Slashdot Mirror
Researchers Transmit Optical Data at 16.4 Tbps 2550km
Stony Stevenson writes "The goal of 100 Gbps Ethernet transmission is closer to reality with the announcement Wednesday that Alcatel-Lucent researchers have recorded an optical transmission record along with three photonic integrated circuits. Carried out by researchers in Bell Labs in Villarceaux, France, the successful transmission of 16.4 Tbps of optical data over 2,550 km was assisted by Alcatel's Thales' III-V Lab and Kylia, an optical solution company. The researchers utilized 164 wavelength-division multiplexed channels modulated at 100-Gbps in the effort."
What's that in Library-of-Congresses per fortnight?
Knowledge is how to play a game, intelligence is how to win, wisdom is knowing what game to play.
Would this qualify as 11?
Surely there must be some incredibly processing power behind transmission speeds like this? Anyone one have any idea?
That's a lot of cows.</walken>
Carousel is a lie!
That's just BURST throughput. Depending on factors like time of day, how many other users there are, and environmental conditions, throughput may drop as low as 33kbps. And we do NOT filter bittorrent.
Just check your TOS agreement. It's all right there.
In other news: American telcos wonder how French providers are able to afford research and development without additional funding from a tiered billing billing scheme that is needed to advance the science in the United States.
With the ever-predicted bandwidth crunch always just around the corner, can existing cables be reused just by replacing the signaling equipment at substations with this? If we don't have to lay all new cables - just upgrade the nodes - then upgrading to these bandwidth capacities on our current networks would be a cinch.
Shh.
"164 wavelength-division multiplexed channels modulated at..."
how very Star Trek of them.
You feel sleepy. Close your eyes. The opinions stated above are yours. You cannot imagine why you ever felt otherwise.
Tbps speed, and over 100 Gbps. Something is wrong here.
Slim article... How long would it take to error check that much data?
on Another note... What did they do with all that Pr0n once it got to the other end?
--- Relax, that mass muderer is just trying to reduce our carbon footprint, one fetus at a time...
But what is the latency involved in a system like this? Right now I get about 25000kbps and that is plenty for me, but what I really need is reduced latency for real time simulations.
No matter how much speed they create, they will still be subject to the Law of Diminishing Porn Returns, which states:
For download rate n, my demand for new porn will require me to download at a rate of n+1.
Surely if an ISP adopted this, they'd have people signing up left right and centre. Wouldn't it be awfully attractive to their target audience?
Where I the only one thinking a truck filled with DVD's when the headline said optical data?
Well the optical transmission would go nearly the speed of light, but I'm not sure how fast the transmitter and detector can keep up with splitting and recombining 164 channels of data.
I'm sure this is being looked at as a means of an intercontinental backbone, rather than something coming into your house.
GE/CS/IT d- s: a- C++++$ UL+++ P-- L++++ E W+++$ N+ o? K- w---() !O M- V- PS+ PE(++) Y+ PGP+++(+) t+++ !5 X++> R- t
Consider a standard computer utilizing two channels of ddr2-800. This only has a bandwidth of of 102.4 gbps. You would need to attach this connection directly to your ram in this case. Now computers are starting to utilize ddr3 which can theoretically provide double that in the same configuration. But at those speeds, amount of available ram would be limiting the actual throughput. 100gbps utilizing 128GB of ram would be emptied in 10.24s (1024gb/100gbps). Having to have a storage array capable of sending out 100gpbs continuously would require at least 125 drives at 800mbps(100MB/s is a good number) constantly. This is the reason 100gpbs isn't being considered for lan use. It just isn't feasible at this point.
Jeepers, that's really cool.
Meanwhile, I can't even get FIOS service in Philadelphia, one of our major cities, despite my keen desire to purchase it.
Where can you get 10G Ethernet cards for $50? I'd really like to know and I'm sure there would be a lot of interested people...
This is the reason 100gpbs isn't being considered for lan use. It just isn't feasible at this point.
Stick a thousand machines on each end, and you'll understand why 100Gbps is needed.
Unfortunately both target audiences (that's 1. Supercomputing/clustering big/experimental stuff and 2. being a Tier 1 backbone provider) for this kind of technology don't need their ISP to provide it. Your average SoHo couldn't even deliberately use a single Gigabit link; even well-connected datacentres don't (yet) need 100 Gbps outside connectivity.
