Slashdot Mirror
Cheap Gigabit Ether
Avrice writes "National semiconductors gigbit ethernet is backwards compatible with existing systems and smart enough to fix your wiring screw ups for only $95. Maybe bandwidth (at least on the network) won't be such a problem after all."
All this is great until you realize that you are going to have to do something with that data. It doesn't do any good to pull off 1 GBps over the AGP port only to discover that your IDE drive can't write data anywhere near that fast, or that you can't even push the data down to the disk controller that fast. Heck, even doing some sort of simple calculation with that much data is going to overrun your processor in no time.
I read the internet for the articles.
Anybody making cables without at least some sort of tester deserves to have to spend a weekend searching for a bad cable. :)
No, gigabit (and 100b, 10b...) ethernet refers to the raw number of bits you can spew over the wire. It includes all preamble and postamble.
Time to update your crash course. Gigabit Ethernet uses all 8 wires and a form of encoding/compression to achieve its speed.
you are, however, correct in terms of 10bT and 100bT networks. :-)
Gigibit ethernet is a trick, you're right, but in 99% of cases it is not connector problem, it's usually the cable itself went bad for one of several reasons. If you're using good grade cable and your crimper and ends are of good quality, your cable will be fine.
You can't stop the cable from radiating by crimping any better (unless you really blow at making cables). The cable will radiate more than spec allows if you've got sharp bends, mismatches pairs or poor twist. Your fancy-schmanzy cable tester doesn't test for one of the biggest causes of cable failure: stress. Binding cables with tie-wraps too tightly or bending them too sharply often gives you the problems at gigabit speeds that you refer to. Note that fiber has the same "bend radius" problems that copper has, but for different reasons.
From a techical standpoint, there is only so much to go wrong with a cable connection. As long as you're crimping right you'll be mostly safe. Far greater problems come from the way the wire is treated when installed, as mentioned above.
As far as searching all weekend for a bad cable: how the hell are you doing your installs? Computer A can talk but computer B is flaky. Well it sure ain't the backbone connections, check the connection from the switch to the computer in question. Use a network analyzer. It's not difficult..
Gigabit ethernet will give the average-joe cable maker headaches beyond his wildest dreams if he doesn't learn why it's different.
1024/8 == 128MB/s
My ATA hard dive bursts up to 33MB/s (13MB/s sustained).
Perhaps it's most useful when used in confunction with a busy file server.
I say the entire market should drop copper based products and go 100% fibre optic. Start massively mass producing it to jack prices down and make it cheap enough to have in the home.
And while they're at it, replace all phone and cable networks (to our homes) with fibre too. It'll be necessary if we're ever to have the massive multi-media "global village" corporations like so much to advertise about. Some of us are still on dial-up, dammit.
Phone and Cable corps. could get together (it'll be a cold day in hell) and split the cost for this, then compete for our business over shared lines.
Do I know what I'm talking about? No. But it sure as hell would be nice.
-kidlinux.
It is a bit like saying SCSI maxes out at 10 MB/s. It is not generally true.
-jwb
Here's a thought: How about informing the user their network admin #$@!'d up the wiring and refuses to run along with a detailed description of WHY it doesn't run. We should not be letting things like network wiring be done improperly ... it leads to sloppiness and ignorance.
According to the 3COM white paper, all four pairs are used. Hybrids are used, like in a telephone, to allow simultaneous transmit and receive on each pair.
Mea navis aericumbens anguillis abundat
This is something I've been thinking about for a while. How fast would a network have to be before it becomes faster for one system to swap to another's physical memory than to a local disk?
:)
Heh heh...my roommate better start paying close attention to his memory usage...or it might start disappearing
The best part is, with these chips you won't even need the null cable. Just use a regular patch cable, and the chip will fix the "wiring mistake". Kind of cool. Now, if only the auto-negotiate doesn't suck...
As a quick guess. I expect to see them come out at about 3-5X the chip price till one gets over 10,000+/Month volume productions. The prices will drop to about 2X. Companies need their profit.
Actually, apart from the fact that token ring works rather differently (it relies on a virtual token being passed from host to host), this could be arranged in a star shape to, provided you have enough pci/agp/whatever slots.
Yeah.. I remember....
except, Gigabit ethernet = 1000000000 bits/second. and is not based on a power of 2.. so why change it?
Powers of 2 only apply, generally, to memory.
If your existing wiring could barely handle 10Mbps, it wasn't cabled to Category 5 standards. In short, it must have been crap.
