS" kiddie-mentality of movies-per-second ignores the whole issue of last-mile distribution. And PC buses. And practility. And economics.

Youd think slashdot would have better things to post than PR releases.

Re:ya but..

[Ogun]

Fastest backbone router that I know of is the Cisco CRS-1. It can scale to a system capacity of 92 Tbps in total, using 72 42U rack units as one large router. Still, the fastest interfaces on that machine is OC-768 at roughly 40 Gbps.

Re:ya but..

[Kjella]

The practicality and economics is that in all larger construction projects here in Norway today, whether it's apartment blocks or new fields of housing they lay fiber connections. There are approximately two million households and about 150,000 (7,5%) of them can get fiber connections. Each year 30,000 new houses are built and many of them will have fiber connections, though lone houses don't qualify. If we say 25,000 a year (30,000 less lone houses plus retrofits) then over the next decade I expect that to rise to 150,000 + 10*25,000 = 400,000 (20%) for a conservative estimate. Oh yeah and we're considerably more sparsely populated than the US. Fiber has good end-mile economics as long as you're putting down cables anyway. Now, that wouldn't make it useful with a terabit last mile but if you want real capacity and not US "unlimited" capacity, then it's really nice if actually delivering is very cheap. And a few thousand people on gigabit connections add up to terabits quite fast...

Re:ya but..

[funkboy]

Probably because you haven't seen a Juniper T1600. It has 2.5x the per-slot bandwidth of the CRS-1. The Cisco marketing literature may go to 92tbps, but I challenge you to show me a production CRS multishelf system with more than one fabric shelf. Once T1600 modules are available for the TX Matrix the system will provide 6.4tbps in two and a half racks, using far less power than the equivalent real estate worth of CRS hardware (2.4tbps max), at about the same cost. BTW a fully configured 72-rack CRS-1 would require about .8 megawatts of power and belch about 2.5 million BTUs of heat per hour...

Erm, not that that's a biased viewpoint or anything (heh)...

Anyway, IMHO far more important to router scalability is the per-slot and per-watt bandwidth, not how many systems you can chain together (as long as you can chain some reasonably useful number, but I don't see a need for more than 8 chassis in a system). The CRS-1 won't be able to handle 100gE without a system-wide fabric upgrade or double-width cards or something. The T1600 (and for that matter, the Foundry NetIron X series, though not in the same class of capabilities or scalability as the Juniper) will be able to slot in 8 100gE linecards the day they're available.

Re:ya but..

[funkboy]

> True, but the routers and repeaters on the backbone have buses don't they?

The 750hp 2.4L V8 engine in an F1 car produces about 3-4x the amount of power of a production car engine of the same displacement, but you don't see even high-end mfrs like Porsche putting that sort of thing in street cars (for reasons I hope I don't need to explain).

The data plane in high-end routers have custom-designed switch fabrics, which technically are not buses and operate in a different (more scalable) fashion. The wiki article is actually on fibre channel, but the concept is the same. Cost alone precludes use of such components in PC hardware, not to mention various other factors.

That said, PCI Express is pretty damn nice when you start talking performance vs. cost (both per $ and per watt) when the number of high-bandwidth devices on the bus is low, and the existing plethora of 8 & 16 lane devices & motherboards and the potential to scale to 32 lanes (64gbit/sec) in the future mean that the bus in a modern COTS PC is not the bottleneck in high-performance networking on such hardware. The two things that are:

  - The ability of the operating system & host processor to handle the load offered by the networking stack at such speeds. Mitigated by techniques such as TOE and interrupt mitigation & hardware polling. Done in hardware, getting cheaper, widespread implementation in common NICs not there or crappy (ahem, Realtek).

  - The bandwidth to the user's machine, which is what TFA is about...

I bet...

[ceeam]

...someone in MPAA just shat himself.

the vision

[Tumbleweed]

Actually, this opens up some interesting possibilities for people like the RIAA and MPAA. When you can download a whole CD or DVD in seconds, there's no longer much point to someone who's system is connected, in having physical media, or even a copy of the media, on their own machine. Whatever type of business model they'd wind up with could take that into account, and they could come up with a Netflix-type model, or something new and appropriate to the new reality (when have they ever done THAT, though?) - pay $x/mo, or $x/mo/bitrate/resolution, or whatever. The online rental business could be huge.