If you really need greater than 10Gbps then go with Infiniband as you can get 12x HCA's that will do 24GBps (48Gbps full duplex). But if you're paying $50 for 10Gbps ethernet you're not getting offloading and your CPU's are probably swamped of your TCP/IP stack is the problem. I would suggest getting a pair of offloading 10Gbps cards and seeing if you don't see a huge improvement.
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
I'm running Linux on a Playstation3 with SPU video drivers in its Cell uP that can run at over 150GFLOPS. Since the PS3 has only 512MB RAM, it needs to be fed by the LAN and just buffer the LAN in its RAM. Even if SATA drives are delivering only 1.2Gbps, there's no reason I can't have multiple parallel drives on independent servers (if a single server's IO isn't fast enough for multiple SATA at full bore) on my SAN delivering multiple streams through my switch all to my Playstation. Now, the PS3's 1Gb-e is as hardwired to it as is its 512MB RAM, but the point is that there are already machines that can use that bandwidth. The total bandwidth doesn't have to reach 100Gbps, but only exceed 10Gbps, to require faster than 10G-e, which only 8-10 SATA drives in parallel could do today.
So the bottleneck is 10G-e. There are already supercomputer clusters using multiple parallel Cells, so I'm disappointed that they're not already widening the pipe.
--
make install -not war
There is another problem, and is actually the bottleneck of transmitting packets at high rates.
It doesn't really matters (yet, and considering Ethernet technology) if the BW of the fiber is a zillion Petabits/sec.
The problem is now at 1Gbps and 10Gbps in Ethernet technology, and is because the processor overloads with the amount of hardware interrupts. The processors that are general purpose have to waste too many clock cycles processing that much interrupts, the processors nowadays are superscalar [ http://en.wikipedia.org/wiki/Superscalar ]and every time the processor have to change the context (to attend an interrupt) has to do lots of things like unloading the registers, saving the context, loading the registers of the new process, and has to drop something out of the pipeline [ http://en.wikipedia.org/wiki/Pipeline_(computing) ] loosing performance.
Ethernet tech has a huge latency [ http://en.wikipedia.org/wiki/Latency_(engineering) ] and a stack that makes processing not so easy (if you look at te code of a linux network device driver it handles pretty much everything including writing the mac address that is only copied when the driver initialize).
That is why there are some relative new things (NAPI in Linux) that try to make lessen the overload, there are new network devices that handle layer 2 and 3 (or at least parts of those, for example, is used to be handled the checksum algorithm) to avoid doing it in the processor. There are some white papers (one from intel, another from NetXen, I'm sorry I don't have the links now) that explain the problem and some approach to a possible solution.
Yes, I know, there is something I have not said, and is that the main switches or routers have to deal with that and have hardware specially designed to do heavy network packet processing, and that is the point, the network cards will have to do that (and are already starting to), neither is an easy job for hardware designers, nor for the market, is easier and cheaper to have a machine that you can change the behaviour only changing the firmware or changing settings from a program (routers have an operating system, and lots of those are a general purpose microprocessor with a linux kernel and a web server to configure it, for example home routers).
There is much to say yet in this field.
"Everyone" is waiting for either 40Gbit/s or 100Gbit/s Ethernet.
The first one is what server-people push (they claim they do not need more, that's why 40Gbit/s was put into Ethernet standard),
while network people want full 100 Gbit/s.
> But what about all the LAN vendors, which have a real market for 100Gbps
They don't.
There seems to be market either for 40Gbit/s in LAN/local connections or for 100Gbit/s for core/long haul. At least judging but what happened with high-speed ethernet standard.
the title says 100 Gbps
the article says approx. 16 Tbps
and the last sentance says how close we are to creating 100 Gbps ethernet and describes how the terabit link was created using multiplexed signals at 100 Gbps.
so what the heck am i missing, because im confused as hell
feh, lots of things are pointless, this one too
I had found with Froogle a Sun 10G-e card (probably refurbished) for $35, but it's not available now on searching again. It does seem that 10G-e cards do cost something around $500 at least, and plenty up around $900+.
But the point is that I can't buy anything faster for even 10x or 100x, except multiple cards. And maybe some really exotic interconnect that's not ethernet, so apps have to be recoded to use it.
--
make install -not war
Infiniband isn't ethernet.
--
make install -not war
Interrupt loads can be greatly reduced by switching to a polling-driven architecture. See FreeBSD.