A lot of problems cropped up when people started trying to do 100Mbps with shitty cat5 installs (ie: using category 5 cable, but not installing properly) or older category 4 (or 3, I forget) cable... thinking 'it should work, the plug is the same'.
If your existing wiring could barely handle 10Mbps, it wasn't cabled to Category 5 standards. In short, it must have been crap.
A lot of problems cropped up when people started trying to do 100Mbps with shitty cat5 installs (ie: using category 5 cable, but not installing properly) or older category 4 (or 3, I forget) cable... thinking 'it should work, the plug is the same'.
Also, unless I am mistaken, the collision detection is *still* the same as 10/100 networks. The mechanism doesn't change, though the timings do, and the distance requirements change, I bet, probably require a shorter segment again.
The inter-frame gap will be very large compared to the frame size.. hence the maximum speed between any 2 hosts (say, even through a crossover) will probably be a good chunk less than a gigabit, say, 750Mbps....
And the backoff mechanism will still be binary exponential backoff.... so the behaviors are the same.. just different timings.
This is not surprising. I dont' have all the timings in front of me, but the general gist is this.
First, though it looks like a star to you, it's a bus network. I guess if there were a switch instead of a hub, you might get away with calling it a star.
As for your 8Mbps.. that's actually about as high as it's theoretically possible to go. Maybe a wee bit faster, and here's why.
the 100 in 100base (and the 10 in 10base) both describe the signalling rate (or bit rate) of the BASEband medium (the ether in ethernet...). This is different than describing the rate at which 2 hosts can transmit.
What this means is that the ethernet, as a single baseband channel, has bits clocked onto it at precicely 100Mhz (or 10), one bit per cycle.
Now, as part of that standard, there is a mandatory delay any transciever must obey after putting a frame on the channel. In 10Mbps, this 96 bit times, or 9.6 microseconds. I may be a bit off here, but in 100base, this number is *still* arond 9.6 microseconds, as it is a number based in the time it takes for packets to traverse the network from one end to the other; in the case of 100base, 9.6 microseconds = 960 bit-times.
Each ethernet frame consists of an 8 byte 'preamble' (used to synchronize the receiver), the frame header (6 byte source, 6 byte dest) the type/length field (2 bytes) and in the end, a frame check sequence, like a checksum, of 4 bytes. That makes 26 bytes of information, not related to the ethernet data payload, plus a 120 byte inter-frame gap (remember, each bit takes the exact same amount of time on the ethernet, so we can use bits/bytes to reference time).
That makes a total of 146 bytes of non-data. If we add to that, say, the IP header, and a UDP header (assuming we are streaming video, with no handshaking, like TCP, as that would mean the response packets would *also* tie up the channel further), you can see that, given the maximum ethernet data payload is 1500 bytes, we are at over 10% of that as overhead.
This would put the theoretical maximum at around ( I calced it once..) 89%.
Of course, if there is *any* other activity *at all* on your ethernet, this number goes down even further. If you are doing FTP or something, it goes way down....
Now.. I did all this from memory, i'm not 100% sure about the Inter-frame gap on 100base, though I'm sure in 10base. This, of course, doesnt' take into account full-duplex ethernet either...
They quote the transcievers (CHIPS) at $95/ea in quantities of 1000...
This is not at all the same as saying the *retail cards* will be anywhere near that price.
As such a new thing, it wouldn't surprise me if the cards were hundreds of dollars..
THe answer to that is.. it depends.
ON a mixed 10/100 network just now we use 3Com hubs and switches. If you attatch a 10Mb card to a port then the card and the port run at 10Mb. If you then attach a 10/100 card to another port then that port will run at 100Mb (assuming things are configured correctly). The two machines can still communicate, the hub does the rate conversion. Obviously the maximum transfer rate between the two machines is governed by the slower NIC, but if you had a second 10/100 on there then the two faster machines will communicate at 100Mbps despite the presence of the 10Mb NIC on the same segment. First time I saw this working was a real "wow" experience Don't ask how it wall works, I have no idea, but it simply does. I would guess 3com must have some bridging logic for each port, after all the 10Mb NIC could never get to see all the traffic between the two 100Mb NICs running at full tilt, but I have never seen any problems caused by this, and our network is 50:50 10:100. Presumably 3Com could manage the same trick with 1000Mb.
3Com is simply a vendor whose equipment I know from personal experience, I'm sure some other vendors equipment can do the same trick.
Stealing a rhinoceros should not be attempted lightly.
