Slashdot Mirror
New Ethernet Standard — Both 40 and 100 Gbps
Artemis recommends a blog entry that does a nice job of summarizing the history and current state of the Higher Speed Study Group and the IEEE's next-generation Ethernet standard. "When IEEE 802.3ba was originally proposed [there] were multiple possible speeds that were being discussed, including 40, 80, 100, and 120Gbps. While there options were eventually narrowed down to just two, 40 and 100Gbps, the HSSG had difficulties [deciding] on the one specific speed they wanted to become the new standard... [T]wo different groups formed, one which wanted faster server-to-switch connections at 40Gbps and one which wanted a more robust network backbone at 100Gbps... Unable to come up with a consensus the HSSG decided to standardize both 40Gbps and 100Gbps speeds..."
Major telcos has increased the upload speed to 800k at a cost for only $70.00 a month.
40Gbps can be 1 meter long on the backplane, 10 meters for copper cable and 100 meters for fiber-optics. The 100Gbps standard includes specifications for 10 kilometer and 40 kilometer connections over single-mode fiber.
I'm seeing the 100Gbps used for infrastructure with its larger bandwidth and longer cable length while the 40Gbps would be used for datacenters, server rooms, etc. with its faster "connect" speeds (clarification on what exactly this would mean?).
><));>
I'm normally not one to do this, but the article linked is nearly identical to the coverage over at Ars Technica. It seems that only a few words were changed and without even a link to the original ars article.
Slashdot: where repeating an article in a post is "+5 Insightful"
You misunderstand; one group said "We need to connect our servers to the switches with a faster connection." The other group said "we need to make our network backbone more robust by adding faster connections between buildings and such." The group that needed faster server-switch speeds don't need 100Gbps, they just need better than what they've got. The group that needed faster building-building/infrastructure links didn't believe 40Gbps is fast enough.
Adding both takes care of both groups of people.
"Victory means exit strategy, and it's important for the President to explain to us what the exit strategy is." G.W.Bush
High Speeds Standards Group. How hard is it to read the summary? Slashdot: where people don't only RTFA, they don't RTFS.
Well, if you think about it, Beowulf and similar Linux clusters take advantage of network speed to distribute processing load. This isn't really a case where the network does the computing but with 40 GBs of bandwith, you can perform some serious parallel processing.
The big problem with ethernet's design was its "spew everything to everyone" mentality. In practice, this was fixed by good switches becoming almost as cheap as hubs.
The main alternative to ethernet was token ring, which works much like a meeting where you have big stick that's passed around, and only the person with the stick can talk.
Not a typewriter
If you want all the gory details rather than a copy of a summary of a summary, here is a link to all the presentations at the meeting.
. html
http://www.ieee802.org/3/hssg/public/july07/index
Read through the minutes (warning PDF) to get a summary.
Motion #4: Move that the HSSG adopt the following objectives in replacement of
existing HSSG objectives:
o Support full-duplex operation only
o Preserve the 802.3 / Ethernet frame format utilizing the 802.3 MAC
o Preserve minimum and maximum FrameSize of current 802.3 standard
o Support a BER better than or equal to 10-12 at the MAC/PLS service interface
o Provide appropriate support for OTN
o Support a MAC data rate of 40 Gb/s
o Provide Physical Layer specifications which support 40 Gb/s operation over:
- at least 100m on OM3 MMF
- at least 10m over a copper cable assembly
- at least 1m over a backplane
o Support a MAC data rate of 100 Gb/s
o Provide Physical Layer specifications which support 100 Gb/s operation over:
- at least 40km on SMF
- at least 10km on SMF
- at least 100m on OM3 MMF
- at least 10m over a copper cable assembly
I wonder if it has something to do with latency. Maybe the 40Gb connections are faster because they have a simpler routing protocol or they use smaller packet sizes with no CRC. I haven't been able to get through to the actual proposed spec yet, so it's hard to say...
Just junk food for thought...
What does the 40Gb standard have that the 100Gb standard doesn't cover?
In one word: cost. The 100Gb connection is limited to fibre optics, whereas the slower connection support copper. Fibre optics are still more expensive than copper. It should also be noted that backbones deal with more traffic than non-backbone networks. Think of the difference between inter-city high ways and local back streets and you should get the picture.
Jumpstart the tartan drive.
You must be... no, you're definitely new here.
True confidence comes not from realising you are as good as your peers, but that your peers are as bad as you are.
There is already precedent for this at 10G. The LAN people wanted the data rate to be 10.0 Gbps so that it was exactly 10x 1G while the WAN people wanted something compatible with SONET OC-192 wide area transport gear (9.95328 Gbps including framing bits). So they adopted both. They're not compatible at the physical layer so you'd never plug one into another. However they use compatible layer 2 formats so it's easy to switch packets from one to another.
>High Speeds Standards Group. How hard is it to read the summary? Slashdot: where people don't only RTFA, they don't RTFS.
No. It's High Speed Study Group. In IEEE 802 this makes a huge difference.