Mea navis aericumbens anguillis abundat
Part of the Ethernet spec is to wait 9.6 microseconds after the medium appears to be idle before sending, then resend if it collides. Light moves about 3 kilometres in that time. Making an Ethernet of 2550 km pratically guarantees nothing but collisions. So while this is a hunkin' heap of net, it's not Ethernet.
You're apps don't talk to Ethernet, they talk to your TCP/IP stack! There is an IP over Infiniband standard which is supported by most HCA's and it's even fully offloaded to the HCA processor. IPoIB isn't as efficient or low latency as native Infiniband protocols but at least you don't have to recode your apps =)
And the reason there isn't anything faster is that it's such an incredibly niche market, the number of sites that need greater than 1GB/s on a single link are very, very small. Heck the storage for even big servers is generally only 4Gbps per link as 10Gb FC is still astoundingly expensive (Cisco 10Gb FC blade for instance is $60K for 4 ports) so it's cheaper to just put a couple 4Gb cards in and get additional redundancy for free.
P.S.
This is post #5,000 for me =)
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
Well, it turned out that 10Gbps interfaces actually cost more than $50, probably more like $500 per port. But they don't require recoding apps to use the Infiniband fabric.
The point is really that there's a huge jump from a $15 1Gb-e (or $500 10Gb-e), if that jump can be made at all, while there's no 100Gb-e at any price. And instead of rolling out 100Gb-e that works for LANs, the industry is evidently waiting until it's good for continent-spanning WANs.
--
make install -not war
In my experience I have great demand not for 100Gbps Ethernet, but Fibre Channel. It is a common problem is replicating data between data centers. The high end approach to this is to get two DWDM switches and dark fiber (if you can afford it).
The main problem with transporting large data segments over Ethernet is quite simple. Data is read off disk (typically) at 8KB chunks, then converted to 1.5Kb Ethernet packets. This "fragmenting" and additional overhead is what causes allot of the resultant lag.
You keep talking about recoding apps, YOU DON'T NEED TO. There is a standard for IP over Infiniband. It's not as performant as native Infiniband protocols but it DOES exist and is in use. Besides without a dedicated specialized processor and gobs of buffer memory your typical server isn't going to be able to keep up with 10Gbps Ethernet let alone anything faster.
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
How can it be a few years until "any breakthrough" in >10Gbps when we're discussing successful tests of 100Gbps across 2550Km already? Sounds like the breakthroughs have already arrived, but telcos are aiming for a totally different market than LANs. If it's working even most of the time at 2550Km, then it should be pretty close to commercial availability at 10m or 100m.
--
make install -not war
I don't see why a fast ethernet card can't have its own DSPs and/or FPGAs for decoding packets, then sending them across PCI-e to XDR memory. For $500 or $1000, 100Gbps HW should be able to deliver.
--
make install -not war
Hilarious... The fortune that I had at the bottom of the comments for this artcle:
There is more to life than increasing its speed. -- Mahatma Gandhi
Did someone plan that?
Well, "100"Gbps is an arbitrary number. So long as it's significantly bigger than 10Gbps. 40Gbps would be good, too, but it's not available either.
The advantage of ethernet is that apps are already coded to use it. Including all the network mgmt apps.
--
make install -not war
>Interrupt loads can be greatly reduced by switching to a polling-driven architecture. See FreeBSD.
New API (NAPI) takes a mixed approach, read: [ http://en.wikipedia.org/wiki/New_API ], and for more information: [ http://www.linux-foundation.org/en/Net:NAPI ].
So, you want to create an SMP type cluster where each CPU is connected with another access a 100Gbps switched network? And your application sounds like it needs SMP to work but you want to go "on the cheap" and instead of getting a SMP supercomputer or even renting space in you, you just want to have a rack of playstations instead with 100Gbps network adapters?
I suggest you fix your algorithm or get/rent a machine that can handle your data instead of guessing playstation will ship with 100Gbps ports for you anytime soon.
Hmm, does IP over Infiniband run under Linux on a PC with PCI-e? Or any other <$5000 hardware?
--
make install -not war
PS: That post my 16,362nd. Probably fun to post #16384. My UID is (nearly exactly) 1/3 yours, so we were at equal rates when I posted #15313 - you're at 93.6% of mine. We both probably need a faster connection just to keep up with our posting rate ;).
--
make install -not war
Can You Say Linux? I Knew That You Could.
40 SATA drives is no big deal. I've got over a dozen SATA drives running in my home, mostly to store my media collection. And I haven't even begun collecting video, because 750GB-1TB drives aren't that cheap yet. Once drives drop below $0.10:GB, and there are more HD video titles (especially if for download or TiVo), 40 SATA drives will be fairly common in many homes, to say nothing of the media servers that are servicing them.