Boy, isn't that the truth. We use roaming profiles here where I work, and our network is S-L-O-W as molasses in January.
Hey Rob, Thanks for that tarball!
"Going to war without France is like going deer hunting without your accordion." - Jed Babbin
264MBytes/sec is more than double 1000Mbits/sec.
Just to review:
gigabit ethernet = 1000000000 bits per second = 1000mbits/sec = 125mbytes/sec
standard pc pci (33mhz 32-bit) = 33333333 transfers per second (+/- 1%ish) * 4 bytes per transfer = 133ish mbytes/sec
mac pci (66mhz 32-bit) = 66666666 transfers per second (+/- 1%ish) * 4 bytes per transfer = 266ish mbytes/sec
alpha and others' pci (66mhz 64-bit) = 66666666 transfers per second (+/- 1%ish) * 8 bytes per transfer = 533ish mbytes/sec
(I say +/- 1% because the clock chip on the average PC isn't at all accurate - your Celeron 466 might actually be running at 463 or 470mhz.)
Furthermore, you don't have to have a computer that can soak the ethernet to get an improvement in speed out of gigabit ethernet over 100mbit ethernet. You just need a computer that can push the packets out faster than 100mbits/sec.
It seems pretty clear that the average celeron box is capable of this.
I remember that a lot of problems cropped up when trying to do 100Mbit Ethernet on existing wiring, which only barely could manage 10Mbps. What will happen to most of the wiring already laid out. Will it have to be thrown out? I remember hearing about a Cat-6 cable. Will we have to upgrade our networks?
It's my understanding that the 802.3ab gigabit over copper standard is intended to work on standard Cat-5 cabling, so there shouldn't be any need to replacing your existing cables. On the other hand, it does use all four pairs of the cable, so faults that may not have been evident beforehand might turn up...
Cat-6 cable certainly exists, and I believe there's also a Cat-7 standard (with individual routing channels through the cable for the individual pairs?), not to mention Cat-5e. I think quite a lot of the demand for this cable is drummed up by the vendors and installers of cable plant.
Also, anybody got information on how collision handling is done on this new architecture? I would suppose that, being a gigabit ethernet, it would surely see much more usage than a 100Mbps one, and being also much higher speed, there should be more collisions.
Collision handling in gigabit ethernets is a functional irrelevancy; although the 802.3z standard does have provision for shared-media networks, the last time I checked there were no products (nor any scheduled) that supported it.
Basically, if you're looking at gigabit ethernet, you're looking at a full duplex, switched network.
For what it's worth, from a technical perspective, I seem to remember that the collision detecting version uses a carrier extension to allow the network to have a useful radius (i.e. in order to avoid late collisions). The carrier extension was for some moderately significant number of bit-times, which could (theoretically) lead to pretty trashy performance with small packet, high load networks.
But, as I said before, it's not like you care, as your gigabit ethernet network is all going to be switched.
-- O improbe amor, quid non mortalia pectora cogis!
This is four wires we're talking about here, not a programming language.
Crash course in network wiring:
1. There are two possible types of twisted-pair Ethernet cables: 1) straight through (for connecting a card to a hub), and 2) crossover, or null (for connecting 2 NICs together, this is the same thing as crossing over Rx and Tx to allow two DTEs to communicate)
2. This new chip can automatically detect which cable is being used, and set itself up automagically. Now you can use both correct types of cable interchangably. Eventually (hopefully), we can simply just buy straight-through cable all the time, for all situations.
This also makes upgrading from a PC-to-PC network to a hub network very simple as you won't have to completely recable.
It's more of an interoperability thing if you ask me.
get an OC49
According to the Gigabit Ethernet group draft standard IEEE 802.3ab, Gigabit copper 1000base TXshould run on all decent Cat 5 installations.
It does this by running single duplex over all four pairs at 125 MHz. The coding is changed to increase the number bits per symbol from 0.8 to 1.25. Simple wiring screw-up like mixing-up tip & ring are already handled by most 100baseTX ethernet transceivers. But crossover-vs-not isn't, and split pairs are unfixable.
Your Cat5 working 100baseTX is supposed to run 1000baseTX just fine. But it won't if you've left pairs unconnected, or stole them for a second run or phone. Poor crimping might also hurt.
That said, the real question is what you can do with all that bandwidth. Most hard-disks cannot sustain even 10 MB/s that 100baseTX provides. And it's hardly a high spped internet solution. It only runs 100m from the hub. The real problem with internet has always been interbuilding: the last mile between cable heads and user buildings.