A study group studies and recommends what standards are to be written by a Working Group (in this case, the WG is 802.3). They do this by arguing for a while then drafting a scope and purpose for the new spec (you'll find this in the first few pages of each IEEE spec). This is sent up the hierachy (the IEEE 802 EC (executive committee) and IEEE SA NESCOM (IEEE Standards Association New Standards Committee)The Working Group then goes off and writes the spec if the EC and NESCOM approve the PAR (Project Authorization Request).
So the HSSG is not a standards writing group at all, it is a bit of pre work to decide what work is going to be done. Arguing over link speeds is exactly the sort of arguing it is chartered to do.
Evil people are out to get you.
>Yep, Token Ring was indeed more efficient. Good luck reviving it.
Token Ring (spitting) was only more efficient as compared to the original ethernet specification, with all of its collisions. Once we went to a switched architecture and reduced all conversations to two participants that advantage evaporated.
Remember this, being deterministically bad is still bad. Have you ever been on a ring with > 200 nodes? Don't.
Ethernet won because it was cheap. It beat token ring to switching. It beat everything else to get to 100Mbps. Now with 1Gbps and 10Gbps firmly entrenched in the market I look forward to deploying 100Gbps links.
Ethernet is (and was) better.
Dennis Dumont
P.S. I've already scavenged all of my lobe cables for their copper.
Why have one standard when you can have two instead! This strategy has worked so well in the past...
Ethernet is useful because it's cheap, I can attach a 10bt host to a switch and have it transmit the same frame over 100kbt with very little work. I have clients that love Ethernet it's orders of magnitude cheaper than it's main alternative Packet over Sonet. So pretty much it's good enough for most and cheap. In the PC server world the marketing guys want to say they have the latest and greatest copper Ethernet built in and supporting every old standard back to 10bt. This means they ask there chip suppliers to build it and make it cheap. Scale and cutting every corner possible drive down the costs so that it's a couple bucks to add multiple ports of 1000bt today and 10kbt is getting cheaper and cheaper.
No sir I dont like it.
With the 12x QDR InfiniBand spec, 96Gb (after factoring the protocol's overhead) is already on the table and at much lower latencies. This is more helpful for parallel applications (though it really depends on the properties of your application). I've not even worked with 12x nor any applications that would benefit from it. We currently run a 4x SDR setup (which will soon be upgraded to DDR) and it is ample for most of our needs. A cheap 40Gb ethernet solution would be killer for consolidating node management and a storage pathway onto one network. Our current storage solution over 10Gb leaves us with a 25:1 oversubscription ratio which will work quite well for our current crop of applications and how they are used, but it could become a bottleneck in the future.
I think having 40Gb will be really nice once pNFS implementations start to take off. Imagine a pNFS cluster of 32 fully loaded x4500's with 40Gb links between hosts and a 100Gb copper uplink to feed an army processing nodes. Getting close to 1PB of really really fast storage... over NFS and with today's capacities, no less.
Those of us in security are dreading this. IDS/IPS companies are only now dealing efficiently with multi-gigabit solutions for a reasonable price, and no one that I have talked to will do line-speed 10Gbs processing (some boxes can use parallel processing to handle streams from multiple inputs going up to 10Gbps, but not from a single line through a single processor to ensure that attack streams are properly reviewed). I shudder to think of what a 40Gbps stream will be like to monitor.
You can never go home again... but I guess you can shop there.
So does that mean that their either coated in ice or being dug up by MN/Dot?
This sig isn't original enough, it's time to come up with something witty...
When 10Mb Ethernet came out there was widespread debate about its performance, because computers weren't fast enough to saturate it. It was probably the same for 100Mb, and I know the early 1Gb NICs could only handle ~700Mb.
Probably for quite a bit. The biggest hurdle with ethernet is dealing with half-duplex connections and all the collions/detections. These new standards dont even do half-duplex. Everything is full duplex, thus requiring a switch. You've tossed out your biggest setback right there.
Ethernet still is pretty lean. I can imagine an alternative to it, but it might not be worth the trouble, like the anyLAN stuff from a while back. We also still used TCP, but really dont need all the overhead it generates.
You are correct Broadband=2Mb/s+, all other contrary comments are silly marketeer-spin for politicians and corporatist.
Also, the USA ranks 20+ in telecommunications (we ain't #1), because of corporatist marketeer-spin to silly politicians.
AAMOMFF, the USA ranks #1 in international debt only. We're #1, We're #1, We're #1 in debtor nations. THANK GOD and POLITICIANS!
!HAVEFUN!
Unaccountable leaders are masters, and unrepresented people are slaves. How do US and EU fare?
The scaling issue had to do with CSMA/CD, collision detection. To detect collisions, the network propagation diameter/delay must be at most the slot time.
These newer versions of Ethernet apparently don't bother supporting CD. All links must be switched through a hub, period. The hub saves up your packet and prevents collisions, and forwards your packet onto the next link. The "Ether" and "Like Talking" aspect of Ethernet has been lost. Ethernet has become just another framing choice other than SONET, for optical fiber.