--
make install -not war
Yes, yes it does.
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
The movie industry just shat a brick.
The part you missed was where I said "the point is that there are already machines that can use that bandwidth". My TV's HDMI port already does 10Gbps. When I want to mix a couple PiP streams with incoming videophone and a peek at the front door for the pizza delivery, all in HD swapped around my house to different TVs to get someone to answer the door so I can stay on the phone (or some other common combination), that kind of bandwidth won't seem so exotic.
Of course I want to go "on the cheap". Everyone wants to go "on the cheap": the lowest price possible (even if no lower than possible). Even before I get that kind of home media distribution, I do indeed have apps (like VoIP) that I'd like to invest $5000 on some Cell blade processor rather than $50K on just the networking.
Bandwidth demands are exploding. Even telcos don't have 100Gbps over 2550Km yet, as the demo we're discussing demonstrates. "640K ought to be enough for anybody" should be warning enough. When "on the cheap" is near the limits of what what's available at any price, we need to hurry up and extend the horizon further. Or we'll soon fall off the edge.
--
make install -not war
swift kicks in the groin, when all you get at the house is a "blazing" fast speed of 26.4 Kbps on dial up. Thank you Verizon, fucking thank you... ISDN is not even an option where I live.
I've got a very large collection of live music recorded (legitimately) by people in the audience, like Deadheads, over the past 30-50 years. And a solid percentage of that by bytes is video.
A 180GB drive holds only 45 DVDs. When you're talking about hundreds of shows on video, you're talking about dozens of drives.
It's not yet at the point where I need more than 1Gbps, but that's mainly because I don't have the SW quite running seamlessly for real home multimedia. But I'm still accumulating content, including a Blu-Ray drive (40GB movies, not 4GB). By the time >10Gbps ethernet is <$100 a port, I'll be able to justify it with my personal mediasphere.
Yes, we do live in an age of miracles. "Fast, cheap and good", even it's all in the middle range of those parameters, is available, and the peaks available mean even the middle is quite nice. But that's the reason to push the peaks, to raise the middle.
--
make install -not war
I already described in this thread how a single host can saturate 10Gbps. And multiple cards have sync problems. De/muxing 50Gbps across 5-8 cards is a huge problem, in sync and just CPU power that's better off in an ASIC/DSP/FPGA in the network controller.
--
make install -not war
That MHGA28-1TC is very cool. At about $800 for dual 20Gbps, that's $200 for the equivalent of 10Gbps, which costs $500 standalone. But a much more usable package, without de/muxing to get 20Gbps, and probably better ability to "bond" (or equivalent) across both 20Gbps ports for simulated 40Gbps. If only there were more competitors, it might be closer to $100 per each 10Gbps, which is cheaper than even standalone 1Gbps now, and of course much more integrated (lower power, more manageable, less HW...).
:).
Since the product is on the market, there's probably not that much demand right now, or there would be more competitors.
Thanks for clarifying this entire discussion. I rarely learn this much from just whining on Slashdot
--
make install -not war
Let me see if I got this right... They made 164 lights of different color blink very fast and very far. Why can't Slashdot summarize things better?
alias possession='chmod 666 satan && ls
DDR using 102.4 gbps?
Wow, I didn't know you could dance that fast.
Qlogic makes cards too, their Silverstorm 7000 is similar to the MHGA28-1TC in that it supports dual 4x links, but because they are basically a duopoly there isn't much price competition. Actually it's quite similar to FC where there's Qlogic or Emulex.
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
The problem is now at 1Gbps and 10Gbps in Ethernet technology, and is because the processor overloads with the amount of hardware interrupts.
That's not the point of the article. 100 Gb/s technology is not being invented to hook up just 2 or even 100 computers to each other. The 164 wavelengths each carry a different 100 Gb/s stream. This is the type of technology you use when are trying to connect a chunk of Boston to a chunk of Baltimore.
In order to process streams that fast, the first thing you do is demultiplex the streams into wide buses. For example, current 40 Gb/s streams are first demuxed into 16 x 2.5 Gb/s streams, which are then sent to a specialized processor to route the packets. (BTW, there are no interrupts at that level. It's all synchronous, and the processor is dedicated.)
I will tell you that it is a thing of beauty to watch those things in action.
And congrats to the folks at ALU for the result.
JP