10 Terabit Ethernet By 2010

Eric Frost writes "From Directions Magazine: 'Because it is now impossible to sell networking unless it is called Ethernet (regardless of the actual protocols used), it is likely that 1 Terabit Ethernet and even 10 Terabit Ethernet (using 100 wavelengths used by 100 gigabit per second transmitter / receiver pairs) may soon be announced. Only a protocol name change is needed. And the name change is merely the acknowledgment that Ethernet protocols can tunnel through other protocols (and vice versa).'"

Comments

Good stuff

[mao che minh]

From the article: "iSCSI (Internet SCSI) over Ethernet is replacing: *SCSI (Small Computer Systems Interface..."

iSCSI is far superior to stadard SCSI for obvious reasons, and its widespread adoption will really spark a massive gain in the SAN (Storage Area Network) market. The technology is there, now we just need more major vendors of SCSI devices (especially storage and image filing systems) to make more SCSI devices that support iSCSI natively and applications that take advantage of it. Combined with practical solutions from vendors of network storage software like Veritas we could see some major spending in IT. And more money being spent on IT is always a good thing.

I don't keep up much with the progress of the Ethernet technologies at hand, so is it realistic to suppose that the practical implementation/creation of 100 Gigabit Ethernet, 1 Terabit Ethernet, and 10 Terabit Ethernet will be seperated by merely two years each?

"Because it is now impossible to sell networking unless it is called Ethernet". Incorrect. You can easily sell network gear that is tagged with the "WiFi" designation.

Re:Good stuff

[prrole]

iSCSI is NOT far superior to SCSI, or fibrechannel. iSCSI has massive issues related to deterministic latency, and computational cost of processing TCP/IP at gigabit speeds. You may see some growth in the of iSCSI in the workgroup segment, but I don't see iSCSI replacing fc/scsi in the near future for mission critical computing.

Re:Good stuff

[4of12]

iSCSI

A really nice development.

Yet more big advantages to iSCSI are the ability to keep the

• large,
• noisy,
• power-hungry,
• heat-generating,
• unsecure

disks out of workstations in workers' offices and down the hall in a

• sound-proof,
• secure,
• air-conditioned,
• UPS'd server room with
• mirrored images,
• archival backups

.

Next thing you know, GPUs will come with on-board Ethernet controllers and USB plugs for keyboard and mouse, and be built in to the back of an LCD monitor.

Re:Good stuff

[Austerity Empowers]

I don't keep up much with the progress of the Ethernet technologies at hand, so is it realistic to suppose that the practical implementation/creation of 100 Gigabit Ethernet, 1 Terabit Ethernet, and 10 Terabit Ethernet will be seperated by merely two years each?

I think not. 10 GbE hasn't exactly taken the world by storm and it's been around for over a year now. I know of products that have 10 GbE ports, but I have not witnessed an abundance of demand. To be nice the author of this article is just a little facetious in his claims.

In reality if you read the article it's hard to even take him seriously. To say that Nortel's DWDM system is ethernet is like calling your 56k modem ethernet. Yeah, so you can pass ethernet frames on it. It's not standard, it's not documented anywhere in IEEE 802.anything (esp with regards to conformance), so it's NOT ethernet. Just passing ethernet frames does not make you an ethernet device. I'm honestly not really sure what the author's point is except that he seems to think 1) ethernet is increasingly popular, 2) everyone should want to carry ethernet frames, and 3) people want bigger and bigger pipes. The first 2 are true, the third is less true now than it was 3 years ago.

So the answer is, it wouldn't surprise me if we see 10 Terabit links by 2010, I doubt very, very much that we'll see a 10 Terabit ethernet port on a single chassis ethernet switch with 100 Terabits of switching capacity. I could be wrong, I hope I am, but it doesn't seem reasonable.

And what am I going to do with 10TB ethernet?

[raehl]

Is there going to be storage that can read/write that fast by 2010 too?

Re:And what am I going to do with 10TB ethernet?

[mirko]

I guess there will only be one computer, at this time : a virtual computer distributed over millions of physical nodes, so the storage will might be each of these nodes' memory... Like Freenet but also aimed at distributing workload.

Re:And what am I going to do with 10TB ethernet?

[Abm0raz]

Will it be meant for actual LAN usage? I think it's being designed more for back-bone-like reasons. I can't even get my drives to transfer large amounts of data back and forth in a reasonable time, so I can't see the need unless we go to entirely solid state drives.

but ... imagine a beowulf cluster interconnected by lines of these ... ;)

-Ab

Re:And what am I going to do with 10TB ethernet?

[default luser]

Actually, bandwidth and route have everything to do with latency.

The efficiency of the routers / backbones you encounter is always a factor, and if one router in the chain takes forever to respond, it's going to kill your latency.

Your packet has a certain size, and the time it takes to completely transmit that packet and complete the ack is your latency. Distance and bandwidth are the prime factors.

Sure your packets travel fast on a fiber backbone, but if your last mile connection is several orders of magnitude slower ( broadband or dialup ), it's going to cause a significant increase in your latency.

Even high bandwidth cannot save you from real distance. You try to play a game on the other side of the US, you're going to add a sizeable delay even with those high-bandwidth backbones. Gaming with a server on another continent? It becomes largely unplayable.

Hell

[Dachannien]

...is there going to be a bus on desktop machines that can read or write that fast?

Probably not. But I could definitely see it being useful for top-end server systems with hugely parallel storage and memory access.

Re:boy! If you could build a Beowolf Cluster of th

[DeathPenguin]

Interestingly enough, if you did it wouldn't be a very big success because the internal PCI or PCI-X bus in the system would bottleneck the interconnects. The NICs would need on-board processors to scale with their enormous bandwidth potential so that they could solve problems like matching checksums and other package management tasks and not have to pound on the system bus so hard.

It wasn't long ago that we really started exploiting video chipsets for rendering graphics, either...

packetengines

[joeldg]

I am sure packetengines (http://www.scyld.com/network/yellowfin.html) is all over this.
These guys had gigabit routers four years ago when I was helping to set up the AFN (ashlandfiber.com)

Cool to see.. mo'faster is mo'betta

Re:What about latency?

[Brahmastra]

100 ms latency would affect a 10Mbps network and a 10 Tbps network almost equally if a clustered application is using very small packets to communicate. Only if the application is using very large packets, the bandwidth will overcome the latency. At small packet sizes the latency will largely overshadow the bandwidth. And considering that a lot of scientific applications use small payloads, latency is very important. If ethernet wants acceptance in the High-Performance-Computing-Clusters world, something has to be done about the latency.

Better question...

[siskbc]

Will there be a computer with a bus that can transmit data that fast? To hell with read/write - I'll concede it's all in memory. I don't think computers will be able to do (10^13)/64 bus cycles by 2010, assuming Moore's is loosely adhered to. As I calculate it, 7 years at 1.5 years/doubling cycle leaves 4.8 doubling cycles. Assuming a top speed of 3 GHz and 64 bit architecture, 1 get 1E13bits/(64bits/clock)/((2^4.8)*3E9clocks) = 1.87, or 87% overcapacity.

And that assumes that transfer occurs at chip speed, which it doesn't. Assuming a modest clock multiplier of 8 between system bus and chip, that's a 15x overcapacity, even if the entire computer were used to transmit.