There's also the benefit of being able to do real-time offsite storage. The people who would care about needing massive amount of storage for their movie collection - no longer need to store their movies locally. Your whole machine could wind up being nothing more than an online access point with it being customized to be the HCI that you prefer: curvy keyboard (w/ or w/o lights) or not, big-ass widescreen display ... or not, your choice of colour, and a big honkin' net connection. Lots of RAM and a SSD boot drive, and something (magnetic card, keyfob, whatever, or nothing - just swipe your retina across a scanner or something) you can take with you to plug into whatever other machines you use to let that machine know it's you and to configure to your preferences. And nothing more. No moving parts other than the keyswitches and GP/CPU fan.

This is the kind of technology advancement that can change almost everything in its field if enough people with vision can take advantage of it and work together to make it seamless.

Academic work

[Bananatree3]

multi-terabit connections are an absolutely wonderful thing to have in some academic research fields. Science research, computing research can all benefit. For some dude downloading movies and music? A 100mbit would be absolutely wonderful and gigabit would be more than enough.

Re:Academic work

[porpnorber]

You go over the 1G mark just by doing uncompressed HDTV, and uncompressed is good; for teleconferencing applications, codec latency is the killer, since your brain is hardwired with estimates of other people's response times. Now, you may think that HDTV is good quality, but if the future offers me 64Mpixel HDR images in stereo (or better, with full depth representation) at 100fps, I for one am not going to complain. Multiply it out; that's approaching the terabit per second, and I didn't even have to choose any outrageous numbers—2*8k*8k*3*16*100 is pretty conservative for a convincing virtual French window. Contemporary video, even HDTV, is not enough like being there, as you come to realise once you've had a chance to play with high-end systems (my stuff: http://ultravideo.mcgill.ca/activities.html; my friends': http://www.hp.com/halo; both a few years old by now).

So, yeah, what you really want the terabit network to your home for—is chatting with your mum.

I wish I could show you even current research teleconferencing systems in operation... and they suck compared to what I'd like to be doing.

(I'm not, by the way, suggesting that there are no useful low-latency techniques providing moderate compression for when you don't have gigabandwidth—of course there are. I'm just pointing out that these numbers are not unimaginable, and that if the pipe were provided, there would indeed be end-user applications for it.)

Re:Nice, but what for?

[stormguard2099]

Please let's not start that debate again. I know it started a long time ago with "no one needs an abacus, who's going to count over ten?" but please no more debating on what's sufficient and what's not as far as computing, etc. It comes up everytime there is talk of major increases in x aspect of computing. We don't need anymore of it.

Botnet, anyone?

[Myria]

Just wait until someone with one of these gets Trojaned and the controller starts DoS-extorting Google.

Whoa

[NIckGorton]

If its "Terabit-Per-Second Class Connection" I wonder what a first class connection gets you.

separate channels

[j1m 5n0w]

True, but the routers and repeaters on the backbone have buses don't they?

The way a lot of telco hardware gets around the limitation that no computer exists that's fast enough to process the full available throughput, is that the connection is split into hundreds of separate channels, each one on a separate wavelength. A particular router interface need only deal with one channel, not all of them at once. (A single channel might be an OC-192, which runs about 10 gbps.)

The channels are combined and split apart by a dense wavelength division multiplexer; I don't really know how they work, but if you think of it as an expensive prism you're probably not far off.

Re:QAM (how to fail physics 101)

[Crypto Gnome]

They're been doing way more than QAM in the last decade, they're doing 64-way amplitude modulation, with frequency spectrums (cable) for ages How the fuck are they using multi-frequency modulation techniques on light rays (fibre) ?
Are you aware that "radio waves" and "light rays" are fundamentally the same thing?

<Massive generalization> anything we have worked out how to do "with radio" is something that there is no fundamentally intrinsic reason why we should not (one day) be able to work out how to do "with light"</Massive Generalization> (and don't bother saying things like passing 'radio" through a sheet of cardboard which obviously blocks "light" - I'm talking about *uses* ie modulation/signalling techniques, not "modifying the laws of physics" issues)

Or do you think that a 1kHz audio wave is in some *magic* way fundamentally and intrinsically different from a 5kHz audio wave? or a 25kHz wave?