RCN is doing just this. They are laying fiber out to people's houses... I think the only copper part is between you and the box, and a single box only servers some 100 homes. I believe that each box gets 12 pairs of fiber.
:b
They are going to provide a single solution for everything - TV, phones, internet, etc. It is expected to be very very fast.
I know this because they bought the company I was working for last summer (an ISP here in the SF Bay Area) and this is what they told us. But don't worry, they said it's fine to tell the world.
IIRC, this service is going to be availible here (in the bay area) as well as in the Boston area.
I know exactly how to make a crossover cable. From memory:
:-)
Brown - Brown White, Green, Blue White - Blue, Green White, Orange - Orange White
Other Side:
Green - Green White, Brown, Blue White - Blue, Brown White, Orange - Orange White.
Splitting the center pair, and keeping the blue in the middle reduces crosstalk, and the pairs are matched with transmit and recieve.
I still sometimes only 10 Mbit. Maybe some of the ethernet cards I've used are crap. I'll take your word for it and try a little harder next time.
--- "So THAT's what an invisible barrier looks like!" - Time Bandits
On 100 Mbit ethernet, a crossover cable will usually only give you 10Mbit. I assume the same will be true with gigabit. You probably won't get full speed out of a crossover cable.
I believe the top speed for the firewire spec is 400 Mbit. I'm not sure if all devices, or ports, support 400 Mbit, but that's what's in the spec.
Also, the EV6 bus for the Athlon is 200 Mhz, with separate switching for RAM and PCI. I wonder if that might be a good solution.
--- "So THAT's what an invisible barrier looks like!" - Time Bandits
Let's call a dual Pentium 2/300 a mid-range SMP box, shall we?
A Pentium 2 processor has 512K of level2 cache (running at clk/2 - 150 MHz) on the cartridge. Are you suggesting:
Trees can't go dancing
So do them a big favor
Pretend dancing stinks!
Also, anybody got information on how collision handling is done on this new architecture? I would suppose that, being a gigabit ethernet, it would surely see much more usage than a 100Mbps one, and being also much higher speed, there should be more collisions.
Yeah, and the faster those packets move, the more likely they'll be damaged when they smack into each other!
Trees can't go dancing
So do them a big favor
Pretend dancing stinks!
For just doing point to point stuff like this it seems more efficient to just use a 100 or even 10 megabit switch.. And hook THOSE together with a nice fat gigabit uplink.
Trees can't go dancing
So do them a big favor
Pretend dancing stinks!
Actually, U2W tops out at 80MB/s. That's a limitation of the SCSI bus, NOT the PCI bus. And throughput tends to be non-linear as you increase the number of devices on a SCSI bus. I've seen graphs of some controller tests as the number of drives increased -- most controllers started to suck at 5 drives. (Of course, this was several years ago -- long before U2W, LVD, U160, and fiber channel.)
The advantage of 64bit/66MHz PCI is for hardware (read: very large cached) RAID controllers [the Mylex extremRAID 3000 comes to mind] and multiport Gigabit ethernet cards.
I want.
Seriously, Cheap Gigabit ethernet could really help out in the office setting. With Buttloads of server space, you could actually implement those roaming profiles on NT, for instance, without clogging your network to bits.
---
DO NOT DISTURB THE SE
Finally, something I can actually say "Wouldn't it be cool to build a Beofwulf system with these puppies!" about... :-\
Jack
- -
Are you an SF Fan? Are you a Tru-Fan?
This type of low cost high-speed connectivity could bring the benefits of a SAN (Storage Area Network) architecture to those who can't afford a FibreChannel based system.
The storage server can be based on PC architecture with a stripped-down linux kernel, emulating FibreChannel over gigabit ethernet. It has no notion or filesystems, users or anything like that - it is optimized to just ships disk sectors to the network at maximum performance.
The application servers can be diskless or use their local disks only for swap and caching. One ethernet interface will connect to the internet and another will support access to the SAN. Replacing or upgrading such servers is easy when they store no state information.
XFS is capable of letting two or more systems share access to the same disk at the sector level.
I don't know if the linux port of XFS will support this feature, but assuming it does this could be very useful for this kind of applications.
----
Stop worrying about the risks of nuclear power and start worrying about the risks of not using nuclear power.
Sure, you can get a fairly inexpensive gigabit ethernet card, but how much is the hub gonna cost. You can only connect 2 computers through a Null Cable (cross over).
(First?)