10Gbit Ethernet Half-Duplex

Any reason why 10Gbit Ethernet at Half Duplex is dis-allowed?

When gigabit was new, before switches were easily available, I had thought it would be useful to have NICs and repeaters that didn't do carrier extension for clusters of machines in a single room. Especially when a large amount of communication is broadcast (such as loading the same code into many machines) it would have seemed useful. Now with gigabit prices low, it is hard to argue that, but it might be for 10Gb.

-- glen

But again, what advantage is there to going with half duplex? With HD, you've got all the problems with collision detection, reduced throughput etc. While there were techniques developed for gigabit, to allow HD with reasonable segment lengths, the problems are 10x worse at 10 Gb.

Surely the real question is why 1 Gbit Ethernet at Half Duplex WAS allowed ;-)

John

Here is a complete explanation, excerpted from Chapter 10 of my book "Gigabit Ethernet" (Addison-Wesley, 1998):

----begin excerpt----

Do we even need a half duplex (traditional) Ethernet MAC at gigabit rates? This is a reasonable question, especially since it is clear that we had to make some significant changes to the Ethernet design in order to support half duplex operation at this speed.

There is only one advantage of half duplex over full duplex operation: Half duplex (shared) LANs can use a repeater hub rather than a switching hub, thereby potentially saving some system cost. A switching hub will always cost more than a repeater hub (on a per-port basis), however, over time the difference shrinks considerably with competition and availability of high-integration silicon switching components. The cost of the station attachment (NIC) is identical whether using shared or switched hubs.

So, to the extent that the hub constitutes a cost factor for the system as a whole, the system cost will be greater when using full duplex vs. half duplex operation, since full duplex LANs require switching hubs. There is a premium paid to use full duplex mode, but it is not very large, and the premium decreases over time. In return for this price premium, we can avoid:

- Changing the Ethernet MAC algorithms,

- The performance degradation of carrier extension (for short frames),

- The performance limitations of CSMA/CD in general, and

- Any need to reduce network extent (distance).

[So] why did the industry bother to develop a half duplex Gigabit Ethernet standard at all?

The answer is more political than technical. Gigabit Ethernet was developed under the auspices of the IEEE 802.3 Working Group. By definition [at that time], 802.3 networks must include the capability of CSMA/CD operation. If Gigabit Ethernet offered a full-duplex-only solution, it would have been difficult to justify its development within the IEEE 802.3 Working Group. This would have resulted in some important political difficulties:

(1) A new Working Group would have to have been formed within IEEE

802; this takes considerably more time than a new project within an existing Working Group and would have delayed the development of the standard.

(2) The resulting standard would have had difficulty calling itself "Ethernet", since it would neither use CSMA/CD (even as an option), nor would it have been developed as part of IEEE 802.3, the recognized "owner" of the Ethernet name.

So, rather than fight the system, the IEEE 802.3z committee chose to work on both half duplex and full duplex Gigabit Ethernet, making modifications to the CSMA/CD algorithm to support reasonable distances in half duplex mode. The real rationales for the specification of half duplex Gigabit Ethernet are standards-committee machinations and market positioning. Few industry observers believe[d] that there will be much (if any) half duplex Gigabit Ethernet deployment.

----end excerpt----

(C)1998, Addison-Wesley Longman and Networks & Communications Consulting. All rights reserved.

-- Rich Seifert Networks and Communications Consulting 21885 Bear Creek Way (408) 395-5700 Los Gatos, CA 95033 (408) 228-0803 FAX

Send replies to: usenet at richseifert dot com

"It ain't Ethernet if it don't do CSMA/CD?" :)

rick jones

The VG-AnyLan trauma? ;-)

SCNR

Nah, I think I'm mostly over that one by now. Just feeling irreverant today :)

rick jones

(snip)

But I am specifically interested in the case where the segment lengths are known to be short. That is, even 10Gb without carrier extension.

And yes, in a fairly short time 10G switches will be cheap, but that day isn't here yet.

How many clusters of machines packed together as close as they can be exist today that could use faster network interconnects? How much faster would 10Gb half duplex be than 1Gb full duplex?

Those are the questions. I am not interested in carrier extension and long segments.

-- glen

But again, why bother? What possible advantage is there, that you'd want to go backwards? If you want faster, why waste time with HD? You'd be back to the overhead of collision detection, collisions, cutting the bandwidth in half etc. If you want to maximize bandwidth, FD is the way to go.

I think perhaps that his reasoning is that 10G HD would be cheaper to implement than FD but still give higher performance than gigabit. Which, while it might be true, is mostly a matter of degree--10G right now is going to cost you whether it's full or half and using a repeater instead of a bridge wouldn't be likely to save that much. I suspect that the same performance level as 10G HD would provide can be achieved for less cost with trunked gigabit.

On the other hand the cost of half duplex (collission detection) will follow you even if only full duplex will be deployed. So in the long term it will be more expensive. Although i have no real idea of the relative costs of collision detection to rest. But another interesting point is the delay, an repeater will introduce. A switch is allowed to have a delay greater than the time of a collision domain...

Ciao

I think the size of the padding in each frame would be detremental to using half duplex especially at the distances it would be implemented over. The cost thing might be only partly responsible for it. But both seem worth commenting on.

- Thanks everyone.

I would be interested in one where the maximum link distance was 2m.

Consider clusters of machines, MPI or PVM, for example. You can put a lot of machines within a 2m radius of a repeater.

-- glen

Rich,

I hope you agree that this is a very minor comment. From what I have read on this and other newsgroups your work seems to me to be utterly outstanding.

Rich says: "A switching hub will always cost more than a repeater hub (on a per-port basis), however, over time the difference shrinks considerably with competition and availability of high-integration silicon switching components"

As far as I know (which is not far at all) the cost of LSI components is strongly related to the production volume and is much less closely related to the complexity of the chips (until the yield falls off anyway).

So:- I contend that your comment "A switching hub will always cost more than a repeater hub" is only the case for equal production volumes and since GBE repeater hubs are practically non existent, your comment is not strictly applicable.

It would clearly be applicable if neither switching hubs nor repeater hubs were yet in production.

I have only made this comment since I was hoping for a second edition:)

Thanks.

Which was the case in the emerging market for Gigabit Ethernet (which, I believe, was the context for the statement); of course, you are correct that I am assuming at least comparable, if not equal, volumes.

Not likely to happen anytime soon, I am sorry to say. My work has taken a very different direction.

-- Rich Seifert Networks and Communications Consulting 21885 Bear Creek Way (408) 395-5700 Los Gatos, CA 95033 (408) 228-0803 FAX

Send replies to: usenet at richseifert dot com

Rich Seifert wrote:

That's too bad (the book part, not your work part! :)

It's hard to find good technical books.