10-Gigabit Ethernet Standard Approved

A little birdie brings news that that 802.3ae standard for 10 Gigabit/second Ethernet has been approved. Everyone out there with Gigabit Ethernet - you are now officially obsolete. The new standard is fiber only, no more of that nasty copper stuff.

Comments

As if 1000BaseT didn't suck enough CPU cycles

[green pizza]

I'm guessing 10GbitE will be used for inter-switch and inter-router connections long before it gets to the desktop. Ever looked at performance comparisions between 100BT and 1000BT between just two PCs? A couple years ago the difference wasn't much... NICs weren't efficent enough and the host PC's didn't have enough CPU power to handle that many tiny packets per second. Jumboframes and faster CPUs have helped a lot since then, but we're still a long ways away from even 90% utilization between two PCs with 1000BT. And here we are with 10GigE, with 10x as many packets per second.

I'm I the only one that thinks the only efficent 10GigE NICs are going to be PCI-X cards with an onboard 2.6 GHz P4 co-processor and 512 MB of buffer?

Re:As if 1000BaseT didn't suck enough CPU cycles

[Enry]

There was an article in the Linux Journal a few months ago (February issue I think) that talked about intelligent network cards. They had an onboard XScale CPU and its own OS and TCP/IP stack.

What would happen is the OS (Linux) would get intercepted at the socket layer and pass the data to the network card. The card would then handle the process of building the packet and all the remaining layers of communication.

This allowed for a high amount of main CPU time left over for actually doing processing while the network card CPU was focused on handling the TCP/IP packet work. IIRC, you could saturate a 1Gb line with data at only 5% main CPU usage.

Re:As if 1000BaseT didn't suck enough CPU cycles

[bluGill]

My company once did this for a 25Mhz dec machine. We discovered that you now need a new protocol to the adaptor card, and the overhead of that protocol is equal to a well tuned tcp/ip stack. So if they can actually make this work, what it really means is that linux hackers should spend some time to tune the stack.

Note though that tuning the stack may come at the expense of maintainability, or flexability.

Re:With this annoucement

[stevelinton]

Actually, burying pipe is the most expensive part..... I worked briefly for a university computing service a few years ago and they spent an absolute fortune to buy a network of yellow plastic pipe connecting all their buildings. A relatively trivial incidental expenditure was to pull some cable through it. When that sort of cable is obsolete, a further trivial expenditure will replace it, etc....

10Gb over copper? Won't happen!

[xt]

The reason is material properties.

Six months ago, I had the chance to talk with the 3Com technical manager who was on the board drafting the spec.

What he said was very simple; all tests indicated that the only way to have 10Gb over copper is to limit the connection distance to centimeters!

1Gb already pushed the envelope for copper, using all pairs, multiplexing, and error correction; 10Gb is just not possible.

Anyone knows the MTU?

[Oestergaard]

Does anyone know how big packets one can send thru such a pipe ?

100MBit maintained the same MTU as 10MBit, 1GBit maintained the same MTU too - leading to severe problems with performance. It's bad enough on 100Mbit, it's horrible on 1Gbit, to think that they maintained the 1500 byte limit on 10Gbit gives me the shakes...

Yes, I know about "jumbo frames", and I challenge you to find an affordable 1Gbit switch that actually supports it.

Anything below 64KByte packets would be insane as I see it.

Anyone knows ?